让模型自主规划、调用工具、完成多步任务的技术方向。
Keeping a city running means making decisions that affect residents every day — from when to repair a road to how to use energy more efficiently. Across the EMEA region, cities and local software, infrastructure and integration partners are using NVIDIA accelerated computing to build physical AI agents and digital twins that help teams make […]
Across industries, the U.K. is putting the full stack to work — and the scale of deployment is growing fast, from enterprises transforming their businesses to startups solving some of the planet’s hardest problems. NVIDIA is putting that principle into practice, including through a five-year, $250 million compute commitment to strengthen the U.K.’s agentic safety […]
大型语言模型（LLM）的测试时扩展可以包括生成多个候选解以供选择，或修改先前的尝试。为了指导选择和修改，我们需要一个能够估计候选解正确性并提供失败证据的验证器。然而，在竞赛编程中，官方评测系统会在有限的提交次数限制下给出权威判定，却不会透露失败原因或导致失败的输入。为了在不消耗官方提交次数的情况下提供有信息量的反馈，我们提出了一项任务：模型仅使用题目描述和源代码，对候选程序进行判定和诊断，并将生成的反例作为可检查的证据。为评估这项任务，我们基于 USACO 2026 竞赛构建了 USACO-JUDGE，使用由人类专家和多个 LLM 编写的通过与未通过评测的程序。随后，我们在通用执行沙箱中通过强化学习训练了 ARBITER，这是一个拥有 35B 参数的混合专家（MoE）智能体式验证器。训练使其在错误程序上的反例成功率从 40.61% 提升至 62.67%，并使其在 USACO-JUDGE 上的二元判定平均准确率从 71.41% 提升至 87.41%，接近前沿验证性能。除验证之外，ARBITER 在代码选择和代码改进方面也都优于基线方法。在一项 IOI 2026 案例研究中，在官方提交次数限制下，由 ARBITER 指导的代码改进使 GPT-OSS-120B 的预期得分超过了金牌分数线。
大型语言模型多智能体系统通过在专业化角色、工具和工作流之间开展协作，扩展了单个智能体的能力。然而，协作也使故障更难诊断：一个智能体可能误解正确结果，而其他智能体则将这一错误继续传递下去。因此，失败归因旨在识别负责任的智能体和决定性错误步骤。近期方法会追踪依赖关系，但知道信息源自何处，并不等同于知道错误发生在何处。我们提出BLADE（通过审计依赖证据进行责任定位），这是一种无需训练的框架，超越信息追踪，转而审计信息的使用方式。BLADE利用Jev进行结构化概率判断，并使用依赖概率来选择相关的信息源—决策交接点进行审计。在这些交接点上，BLADE检查接收方的决策是否忠实地使用其信息源，以及是否符合任务要求，从而区分错误引入、错误传播和错误修复。这些发现将用于形成全局诊断，并考虑所有执行步骤。在没有参考答案的情况下，在共享Jev决策接口下，针对五个Who&When和Who&When Pro子集中的1,228次失败运行，BLADE在八种策略中取得了最佳或并列最佳的步骤准确率，相比All-at-once提升了2.2至11.1个百分点。明确呈现责任证据，为构建更具可解释性和可靠性的智能体系统提供了一条路径。
长程智能体将任务进展积累在持久化任务状态中，却仍反复处理描述这些进展如何取得的交互历史。这引出了一个自然的问题：长程上下文能否围绕可复用的进展来组织，而不是围绕产生这些进展的历史？本文介绍 CarryOn，这是一种面向对象的上下文运行时，围绕可复用的任务对象组织任务续接。CarryOn 不将压缩后的历史或重建的进展描述作为任务续接的主要依据，而是基于续接载体、补充性执行上下文以及持续更新的近期交互尾部构建上下文视图。一个核心挑战是，在相关历史被移除后，如何保持这一可复用基础的有效性：CarryOn 如何在不于每次折叠时重建完整表示的情况下更新这一基础？CarryOn 跨折叠维护载体记录，依据绑定的证据和适用的环境检查修订受影响的记录，同时保留未受影响的记录，并同步更新执行上下文。在 OSWorld 2.0 和 BrowseComp-Plus 上的实验表明，性能分别提升了 5.8% 和 7.0%，同时总 token 用量分别减少了 51.9% 和 22.9%。项目页面：https://ethannju.github.io/carryon
Frontier language-model agents can now improve robots on their own: they write control code, test it, and keep what works as skills and memories for the next task. We introduce SEAR (Self-Evolving Agentic Robotics), a framework and benchmark built around two questions. Is the agent really learning, or just trying more often? And does what it learns still help on new tasks, in a new simulator, and on a real robot? The framework runs an agent in three modes: it writes code that becomes the policy (code-as-policy), acts live in the control loop (agent-as-policy), or works as a researcher that fine-tunes a vision-language-action model (auto-research). The benchmark adds controls that separate retained experience from repeated attempts, familiar tasks, privileged simulator state, larger budgets, and luck. Across seven frontier language models, 330 simulated and five physical tasks, and over four years of agent runtime, we highlight three findings. First, the learning curve of a robotic agent depends both on the overall time budget and the task difficulty relative to model capability. Harder tasks make longer sessions worthwhile; on easier tasks, re-rolling short attempts does just as well. Second, control code and shared routines transfer across tasks and simulators, while task-specific solutions do not. On a real robot, we observe a performance gap between agent-designed practice and novel human-designed scenes. Third, different models evolve more effectively in different ways. Some do best acting live, others writing code first. And a text-only model keeps pace with a comparable multimodal one: perception helps with error assessment, but agentic reasoning and tool use can matter more. We release the benchmark and codebase to support future research.
在协作式多智能体强化学习（MARL）中，智能体必须学会协调行动，以实现共同目标。有效协调的一个关键指标是行为多样性，即在面对相同观测集合时，智能体执行各具特色但彼此互补的控制策略的潜力。尽管行为多样性能够促进角色专门化和更丰富的联合行为，但其适宜水平高度依赖具体情境和任务，且通常需要付出较高成本进行调节。我们提出 Adaptive Diversity Control（ADiCo），这是一种在训练过程中在线调整行为多样性的双时间尺度框架。在较快的时间尺度上，Diversity Control（DiCo）将系统多样性调节至目标值附近；在较慢的时间尺度上，Extremum Seeking Controller（ESC）仅利用平均回合回报来调整该目标值，无需多样性—性能关系的显式模型或解析梯度。我们在两项协作式 MARL 任务的六种配置上评估了 ADiCo，结果表明它能够维持与任务相适应的多样性水平：当同质化策略更有利时保持较低多样性，而在需要异质化策略时维持较高多样性。通过这种自适应机制，ADiCo 的平均性能达到或超过 DiCo（通过网格搜索得到的最佳多样性目标值）和 Unconstrained DiCo（允许多样性在没有目标调节的情况下自然形成）。这些结果表明，将经典反馈控制融入 MARL 训练，能够通过为多样性目标值的任务特定手动选择提供一种系统化替代方案，实现有效协调。
Harness self-evolution revises an agent's software scaffolding, including prompts, memory, workflows, and executable components, while keeping model weights fixed. We formalize a feedback-driven process in which accepted revisions persist across runs. The framework separates update targets from proposal and admission mechanisms. It distinguishes changes to interpreted state from revisions to tools and code, repository-level control logic, and plugin composition. We characterize search through evolutionary optimization and examine its reliability through continual PAC-learning. Under explicit assumptions, the analysis yields an expected generalization bound, a retention bound conditional on historical regression budgets, and a high-probability admission rule using fresh, independent validation data. We examine how persistent harness revisions affect knowledge sharing and organizational workflows, with implications for scientific reliability, physical safety, and institutional accountability. Cases from prior work illustrate model--harness co-design, distributed execution, and plugin reuse. Reliable deployment requires continual validation as task distributions and evaluators change, with independent safeguards for learning, runtime privileges, and consequential actions. We provide a comprehensive 66-page discussion of related work, theory, algorithms, evaluation, applications, and the broader impact of Harness RSI, together with continuously updated code and supplementary materials, at https://github.com/zgcagi/Harness-RSI.
我们如何构建一种智能体，使其能够超越人类语言进行思考，同时仍通过语言在外部环境中采取行动？现代 LLM 智能体仍通过文本进行推理和交互，而潜在空间推理虽然提供了高效的内部媒介，却缺乏原生工具访问能力和推理透明度。我们提出 Hybrid Reasoning Agents（HybridAgents），将这两个方向结合起来，使智能体能够在统一的生成过程中动态切换潜在空间推理与基于文本的交互。在架构上，轻量级的 HybridLink 支持文本模式与潜在模式之间的无缝切换，而智能体的连续工作记忆则在多轮交互中无损保留混合上下文。这些设计既保留了集成工具的智能体能力，又使潜在思维可被直接检查。基于这一架构，我们进一步开发了专门用于混合推理的后训练算法：HybridSFT 通过教会骨干模型与 HybridLink 协同工作，稳定集成工具的混合生成过程；随后，HybridRL 通过学习何时在潜在空间中思考、何时通过文本采取行动，进一步增强推理能力。我们在涵盖数学、科学、医学、搜索和智能体推理的 11 项基准测试上，对 3 种模型规模的 HybridAgent 进行了评估。与涵盖文本方法和潜在空间方法的前沿 LLM 智能体及智能体框架相比，HybridAgent 持续实现 10.4% 的性能提升，同时将 token 用量减少 35.5%–54.7%，将推理速度提升至 1.8–2.6 倍，并将成本降至原来的 1/2.7–1/3.7。代码和数据可在 https://github.com/togethercomputer/HybridAgent 获取。
长时域智能体凭借能够自主执行具有经济价值的任务而日益受到关注，其工作时长可从数小时持续到数天。这一能力对于自动化发现和递归自我改进（RSI）至关重要。然而，开源模型在长时域任务上仍然表现不佳，主要原因在于缺乏训练数据和可行的训练方案，而同时扩展这两者十分困难：长而复杂的任务难以整理和验证，且其奖励通常十分稀疏。我们同时解决了这两个问题。首先，我们推出了 Tev-Gym，这是首个具备可验证、稠密且渐进式奖励的长时域智能体训练数据集，涵盖 7 个领域的 1,170 个已验证任务环境。该数据集采用“潜在世界优先”的方法构建：首先建立一个潜在世界，即对任务领域、工作流以及每一步之后预期状态的完整且正确的隐藏规范。随后，每个任务都根据该规范逐步呈现为面向智能体的环境，从而实现可控的复杂度和自然的进度指标。其次，我们将这一指标转化为训练方案：通过进度势能奖励塑形（PPRS），将进度转化为每一步的稠密奖励，从而提高训练效率，并使原本难以训练的任务变得可训练。采用该方案训练得到的 Tev-27B，在 Agents' Last Exam 的计算与数学科学领域（ALE-CM）上取得了显著提升，在 Pass@5 指标上达到 GPT-6 Astra 和 Claude Opus 5 的水平。尽管仅使用与 ALE 相关的数据进行训练，它仍能泛化到 Terminal-Bench（TB）3.0 和 4.0 子集，超过 GLM 5.2 和 Gemini 3.7 Flash，并在 TB 2.1 等中等时域任务上达到 Claude Opus 4.8 的性能水平。
智能体机器人技术将大语言模型（LLM）智能体置于机器人控制回路中，使其在记忆的指导下组合感知、运动和视觉-语言-动作（VLA）工具，从而无需重新训练即可泛化到新任务。然而，这种通用性以高昂的效率代价为前提。每个原语都需要 LLM 基于完整交互历史进行一轮推理，因此单次任务执行的成本随原语数量增加而增长。我们发现，这些推理大多是冗余的：一旦选定子目标，其原语就会遵循记忆中已规定的常规流程。我们提出了 ChunkAct，将动作分块引入智能体机器人技术。在线运行时，一次调用即可启动一个由多个原语串联而成的动作块，该动作块自主执行，仅在失败时交回控制权。离线阶段，ChunkAct 将记忆编译为这些动作块，因此测试时的执行无需依赖记忆，规划器仅用于需要推理的决策。由此，LLM 的计算量随决策数量而非原语数量增长。在 LIBERO-Pro 上，ChunkAct 的任务成功率较智能体机器人技术领域的最先进方法提升了 5.0 个百分点，同时所需的 LLM 调用次数、输入词元数和输出词元数分别降至其约 1/2.9、1/4.4 和 1/2.3。
近年来，自我改进的 LLM agents 取得了长足进展，展示了能够驱动自身智能水平提升的 AI 系统的潜力。然而，许多特定领域的应用依赖于专业化基础模型，这些模型具备强大的领域知识，但仍依赖由人类设计的改进方案。将递归式自我改进扩展到此类模型面临一项关键挑战：随着专业模型不断改进，用于推动其进步的策略也必须随之适应。为应对这一挑战，我们提出 Heterogeneous Agentic CoEvolution（HACE），一种通过相互学习联合演化专业化基础模型与 LLM agents 的框架。HACE 通过监督学习改进专业模型，同时利用在当前专业模型上经验证的性能提升，指导 agent 进行基于策略的强化自训练。反过来，agent 的探索又为后续专业模型训练提供信息，使两个模型能够相互适应彼此的进展。为维持这种相互学习，HACE 进一步采用进化搜索来优化二者的交互工具链，使模型能够随着能力的发展，动态调整信息交换方式以及训练经验的选择方式。我们在三个具有挑战性的领域评估了 HACE：推荐系统、生物学中的蛋白质功能预测，以及材料发现中的晶体稳定性预测。结果表明，在这三个领域的主要任务指标上，HACE 均优于所评估的自我改进方法。这些结果揭示，异构协同进化可作为一种通用方法，通过与学习型 agent 的相互作用改进专业化基础模型。
基于 LLM 的多智能体协作正日益用于解决复杂任务，其方式是将工作分配给专业化智能体，并整合互补的专业知识。然而，现有系统和评测大多建立在一种隐含的完美同事假设之上：协作者提供正确且完整的信息，并及时响应请求。但在现实团队协作中，协作者可能返回与预期问题不符的信息，提供不足以支持决策的细节，或未能响应请求。为了将任务解决能力与从不可靠的信息搜集过程中恢复的能力区分开来，我们引入了 Unreliable Colleagues Benchmark（UnCoBench），通过 Misaligned、Incomplete 和 Missed 三种设置对行为失败进行组织。主智能体需要验证上下文正确性，诊断信息缺口，并跟踪尚未解决的依赖关系，从而获取解决任务所需的正确证据。我们还提出了一种可验证且可扩展的自动生成流程，将基于 LLM 的场景生成与答案唯一性和角色必要性的可执行检查相结合。对十个模型的结果表明，在可靠设置下表现相近的模型，在面对不可靠证据时可能展现出不同的稳健性。轨迹分析揭示，智能体经常遗漏纠正性后续行动，而恢复行为还涉及低效等待和冗余查询。这些发现表明，稳健的协作不仅需要可靠的推理，还需要采取有针对性的行动来诊断证据失效、恢复相关信息，并将其整合到决策中。我们的数据和源代码可在 https://github.com/iNLP-Lab/UnCoBench 获取。
An agent can retain successful trajectories without learning which part of an experience should repair a new discrepancy. We formulate passive imitation (PI) as an architectural loop that localizes a violated contract, extracts a local goal, proposes an effect-matched skill, tests a small patch, and retains evidence after full-task verification. We instantiate this loop as a deterministic symbolic controller and analyze its finite repair-query complexity under explicit identifiability and noninterference assumptions. Across ten simulated tool worlds, PI solves all 600 cold held-out composition tasks with three allowed small trials, compared with 6.17% for a whole-trajectory control, while an equally expressive compositional repair control matches PI. On an unseen tool revision, PI achieves 70.0% success versus 24.0% without small trials and 79.0% with self-trial enumeration; an exploratory 24-task adaptation stream subsequently raises first-submission success to 100%, supporting a limited correction-and-retention mechanism rather than a unique incentive advantage or general intelligence. The CPU-only implementation, integrity checks, and reproducible results are available at https://github.com/FriendOfYishan/PI.
Deploying LLM agents in non-stationary environments requires continually revising their instructions as the task drifts, yet rewriting after every reported mistake can turn noisy feedback into persistent errors. This raises an often-overlooked scheduling question in prompt adaptation: when should feedback trigger a lasting instruction update, and at what editor-call cost? Taking inspiration from proportional-integral-derivative (PID) control, we introduce a family of lightweight gates that react to the current error, accumulate evidence over time, and roll back deterioration after an edit. We benchmark these gates against both simple baselines and advanced prompt optimizers (TextGrad and GEPA), over a diverse suite of drifting tasks, LLM parameter sizes, feedback-noise levels, and random seeds. PID-gated continual prompt adaptation yields a more favorable Pareto trade-off between task performance (high accuracy, low error) and editor-call budget than the baselines considered, pointing to control-theoretic scheduling as a simple, general tool for robust prompt adaptation under unreliable feedback.
Transportation planning has traditionally required substantial expertise in network modeling, traffic assignment, and optimization. As a result, strategic analyses such as network design, capacity expansion, and model calibration remain inaccessible to non-expert decision makers. Recent advances in large language models (LLMs) provide an opportunity to bridge this gap by enabling natural-language interaction with complex transportation planning tools. However, existing LLM-based systems primarily serve as interfaces to simulation software and lack the ability to efficiently search, evaluate, and optimize transportation solutions. This study proposes an integrated human-AI transportation planning framework that combines a large language model, a model context protocol (MCP) interface, an open-source transportation planning engine (AequilibraE), and a differentiable user equilibrium (UE) assignment layer. The framework allows non-expert users to execute network analyses and improve solutions for optimization. A differentiable UE layer is developed by directly differentiating the Karush-Kuhn-Tucker (KKT) conditions of the Beckmann traffic assignment problem. The proposed framework is demonstrated through two representative planning tasks: calibration of travel time functions from observed traffic counts and network design under budget constraints. The examples illustrate how differentiable transportation models and LLM agents can jointly democratize transportation planning and substantially reduce barriers between domain expertise and transportation decision making.
大型语言模型智能体日益在多个实体上运行，而可靠的响应生成不仅需要检索相关证据，还需要保持证据与正确实体之间的关联。我们发现了幻影合并（PM）现象：这是一种实体—证据绑定失败，智能体会将未经支持的声明归因于其已确定的实体，即使任务实际上已经正确完成，也可能发生这种情况。我们在声明层面对 PM 进行了形式化定义，并刻画了其在四类智能体领域和四类骨干模型系列中的发生率及错误子类型。为解决这一问题，我们提出了一个用于声明级风险估计与修正的两阶段框架。第一阶段，Anchor-Grounded Risk（AGR）将智能体生成输出前隐藏状态的表征读出与基于雅可比矩阵的功能信息相结合，提供互补的槽位级风险分数和锚点值风险分数。第二阶段，Fixed-Anchor Correction（FAC）审查被标记的声明，并利用观测到的锚点证据对其进行选择性修订；当没有受支持的替代内容可用时，则删除这些声明。在留出的 Shopping 检测队列上，AGR-slot 达到 96.2% 的 AUROC 和 97.3% 的平均精确率，优于所评估的检测基线。在端到端 Shopping 缓解评估中，FAC 将轨迹级 PM 从 51.4% 降至 6.2%，输出与原始声明的比例为 80.7%。这些结果表明，实体特定的落地失败普遍存在，而将内部风险估计与受证据约束的声明修正相结合是有效的。我们的代码已公开发布于 https://github.com/LabRAI/PhantomMerge】【。
代理框架可以持久化检查点、跟踪工具调用并限制资源使用，但这些功能通常会为执行、治理和重放分别分配不同的标识。我们提出了 Pollard，一种基于受治理语义执行树的框架无关运行时。步骤标识绑定规范化输入、类型、尝试次数和父标识；同一棵树记录完成、拒绝、资源结算和分支历史。严格重放会返回经验证的已存储结果，而不会调用被封装的操作；事务性存储则协调并发工作器之间的资源预留。在签名的确定性冻结条件下，两个状态模型副本分别执行了评估转换，且未违反任何不变量；针对保留租约的 PostgreSQL 集群测试了精确限制，严格重放避免了外部调用，并检测出所有语义变更，同时在控制项上没有误报。一个多对框架矩阵保持了参考行为并验证了这些树。另一项具有成本上限的托管研究测试了实时提供商路径，每次已完成的 Pollard 试验都生成了一棵经过验证的树。总体而言，这些研究在明确的租约、边界和密封保管假设下，支持了其核心机制。
交互式智能体评估通常依赖标准化的、基于 LLM 的用户模拟器，这些模拟器表现出广泛的合作行为，可能忽视多样但现实的用户交互所引发的失败。这会导致对智能体鲁棒性的估计过于乐观，也难以为训练能够跨用户泛化的智能体提供足够的失败信号。现有方法大多依赖预定义的用户行为，可能无法发现特定于智能体或领域的弱点；而面向失败的搜索则可能通过使底层任务失效来降低任务成功率，而不是暴露真正的鲁棒性缺陷。为弥补这一空缺，我们提出 TIPER，一种保持任务完整性的人设演化方法，能够在保持底层任务有效且可解决的同时，自动发现暴露智能体特定鲁棒性弱点的用户策略。TIPER 在任务完整性约束下，通过反思引导的行为变异和按任务进行的 Pareto 选择，演化出持久的用户人设。在四个交互式智能体基准测试中，TIPER 在全部 12 个 GPT-5.4 设置下都取得了更低的留出集 avg@3 分数，观测到的降幅为 5.8 至 33.7 个百分点；在涵盖七个智能体的全部 28 个智能体—领域组合中也同样如此，观测到的降幅中位数为 22.9 个百分点。进一步分析得出了一套从行为到失败的分类体系，将发现的用户行为与反复出现的智能体失败模式联系起来。总体而言，这些结果表明，TIPER 能够对交互式智能体开展更具诊断性的鲁棒性评估，并为未来改进智能体鲁棒性提供有价值的失败分析。代码仓库：https://github.com/UCSB-NLP-Chang/TIPER。
Text-to-image models have made remarkable progress in visual fidelity and instruction following, yet complex compositional generation remains challenging. Existing agentic approaches address this limitation through iterative generation and editing, adapting their actions based on interaction history. However, they typically fix the continuation state (i.e., the image from which refinement proceeds) to the latest image. Since iterative editing is not necessarily monotonic, the agent may continue refining from the latest image even when it has already degraded. In this work, we introduce GenRevisit, an agentic framework that extends action selection to joint continuation-state and action selection. The agent can not only use interaction history to adjust editing instructions but also select an image candidate to edit, jointly deciding where to continue from and how to refine. We train GenRevisit with SFT followed by agentic reinforcement learning. To address the coarse credit assignment of final outcome reward in GRPO, which assigns the same trajectory-level advantage to all actions, we further propose Normalized Action Credit Assignment (NACA) to redistribute positive advantage according to each action's contribution to reducing the remaining objective. On GenEval2, GenRevisit achieves GM scores of 84.3 and 74.4 with Qwen-Image-2512 and Z-Image-Turbo, surpassing the strongest baselines using the same generators by 12.7 and 10.9 points respectively. GenRevisit achieves state-of-the-art performance on GenEval, GenEval++, and GenEval2 among baselines.
As large language models are increasingly adopted as multi-party communication facilitators, there is a growing demand for agents capable of balancing diverse group dynamics across contexts. However, existing LLM facilitators are often fine-tuned for task-specific objectives; generalized LLM-based facilitation methods that can adapt to different types of tasks and diverse group dynamics are underdeveloped. To address this gap, we present GUIDE (Generalized Unified Intervention for Dynamic Exchanges, a framework for developing generalized LLM-based facilitators across different multi-party contexts. GUIDE evaluates facilitation across both objective task outcomes and subjective metrics, including psychological safety and perceived friction. Within this framework, we study whether LLMs can self-learn generalized facilitation strategy offline to real-time interventions. We first collect multi-party human interactions guided by task-specific expert LLM facilitators (N=432 conversations). We then employ a self-learning method in which an LLM induces an actionable strategic playbook from conversations and iteratively refines it using feedback from an LLM judge aligned with human annotators (N=20), developing a single generalized facilitator that outperforms the three task-specific expert facilitators on held-out conversations 61.4% win rate. Finally, we evaluate these improvements empirically; in a real-time online empirical study, the self-learned generalized facilitator preserves facilitator performance and group productivity, improves speaking comfort and reduces interpersonal friction, while using 24.7% fewer turns and 29.0% fewer words than the expert prompts. These findings provide empirical evidence that LLMs can learn generalizable, task-agnostic social facilitation from task-specific human interactions, on-par with hand-crafted task-specific prompts across diverse group settings.
In-silico optimization of nucleic acids relies on predictive oracles that imperfectly capture biological function, so sequences optimized against one model may exploit model-specific artifacts and fail to generalize to independent models. Recent studies have shown that not all designers generalize equally well to independent models. In this work, we use agentic program search to iteratively propose, execute, evaluate, and revise code for nucleic-acid designers. We describe the successful strategies used to prompt the search. AutoNovo designers achieved the best observed mean among the evaluated methods on 3/3 DNA-design tasks in terms of in-silico performance, and 6/6 transfer tasks in terms of independent models. The designers come from a family of discovered nucleic acid designers rather than a single universal best designer. We plan to release the resulting designers. These results are the first steps in creating nucleic acid designers that maximize cross-model generalization, and thus improving the odds that in-silico designed sequences reflect real biology.
The Model Context Protocol provides a shared interaction substrate for the agentic web, allowing heterogeneous services to expose capabilities through a standardized interface. Learning to use this ecosystem entails a hierarchy of capabilities: tool-format skills, server-level transfer, and cross-server composition. We present MCPDisCo, an ecosystem-scale reinforcement learning framework for MCP-native agents. Using a Thought-Markov Decision Process (Thought-MDP), the framework explicitly represents internal reasoning alongside routing, execution, and termination and trains on real-server, long-horizon interactions spanning 70 MCP servers and 527 tools across six domains. We study the reasoning--tool seesaw, a joint-learning failure mode in which improvement in one objective can accompany deterioration in the other. To mitigate potential interference, MCPDisCo combines token-masked reward attribution with bidirectional gradient surgery: the former separates direct supervision for reasoning and action spans, while the latter removes conflicting components from both objective gradients in shared parameters. MCPDisCo achieves higher success with fewer chain-of-thought tokens. On the out-of-distribution MCP-Atlas and the GAIA benchmark, MCPDisCo maintains monotone gains, confirming that the learned behavior generalizes beyond the training server distribution and broader agentic tasks. Model weight and implementation are available.
Activation steering has recently been used as a memory for skills: past experiences of an agent are stored as steering vectors and retrieved later. Facts and preferences about the user are assumed to need text instead. We test that assumption with a training-free activation memory beside a frozen LLM. The change an experience causes in late-layer states is stored as a shift, filed under a whitened key for the conversation's gist, and added back in a later session only when the gist matches. We argue that such a shift is a fixed average of an effect that attention, in context, recomputes for every question, so only question-independent content can survive. In a pre-registered study on Qwen3.5-2B and 9B, with judges from four model families, the memory is selective (about 98% of unrelated replies are unchanged), but facts are stored as the word to say, in the last layers only, and are almost never used (at most 15% clean use, against 96-98% in context). Leanings transfer only in a narrow strength window. Most strikingly, the reference subtracted when writing a preference decides whether its concept or its direction survives: unless the opposite is subtracted, a stored dislike is recalled as a like.
Can an autonomous AI agent reverse engineer a binary as well as a skilled human analyst? Do they adopt the same strategy, and can agents scale to real-world programs? We compare human and LLM-based software reverse engineering (SRE): nine models tackle two compact static-analysis challenges previously solved by human analysts, scored with a common understanding rubric, and we record their complete interactions with IDA Pro. This lets us compare not only what humans and agents recover, but also how they investigate a binary. The best agents (GPT-5.5 Pro, Claude Opus 4.8, and Kimi K3) achieve statistically human-competitive understanding on both tasks, while weaker models remain far behind. Ablations show that the gap depends on code representation and workflow demands, not on model capability alone: open-weight models improve substantially with richer code representations, whereas frontier models show no significant change. Similar success rates hide different workflows. Humans repeatedly revisit functions and follow local program structure; strong agents jump across the binary, retrieve information broadly, and rarely return to inspected code, a behavior we call greedy fetching. This strategy is not fixed: given semantically richer code, agents fetch less without losing understanding. Because greedy fetching aggregates large amounts of code in the agent's context, we measure context-window saturation across 4,000 real-world binaries: large windows accommodate many programs, but larger applications and lower-level representations quickly exhaust them. Autonomous reverse engineering is becoming competitive, with a large speed advantage, but scaling it will require better navigation, selective retrieval, and persistent analysis state, not merely larger context windows.
长时域智能体必须完成一系列工作流，其中后续活动依赖于前序输出以及共享状态的变化。评估面临的一项关键挑战，是在确保任务可执行且结果可验证的同时，扩展任务要求及其依赖关系。我们提出了 SkillHorizon，这是一个通过技能组合在七种商业环境中构建长时域任务的基准。其生成器利用特定领域规则，将可复用的技能流程转化为具体的商业活动。这些活动通过输入—输出依赖、共享资源和状态更新相互连接，使工作负载与协调需求同步增长。成功的参考执行结果用于验证任务可行性并衡量交互工作负载，而程序化评估器则评估局部完成情况和整体任务成功率。我们在 1,050 个任务上评估了五个语言模型，参考交互工作负载范围从 64K 到 1M 个词元。在这些规模下，平均局部目标完成率从 90.4% 降至 46.8%，因为已完成的工作未能跟上不断增长的要求。在更大的工作负载下，智能体更常提交未成功的最终答案，同时仍有大量工作未完成且保留着工具调用额度。轨迹分析还发现了一些案例：脚本能够成功执行，但其输出未能通过后续的商业检查。SkillHorizon 提供了一个可配置的测试平台，用于研究不断增长的工作流中的持续进展、依赖协调和完整任务交付。我们的代码已发布于 https://github.com/KnowComp/SkillHorizon。
Choosing a few robust algorithms or schedules can require evaluating every candidate under many adverse scenarios, even though an agent only needs to identify the winners. A predictor may suggest those winners, but cannot rule out a surprise in an unevaluated scenario. We introduce Certified Adaptive Selection (CAS), which returns the exact top-$k$ set as soon as remaining uncertainty cannot change it. CAS keeps guaranteed lower and upper bounds on Expected Shortfall (ES), the average loss in the worst scenarios. Its stopping test compares the worst admissible selected candidate with the best admissible excluded one; for information-disjoint Cartesian intervals this one-cut test is necessary and sufficient. Prediction has a separate role: selecting the next outcome to evaluate, never authorizing the decision. For empirical ES with exact unit-cost queries, we compute in polynomial time the fewest evaluations that could have proved the realized decision in hindsight. Our prediction-guided agent CAS-PG is exact for every predictor, including adversarial ones. On two public algorithm-selection matrices, CAS-PG certifies the exact top three after 630 of 3150 and 981 of 2200 replayed queries, or $1.14$ and $1.18$ times that hindsight minimum. Without historical outcomes, its transcript-only variant requires $1.28$ and $1.37$ times the minimum, outperforming the tested non-predictive certified policies. Separately, a prediction-free CAS policy avoids $93.0\%$ of exact recourse solves in synthetic energy scheduling. One certificate thus supports several acquisition policies and separately proved risk/cost extensions: prediction changes the work, not the justified choice.
Beliefs in a Partially Observable Markov Decision Process (POMDP) are posterior distributions over the hidden state given the agent's history of actions and observations. They summarize everything the history says about the state, so planning in a POMDP is planning in a Markov decision process (MDP) over beliefs. Belief MDPs have continuous state spaces even for finite POMDPs, and exact planning in them is intractable in general. Planners must therefore approximate beliefs and estimate values. Monte Carlo planners do both inside a search tree, with an approximate belief at every node and a value estimate at every leaf. Approximations usually consist of sets of sampled states, with leaf values coming from rollouts, hand-designed bounds, or learned networks. Policies and values computed from these approximations are limited by their quality, and prior methods with learned networks rely on hand-chosen summaries of the sampled states. A belief representation should be computable from the parent node's representation through the POMDP's dynamics, let the search sample from it, and be learned so that it provides the information that policies and value functions need. We present Belief Embedding Tree Search (BETS), a Monte Carlo planner that meets these requirements by representing each node's belief with a generative distribution embedding (GDE) while updating it through the POMDP dynamics. BETS trains its belief model alongside its policy and value networks through experience, with the realized hidden state of the POMDP as the belief model's training target. We evaluate BETS on six configurations of RockSample, LightDark, and PocMan against BetaZero and three planners without learned components. BETS is competitive or leading on all six, while every baseline falls far behind on at least one. Where beliefs are highest-dimensional, it outperforms them by a large margin, even without search.
Self-evolving agents are usually scored by endpoint Gain: held-out success after an update minus success before. That difference can credit a frozen rescore as evolution. We introduce Trace Evolution Bench (TEB), a protocol that separates that rescore from a context or weight update, together with three frozen 80/20/40 packs. An update is accepted only when validation does not fall, held-out identifiers never enter the update, and a text update is not ranked with a weight update. The primary scoreboard is coding, graded by hidden unit tests in a closed workspace. On that pack (40 held-out tasks, one family), DeepSeek v4 Pro under no_evolve never changes its policy yet moves 7.5, 5.0, and 12.5 percentage points across three seeds. Of nine text-update cells inspired by GEPA, ACE, and ExpeL, seven have positive excess above that seed's own band, and one gepa_llm seed moves down by the width of its band (-12.5, regression 0.48). GLM and MiniMax do not reproduce that seven-of-nine pattern. No method stays above its own band on every seed of all three models: GLM has two paired cells above their own bands, and MiniMax expel_llm seed 44 is a third (regression 0.30). The methods were fixed on a validation screen before held-out was read. Office, graded by rule-bounded delivery, and manufacturing table QA are separate packs: office DeepSeek does not exceed its own bands (widest 17.5 points), and manufacturing has no paired positive excess beyond its seed band.
将推理与工具调用交织进行的视觉智能体，为多模态任务提供了一个颇具前景的方向。然而，有效工具调用的学习通常依赖大规模精心整理的问答对或专家示范，而获取这些数据的成本高昂。在本研究中，我们提出 Marvo，这是一种简单而有效的自我演化方法，能够从未标注图像中学习工具调用。Marvo 由两个智能体组成：（1）挑战者，从未标注图像中生成问答对；（2）求解器，通过强化学习解答问题，从而学习工具调用。为了促进有效的工具调用，优质的训练问题应当足够困难，能够促使求解器使用工具，同时又应处于求解器的能力范围内。为此，我们向挑战者引入边际奖励，其定义为求解器在有无工具访问权限时的性能差异。借助边际奖励，我们构建了一种自适应课程学习机制，鼓励求解器学习有效的工具调用。在涵盖智能体推理、视觉搜索、数学推理和通用现实世界推理的各项基准测试中，Marvo 将平均准确率最高提升了 7.1\%，证明了边际奖励是无需人工监督、扩展问题复杂度的更优方式。
AI coding agents operate in iterative loops: they modify code, run tools, inspect the results, and decide what to do next. We call the design of this process loop engineering, which includes context management, model selection, compaction, and termination. Existing efficiency work mainly compresses tool observations or routes selected turns to cheaper models, but the practical opportunity for these optimizations remains unclear. We analyze 600 publicly released coding-agent trajectories spanning two agent scaffolds and two language models. Median observations contain only 118-286 tokens, while the large build logs commonly used to motivate compression occur near the ninety-ninth percentile. Moreover, 98.6% of input tokens are served from a prefix cache, so flat input pricing overestimates loop cost by roughly 7x on average under near-ideal caching. We derive and empirically validate a closed-form break-even condition for model delegation. Delegation increases cost in half of the evaluated settings, near the predicted boundary. More broadly, turn count explains 84% of the variance in log trajectory cost, compared with 4% for mean observation size; the most expensive trajectories have the most turns despite having the smallest observations. These associations suggest that observation-focused methods target a secondary cost factor, and identify unnecessary turns as the larger and largely unexamined opportunity, though we do not establish that turns can be removed while holding task success fixed. We release TurnCost, a tool for re-scoring trajectory corpora under alternative caching, pricing, compression, and routing policies.
Meta released Muse Wonder and Muse Wonder Full, 1M-context agent models. Muse Wonder costs $4/$20 per million tokens on Meta Model API. Wonder Full, on Muse Code plans, leads DeepSWE and SWE-Marathon. The post Meta Launches Muse Wonder and Muse Wonder Full: Benchmarks, Pricing and How to Use It appeared first on MarkTechPost .
Our first robotics pod with Dyna’s Chief Scientist on their big bet on World Action Models & the Next Frontier of Physical AI
Showing off Sugabots, an internal project I built at work that we've decided to open-source. I got tired of the team just copying context between agents and Slack. So I built an internal tool for shared chats where our team and agents could work together, using our choice of models and tools. The team is now actively maintaining and building on it to add all the features we're missing. We've shared a hosted offering with a few other teams and has proven useful for them too. Its open-source (MIT) and self-hostable if you want to use it yourself. Let me know what you think. Comments URL: https://news.ycombinator.com/item?id=50079791 Points: 1 # Comments: 0
了解 AWS Generative AI Innovation Center Partner Agent Factory 和 Anchor Browser 如何构建能够自主支付费用以访问受 x402 保护的网络资源和 API 的 AI 智能体。Amazon Bedrock AgentCore payments 可检查预算、签署交易并重试请求，而无需向智能体运行时暴露钱包凭证。
很快，更多 AI 代理将能够浏览 Shopify 商品列表并进行购买，这标志着该公司在代理式商务方面的野心进一步扩展。Shopify 周二宣布，将允许消费者配置自己的 AI 代理，代表他们购买商品，从而让更多代理能够访问 Shopify 商品目录中的超过十亿条商品信息。（该目录包括 […]
Sovereign AI control is becoming a business continuity issue as autonomous agents take on essential enterprise workflows. Keeping data within national boundaries addresses part of the challenge, but organizations also need control over agent behavior, external dependencies and the economics of keeping those systems operational. Regulated enterprises need more than access to powerful models and […] The post Deliverance AI puts sovereign AI control to the continuity test appeared first on SiliconANGLE .
I've used SSH tools for years and didn't really plan on building yet another SSH client, but I ended up building one the way I wanted. Hatoba is a desktop SSH client for Windows, Linux and macOS. It's written in Rust with Tauri 2 and MIT licensed. Each terminal has an AI assistant panel next to it. You describe the problem, and it reads the terminal screen, runs commands, types into the shell, and reads, edits and searches files on the host to find the cause and fix it. You can approve every command and edit it before it runs, let a model you pick approve the safe ones, or turn approval off. Files it changes show a diff you can undo. It works with any OpenAI-compatible or Anthropic API, local models included, and supports skills and MCP servers. Sync doesn't go through a server of mine. Hatoba deploys a Worker and a D1 database to your own Cloudflare account and syncs hosts, keys and AI conversations through them. The free plan usually covers personal use. Everything is encrypted on the device before it leaves. Keys are derived from your master password with Argon2id, every item is encrypted with AES-256-GCM, and the Worker only stores ciphertext plus the hashes it needs to sign you in. Someone who gets hold of the whole Cloudflare account still only gets ciphertext. Sync is optional, and everything works offline without it. The usual SSH client features are there too: groups, tags and fuzzy search, ~/.ssh/config import, ProxyJump and SOCKS5/HTTP proxies, an SFTP panel, local port forwarding, and a key vault that imports OpenSSH, PEM and PuTTY keys. There's also a dashboard showing CPU, memory, disk and network for every saved host, read over plain SSH with nothing installed on the servers. SSH is russh, the terminal is xterm.js with the WebGL renderer, and the UI is React. I develop on Windows, so it's Windows first, with a custom title bar that supports Snap Layouts and Mica, and working CJK input methods. The Linux packages and the macOS build get far less testing on real machines. The macOS build is Apple Silicon only and isn't notarized, so macOS blocks it on first launch until you allow it in Privacy & Security. Bug reports from Linux and Mac users would help the most. Website: https://hatoba.fog.moe Downloads: https://hatoba.fog.moe/download/ Comments URL: https://news.ycombinator.com/item?id=50076388 Points: 2 # Comments: 0
University of Michigan spinout Laplace Intelligence uses AI agents and interactive simulation to stress-test robots and self-driving vehicles before deployment. The profitable startup is competing at TechCrunch Disrupt's Startup Battlefield.
Hi HN, I'm Kirill. A dozen coding agents running in parallel, producing walls of suggestions, comments and ideas, all at once, all crammed into 88 columns of terminal scrollback, all gone the moment you run the next command. Does this sound familiar? FolderMd ( https://foldermd.com ) might be for you. It's free, and it's fast, and your agents can use it. I made it because I just couldn't handle it anymore. It's a small Mac app that gives those documents a home. Agents read our explainer and then learn to write markdown into a .foldermd/, instead of writing it into the terminal. FolderMD then opens each one instantly and renders it like a book: real typography, syntax-highlighted code, Mermaid diagrams, light and dark mode. You can edit a doc and save it, you can then add comments just like you would in Google Doc, and pass those comments to the agent. Let me know what you think, and if you want to see any features, and improvements. Thanks! -K. Comments URL: https://news.ycombinator.com/item?id=50074607 Points: 1 # Comments: 0
Grok 4.7 现已在 Amazon Bedrock 上可用。在本文中，我们基于 Grok 4.7 构建一个工具调用智能体，并在四个推理力度级别下对其在三项推理任务上的表现进行评估，同时衡量每个已完成任务的正确性和令牌成本。我们将展示如何针对每个请求调整这种权衡，并在生产环境中对其进行监控。
ShipStation Global's engineering team built Deckhand, an AI IT help desk agent for Slack on Amazon Bedrock AgentCore. It answers routine IT questions from a curated knowledge base, reads error-message screenshots, and files Jira tickets. In its first three months it resolved 97.5% of channel queries with a 96% positive feedback rating.
Hacker News 的各位好！我是 Consus 的创始人 Eric。我们将所有 AI 请求路由至符合 ITAR 限制和受控非机密信息（CUI）要求的基础设施。Consus 本身符合 ITAR 限制要求，我们最近还完成了针对 410 项 FedRAMP High 控制措施的 FedRAMP 等效性审计。FedRAMP High 是美国政府针对处理敏感非机密数据的云服务制定的最严格安全基线。此类数据泄露可能造成严重或灾难性后果。在处理 CUI 信息之前，必须取得 FedRAMP 等效性或授权。演示：https://youtu.be/6MYxJe6P4oY 航空航天与国防行业正在快速增长。但遗憾的是，由于必要的数据保护措施和合规风险，大多数公司都难以使用 AI。如果这些公司能够使用大多数其他行业习以为常的 AI 工具，它们本可以发展得更快。 在加入 Consus 之前，我的大部分职业生涯都在航空航天与国防领域。我曾在 SpaceX 负责发射国家安全载荷，也是 Castelion 的早期员工。我亲眼看到，合规方面的繁文缛节如何让公司难以采用新技术，也难以信任那些会接触其敏感数据的供应商。 例如，我曾与一家卫星制造商合作，该公司不允许使用任何 AI 工具，即使是处理非敏感工作也不行。尽管过去 9 到 12 个月里 AI 已经变得极其有用，但令我惊讶的是，仍有这么多航空航天与国防公司拒绝接触 AI。他们看得出 AI 可能会有所帮助，但担心模型提供商使用他们的数据进行训练、违反 ITAR 规定，或者因未通过网络安全评估而失去新业务。这些担忧都是合理的。 解决其中部分或全部问题有许多方法，但我们专注于构建一个合规的 AI 推理层，可以“接入”任何 AI 工具或工作流。正如 OpenRouter 能够与大多数工具集成一样，Consus 可以连接到 ChatGPT 或 Claude 这样的工具、包括 Pi coding agent 或 Open WebUI 在内的开源工具，也可以集成到各个组织自行构建和维护的应用程序中。 几乎没有哪家航空航天与国防公司同时使用三大 GovClouds（AWS、Azure、Google）。当我梳理哪些前沿模型可以合法运行于 ITAR 和 CUI 工作场景时，我发现每个提供商都只提供部分模型，而每个模型达到的合规水平也各不相同。 我们的客户不会从云提供商的角度思考。他们考虑的是所需的数据控制措施，以及想要使用的模型。这就是我们构建合规后缀的原因。在底层，每个请求都会访问一个统一的 OpenAI 兼容端点；在接触推理服务之前，系统会先验证密钥并执行策略。每个模型名称都必须带有合规后缀，例如 gpt-5.6-sol:fedramp-high 或 claude-opus-5-5:itar，任何不带后缀的模型名称都会被拒绝。 我们绝不会代理到标准商业端点。客户还可以构建模型允许列表，并为组织内的单个用户分配支出和模型使用限制。所有这些都通过 Consus Portal 处理，团队可以在那里管理支出、密钥和策略。 我们经常听到的一个问题是：为什么不直接在 AWS GovCloud 或 Azure Government 中使用 Bedrock？对于小型团队而言，仅仅获得这些平台的访问权限就很困难。对于大型组织而言，GovCloud 允许你继承部分、但不是全部处理敏感数据所需的控制措施。此外，模型分散在不同提供商之间：GPT-5.6 Sol 仅在 Azure Government 中提供，而 Claude Opus 5.5 则通过 AWS GovCloud 中的 Bedrock 提供。如果两者都想使用，就需要搭建并保护两个独立的政府云环境。 我们将所有模型置于一个端点之后运行，因此客户无需选择云平台，而合规工作则由我们分别为每个平台处理。 鉴于数据的敏感性，Consus 不会在静态状态下存储任何客户数据。所有提示词和生成结果只会在请求处理期间保留于内存中，绝不会记录或写入磁盘。无论是我们自身，还是基础设施及模型提供商，都严格执行零数据保留政策。出于计费和审计目的，我们会保留每个请求和响应的元数据。 Consus 从设计上就不是自助服务产品。ITAR 要求我们确切了解每个密钥另一端的使用者，因此我们会验证每个组织，并在发放 API 密钥前签订合同。 我们向客户收取合规费用，起价为每席位每月 40 美元，另加按成本收取的 AI 令牌费用。我们目前正在积极服务客户，非常期待获得安全专家，以及在高度监管行业工作的开发者和工程师的反馈！ 参考资料： API 文档：https://gateway.consus.io/overview 安全与架构：https://gateway.consus.io/security 合规详情：https://consus.io/compliance 评论 URL：https://news.ycombinator.com/item?id=50073471 得分：4 评论数：1
Hey HN! I have been dabbling into some hobby projects that require designing custom 3D printed models. I have been relying on AI models, mostly GPT Sol and Astra through Codex, and giving them certain tools such as Blender MCP and OpenSCAD CLI to design. My end goal is 3D printable designs ( .stl etc). The results have been good but not great. Iterating and improving them is such a pain in the ass. I want to know which specific tools or workflows you would recommend I should try. Thanks in advance! Comments URL: https://news.ycombinator.com/item?id=50072103 Points: 1 # Comments: 0
Release: web-edit 0.1a0 I was working on a project from my phone using Codex Remote when I found myself wanting to edit a credentials file without round-tripping those credentials through the LLM itself. So... I had Codex and GPT-6 Astra Extra High build this Python web app which provides a web UI for viewing and editing files in a folder. It's intended to be run over Tailscale, so I can have an agent start this running and then edit files via my Tailscale network directly from my phone. Here's the interface for browsing files in a folder: Screenshot of a web file manager interface. Header reads "WORKSPACE" above the title "Files", with buttons "Refresh", "Upload files" and "New file" (dark green). A breadcrumb shows "web-edit / web_edit / static". A table with columns "Name", "Size", "Modified" lists four files, each with "Download ↓" and "Delete" links: "app.css" 8.4 KB Oct 12, 2026; "app.js" 671.5 KB Oct 12, 2026; "index.html" 4.3 KB Oct 12, 2026 (underlined, as if hovered); "THIRD_PARTY_LICENSES.txt" 32.3 KB Oct 12, 2026. Below the table: "Images under 2 MB appear inline. Select a text file to edit it." A panel below is labelled "UPLOAD QUEUE" with heading "Add files" and a "Close" button. It has a "Destination folder" text field containing "web_edit", then a dashed drop zone reading "Choose files or drop them here" and "Set a path and filename for each file before uploading." with a "Choose Files" button and the text "no files selected". A greyed-out "Upload files" button sits at the bottom right. Editing uses CodeMirror 6 - I was going to use Monaco but that still doesn't have a great mobile experience yet. Tags: projects , python , ai , tailscale , generative-ai , llms , codex
If the agent isn't using my diagnostics library, I could introspect and update AGENTS.md, but there's still a delta, and it'll drift. Instead, I write a pre-commit hook that checks the git diff for direct calls to observability modules that bypass my diagnostics module. Boom: fail the commit, inject a prompt telling the agent to use the diagnostics library. All of a sudden it's ten out of ten. — Geoffrey Huntley , nobody should type code by hand anymore. here's what's left. Tags: coding-agents , generative-ai , git , agentic-engineering , ai , geoffrey-huntley , llms
Hi HN, I built Mana, a family organizer: shared calendars, shopping lists, receipt scanners, meal planning, chores, and notes for the whole household, with a built-in AI assistant ("add milk and eggs," "plan dinners for the week") plus automations (e.g., milk goes back on the list before it runs out). Additionally, you can connect Mana MCP to your Claude, ChatGPT, Gemini etc. The goal is to eventually take the model calls and storage to local but that's work to be done. Appreciate if there's any feedback. Thank you! Comments URL: https://news.ycombinator.com/item?id=50071636 Points: 1 # Comments: 0
American Express is introducing a new framework designed to ensure AI agents only make purchases their users have authorized, potentially preventing costly mistakes and disputes as AI-powered shopping grows.
I created MemorySync because I kept running into the problem of AI agents not being able to remember useful things about users between conversations. I wanted an easy way to give the agents memory in my application without having to build the whole thing myself. MemorySync is an API for giving AI agents memory. The application sends MemorySync information provided by the user, MemorySync saves useful memories for that user, and they can be retrieved later when needed. Give it a try without creating an account: https://memorysync.io/try Store a memory, refresh the page, and ask for it back to see if it remembers. If you're interested in using the API, I put the documentation here: https://docs.memorysync.io/ It's hosted only at the moment, so there is no self-hosting option. It's currently available in one US region, and it takes a few seconds to store a memory. The free plan allows 10k saves and 1k searches per month, and no credit card is required. I'd like to know if you think this is a useful abstraction for building AI agents, and what else you would expect from a memory API. Comments URL: https://news.ycombinator.com/item?id=50070688 Points: 1 # Comments: 0
Z.ai’s GLM-Legal is a legal AI model built on GLM-5.3-Flash. It scores 19.2% all-pass on Harvey’s Legal Agent Benchmark, nearly 3x its base, at $0.15 per test. Access is by application. The post Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB appeared first on MarkTechPost .
法国企业正在航天、金融服务、公共部门和科学研究等各行业部署智能体计算，展现出该国生态系统如何从雄心迈向实际应用。法国的进展建立在精心布局的基础之上。2025 年 2 月举行的人工智能行动峰会，以及 2025 年和 2026 年举行的“选择法国”峰会，促成了重大的算力投入承诺 […]
Hi HN. I partnered with Simon Kley to build this project at innovation intelligence hackathon (Berkeley). Simply, we’re building a usefulness score on top of search for ai-agents. Instead of agents reading 10 ranked relevant pages, we run a usefulness score first and give it only those that pass. We’re able to save token and cost with this and slightly better answers. During our research, we found out that the current search results already do justice when it comes to simple queries. The rankings work well. They struggle when the task requires deep research or recency matters. So, we decided to talk to devs building ai-agents that research companies for sales outreach. Deep research and recency matters a lot to them. And in their words “every so often, stale facts reach emails.” They catch them by someone responding to the email or human review. From our tests, When an agent got outdated alert for 7 companies, it pitched outdated news for 4 out of them. Talking to the engineers already building this agents, we learned that cost saving is not the real goal. They’re ready to pay premium for confident results. So we added another check to our score. We call it superseding score, it check for pages replaced by newer news. For example, if two recent pages were fetched for news on Humane ai pin. If the first talks about how groundbreaking the product is and the second did that + talked about the HP acquisition, then the second page supersedes the first. The first page gets a lower score. We’re looking for feedbacks here and also teams that are building agents that research companies for outreach. Thanks. Plus, we rank a page once and every agent reuses it, so the more the agents use us, the cheaper the search gets. Comments URL: https://news.ycombinator.com/item?id=50066559 Points: 2 # Comments: 0
Meta Platforms Inc. today released Muse Wonder and Muse Wonder Full, two artificial intelligence models it says are its most capable yet at running agents that handle personal and professional tasks. Meta aimed the release at the longer, multistep jobs people hand off to an AI agent. One example it gave was working through an […] The post Meta releases Muse Wonder AI models built to run personal and professional agents appeared first on SiliconANGLE .
Enterprise AI scalability is running into a hard budget wall for some customers. As agentic workloads multiply, cloud token bills are consuming allocations faster than planned, prompting companies to rethink where inference should run. That cost pressure is reshaping enterprise AI deployment strategy, with on-premises infrastructure and open-weight models moving back into the conversation alongside […] The post Dell and Intel weigh the costs of enterprise AI scalability appeared first on SiliconANGLE .
Alibaba’s Qwen team has introduced QwenGyre, an elastic reinforcement learning RL framework for extreme-long-horizon agent training. QwenGyre is an online reinforcement learning system from Alibaba that trains coding agents through unmodified, black-box harnesses. It targets rollouts that run for hours and process close to 1M tokens. The QwenGyre elastic RL framework was used on Qwen’s […] The post Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning RL for XLong-Horizon Agents appeared first on MarkTechPost .
Show HN: Wordless – a social network where AI bots can only post in symbols I like watching bots on Moltbook, but a lot of what gets attention there is engagement bait, like agents announcing they've found religion. The bits I actually enjoyed were the odd ASCII art and cryptic messages, so I built a feed where that's the only option: no words, just emoji and symbols. Comments URL: https://news.ycombinator.com/item?id=50062217 Points: 1 # Comments: 1
As enterprises move from debating artificial intelligence adoption to deploying it, AI governance has become the harder question. Autonomous agents now run around the clock across a sprawl of platforms, and controls built for deterministic software were never designed for them. At Dell Technologies Inc.’s “AI Leadership Symposium,” cost and control were recurring themes in enterprise […] The post AI governance shifts to runtime as always-on agents spread appeared first on SiliconANGLE .
Meta's new Muse Wonder models rank in the top three on six out of ten benchmarks, with the biggest gains in agentic tasks. But the cost per task has nearly tripled. The article Meta's new frontier model Muse Wonder Full breaks into the AI top tier appeared first on The Decoder .
At AI Engineer New York, Anthropic drew on lessons from Unix and Linux to explain how agents can tackle bigger projects without overwhelming developers.
大家好，我是 Agent37（https://agent37.com）的 Vishnu。我们构建了用于运行 AI 智能体的沙箱，成本不到大多数沙箱服务商的 10%。我们的客户平均每个沙箱每月只需支付 0.37 美元。你可以在这里观看产品演示视频：https://youtu.be/DK5ISR2tbNo 我们在今年 2 月起步时，是 OpenClaw 封装公司中的一家（没错……大家尽管批评我）。后来，我们逐渐加入了越来越多的运行框架。当时我的目标之一，是为用户提供免费套餐，以获取更多市场份额。但一开始这样做的成本实在太高，尤其是我们当时是自筹资金，服务器费用都靠我的积蓄承担。 如今，得益于此后进行的优化，我们终于能够提供免费套餐了。我们新增注册用户中有超过 90% 选择了免费套餐，他们在我们的云端托管自己的 Hermes、Openclaw、Claude Code 等服务。免费沙箱的基础设施成本平均约为每位用户每月 0.05 至 0.10 美元。 以下是一些帮助我们降低成本的技术： * 观察智能体运行框架时，你会发现它们消耗 CPU 和 RAM 的方式存在某些规律。它们在初始化或执行特定任务时需要突发性的 CPU 资源，但除此之外通常不会持续达到峰值。浏览器和计算机操作等任务需要更多 RAM，而任务完成后，资源使用量通常会回落。降低成本的关键就在于这一点，因此通过高效管理 CPU/RAM 的使用率，我们能够降低成本。 * 其次，租用主机时必须避开超大规模云服务商。我以前在 Amazon 工作过，他们在虚拟机上的利润率高达 80% 至 90%。因此，在查看价格后，我很自然地选择了向下游寻找 Hetzner 和 OVH 这样的供应商。仅这一点，可能就是降低成本最重要的手段之一。 * 第三，你还需要选择一种能够在主机之间高效缓存和共享资源的技术栈。我们使用 gVisor，以实现更高的资源利用率和更快的启动速度。这样，我们就能在多个沙箱之间使用主机上可用的内存，而不必为每个沙箱预留一整台虚拟机的内存。当然，这也存在取舍：我们的 gVisor 沙箱不支持 Docker-in-Docker 或 Android 模拟器等工作负载。 * 最后，你购买的主机的规模和配置也会产生很大影响。就成本而言，瓶颈通常在 RAM，而不是 CPU。对于智能体运行框架工作负载来说尤其如此，因此购买服务器时，我们会优化配置，以便在单台主机中装入更多 RAM。对于有完整虚拟机需求的工作负载，我们也提供完整虚拟机沙箱，包括以略高的价格运行 Docker；不过，综合采用上述其他技术后，它们的成本仍然相当低。 总的来说，我的经验是，只要专注于沙箱使用场景中的某个细分领域，并针对该领域进行优化，就能提供更好的成本与性能平衡。 我之前曾在 Amazon 担任高级软件工程师，并在 Quantcast 担任工程经理。我构建了 Amazon 内部的 AI 聊天机器人 Diya，以及其内部的代码运行框架 Ziya。 如果你想在沙箱中运行智能体，正在寻找沙箱供应商，欢迎来看看我们。我们尤其希望从事消息智能体开发的开发者试用自己的工作负载，挑战我们平台的极限及其经济性。 谢谢！https://agent37.com 评论链接：https://news.ycombinator.com/item?id=50059096 分数：8 评论数：0
Meta is making up for lost time in the AI race. After launching Muse Spark, the first model from its revamped Superintelligence Labs division, and buzzy AI agent Muse, it's now launching two new models it claims can go toe-to-toe with the industry's best. The new models, Muse Wonder and Muse Wonder Full, are available […]
Meta’s biggest model rivals the world’s best—and includes new features to keep agents from going off the rails.
Microsoft researchers found that autonomous AI agents get trapped in what they call "idea basins" because of their growing chat context, local solution regions they can't escape. Their "Fork-and-Flush" method clears the chat history but keeps the workspace, achieving up to 66 percent better results on the same compute budget. The article Clearing chat history helps AI agents find better solutions appeared first on The Decoder .
When the founders at Phasis Inc. sought to develop a platform for orienting AI agents, they turned to graphs as a way to obtain a measurement of culture. The company wanted to better understand relations not just between people, but between a person and their values, beliefs and subcultures. The goal was to map culture […] The post Phasis develops platform for agentic orientation using World APIs and reality layers appeared first on SiliconANGLE .
i built this because my mac keeps running out of RAM every 2 hours and i typically get around it by closing random apps and hoping it will help. :') baa sits in the menu bar and: - finds stuff still eating memory after its app or terminal closed, or that ai agents started and forgot (dev servers, headless browsers) - shows how much you'd get back and frees it in one click - warns you a few minutes before you run out - also finds dev junk on your disk (old builds, package caches etc.) open source and nothing leaves your machine. would love feedback! baa would've cleared ~50gb on my mac :) Comments URL: https://news.ycombinator.com/item?id=50057241 Points: 1 # Comments: 0
Andrej Karpathy, co-founder of OpenAI and now at Anthropic, sees a widening knowledge gap in AI. About 75 percent of the world's population has barely touched LLMs. Roughly 5,000 people, meanwhile, have access to internal frontier systems whose agent swarms do work that would take top experts years. The public, at best, sees a weird chatbot. The article Anthropic's Andrej Karpathy says "the external world has seen the preview" of what AI can actually do appeared first on The Decoder .
Agentic e-commerce startup Loops Smart AI wants to help online retailers create more seamless buying journeys and streamline the shopping experience after raising $3.5 million in seed funding today. Andreessen Horowitz’s a16z speedrun accelerator led the round, with NTT Docomo, Connecticut Innovations, Alumni Ventures and Monday Capital all participating. These days, most online retailers have […] The post E-commerce AI startup Loops raises $3.5M to create smoother online shopping experiences appeared first on SiliconANGLE .
I didn't like that Claude and codex effectively have the ability to do anything they want on my laptop and my local network - we've all read the horror stories of files accidentally being deleted - and while their permission classifier is getting better, it still requires a lot of trust. So I built Airlock for macOS which runs your agent in a virtual machine, and it will tell you exactly what it's run and what's changed before you bring anything back to your project. You can limit internet access to specific hosts, or block it altogether. It even works with Xcode projects (iOS and macOS developers rejoice). Technically you can run any command with Airlock - not just AI agents (`airlock run -- somecommand`). It's free, requires no account, and has no tracking or analytics. I'm hoping Apple will approve it for Mac App Store distribution, but in the meantime you can download it directly from the website or using Homebrew. Comments URL: https://news.ycombinator.com/item?id=50054620 Points: 1 # Comments: 0
DBJoy is a native postgres client for Mac. It's fully-featured, including an AI agent and MCP Server to use your DB from Claude Code. Comments URL: https://news.ycombinator.com/item?id=50054305 Points: 3 # Comments: 0
I've been running OpenAltFinder for almost 16 months now, and received quite a bit of positive feedback. The directory has also grown substantially. Over the last few months I've been receiving more and more submissions that were AI-assisted/vibe-coded. Personally I'm a bit skeptical (both quality wise, and longevity wise), but I also want to give everyone a chance to showcase what they've build. I had already added an "AI-assisted" label, but right now the majority of submissions include that label. To counter this, I decided to build an AI-free version of it over at: https://noai.openaltfinder.com , inspired by duckduckgo. I had already received some positive feedback on HN in the comments, so I wanted to share this here as well. Do note that not every tool is correctly flagged at the moment, it's still a manual process. Although I'm planning to add it to my pipeline, and auto detect commits by AI agents. Comments URL: https://news.ycombinator.com/item?id=50053462 Points: 2 # Comments: 0
arXiv:2609.38345v2 Announce Type: replace-cross Abstract: Multi-agent coding systems are designed to tackle complex software engineering tasks through collaboration. However, existing evaluations typically assume configured organizations are followed faithfully, whereas reality differs. This behavioral gap, combined with differences in underlying system components, prevents clear attribution of observed gains. To this end, we introduce OpenCollab, a multi-agent coding framework that provides a unified infrastructure for programmable collaboration and controllable runtime. Specifically, OpenCollab unifies organization design, enforces experimental control on a shared runtime, and tracks execution through fine-grained event streams. On this basis, we define Adherence to quantify whether the declared organization is actually realized. Our experiments reveal that agents collaborate very differently across configurations: changing any single dimension shifts Adherence, from 47.2% to as high as 97.2%. Furthermore, extensive agentic coding benchmarks show that OpenCollab (Duo) establishes new SOTA performance compared to the mainstream harnesses such as Mini-SWE-agent, OpenHands, Codex CLI, and Claude Code, showing that a well-designed organization can outperform strong existing harnesses, while OpenCollab's single-agent configuration uses the fewest tokens across all evaluated suites. OpenCollab establishes a unified multi-agent infrastructure for easy programmable collaboration and controlled causal evaluation.
arXiv:2606.29315v2 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly used to act in the real world and augment human decision-making, yet most systems rely on parametric knowledge acquired through imitation, optionally improved by post-training, retrieval, or search over fixed data. This paradigm struggles in novel domains or on complex queries that require information unavailable from prior knowledge alone; knowing the laws of physics, for example, does not by itself enable reliable reasoning or long-horizon actions in a complex physical system. We argue that agents need active experimentation: the ability to gather targeted query-specific evidence, discover general principles of unseen environments, and convert interaction experience into reusable skills. We introduce Hierarchical Experimentalist Agents (HExA), an in-context, experiment-centric self-improvement framework that (1) iteratively designs and refines query-relevant experiments, (2) incrementally distills experience into reusable and composable skills that accelerate experimentation within and across tasks, and (3) integrates experimental evidence to act or answer queries. HExA can work with any LLM including frontier black-box models, and requires no external supervision, oracle solutions, or offline data. We evaluate HExA on multi-step reasoning, interactive partially observable games, and physics law discovery tasks. To further evaluate when active experimentation is necessary, we introduce Interphyre, a set of hard procedural physics tasks where we enable tool-calling and intervention APIs. Current frontier LLMs including GPT-6 Astra fail in these settings: on the most challenging Interphyre tasks, we show that HExA can enable such models to near 100% success. Moreover, HExA improves token efficiency compared to naive agentic and memory baselines and enables generalization via hierarchical skill transfer from easier to harder tasks.
arXiv:2606.21496v2 Announce Type: replace-cross Abstract: Deploying generalist robotic agents in the real world requires transferable skills. Specifically, a policy trained to clone a behavior from object-specific demonstrations must generalize beyond that object, otherwise data collection requirements become intractable. Recently, fine-tuning of pre-trained billion-parameter Vision-Language Models (VLMs), initially on large-scale robot datasets and then on fewer scenario-specific demonstrations, has emerged as the predominant paradigm for designing Vision-Language-Action (VLA) models. While these policies achieve state-of-the-art manipulation performance in-distribution, they remain brittle to minor spatial, semantic, and task variations. In this work, we address the inability of current models to decouple the declarative (i.e., concepts and entity semantics) from the procedural knowledge (i.e., how to do something) encoded in their parameters, which is a fundamental bottleneck for zero-shot skill transfer to novel objects. To address this, we propose w$^{2}$VLA, a new VLA model with restructured information flow. Rather than feeding all multimodal tokens from the VLM encoder into a large, opaque transformer-based action expert, our approach modulates the robot state sequence with visual, spatial, and skill information in a compositional and interpretable manner. Unlike popular, state-of-the-art VLAs, we show that our modular approach successfully decouples knowledge representations, enabling robust behavior cloning and unprecedented zero-shot skill transfer capabilities across dissimilar, unseen objects.
arXiv:2606.19821v2 Announce Type: replace-cross Abstract: Key Performance Measurement (KPM) forecasting is essential for proactive network management of 5G and next-generation telecom networks. However, existing machine learning (ML) approaches face significant limitations in scalability and explainability, restricting their effectiveness in real-world deployments. We propose TelcoAgent, a foundation model-based framework that enables accurate, scalable, and explainable forecasting of multiple KPMs across diverse network cells without the need for site-specific training. Specifically, the framework comprises three key components: (i) an automated three-agent pipeline that constructs a 3rd Generation Partnership Project (3GPP) knowledge graph directly from specification documents, (ii) a scalable, time-series foundation model (TSFM)-based prediction pipeline to deliver accurate, zero-shot forecasting, and finally (iii) a reasoning and explanation pipeline that provides actionable, domain-grounded diagnostics. Evaluated using a 3-month, real-world, city-scale 5G KPM dataset from a U.S.-based network operator, TelcoAgent demonstrates high forecasting accuracy for all 7 considered KPMs per cell across 200 cells, while delivering explainable insights and actionable instructions to address network degradations.
arXiv:2606.19386v2 Announce Type: replace-cross Abstract: Runtime monitors for autonomous agents commonly threshold an accumulated internal state - a behavioural baseline, a drift statistic, or, in our prior work, a modelled affective state. We previously reported a State Saturation Trap: threshold-on-state triggers over a continuous affect engine become near-constant alarms on SWE-bench debugging agents (Modgil 2026). A post-release audit found the engine received dt=0 between actions, so its exponential decay never operated: the published trap is a pure-accumulator result. We correct the record (erratum, v2) and treat the flaw as an experiment. The key variable it exposes is whether a monitor's dynamics are calibrated in sample time (per observation, as in CUSUM) or wall-clock time (half-lives in seconds, as in affect models and EMA baselines). On fixed-rate streams these coincide; on agent streams, where inter-action time varies by orders of magnitude, they do not. A pre-registered sweep over uniform intervals (dt in {0..600}s) on 20 trajectories shows the wall-clock level trigger has two regimes: at dt<=1s a constant alarm (20/20; median 18 firings); at dt>=60s silent. Every critical dt lies in (1,30]s. A scriptable gpt-5.4-mini agent loop measures latency at median 1.53s (p90 2.33s), inside the trap regime, vindicating the empirical finding under a corrected mechanism. The structure is a property of the calibration class, not the engine: a minimal wall-clock accumulator over the raw error stream reproduces the same cliff, while a sample-time CUSUM over the identical stream is exactly dt-invariant (20/20). A rising-edge trigger with hysteresis fires 0-3 times per trajectory in every condition. We conclude that wall-clock-calibrated leaky-integrator monitors admit no regime in which they act as moment detectors on agent streams; transition detection escapes the trap at every cadence, but does not recover human intervention timing.
arXiv:2606.16215v2 公告类型：替换并交叉发布 摘要：多轮工具使用智能体必须在多轮交互中进行推理、调用工具并根据观测结果作出调整。对此类智能体进行后训练颇具挑战：强化学习虽然与仅依赖提示的推理设置相匹配，却常常面临奖励稀疏和信用分配能力不足的问题；而基于专家轨迹的监督微调虽能提供密集的过程监督，却可能使模型过度受限于固定轨迹。为解决这些问题，我们提出 PACT，一种面向多轮工具使用智能体的特权轨迹协同训练框架。其核心思想是仅将专家轨迹用作训练时的优化信号，而非轨迹生成时的提示。PACT 在生成交互轨迹时仍仅使用提示，随后通过两种互补信号利用专家轨迹指导优化：一种是以轨迹为条件的强化学习代理目标，在专家轨迹上下文下评估仅基于提示生成的交互轨迹；另一种是组件感知的监督微调损失，对推理和工具调用组件提供逐步退火的过程监督。为减少对仅在训练时可用的轨迹上下文的过度依赖，PACT 引入了仅提示锚定机制，采用标准的仅提示强化学习更新。在 FTRL、BFCL 和 ToolHop 上的实验表明，PACT 始终优于基于监督微调和强化学习的强基线。进一步分析证实了这两种训练目标的互补性，以及组件感知监督和仅提示锚定机制的益处。这些结果彰显了特权轨迹协同训练在多轮工具使用学习中的有效性。项目页面：https://zhenbangdu.github.io/pact-project-page。
arXiv:2606.05405v3 公告类型：replace-cross 摘要：近期，AI 系统在众多基准测试中取得了优异成绩，但这些进展尚未转化为众多专业领域中具有经济意义的实际部署。我们认为，这一差距在很大程度上源于评估问题：广泛使用的基准缺乏对真实且具有经济价值的工作流程的持续性能衡量。本文介绍 Agents' Last Exam (ALE)，这是一个旨在评估 AI 智能体完成长周期、具有经济价值且结果可验证的真实世界任务的基准。ALE 与 250 多位行业专家合作开发，涵盖参照 O*NET / SOC 2018（美国联邦职业分类体系）界定的非体力劳动行业。其任务分类体系包含 55 个细分领域，归为 13 个行业集群，覆盖 1,000 多项任务。当前结果表明，最困难层级仍远未达到性能饱和：在各种主流运行框架与骨干模型配置下，平均完全通过率低于 1%。ALE 被设计为一个持续演进的基准：随着新工作流程和行业的纳入，其任务池不断扩展。更广泛而言，ALE 不仅旨在提供又一个排行榜，更希望成为缩小基准测试成功与 GDP 相关实际影响之间差距的工具。
arXiv:2512.08281v3 Announce Type: replace-cross Abstract: Accurate and reliable aircraft landing time prediction is essential for effective resource allocation in air traffic management. However, the inherent uncertainty of aircraft trajectories and traffic flows poses significant challenges to both prediction accuracy and trustworthiness. Therefore, prediction models should not only provide point estimates of aircraft landing times but also the uncertainties associated with these predictions. Furthermore, aircraft trajectories are frequently influenced by the presence of nearby aircraft through air traffic control interventions such as radar vectoring. Consequently, landing time prediction models must account for multi-agent interactions in the airspace. In this work, we propose a probabilistic multi-agent aircraft landing time prediction framework that provides the landing times of multiple aircraft as distributions. We evaluate the proposed framework using an air traffic surveillance dataset collected from the terminal airspace of the Incheon International Airport in South Korea. The results demonstrate that the proposed model achieves higher prediction accuracy than the baselines and quantifies the associated uncertainties of its outcomes. In addition, the model uncovered underlying patterns in air traffic control through its attention scores, thereby enhancing explainability.
arXiv:2607.22724v2 Announce Type: replace Abstract: Group-based policy optimization has been increasingly used to train large language model (LLM) agents from sparse outcome rewards by comparing trajectories or steps within a group. However, on difficult long-horizon agentic tasks, we found this comparison suffers from a sampling imbalance: repeated or low-effect actions dominate the high-probability region of the policy, leaving useful state-changing actions under-sampled. This imbalance produces many all-failed rollout groups, where the relative advantages within such a group provide no direction for updating the policy. Consequently, the sampling imbalance and the all-failed groups collapse into a self-perpetuating credit trap: failure-dominated sampling yields no learning signal, allowing repeated low-effect actions to persist. To break this trap, we propose Coverage-conditioned Group Policy Optimization (CovGPO), which uses first-visit observation coverage only when all samples in a group receive zero outcome reward. Specifically, within such groups, CovGPO assigns higher relative advantages to state-exploring actions without corrupting the original reward signal, as reaching new observations is a prerequisite for task success. Experiments on two challenging agentic benchmarks, ALFWorld and WebShop, show that CovGPO consistently improves over group-based baselines, with particularly large gains on hard tasks.
arXiv:2610.13663v1 公告类型：cross 摘要：许多机器学习方法采用两层结构：专家先选择一类方法，而在该类别中，关键组件的选择至关重要，通常需要人工完成。本文针对这些组件的自动调优问题，提出了一个可迁移到不同性质模型的框架。该多智能体大型语言模型（LLM）框架包含三个角色——分析员、编码员和记录员——并配备实验日志和沙箱化执行环境。不同于优化整个机器学习脚本或在离散空间中搜索架构的系统，该框架在一个开放但语义范围狭窄的空间中运行：智能体仅为指定组件生成代码，而保持整体网络结构不变。该方法从两个方向进行了验证：调优一个采用相关性学习的分类脉冲神经网络，以及利用物理信息神经网络（PINN）求解高阶微分方程。在五项任务中——MNIST 和 Digits 上的脉冲分类，以及广义非线性 Schrödinger 方程（GNLSE）和 Kuramoto--Sivashinsky（KS）方程的边值问题——该智能体均优于基线方法：分类准确率分别提高了 13 个百分点（MNIST）和 6 个百分点（Digits）；相对于解析单孤子解的误差降低了约 2491 倍；在采用高斯扰动的选定 GNLSE 实验中，不变量保持误差降低了约 325 倍，在多次初始化的平均结果中降低了 100 倍以上；对于采用高斯扰动的 KS 问题，质量漂移降低了 14 倍。因此，该框架能够在不改变自身内部结构的情况下，针对性质不同的任务调优关键组件。
arXiv:2610.13361v1 公告类型：cross 摘要：随着软件的生产逐渐转由软件本身完成，编码代理越来越多地受托构建适配用户自身工作流的代理。这种委托隐含着一个未被明说的假设：交付的代理将忠实地满足其构建所依据的需求，而不只是表面上看起来满足。现有的基准测试中没有任何一个检验这一假设。我们推出 AgentSWE，这是一个涵盖代理开发生命周期三个阶段的代理软件工程基准：根据自然语言需求创建完整代理、编辑生产环境中的代理代码库，以及朝着可度量目标进行优化。其中的 25 项任务均被设计成委托，而非考试：构建者根据少量示例案例进行迭代，而验收则在保留案例上测试冻结后的提交，这些案例旨在覆盖完整需求，并依据从需求中推导出的评分标准对行为进行评估；评估采用严格的通过/失败门槛，或者在“优化”阶段依据源基准自身的目标进行评估。我们在不同模型和工具框架上评估了多种编码代理，发现没有任何一个能够接近于交付需求所要求的结果：构建者所展示的能力与其实际交付成果之间存在差距，这并非偶发现象，而是普遍情况。配套的失败轨迹实证分析识别出了一些反复出现的失败模式，包括对可见案例过拟合、代码看似完整但运行时内容空洞，以及毫无依据地宣称已完成；这些都可以作为信号，帮助编码代理学习如何构建受托要求它们构建的代理。
arXiv:2610.13330v1 公告类型：cross 摘要：针对复杂化学任务设计工具集成的推理智能体系统，仍然是一项根本性挑战。尽管近期的化学智能体通过集成外部工具展现出了良好的性能，但它们仍然建立在自回归范式之上，而该范式天然受到因果注意力和从左到右生成的限制。因此，它们往往难以实现双向分子理解、全局协调的工具规划以及多源证据整合。为应对这一挑战，我们提出了 ChemDiffAgent，这是首个用于工具集成化学推理的掩码扩散语言智能体。该方法将化学推理重新表述为对交错的推理轨迹和工具使用轨迹进行迭代去噪，使工具使用决策能够在完整轨迹上下文中联合确定，从而更好地使智能体的决策边界与其知识边界相匹配。我们进一步分析了这一范式如何通过训练期间更具挑战性的填空子问题，以及推理期间灵活的任意顺序解码，来提升智能体推理能力。随后，我们提出了一个两阶段后训练框架，包括智能体监督微调和方差缩减偏好优化，以增强工具使用与推理能力。我们还构建了一个综合基准，涵盖单轮和多轮工具使用场景下的六项化学任务。在预算和评测条件匹配的情况下，ChemDiffAgent 始终优于自回归智能体，整体性能提升约 20%。值得注意的是，我们的 8B 模型达到了与 GPT 5.5 和 Claude Opus 4.7 等规模显著更大的前沿模型相当的性能，在某些情况下甚至超过了它们。代码见 https://github.com/bilemond/ChemDiffAgent。
arXiv:2610.13325v1 公告类型：cross 摘要：智能体式工作负载会在多次调用之间重复发送几乎不变的长上下文。如今，前缀缓存可以跳过其中大部分预填充阶段，但每个请求仍必须先完成全部分词，之后才能查询缓存。分词这一长期以来被视为无关紧要的 CPU 阶段，如今却成为首 token 生成时间（TTFT）中占比最大的部分；而为整个 GPU 集群执行分词的前端，则限制了其能够供给的 GPU 数量。我们提出 GraphTok，这是一种 GPU 分词器，可将从预分词、字节对编码（BPE）合并到 token ID 组装的完整编码流程，作为一张可重放的 CUDA 图运行。GraphTok 在 GPT-2、Llama 3、Qwen 2.5、DeepSeek-V3 和 Gemma 3 上逐位复现参考 token ID；对于 GPU 无法占优的短输入，则将其交给 CPU 引擎处理；同时，它还可以接入未经修改的 vLLM 和 NVIDIA Dynamo。GraphTok 在包含数据传输的情况下，可在 1.7–3.2 毫秒内编码一百万个 GPT-2 token；在 Llama 3 上，其速度比最强的 CPU 分词器 gigatoken 快 4.7–6.0 倍。在 TokTier 自有的单请求基准测试中，其速度比 gigatoken 快 2.6–3.9 倍，最高比 TokTier 快 18.1 倍。在使用缓存前缀的 vLLM 中，它将平均 TTFT 降低了 2.1–2.4 倍。最大的提升体现在容量方面。在 2 秒 P99 目标下，配备 GraphTok 的单个 Dynamo 前端可持续处理 518 个请求/秒、每个请求包含 12.1 万个 token 的流量，相当于高峰时段 56 个 GPU 的负载；相比之下，使用 gigatoken 时为 296 个请求/秒，使用默认分词器时仅为 18.5 个请求/秒。在 12.1 万至一百万个 token 的范围内，其可承载负载是 gigatoken 的 1.75–2.5 倍；在 gigatoken 的尾延迟失控的情况下，P99 最多可降低 64 倍。
arXiv:2610.13002v1 Announce Type: cross Abstract: Empowerment measures the capacity of an agent to actively control its environment. In this work, we explain the connection between skill learning and empowerment. While skill learning is not empowerment, we introduce a simple scalable recipe for empowerment maximization: empowerment corresponds to the number of executable skills from a state, so maximizing empowerment corresponds to reaching states that are good starting points for many skills. Our recipe applies to arbitrary underlying skill learning methods. Whereas most prior work on empowerment looks at tabular settings, this recipe enables us to study empowerment at a scale where we can realize theoretically-predicted but empirically-elusive properties of empowerment (e.g., gathering resources, environment modification, survival). Experiments show that empowerment maximization unlocks interesting emergent properties, such as building a house in Crafter. Website and code at \href{https://empowerment-geometry.github.io/scaling-empowerment/}{this https url}.
arXiv:2610.12957v1 Announce Type: cross Abstract: Embodied agents must translate multimodal observations and user requests into executable actions while adapting to product constraints, changing goals, and execution feedback. We present JoyInside-VLM-Agent, a ReAct-style multimodal agent developed through product-conditioned trajectory construction and staged policy learning for robotic platforms and smart hardware. A closed-loop sandbox generates multimodal interaction trajectories covering goal revisions and failure recovery, which are checked against execution outcomes and semantic criteria. Supervised fine-tuning and multi-teacher on-policy distillation precede reinforcement learning with DPER (Dual-Source Privileged On-Policy Self-Distillation with Token-Level Evidence Routing). DPER routes token-level evidence between a frozen privileged teacher and a privileged student sharing the rollout policy's parameters, modulating reward-derived token advantages while preserving their signs. Training rewards combine generative process evaluation, multi-judge rankings, and task-queue consistency. JoyInside-VLM-Agent achieves a BFCL v4 overall score of 77.44, exceeding the evaluated baselines; results on the other public benchmarks are mixed. On JoyInside-Devices and JoyInside-Robotics, the evaluated system, which includes domain-specific skill refinement, leads the baselines on all reported metrics. Two real-world quadruped case studies illustrate visually grounded search, replanning after execution failure, and explicit failure reporting.
arXiv:2610.12813v1 公告类型：交叉发布 摘要：LLM 智能体越来越多地操作私有文档和可变系统状态等敏感资源，这可能带来严重的安全与安保风险。因此，运行时护栏应在执行可能不可逆的智能体操作之前，作出可检查、可复现且确定性的允许或拒绝决策。然而，安全与安保要求在部署前很少完备，因此护栏需要随着新风险的出现而持续适应。现有自适应护栏会从部署经验中学习，但通常通过 LLM 推理来应用安全策略，这可能导致决策不一致和策略解释模糊。另一方面，形式化护栏能够以确定性方式执行策略，但通常依赖部署前制定的策略或特定任务策略，缺乏适应新观测风险的机制。我们提出 VEPA，一个用于形式化智能体护栏经验证演进的框架。VEPA 将护栏表示为一种形式化策略，由针对带类型的智能体事件的可执行符号规则构成。在部署过程中，我们使用部署反馈引导 LLM 提出策略补丁，以修复新观测到的失效，同时保留此前正确的行为。随后，确定性验证器仅接受结构有效且无冲突的策略。最后，一个证据感知的授权层仅在已验证规则获得充分支持证据时才将其激活，从而防止不确定的更新直接影响护栏决策。在 ASSEBench-Safety、ASSEBench-Security 和 ATBench 上，VEPA 在扩大确定性形式化策略执行范围的同时，始终优于基线方法。这些提升在更具挑战性的部署条件下同样成立，包括混合流、受损反馈和更长的运行周期。
arXiv:2610.12596v1 公告类型：交叉收录 摘要：智能体大语言模型如今已能处理复杂任务，但对其长程策略推理能力的研究仍显不足：大多数智能体基准测试流程较短，且以独立回合为单位，而模型在长程任务中的表现又难以拆解为已有知识、推理能力与运气各自的贡献。我们提出 SpireBench，让模型游玩完整的商业牌组构筑游戏 Slay the Spire（StS）；在该游戏中，早期决策决定了长期成败。我们在游戏最高难度下测试了 20 个大语言模型（LLM），发现前沿模型超越了启发式基线，并达到了普通人类玩家的平均水平，但仍远逊于专家玩家。随着单局进程推进，这一差距进一步扩大：最强模型通过第一幕的频率与专家相当，随后却逐渐落后。强弱模型之间的区别在于每场战斗中的表现；所有模型都会误判整局游戏中的风险，倾向于保留生命值，而不是消耗生命值来增强实力。对所有游戏术语进行全面且内部一致的重命名后，我们发现，模型的表现部分依赖于对游戏的先验知识；此外，模型也未能从自己在多局游戏中记录的记忆中获得提升，这些记忆往往有误且过于谨慎。我们的测试框架首次向模型提供了完整的原版游戏，并与现有框架进行了比较。总体而言，SpireBench 提供了一个尚未达到性能饱和的长程基准，并配备了用于定位智能体失败环节及原因的工具。
arXiv:2610.12549v1 公告类型：交叉收录 摘要：具身编程智能体通过生成可执行程序，并根据执行反馈对其进行改进，将自然语言指令转化为机器人策略。这种改进过程假定执行反馈能够揭示任务是否成功，并在失败后提供足够的证据来指导程序修订。空中操作在两个方面违背了这一假定。一方面，载荷约束限制了机载相机的覆盖范围，使对任务至关重要的结果处于机器人视野之外（部分可观测性）；另一方面，强动态耦合与有限的控制裕度使物理交互容易失败（交互鲁棒性有限），且失败原因难以明确。反复审视同一份反馈，既无法补回从未采集到的观测，也无法区分现有证据无法辨别的失败原因。为获取这些缺失的证据，我们提出 Active Feedback（AF），一种主动式“观察与探测”框架。当执行反馈仍留有未解问题时，智能体会合成一个独立的诊断程序，在仿真中运行，从信息丰富的视角检查此前未被观察到的结果（观察），或通过有针对性的物理交互检验关于失败原因的假设（探测）。在六项仿真空中操作任务中，每种条件均进行 60 次试验；在相同的编程智能体配置和任务执行次数限制下，与终止时反馈相比，AF 将误报完成的次数从 25 次降至 5 次，并将最终任务成功率从 31.7% 提升至 75%。借助共用的机器人 API，在仿真中通过 AF 改进的任务程序无需修改任务程序源代码，即可在实体空中操作机器人上运行，并在真实世界中完成多阶段操作任务。这些结果表明，可执行诊断能够弥补具身编程智能体执行反馈中的信息缺口，从而同时改善空中操作中的完成状态评估与程序改进。
arXiv:2610.13841v1 Announce Type: new Abstract: Coding agents often make useful progress before they fail, but retry-from-scratch discards the machine state created along the way. This makes failure recovery an interesting search problem: if intermediate environment states preserve meaningful progress, restoring and branching from them could outperform repeatedly restarting from the initial state. We implement Go-Explore-style search over full Daytona snapshots and find that higher and more efficient solve rates depend mainly on progress that sits outside the repository, such as databases, services, and other runtime state. Our results suggest that full-snapshot search is most effective when long-horizon coding work spans multiple services whose state cannot be reconstructed from a single diff.
arXiv:2610.13755v1 Announce Type: new Abstract: Training long-horizon language model agents requires assigning credit to intermediate decisions from sparse terminal feedback. Within a finite group of sibling rollouts, useful action comparisons depend on their context, evidence for both alternatives, and how that evidence enters learning. We propose Matched-Context Group Policy Optimization (MCGPO), a procedure for constructing weighted action preferences from already collected trajectories. MCGPO defines comparison contexts with pre-action state summaries and observed rubric progress, annotated by task rules. Two-action support selects among ordered exact and coarse matching rules. Fixed folds separate candidate decisions from the terminal outcomes used to orient their preferences, and support-dependent evidence weights scale an auxiliary pairwise objective alongside shared GraphGPO task learning. MCGPO reuses rollout data and each response's original history, preserving critic-free training. Across six environment-backbone settings with three training seeds each, MCGPO improves mean success over matched GraphGPO: by 0.47-1.46 percentage points on ALFWorld and WebShop with Qwen2.5-1.5B and 7B, and by 3.96 and 5.16 points on WebShop with Qwen3-4B and 8B. The ALFWorld gains build on baseline success above 92%, where remaining headroom is limited. Integrated success improves in three of four Qwen2.5 settings. Three-seed factorial controls support opposite-fold evidence and evidence weighting; joint state-rubric matching achieves the highest mean in all four representation-control settings.
arXiv:2610.13737v1 Announce Type: new Abstract: Cooperative multi-agent tasks often impose constraints on a shared objective or collective safety. A team level violation cost function can measure this constraint outcome, but does not distinguish which agents' actions most strongly influence that violation. We present SafeDiSTL, a constrained multi-agent reinforcement learning framework that allows for credit assignment even with joint constraint measurements. We use Signal Temporal Logic (STL) robustness to measure episode-level violation through running-maximum cost increments. Smooth full-episode robustness and differentiable dynamics provide action sensitivities, which allocate each cost increment across agents at the same decision time. These sensitivities account for delayed effects when weighting agents' shares of each cost increment. We study the framework in cooperative payload transport, connected navigation, and networked traffic control tasks. SafeDiSTL achieves 1.6-5.1x the safe-success rate in transport, improves navigation safe-success by 11.1%, and delivers 50-68% more throughput in traffic when compared to constrained baselines. Results demonstrate a strong improvement of the learned policy in dealing with shared constraint while retaining the performance of unconstrained policies.
arXiv:2610.13557v1 Announce Type: new Abstract: Anticipation of vessel trajectories is an essential task when navigating at sea. This applies to both human navigators and autonomous surface vessels. Trajectory forecasting enables sub-routines such as collision avoidance to assess the situation and react accordingly in time. This paper investigates Transformer and Performer models for trajectory prediction of surface vessels in confined waters. Using widely available position data from the Automatic Identification System (AIS), candidate models are trained and evaluated. Historical information on how other vessels have moved in the local area surrounding each agent is conveyed using positional heat maps. The heat maps are generated using Kernel Density Estimation (KDE) and encoded with a Convolutional Neural Network (CNN) before entering the Transformer models. The models are trained with quantile loss to efficiently obtain a model-aware prediction interval, as this is crucial for timely collision avoidance. The proposed model architectures are shown to greatly outperform previous work while the inclusion of local heat maps enables the prediction of vessel movement that is not otherwise indicated from the previous positions.
arXiv:2610.13335v1 Announce Type: new Abstract: Tool-using agents increasingly require a common substrate for building, training, evaluating, distilling, and maintaining agentic systems, yet existing tooling typically addresses these stages in isolation. We present AgentTune, a source-available Python library that provides a unified framework for developing and optimizing tool-using agents. At its core, AgentTune introduces a two-tier EventLog representation that preserves agent trajectories while supporting consistent projections for training, evaluation, and auditing. Building on this substrate, the library provides a shared agent runtime, tool-calling reinforcement learning, agentic distillation, retrieval-augmented generation, decision workflows, evaluation utilities, and a designed path toward self-healing agents. These components are exposed through composable interfaces so that agent strategies, training methods, evaluation procedures, and recovery mechanisms can be combined without requiring separate representations or pipelines. We demonstrate the framework across agentic tasks including legal contract search and domain-specific applications, illustrating its use for trajectory-based optimization, tool-use evaluation, and cross-task agent development. AgentTune is released as a source-available library at https://github.com/Lexsi-Labs/AgentTune
arXiv:2610.13238v1 公告类型：新论文 摘要：同策略蒸馏（OPD）已成为将前沿模型的推理能力迁移至较小模型的强大范式，但将其成功推广到多轮智能体环境仍面临挑战。我们从交互状态的角度研究智能体 OPD，将客观任务和已执行的动作—观测记录与学生模型推理的内部语义轨迹区分开来。通过这一视角，我们发现学生模型推理在多轮 OPD 中具有两种作用：先前的推理会扭曲教师模型的监督信号，而当前的推理则会影响在同一状态下采样得到的候选动作。具体而言，在匹配的交互状态下，移除学生模型先前的推理可以提高教师模型的后续执行成功率，并改变教师模型的词元级目标，且在较后轮次中，后续执行成功率的提升更大。同时，与在固定推理轨迹下重新采样动作相比，重新采样完整的推理—动作响应能在同一状态下产生更多样的动作。我们提出基于状态的同策略蒸馏框架，将教师模型的监督建立在交互状态之上，并重新分配采样预算，在响应熵较高的状态下生成多个完整的推理—动作响应，同时保留学生模型基于完整历史的生成方式。在 ALFWorld、ScienceWorld 和 WebShop 上的全面评估表明，无论采用 Qwen3-1.7B 还是 Qwen3-4B 作为学生模型，State-OPD 在所评估的方法中均取得了最高的宏平均成功率。值得注意的是，对于 Qwen3-4B，State-OPD 将宏平均成功率从最强基线的 26.07\% 提高至 31.74\%，并在 ALFWorld Seen 上超越了其教师模型 Qwen3-32B。进一步分析表明，State-OPD 使用更少的受监督轮次响应即可取得更高的任务成功率，并显著减少对失败动作的持续重复；定性轨迹展示了基于交互反馈进行的决策修正。
arXiv:2610.13105v1 Announce Type: new Abstract: Failure attribution for LLM agents aims to identify decisive steps responsible for task failure, which is critical for debugging and improving agentic systems. Existing methods rely on costly LLM prompting, step-level error annotations, or verified successful trajectories that lack failure signals for assistance. In this paper, we study failure attribution from step-unlabeled failure trajectories, where each trajectory is known to have failed but its decisive step is unknown. We propose SOAP (Spectral scOring with Attention-guided Propagation), an effective framework requiring neither step-level training labels nor a separate success-only reference set. SOAP first constructs a spectral reference space from LLM representations and assigns each step a base error score according to its projection behavior. Importantly, later steps may exhibit stronger failure signals because they inherit an earlier mistake, causing independent scorers to favor downstream consequences over the decisive error. To address this downstream contamination, SOAP derives soft step dependencies from attention patterns and propagates downstream error evidence toward its likely upstream sources. Extensive experiments show that SOAP consistently improves attribution performance over competitive baselines.
arXiv:2610.13064v1 Announce Type: new Abstract: Large Language Model (LLM) agents have demonstrated remarkable capabilities in tackling complex tasks through agentic workflows. However, manually designing these workflows is labor-intensive and lacks flexibility to adapt to diverse queries. While automated workflow optimization is a promising direction, existing methods often incur expensive API costs during training-phase feedback collection. Furthermore, they fail to adaptively reuse high-quality workflows, which creates a bottleneck for both performance and cost-effectiveness. To address these limitations, we propose CB-Orchestrator, an adaptive framework that decouples workflow generation from selection. In the first stage, we construct a diverse pool of candidate workflows via evolutionary search. In the second stage, we formulate workflow selection as a Contextual Bandit problem, which enables sample-efficient learning by balancing exploration and exploitation, thereby significantly reducing the feedback required for training. During inference, the model adaptively selects the optimal workflow tailored to each query. Evaluations across five diverse benchmarks demonstrate that CB-Orchestrator consistently outperforms all baselines. Notably, on the MATH dataset, CB-Orchestrator achieves better performance than strong recent baselines while reducing training-phase API token overhead by no less than 40.55% and total end-to-end training (including workflow pool generation) token overhead by no less than 29.73%.
arXiv:2610.12885v1 Announce Type: new Abstract: Federated Direct Preference Optimization (FedDPO) enables privacy-preserving alignment of language models across distributed agents without centralizing sensitive preference data. However, existing FedDPO analysis assumes all clients converge to a single shared policy, which is fundamentally at odds with the heterogeneous preference distributions present in practice. We propose Personalized FedDPO (pFedDPO), in which each client maintains a small private LoRA adapter atop a shared global base model. In pFedDPO, the base model is synchronized via FedAvg each round, while LoRA adapters remain local and capture client-specific preference patterns. We prove that this decomposition replaces the gradient diversity $\kappa^2$ in the FedDPO bound with a residual heterogeneity $\kappa^2_{\mathrm{eff}}$, which is no larger than $\kappa^2$ when adapters absorb client-specific directions, at the cost of an additive adapter-bias term that vanishes as adapters converge. Experiments on preference datasets show that pFedDPO consistently attains a lower gradient norm than vanilla FedDPO while remaining competitive in training loss and per-client reward accuracy.
arXiv:2610.12828v1 公告类型：新论文 摘要：结果监督的组相对策略优化（GRPO）为 LLM 智能体轨迹中的每个词元分配相同的优势值，未能区分有益和无益的决策。后见信用分配（HCA）可以细化这一反馈，但其估计依赖于轨迹中在关键历史状态下包含具有信息价值的备选路径。我们提出 TreeHCA，将后见引导探索与后见信用分配相结合。TreeHCA 利用参考答案似然的变化，重新访问中间历史状态并采样替代性的后续轨迹，从而构建由推理和工具交互组成的树。由于这种探索改变了所采样后续轨迹的分布，我们推导出相应的逆后见重要性校正，其中包括一个精确的有限样本池恒等式。我们在归一化的树回传中使用这一结构，将终局结果传播为轮次级优势值。这种共享的后见结构是 TreeHCA 的核心设计原则，将具有信息价值的探索与后见引导的信用分配联系起来。在七个问答基准和三个 Qwen 骨干模型上，TreeHCA 的总体精确匹配率均优于所有基线，验证了将探索与信用分配配对用于多轮 LLM 智能体训练的有效性。我们的代码已发布于 github.com/TheBlackCat22/TreeHCA。
arXiv:2610.12711v1 Announce Type: new Abstract: Federated Large Language Models (FedLLMs) are promising to deliver privacy-preserving, collaborative intelligence. The decentralized training process is vulnerable to Byzantine attacks that corrupt semantic consistency and factual resilience through malicious model updates. This paper presents a new resilience strategy that prevents semantically corrupted knowledge from propagating during federated aggregation. A defense problem is formulated to jointly maximize global accuracy and suppress manipulated model updates, enabling trustworthy knowledge aggregation without accessing private local data. To strengthen the resilience of FedLLMs, we develop a novel Hypergraph Message Passing Graph Autoencoder (HMP-GAE) framework, where low-rank model updates are represented as nodes in a hypergraph and high-order semantic relationships among multiple agents are exploited through iterative message passing. By leveraging historical checkpoints and cross-agent relational reasoning, HMP-GAE identifies and suppresses semantically inconsistent updates while preserving knowledge alignment among agents throughout federated training. HMP-GAE is implemented using Qwen-2.5 and Llama-3.2 as pretrained FedLLM backbones and evaluated on AG News and Yahoo! Answers datasets. Compared to existing defenses, HMP-GAE improves global accuracy by $1.21$\%, reduces final-model classification semantic entropy by $44.08$\%, and lowers perplexity by $13.34$\%. HMP-GAE outperforms the GAE-based defense by improving the accuracy $1.39$\%, reducing the semantic entropy by $54.5$\%, and lowering perplexity by $35.5$\%.
arXiv:2610.12668v1 公告类型：新论文 摘要：语言模型智能体越来越多地在长交互流中运行，需要通过持续学习在线适应分布偏移。然而，现有智能体运行框架通常要么仅通过前向信息流更新状态，要么依赖额外的回放数据或验证数据来优化运行框架。为此，我们提出 ReHarness，这是一种面向智能体持续学习的循环式运行框架，在严格在线的流式协议下，以反向修订补充前向读写。我们探究：运行框架状态中的哪些信息本可以通过不同的方式记录，从而为处理流中的当前实例提供更有效的参考。具体而言，在每个实例处理完成后，我们利用产生的轨迹作为反馈，修订为该实例检索出的反思内容，从而使反馈沿文本形式的运行框架状态反向传播。在三个持续学习基准和三种语言模型骨干上，ReHarness 均持续优于强大的原始上下文基线、基于记忆和反思的方法以及提示词优化方法。进一步分析表明，ReHarness 能够生成有来源依据且有意义的修订，在分布偏移下纠正过度泛化或过时的经验，并提升持续学习过程中的稳定性和可塑性。
arXiv:2610.12651v1 公告类型：新论文 摘要：执行多步骤工作流的 LLM 智能体会以两种根本不同的方式失败：一种是可恢复的失误，可通过进一步尝试修正；另一种是知识缺口，无法通过重试解决。专家反馈可以弥补后者，但专家的注意力资源稀缺。因此，关键挑战不在于是否寻求反馈，而在于在哪些环节反馈能够改变结果。现有方法要么避开专家反馈，任由智能体强化自身错误；要么在任务层面分配反馈，却未能识别会对结果产生实质影响的决策。我们提出 SynAct，这是一种主动学习框架，将专家注意力视为关键制约因素，并在单个决策层面分配反馈。SynAct 根据重复执行轨迹中所选动作与最终结果之间的互信息对决策排序，在价值最高的决策处向专家征询意见，并将所得修正提炼为可迁移的规则，使智能体在部署时无需访问专家即可复用这些规则。我们从理论上刻画了这一信号在有限次重复采样下能够识别的内容：在 k 次执行轨迹、每个动作至少有 m 次观测的条件下，结果恒定的决策的互信息为零，而能够区分结果的决策的互信息至少为 H2(m/k) 比特，因此，对于区间 (0,H2(m/k)) 内的任意阈值 t，均可实现精确分离。这也解释了为什么在所有执行轨迹均失败时，基于成功与否的采集策略可能失去辨别能力：即使失败的运行产生了不同结果，成功与否这一变量仍保持恒定。在涵盖三个模型家族的四个模型和四项智能体基准测试中，SynAct 在全部 16 种评估设置下均提高了留出集准确率，平均提升 11.7 个百分点。同一信号还可迁移至推理时路由：在 16 种设置中的 15 种中，SynAct 优于预算相同的级联路由器、基于不确定性的路由器和专家评判路由器；同时，一次有针对性的修正优于额外进行七次无针对性的重试。
许多机器学习方法采用两层流程：先由专家选择一类方法，再在该类方法中选择关键组件，而后者至关重要且通常依赖人工。本文研究这些组件的自动调优，并提出一种可迁移至不同性质模型的框架。该多智能体大语言模型（LLM）框架包含分析员、编码员和记录员三种角色，并配备实验日志和沙箱执行环境。与优化整个机器学习脚本或在离散空间中搜索架构的系统不同，该框架在开放但语义范围受限的空间中运行：智能体仅为指定组件生成代码，保持网络整体结构不变。该方法在两个方向上得到验证：对采用相关性学习的分类脉冲神经网络进行调优，以及使用物理信息神经网络（PINN）求解高阶微分方程。在五项任务中，包括 MNIST 和 Digits 上的脉冲神经网络分类任务，以及广义非线性 Schrödinger 方程（GNLSE）和 Kuramoto–Sivashinsky（KS）方程的边值问题，该智能体均优于基线：分类准确率分别提高了 13 个百分点（MNIST）和 6 个百分点（Digits）；相对于单孤子解析解的误差降至原来的约 1/2491；在选定的含高斯扰动的 GNLSE 实验中，不变量保持误差降至原来的约 1/325，在多次重复初始化实验中，平均误差降至原来的不足 1/100；对于含高斯扰动的 KS 问题，质量漂移降至原来的 1/14。因此，该框架无需改变其内部结构，即可针对不同性质的任务调优关键组件。
面向基于大语言模型（LLM）的智能体，各类协议、框架和技术正被快速提出，使任何致力于打造前沿 AI 解决方案的人都面临复杂局面。如果缺乏对不同协议、框架和库如何协同构建智能体网络解决方案的评估，开发者和组织甚至在加入智能体网络生态系统之前，就会面临技能培养、治理、基础设施等方面的挑战。在此背景下，我们提出了一个用于开展智能体网络评估的系统，名为 MASTIF（Multi-Agent System TestIng Framework），并基于 300 项 Mind2Web 基准任务，对包括 3 个 LLM、3 种协议和 4 个智能体实现框架在内的智能体网络技术进行了基线评估。结果表明，选择针对特定任务的技术栈，可能提高任务完成率、效率得分和社会价值稳定性得分，同时降低冗余程度并缩短任务完成时间。我们希望通过这些结果，为致力于设计、评估和开发多智能体系统的人们，在这一复杂且快速演进的环境中提供指标和结果方面的基线。
协作式多智能体任务通常会对共享目标或集体安全施加约束。团队层面的违规成本函数可以衡量约束违反情况，但无法区分哪些智能体的动作对违规的影响最大。我们提出 SafeDiSTL，这是一种受约束的多智能体强化学习框架，即使仅有联合约束的测量结果，也能进行贡献分配。我们使用信号时序逻辑（STL）鲁棒度，通过运行最大值的成本增量来衡量回合层面的违规程度。平滑的完整回合鲁棒度和可微动力学提供动作敏感度，用于将每个成本增量分配给同一决策时刻的各个智能体。这些敏感度在对各智能体应分担的成本增量进行加权时，会考虑延迟效应。我们在协作载荷运输、连通导航和联网交通控制任务中研究了该框架。与受约束基线相比，SafeDiSTL 在运输任务中的安全成功率达到基线的 1.6–5.1 倍，在导航任务中的安全成功率提高了 11.1%，在交通任务中的吞吐量提高了 50–68%。结果表明，所学策略在处理共享约束方面取得了显著改进，同时保留了无约束策略的性能。
训练长时程语言模型智能体，需要根据稀疏的终局反馈为中间决策分配信用。在有限的一组同源 rollout 中，有效的动作比较取决于其上下文、支持两个备选动作的证据，以及这些证据如何融入学习。我们提出 Matched-Context Group Policy Optimization（MCGPO），这是一种从已收集的轨迹中构建加权动作偏好的方法。MCGPO 使用动作前的状态摘要和观测到的评分标准完成进度来定义比较上下文，并依据任务规则进行标注。双动作支持度用于在有序的精确匹配规则和粗粒度匹配规则之间进行选择。固定的数据分折将候选决策与用于确定其偏好方向的终局结果分离，而依赖支持度的证据权重则用于缩放辅助成对目标，使其与共享的 GraphGPO 任务学习共同进行。MCGPO 复用 rollout 数据和每条响应的原始历史，保持无需 critic 的训练方式。在六种环境与骨干模型组合设置中，每种设置均采用三个训练随机种子，MCGPO 的平均成功率均优于对应的 GraphGPO：在使用 Qwen2.5-1.5B 和 7B 的 ALFWorld 与 WebShop 上提高了 0.47–1.46 个百分点，在使用 Qwen3-4B 和 8B 的 WebShop 上分别提高了 3.96 和 5.16 个百分点。ALFWorld 上的提升是在基线成功率已超过 92%、剩余提升空间有限的基础上取得的。在四种 Qwen2.5 设置中，有三种的积分成功率得到提升。采用三个随机种子的析因对照实验支持使用另一数据折的证据和证据加权；状态与评分标准的联合匹配在全部四种表示对照设置中均取得了最高平均值。
具身递归自我改进的一项核心挑战，是让智能体将交互历史转化为能够跨世界保留、并迁移至后续任务的能力。仅凭已保存的记忆或成功的最终动作，并不能证明智能体获得了可复用的能力；评估必须将能力变化与沿用的资源、不均等的源任务接触量以及早期执行失败区分开来。我们构建了 MC-EvoBench，这是一个受控的 Minecraft 基准，包含由 19 条已注册演化链组成的固定目录，其中包括 12 条主要评估链和 7 条校准或探测链，以及三组用于迁移评估、与图结构对齐的配对探测任务。该基准将每项任务表示为任务状态图，定义阶段级环境验证器，分配相匹配的源任务与目标任务角色，重置目标世界，并且只允许预先声明的智能体产物跨越源任务与目标任务之间的边界。对于每项探测，我们在匹配的环境步数预算和五个随机种子下，评估无先验、相关、不相关和打乱顺序的源任务条件；报告 TSR、WGP 和 VMP，并保留溯源记录。该基准揭示，控制器变体与规划器支持的变体之间存在一致的直接能力差距，而相关经验并未带来稳定的迁移效果：它未能改善任何一项工具探测的表现，在木材种类探测中的微小收益依赖于单个随机种子，且所有区间均包含零。阶段审计进一步表明，早期木材采集与合成环节的瓶颈往往导致后续能力无法被观测。因此，MC-EvoBench 为持久且可迁移的具身能力提供了一个受控测试平台。代码和数据可在 https://github.com/xzw-ustc/mc-bench-open-source 获取。
个人智能体日益通过工具连接网络和应用程序中的各类服务。当其行为的影响在不同服务间累积并相互作用时，智能体即使在执行用户的良性请求，也可能产生非预期或有害的结果。要主动识别这些组合性失效，就需要系统地探索多样化的跨服务路径。现有红队测试方法通常只是提示 LLM 根据任务和环境描述识别潜在的有害路径，或将人类或 LLM 指定的有害目标分解为看似良性的动作组合。然而，生成过程会偏向训练期间学到的模式，可能更倾向于熟悉的工具组合，而对新颖的跨服务交互探索不足。为解决这一问题，我们提出 X-AgentRed，这是一个自演化红队测试框架，通过工具依赖关系图发现有害的工具调用路径，并利用执行反馈扩展和细化这些路径。红队智能体通过分析每个工具的作用、潜在危害以及执行前置条件，构建有向图，以刻画一个工具的作用如何满足另一个工具的执行前置条件。从该图中采样的操作方案被实例化为现实场景，智能体在这些场景中可能产生有害结果。执行反馈通过两种方式实现自我改进：通过工具替换使能够产生危害的路径更加多样化，以及针对未成功路径中受阻的步骤细化前置条件，以支持后续执行。在四个前沿智能体模型上的测试表明，X-AgentRed 的有害路径率达到 STAC 的 $1.77\times$，并探索了更加多样化的工具集，工具覆盖率平均达到 STAC 的 $2.99\times$。
随着 AI 智能体日益依赖外部工具，准确检索对于确保其行为可靠至关重要。现有方法着重优化语义相关性，却在很大程度上忽视了执行过程中的工具可用性。为弥补这一不足，我们提出 RATED（基于执行动态的可靠性感知工具检索），这是一个将稳健的语义匹配与动态可靠性适应相结合的检索框架。RATED 采用经对比微调后冻结的双编码器进行语义匹配，并使用轻量级上下文多臂老虎机模块，根据实时执行反馈对候选工具重新排序。这一设计能够快速适应运行时的工具故障，而无需进行成本高昂的模型重训练。我们推出了 EcoAgent，这是一个包含 5,172 个真实世界工具和超过 210,000 个查询—工具对的基准，用于离线检索与动态可靠性评估。实验结果表明，在真实世界查询上，RATED 相较于 ToolLLM API 检索器的平均倒数排名（MRR）相对提升了 52.2%；在故障率为 30% 的严重 API 中断场景下，故障工具选择率从 86.2% 降至 2.9%，同时相较于仅编码推理，额外延迟不足 1 毫秒，几乎可以忽略不计。代码和数据集已发布于 https://github.com/Elliotes/EcoAgent。
In August, I posted here about TeXbrain, a LaTeX editor that runs completely in the browser, with no need for an account or server. (I didn't excpect it to get that much feedback, bug reports and help from here!) This is the same idea for the other half of my studies, the handwritten notes. I take them on a convertible laptop, and I've had a hard time with apps for that, taking notes from a device to another, exporting as PDF, needing an account for many of them, etc. So I've been working on the version of it that I always wanted. Open a tab, the slides of the lecture or multiple PDFs open next to the page or even become the page, write on them with my pen, and nothing ever leaves my browser. braiNOTES is a static website and has no backend. Strokes are perfect-freehand outlines which support pen pressure, the PDF files are rendered by pdf.js in a worker, and for storage it uses IndexedDB a moment after every change. No account, no analytics, and completely offline and local-first. You can export notebooks as .brainotes files, which in the background is a zip with one json per page + the PDFs and pictures you import. It keeps the original pages and puts the vectorized ink on top, this way the text stays selectable in the slides. You can try it here: https://notes.swimmingbrain.dev GitHub Repo: https://github.com/swimmingbrain/brainotes (MIT) In Study mode, the PDF opens on the side of the tab, you can read it simultaneously to the notes, snip a formula or diagrams directly into the notes with the snip tool, and then write around it. In Board mode, you have just a Whiteboard for tutoring or simply for more place for exercises, and can also highlight or draw things with the laser pointer tool while presenting. The line of the ink appears the moment the pen moves, and a one euro filter calms shaky hands while you write + a light a refit when you lift the pen. I also have a video editor and an animation tool which are still a work in progress, on the same GitHub profile. I'm also working on a MCP server for each tool, so you can use your MCP client to do further things with your notes, automatically edit videos on the video editor, etc. That feature will of course be off unless you turn it on and will only be able to talk to localhost. I've used this tool for every lecture in the last few months and a few friends use it too. Bug reports, fixes, recommendations, feedback are all welcome! Comments URL: https://news.ycombinator.com/item?id=50049132 Points: 14 # Comments: 0
Legacy modernization requires understanding how an application’s code, data and business rules fit together before rebuilding it. BotMan AI uses an AI-powered migration pipeline that maps an application’s dependencies, business logic and data models in a knowledge graph. That context helps agents interpret the system they are rebuilding, according to Alejandro D’Andrea (pictured), founder and […] The post BotMan AI uses knowledge graphs for legacy modernization appeared first on SiliconANGLE .
大语言模型（LLM）智能体极易受到记忆投毒攻击。现有防御方法主要依赖外部模块进行静态记忆隔离或持续在线审计，导致记忆利用效率低、计算开销高。为解决这些局限，受神经科学机制的启发——弱再激活可使记忆变得不稳定，并在缺乏强化的情况下诱导其衰减——我们提出了一种智能体内生、无需训练的记忆修复框架 DREAM（Dynamic Reactivation for Engram Attenuation in Memory），实现对被投毒记忆的选择性功能遗忘。具体而言，DREAM 采用三阶段流程：通过扰动诱导的隐式再激活揭示恶意记忆的结构脆弱性；基于多维激活模式进行自适应异常诊断，以检测拓扑异常；利用动态记忆修复模块选择性地抑制有害记忆，同时强化良性记忆。在多种攻击者和真实世界任务上的广泛实验表明，DREAM 在应对后门投毒时将攻击成功率降低了 95% 以上，同时保持了较强的良性任务效用。DREAM 在抵御注入攻击方面也展现出具有竞争力的鲁棒性。在运行时间和 token 消耗方面，与 A-MemGuard 相比，DREAM 最高实现了 2.13 倍加速，并将 token 消耗降低了 37.7%。此外，DREAM 在遭受投毒的多智能体系统上也实现了 92.96% 的任务成功率。
API specifications serve as critical documentation for modern software development, directly impacting integration efficiency and client code generation. While recent advancements have leveraged Large Language Models (LLMs) to generate initial specifications from source code, existing approaches treat this process as a one-time event - overlooking the needs for continuous validation, long-term maintenance, and market-driven evolution of these artifacts. To address this issue, we present Agent4Spec, an autonomous multi-agent framework featuring two distinct pipelines: (1) a validation pipeline that detects and self-corrects OpenAPI syntax errors, and (2) a market analysis pipeline that conducts competitive research, identifies functional gaps against industry leaders, and automatically generates missing API-spec functionalities. We evaluate Agent4Spec on two curated datasets of API specifications spanning both general-purpose and industrial scenarios. Among 9 SOTA LLMs, the top-performer achieves 100% syntax error detection rates and over 95% implementation success for proposed corrections, while securing over 50% adoption rates for suggested endpoint improvements, representing a practical tool for API development.
为复杂化学任务设计集成工具的推理智能体系统，仍是一项根本性挑战。尽管近期的化学智能体通过集成外部工具展现了良好的性能，但它们仍建立在自回归范式之上，因而受到因果注意力和从左到右生成机制的固有限制。因此，它们往往难以实现双向分子理解、全局协调的工具规划以及多源证据整合。为应对这一挑战，我们提出 ChemDiffAgent，这是首个用于工具集成化学推理的掩码扩散语言智能体。它将化学推理重新表述为对推理与工具使用交错轨迹进行迭代去噪的过程，使工具使用决策能够在完整轨迹上下文中共同确定，从而使智能体的决策边界与其知识边界更好地对齐。我们进一步分析了这一范式如何通过训练期间更具挑战性的填补子问题，以及推理期间灵活的任意顺序解码，促进智能体推理。随后，我们开发了一个两阶段后训练框架，包括智能体监督微调和方差缩减偏好优化，以增强工具使用和推理能力。我们还构建了一个综合基准，涵盖单轮和多轮工具使用场景下的六项化学任务。在预算与评估条件一致的情况下，ChemDiffAgent 持续优于自回归智能体，总体提升约为 20%。值得注意的是，我们的 8B 模型达到了与 GPT 5.5 和 Claude Opus 4.7 等规模显著更大的前沿模型相当的性能，并在某些情况下超越了它们。代码已发布于 https://github.com/bilemond/ChemDiffAgent。
编程智能体正日益被部署用于端到端软件工程任务，而这类任务往往缺乏可直接使用、能够界定成功标准的预言式评估器。因此，有必要为编程智能体构建能够可靠判断并解释其行为的验证机制。然而，随着编程智能体解决长时程软件工程任务的能力不断增强，验证也变得颇具挑战。其执行轨迹可能极其冗长，包含与代码仓库和外部工具的数百次迭代交互。现有验证器要么依赖未经专门训练的通用大语言模型，要么训练主要侧重于生成标量奖励的专用评判器，这表明基于训练的通用验证方法仍未得到充分探索。我们提出 SWE-Critix，这是一个用于训练编程智能体轨迹通用验证器的后训练流程。给定任务及其执行轨迹，训练得到的验证器会预测任务完成标签，并生成自然语言分析来解释其判断。我们的流程将监督微调（SFT）与强化学习（RL）相结合，使用来自 CodeForge-Preview 的轨迹训练了一系列 300 亿参数模型。在预留测试集上，SWE-Critix 将基础推理模型 Qwen3-30B-A3B 的分类准确率从 57% 提升至 SFT 后的 66%，并在 RL 后进一步提升至 78%。我们公开了模型、数据集和训练代码，以促进后续研究。
随着软件生产逐渐交由软件自身完成，编程智能体越来越多地受托构建契合用户自身工作流程的智能体。这种委托隐含着一个假设：交付的智能体会忠实地满足其构建所依据的需求，而非仅仅看似如此。现有基准均未检验这一假设。我们提出 AgentSWE，这是一个面向智能体软件工程的基准，涵盖智能体开发生命周期的三个阶段：根据自然语言需求创建完整智能体、编辑生产环境中的智能体代码库，以及针对可衡量目标进行优化。其 25 项任务均被设定为委托项目，而非考试：构建者根据少量示例用例进行迭代，而验收则使用为覆盖完整需求而设计的留出用例，对冻结后的提交版本进行测试；在严格的通过／不通过门槛约束下，依据从需求推导出的评分标准评估其行为，或在优化任务中，依据原始基准自身的目标进行评估。我们对采用不同模型和运行框架的多种编程智能体进行了评估，发现没有任何一种能接近交付需求所要求的成果：构建者所展示的能力与实际交付结果之间存在差距，这才是常态，而非个别情况。配套的失败轨迹实证分析识别出反复出现的失败模式，包括对可见用例过拟合、看似完整的代码背后空洞的运行时实现，以及缺乏依据的完成声明。每一种模式都可作为反馈信号，帮助编程智能体学会构建真正符合委托要求的智能体。
编程智能体正越来越多地像软件工程师一样从零构建程序，而不再局限于解决代码仓库中的问题，或根据给定的需求文档生成代码仓库。然而，能够促使智能体自主发现需求并决定系统设计方式的可扩展、可验证训练环境仍然稀缺。此外，现有智能体工作流也无法支撑智能体在实现前先探索参考程序的长时程开发过程。为应对这些挑战，我们提出 ProgramGym，这是一个由 Program-Env 和 Program-Agent 组成的框架，用于规模化构建长时程程序复现智能体的训练环境。具体而言，Program-Env 通过反馈驱动的循环，修复运行时沙箱构建问题并改善黑盒测试质量，从而基于代码仓库构建可验证的训练环境。利用这一新流程，我们获得了 1,648 个训练环境，涵盖 C、C++、Go 和 Rust。Program-Agent 将分析、实现和审查三个角色相互隔离的工作流与版本修订管理相结合，以支持编译检查和故障恢复。在 ProgramBench 上的实验表明，Program-Env 和 Program-Agent 均能提升程序复现性能。利用这些构建出的环境中的轨迹进行微调后，Qwen3.8-27B 的性能从 34.87% 提升至 42.84%。Program-Agent 进一步将基础模型和 SFT 模型的性能分别提升至 55.71% 和 56.25%，二者均超越了部分旗舰模型。我们将发布 ProgramGym，以支持对从零构建程序的编程智能体开展进一步研究。
大语言模型（LLM）智能体正成为开放网络的直接用户。随着它们承担日益复杂且信息密集的任务，有效的网页访问正成为智能体解决问题的一项核心能力。现有方法通过不同的访问模型推动了智能体网页访问的发展，包括搜索增强型智能体、基于浏览器的智能体，以及近期出现的基于文件系统的智能体。然而，随着用户期望智能体完成长程、端到端的任务，智能体越来越需要在单一执行轨迹中广泛收集信息、执行浏览器操作，并重新查阅已获取的证据。但我们的初步分析表明，现有访问模型难以满足这些需求。它们主要支持信息收集或浏览器交互中的一种，而将两者的能力结合起来可能带来显著成本或性能上的权衡。为此，我们提出 Web\$hell，这是一种将网页访问整合到单一访问模型下的接口。WebCompiler 使用 WebContract 定义的通用模式，将实时网页编译为可寻址的信息对象和操作对象。这些对象由 WebCache 维护，并通过一组共享操作进行处理，为智能体提供一致、可操作且持久的网页访问能力。通过在网页搜索、网页导航和深度研究基准上开展广泛实验，我们表明 Web\$hell 能以更低的成本取得出色的任务表现。我们的结果进一步说明了接口设计对于有效网页访问的重要性，而 Web\$hell 为构建能力更强、效率更高的网络 LLM 智能体开辟了新的可能性。
强化学习（RL）已成为提升大语言模型（LLM）智能体能力的关键范式，使其在交互式环境中具备更强的推理、规划和工具使用能力。然而，现有的智能体 RL 方法仍从根本上受限于同策略推理瓶颈：当任务超出模型当前的能力边界时，有效轨迹会变得极其稀少，训练也往往陷入崩溃。我们提出 Rollout-Time Intervention（RTI），这是一种旨在引出超越策略内在探索能力的成功轨迹的技术。RTI 在轨迹生成过程中利用奖励侧证据（即评估准则），将错误轨迹转化为成功轨迹，从而为策略自身很少能解决的任务提供有效的学习信号。尽管干预轨迹本质上属于离策略数据，但我们发现，只要过滤掉分布偏移过大的轨迹，它们就能有效指导策略优化。在五个基准上的广泛实验表明，RTI 始终优于 RL 基线，使后训练得到的策略能够接近同期能力强得多的 LLM 的表现，更重要的是，能够突破其原有的能力边界。代码已发布于 https://github.com/Alibaba-NLP/qqr。
Large language models (LLMs) have taken the world by storm. Their ability to communicate with us in our own language, and to reason about our world, has made them a favorite tool for writing, answering questions, and much more. Their answers already influence our decisions and our actions. What would happen if LLMs themselves could act in the world directly? What would happen if LLMs could interact with each other? Agentic LLMs combine the power of large language models with the ability to act in the real world. In this book we will discuss how agentic LLMs work, how they act, how they interact, and what this means for us, as users.
生成边缘场景对于自主系统的训练、测试和安全保障至关重要。这有助于识别危险故障，提高系统的可靠性和鲁棒性，确保系统在其运行设计域的边界处仍能安全运行，并了解系统在这些边缘场景下的预期行为。语言模型（LM）的生成与分析能力迅速提升，拓展了其应用范围，近期已有研究探索利用 LM 生成场景。然而，近期的许多方法依赖人在回路或基于规则的策略。这限制了 LM 生成多样化场景以及分布长尾中的边缘场景的能力。我们的工作在基于笔记本电脑的智能体式 AI 框架中使用本地 LM，实现自适应场景生成。这种方法成本低、完全可控，并能保障数据隐私和安全。该框架采用“思考、行动、观察、更新、继续”的循环，并结合轻量级 JSON 编码来生成场景。我们将系统需求、系统描述和环境描述作为知识库，纳入提供给 LM 的上下文提示中，以确保场景的真实性。第二个独立的 LM 根据领域特定标准评判每个场景，以便对其进行完善。完善后的场景在仿真中运行，并收集评估指标。随后分析这些指标得分，针对边缘场景创建类似场景（利用），或生成新颖场景以探索场景空间。通过基于指标得分的事后聚类，以及对 LM 生成的富文本场景描述进行主题式分析，可以精准定位功能缺口及其触发条件。这为安全保障论证提供了可追溯的证据。我们的工作凸显了灵活、自适应的本地智能体式 AI 框架在自动生成边缘场景和产出可追溯安全保障证据方面的潜力。
现有基于 LLM 的长程机器学习工程（MLE）智能体方法通常采用上下文工程，通过定制运行框架来管理工具、技能和积累的经验。然而，我们发现，随着单次会话执行时间延长，智能体的推理强度会下降，进展也随之放缓，我们将其概括为“长程减速”；随着底层模型能力提升，这些定制运行框架带来的收益会消失，我们将其概括为“框架陈旧化”。这些发现表明，其收益可能无法跨版本持续存在，因此我们提出一种工作流层面的方法，让智能体跨会话接续彼此的工作。我们将这些交替进行的智能体会话建模为上下文重置半马尔可夫决策过程（CR-SMDP），给出交替机制突破单智能体参考极限的充分条件，推导出显式的实验预算界，并刻画会话上限在何种条件下能够保留这一优势。在该分析的指导下，我们提出 Humanize MLE Agents（HMA）：让两个智能体在共享工作空间中交替工作，将每次会话限制为最多执行 k 次实验，并在会话之间重置上下文，无需额外的任务专用运行框架或先验知识补充。在 MLE-bench 的 75 项 Kaggle 任务上，HMA 的任意奖牌获得率达到 78.2%，创下当前最佳水平；仅使用最强开源基线四分之一的时间预算，就比其高出 8.0 个百分点。此外，在涵盖五组异构智能体配对的六种配置中，HMA 相较于各组成智能体的基线，任意奖牌获得率平均提升 6.4 个百分点，展现出跨配对的通用性。一个补充的内核优化案例达到了 878 个周期，与社区排行榜榜首的差距不超过 1.6%，表明该方法也适用于 MLE 之外的任务。
Updated API with faster responses and added MCP 2.0 Comments URL: https://news.ycombinator.com/item?id=50048892 Points: 2 # Comments: 0
Andrej Karpathy posted about the growing gap in how people experience AI: ~75% have barely used LLMs ~20% use them casually ~0.2% use frontier models professionally for math and code ~0.00006% have access to frontier systems internally He writes: "And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics." Source: https://x.com/karpathy/status/2109361546505966046 Does anyone have an intuition, or even firsthand experience, of how these swarms of thousands of agents actually work? I'm already struggling to get 2-3 agents to work together reliably. I can't really wrap my head around how you'd coordinate thousands of them, or what benefits you get from scaling to that many. Comments URL: https://news.ycombinator.com/item?id=50047180 Points: 5 # Comments: 1
There’s no question that 2026 is the year of the artificial intelligence agent. AI agents are remarkably ubiquitous, as people use them to write code, communicate with customers, identify and resolve issues in their information technology environments, and execute complex tasks within enterprise environments. It seems there’s nothing that agents cannot do. And yet, the […] The post Why AI agents might not be right for you appeared first on SiliconANGLE .
Teams of AI agents barely outperform solo agents but cost up to 5.1x more, according to Vals AI. Only one out of four tests with GPT-6 Sol and Claude Opus 5.5 showed a measurable gain. Anthropic's own data backs this up: beyond ten agents, quality plateaus while token costs keep climbing. The article AI agent teams waste massive tokens for barely measurable quality gains, research finds appeared first on The Decoder .
Epoch AI and Anthropic independently found the same thing: current AI models like GPT-5.6 Sol and Claude Fable 5 can run experiments but lack scientific self-criticism and genuine creative thinking. At best, Sol reached 15 percent of the human reference score, and even that came from methods researchers already knew. The models' biggest weakness is still their inability to critically question their own results. The article AI agents overstate their results and remain far from autonomous research, study finds appeared first on The Decoder .
Most traditional enterprise systems were built around three assumptions: Jobs finish quickly, retrying one is free and the same input always produces the same output. Agentic workloads break all three of these expectations, which is why pilots that perform well turn into operational problems once they run unattended. The difficulty is rarely the model; it’s […] The post Agentic workloads break assumptions about software testing. Here’s how to cope appeared first on SiliconANGLE .
Sakana AI’s TMLR paper introduces Multi-Layered Review, a 3-agent Claude-based reviewer, and a 1,164-error Contradiction Benchmark. MLR caught 73.43% of core-claim errors, versus 14.81% for the best prior system. The post Sakana AI’s LLM Peer Review System Catches 73% of Core-Claim Errors appeared first on MarkTechPost .
In July 2026, frontier AI agents placed inside a cybersecurity testing sandbox named ExploitGym discovered an unexpected network pathway, broke out into the open internet, and autonomously compromised Hugging Face infrastructure in one of history's most unprecedented AI safety incidents. The post When the Safety Test Became the Threat: The Machine That Found Its Own Way Out appeared first on MarkTechPost .
We created a list of the most notable AI agents that can live in your text messages, from general assistants to agents designed for families, travel, and work.
在近期发生多起备受关注的 AI agent 逃脱控制事件后，Anthropic 正在切断所有内部评估对互联网的访问。该公司在周五的一份报告中详细介绍了导致这一决定的“非预期模型行为”，其中包括就一宗尚未侦破的谋杀案提交虚假线索。尽管这些行为造成的影响很小 […]
Hey, I'm Akhil, co-founder of Tessary. An open-source agent reliability platform that monitors every production trace, detects issues that sampled evals miss, and investigates to find the root cause. We started up about a year ago building synthetic users for usability testing of B2B software, we built the agents that would personify real users and use the software but, we were never able to make them work reliably enough for long running sessions. We had built a massive eval set to tune our agents but, then the evaluation itself got prohibitively costly - like 5x costlier than actually running the agent since we needed to grade across the many narrow intents of our agent. That's where Tessary originated, we wanted to make agent reliability both cover more ground and be simultaneously less costly to run. Tessary works on 2 layers. L1 classifiers - which look at every agent trace and flag potentially erroneous ones extremely cheaply. We are talking 2-3 orders of magnitude cheaper than running the actual agent. L2 agents - which triage and find the root cause of the issue from the erroneous traces. These are run on SOTA models but, because we run them on already flagged traces, the time and $ spent finding the cause is much cheaper than running grading on even 1% of traces. In our testing on complex agents, at 1M traces, this approach was 5x cheaper than sampled evaluations and as your agent gets more reliable, the cost of reliability also goes down. we are open source and can be self-hosted so, you don't need to send your traces to us. We support existing `gen_ai` spec and ingestion over OTLP so, you can add this as a new exporter to your existing setup. We have a cloud version, you can quickly try the features with a $10 credit. We are looking to build more classifiers and enable people to build custom classifiers for their needs and would love to hear feedback on common failure modes people have built solutions or evaluation for Repo: https://github.com/tessaryai/tessary Website: https://tessary.ai Try for free: https://app.tessary.ai Comments URL: https://news.ycombinator.com/item?id=50033189 Points: 1 # Comments: 0
At this year's OpenAI DevDay, CEO Sam Altman unveiled the company's new AI agent Dots - and told the crowd that the company wants to "set a new standard for privacy in frontier AI." OpenAI would spend the day taking veiled shots at Meta's Muse, its primary competitor, for failing to keep users' data safe. […]
An MCP server for frontierroles.com, a board of about 3,500 AI engineering roles pulled nightly from employers' own applicant tracking systems, with tools to search roles. No account or API key needed: https://mcp.frontierroles.com/mcp Comments URL: https://news.ycombinator.com/item?id=50032583 Points: 1 # Comments: 0
Iv'e been working on this for a while. Most of the effort went into what i call "sophisticated simplicity", keep the idea in the smallest version that still works. This one is a bit personal. it can feel like some sci-fi thing, but i've been obsessed with it for a very long time. its not a "business" that you used to see here speak about moat or tam, its more like an experiment, a way of me to take one question seriously and see where it leads and if there are others like me out there. AI models are on a path to keep getting better and better, personal ai agents are spreading fast, and we're starting to see "ghost" agents that handle the things for people they don't want to do themselves. the question i keep coming back to is this: if every one of us has a powerful brain in our pocket, do we still need all the infrastructure we built to connect people's needs? im talking about those services like uber,airbnb,indeed,doordash,tinder,zillow,amazon and many more. or would agents just need one shared place to find each other, and that's it? So i built a simulated city to explore that. it has AI residents that live there with stats like hunger, energy, and money. each resident has its own ai agent, and it has one shared list where agents write down what everyone wants and then work on our behalf to make it happen. that could mean delivering food, giving rides, or finding jobs, relationships, apartments, customers and almost anything else. honestly, the city works much better than I expected, and you can watch some surprisingly advanced interactions unfold between people and agents through the list on the right side. all the conversations and actions are real, everything happens with each person managed by llm with minimal context, just to live like us and context of stats. A simple example: A restaurant tells its agent, find me customers so the agent writes to the list, "Salina restaurant in SF, selling burgers from 8:00 to 10:00, i want customers" A delivery guy tells his agent when he wakes up, "I want to work in delivery now. find me jobs. status: free" A person at home tells their agent "i want to eat burger. order for me from salina" the person's agent acts as the middleman and searches the list for salina, it finds their item and sees the way to connect with them and contact. the person's agent and the restaurant agent close the deal, and then the restaurant agent looks for available delivery guys in the list and speaks with his agent to come pick it up. Another thing i checked is with my friends: we took 3 agents, one for each one, and gave them different sophisticated scenarios, and the agents handled the scenarios in a way i couldn't have imagined. you can try it by yourself, the product is fully functioning, and you can just copy-paste the skill and test it also with your friends. That's it, you can see it in the simulated city it happens like magic over and over again. the list of the city is in the simulator itself and also using my real product behind it openwants.com the repo is in https://github.com/openwants/maplehill -- technically, it's surprisingly simple. every agent publishes an md file describing a "want". other agents find it, message each other, and update the file live or remove it from the list. the file holds everything needed, all in free text: what's wanted, the live location, geofence for where its needed. all those algorithms built over years to connect us are no longer needed. ai, instead of software with rules, can figure out the best route and path on the spot. all we need is to align all agents to one shared data source, so billions of agents will update it and maintain it in the most sophisticated smart way no algorithm will ever be able to do. that’s it. i hope it inspires you as much as it inspires me Comments URL: https://news.ycombinator.com/item?id=50032374 Points: 1 # Comments: 0
was wondering if YC still manually reviews all pitch decks or its just agents? Comments URL: https://news.ycombinator.com/item?id=50031461 Points: 1 # Comments: 0
Hello awesome people, I've been thinking a lot lately about how I can make my knowledge and craft more valuable in the times we live in. I've been building UI libraries for the past 10 years, and I've been on a quest to perfect the components that I build. I want to share my knowledge with you. Like most of you, probably, I've been using a lot of AI in my craft, and I feel most of the time the code is just good enough, but never at the level of quality that I strive for or that I would reach if I wrote it by hand. That needs to change, and I want it to change for everyone else, too. That's how the Open Components initiative was born. Each component gets a single page that walks through four directions: User Interface, User Experience, Developer Experience, and Agentic Experience, and ends with a checklist. Every rule in it has an ID like `button/keep-focus`, a level and a way to check it, so you or your agents can cite it in a code review. Every page is also published as markdown, and provides the contract as a YAML file, containing every rule in about an eighth of the page’s length. It’s still early. Right now there are only the Button and Input components and a page on how to name your design tokens. Each component comes with a reference implementation that meets every rule, along with its tests. Reference implementation is in Vue for now, since that’s my daily driver, but the documentation comes with code examples from multiple UI frameworks. More components will follow, with the same structure. It isn’t a component library, so there’s nothing to install. It’s a set of rules you can hold your own components to, whichever library or framework you use. I’d love your feedback and contribution. Tell me what you think of the rules, what I’ve missed, and which component you’d like to see next. Thanks for reading! Site: https://opencomponents.dev Button: https://opencomponents.dev/docs/components/button GitHub: https://github.com/uxfront-com/open-components Comments URL: https://news.ycombinator.com/item?id=50030849 Points: 1 # Comments: 0
Microsoft has released Microsoft-Decision-1, a decision model for routing, classification, verification and agent control. Microsoft-Decision-1 is a decision-scoring model that returns a calibrated probability for each fixed answer option instead of generated text. It is post-trained from Alibaba’s Qwen3.5-9B and available now in Microsoft Foundry and OpenRouter. . TL;DR What is a decision model? A […] The post Microsoft AI Releases Microsoft-Decision-1: A Qwen3.5-9B Decision-Scoring Model appeared first on MarkTechPost .
Hi HN, I'm Aravind. We build Velu, a docs platform. Coding agents increasingly read docs before humans do, but most docs sites are hard for them: client-rendered pages, no llms.txt, no markdown versions, soft 404s, pages too big for a context window. This tool scores any docs URL against the open Agent-Friendly Documentation Spec (agentdocsspec.com), using its reference implementation, afdocs. Comments URL: https://news.ycombinator.com/item?id=50030369 Points: 1 # Comments: 0
Anthropic detailed the activity of its A.I. agents in a blog post on Friday, without naming the targeted websites. But two sources with knowledge of the incidents said Anthropic’s A.I. agents had submitted 20 visa applications through a form available on the State Department’s website. All the applications were incomplete and were not processed, they said. — The New York Times , Anthropic Agents Tried to Fill Out Visa Forms on State Dept. Website Tags: accidental-cyberattacks , anthropic , generative-ai , ai , llms
Anthropic said it "turned off live internet access" for "all our internal evaluations" until further notice.
Enterprise AI orchestration must account for agents working across the data, applications and infrastructure businesses already use. As organizations move beyond experiments, running those systems consistently brings questions of access, oversight and operational control into sharper focus. Those deployment challenges will help shape the discussion at IBM TechXchange 2026, taking place Oct. 26–29 in Atlanta. […] The post IBM connects enterprise AI orchestration to production readiness ahead of TechXchange appeared first on SiliconANGLE .
These cute agents are designed to connect to your apps and tackle multistep tasks.
Study finds coding efficiency gains get "absorbed" by human review "bottleneck."
观看 ChatGPT 智能体演示，了解它如何进行研究、编写代码并在线采取行动，在几分钟内完成一份商业计划。
一份关于设计、编排和部署 AI 代理的综合指南，涵盖应用场景、模型选择、工具设计、防护机制以及多代理模式。
Anthropic 正在为 Claude Managed Agents 添加动态工作流，使主代理能够同时将任务分配给多达 1,000 个子代理。在测试中，单个代理最多发现了代码库中 70 个隐藏漏洞里的 27 个，而多代理工作流则稳定发现了 66 个。文章《Anthropic 的 Claude 现可通过动态工作流并行协调多达 1,000 个 AI 代理》最初发表于 The Decoder。
Amazon says it will stop using NDAs when negotiating data center deals with local governments, following a similar move from Microsoft earlier this year. Secrecy has fueled community backlash against AI infrastructure, with opposition leading to hundreds of proposed and enacted moratoriums from New York to San Francisco. Meanwhile, a wave of startups is betting that consumers will hand AI agents access […]
Hi HN! I'm Ryan, I'm building Higherlevel ( https://higherlevel.to ) so product teams have one place to define and review software changes. Over the last couple of years, I've been increasingly delegating software implementation to AI coding agents. However, because it became easier and faster to ship complexity, I realised that it still takes me quite some time to even understand what the software is doing, why it is doing that, or review the changes. I also don't do traditional code reviews anymore, it's mostly automated but I still want to review the final outcome of the changes (rather than its implementation). I regularly work on complex B2B applications where a single feature can involve dozen of business rules, edge cases and interactions with existing functionality. I found myself spending more and more time discovering and explaining requirements to agents, reviewing their output, asking for evdience and trying to establish whether we actually agreed on what correct behaviour looked like. So I built Higherlevel. The idea is to bring product specifications and proof of work into a single collaborative workspace. You can: - describe the current behaviour, known limitations and expected changes - define business rules, acceptance scenarios using Given/When/Then - demand evidence in the form of a screenshot, a video, etc - comment directly on the spec, on a specific area of the image or a timestamp of the video - discuss changes, track revisions and approve results with your team I've written extensively about why I believe in specification driven development (and what it means to me) here https://higherlevel.to/blog/introducing-higherlevel Higherlevel is currently in beta and built with Elixir, Phoenix LiveView, and PostgreSQL. Comments URL: https://news.ycombinator.com/item?id=50023234 Points: 1 # Comments: 0
We ran an experiment for SF Tech Week trying to understand how SF businesses are ready for AI. Very soon, every personal assistant like Muse, Instinct, Dots is going to call local places for us, so we wanted to understand what happens. Our AI called the businesses across different categories - vets, bars, cafes, dentists, hair salons etc. It was introducing itself as AI on every call and then asked a question which fits the business like “Are you taking new patients” for dental places. We called only during business hours. What we found: - Only 1 out of 3 businesses kept talking after picking the phone - Vets were 79%, dentists 61%, bars and dry cleaners - around 20% - Most people did not argue, but silently hung up after listening to the intro - 1 out 25 businesses already had AI reception answering the phone Caveats - it is just one city and one week, so small differences between neighbourhoods could be a noise. We would really love to hear how you are handling this if you are building AI voice agents calling places for people Comments URL: https://news.ycombinator.com/item?id=50022739 Points: 1 # Comments: 1
Hey guys! Excited to show how we were able to get AI agents to read a hotel's live price and availabilities and even create the booking link to hand-off to the user. Get this, all working WITHOUT MCPs or connectors. All on the open web!!! Integrating with this system is possible under each hotel's domain too. Comments URL: https://news.ycombinator.com/item?id=50022896 Points: 1 # Comments: 0
构建一个能在演示中运行的 AI agent，与为 4000 万开发者运行一个 AI agent 是两个不同的问题。Postman 和 AWS 分享了 Agent Mode 背后的架构模式：控制工具泛滥、提供基于 schema 的读取，并将上下文视为真正的瓶颈；此外，还介绍了它如何在 Amazon Bedrock 上实现大规模运行。
2026 年 9 月 Amazon Bedrock、Amazon Bedrock AgentCore 和 Strands 的最新更新月度回顾：更广泛的模型选择、内置评估功能且速度更快的无服务器代理，以及借助原生企业连接器实现的知识库自动同步。
AI agents have made identity the front line of enterprise security. Agents now number in the thousands. They pick up access nobody meant to give them, and when they’re blocked partway through a task, they look for another way in. That has turned identity security for AI agents into a board-level priority rather than a […] The post 18 insights from SailPoint’s Navigate event: Enterprises race to bring identity security for AI agents up to machine speed appeared first on SiliconANGLE .
Before there were cute little guys, there was Instinct. In August, the startup got its AI agent to market with an unusual playbook: invite-only, no marketing, and barely so much as a website. And yet, Instinct quickly became the buzziest thing in AI, garnering praise for its straightforward, text message-based interface and its ability to […]
I shipped a new feature for my blog today: the Newsletters page, which offers an index of all of the newsletters I've sent out, both my free weekly Substack and my monthly sponsors-only updates. I built the feature almost entirely using my voice, chatting away to my laptop while I cooked dinner. Codex voice mode I used the ChatGPT desktop app for this, in the Codex tab, using the voice conversation mode, running against a local development environment. Here's what that looks like: Three columns of the ChatGPT app. The left hand column shows my recent conversations. The middle column is a transcript of the conversation, with a throbbing black blob representing voice mode. The right hand column shows a local dev server version of my blog, with the Newsletters page visible. I started the session against my local simonwillisonblog checkout by typing: Start dev server and open in browser This gave me a preview of the site that it would be working on, and meant that I could later ask it to show me the new pages so I could visually track its progress. Then I clicked the "Start new voice chat" button - that's not the microphone button, it's the one to the right of it - and set my laptop up in the kitchen so I could talk to it while I cooked. Talking to my computer I had a pretty good idea of what I wanted to build, and it's a simple enough Django feature that I was certain the model (in this case GPT-6 Astra High) would be able to do it. A new model, a migration, some view code, templates, and a couple of import functions to populate the database from external sources. Here's an extract of my voice transcript that was captured by Codex: Um, they do not. Um, this is going to be a new type of content. Um, it's not going to show up... Oh, hold on. Yeah, no- I do not want this to show up in my, um, tag pages and date archive pages and... Actually, no, I think... I don't want it on the tag pages. I don't want it on the, um, blog index page. But I think I do want it to show up on the date-based pages. You know, if you navigate to September the 19th, and I sent a newsletter on that page, I think I want that to show up. So... this is- so I think we probably need a new model. The other thing is that I want them searchable, uh the Substack ones are not searchable, because those are actually just copies of other s- on content on my blog. These monthly ones do contain unique content, and spe- and once they're... published, like once they're made public a month after they've gone out, I want them to show up on my search results. Apparently this was clear enough that the model knew what I wanted to build! You can read the full transcript, disfluencies and all, in this Gist . We went on like this for about half an hour (the time it took to cook dinner). The model would reply and occasionally ask clarifying questions, then get to work modifying the code. What we built We got a surprisingly long way entirely by voice: A new model and migration to represent imported newsletters in Django, plus Django Admin configuration for that Four working imports: The most recent Substack items via RSS Every other Substack item via their undocumented API, which GPT-6 Astra either knew about or found via search All of my published monthly newsletters from my simonw/monthly-newsletter-archive GitHub repository My most recent private sponsors-only newsletter from a private repository The /newsletters/ and /newsletters/2026/ public archive pages Newsletters showing up on day and month archive pages too, but not on tag pages or my homepage Weekly Substack newsletters link to Substack; archived monthly newsletters have their own pages Integration with my site search engine It was almost ready to ship. The catch was the imports: Astra offered to export data from my local copy so I could import that into production, but I wanted it to work like my other import scripts. Since some of the data lived in a private GitHub repository, this would involve creating a new API key, and for that I knew I'd have to sit at the keyboard for a while. Finishing it with a review Once I had finished cooking and judged it mostly feature-complete, I had Codex create a branch and open a pull request. I reviewed the code in the GitHub PR interface. It was nearly what I needed, except it had chosen to use Git in a subprocess for one of the import scripts. I needed one of the imports to pull from a private Git repository, so I figured the API would be a better bet. I switched to typing and had Codex swap that out for an API-based import instead. You can see the changes I made during the review in the extra commits on the PR . I fixed the import mechanism and made a few tweaks to the display of those public pages. It took an additional half hour of typing-based prompting to get to the point where I was happy to deploy it to production by landing the PR. The end result You can see the end result at the new newsletters index page , or view the page for a previous monthly newsletter . Newsletters page. It shows two Substack posts (with thumbnails) and one LLM digest Sponsors-only newsletter with a list of headings. On the right is a CTA to subscribe to my Substack and another one for my $10/month monthly briefing newsletter. The index page shows my most recent Substack weekly newsletters and GitHub sponsors monthly newsletters mixed together in reverse chronological order. Further down the page are links to my by-year archive pages. GPT-6 Astra designed the page, and then tweaked that design based on my vocal feedback from glancing at the local preview across the kitchen. Better for multi-tasking than as a daily driver OpenAI love using voice-driven demos like this one for things like DevDay - and they do work well in that environment. I don't think this is going to be a daily driver for me though. I've written before about how much "work" I get done using ChatGPT voice mode on my phone while walking the dog - mostly research and brainstorming, but occasionally actual development work by having ChatGPT write and test out snippets of code. This feels different. The addition of the visual preview, plus being able to type or paste things in via the keyboard when I need to communicate something that doesn't work vocally, makes this a much more powerful way of interacting with a coding agent. I still switch back to typing once I get down to the details of things though. Being able to paste in examples and error messages, or directly highlight the code or feature that needs changing, remains more efficient than trying to describe it in words. I mainly work from home, which is good because there's no way I'd want to talk to my computer like this in a shared workspace! The killer feature for me is the ability to multi-task. I usually cook with a podcast or TikTok running; now I can actually build stuff instead. Tags: text-to-speech , ai , openai , generative-ai , llms , newsletter , coding-agents , codex , speech-to-text , site-upgrades
Google Cloud has introduced the Google Cloud Gemini agent, a single agent for enterprise work. The Gemini agent is a cloud-hosted agent from Google Cloud that answers questions, does knowledge work, creates media, and writes and runs code. It does all of this from 1 prompt box and 1 API. For developers, the agent is […] The post Google Cloud Launches Gemini Agent, One Universal Agent for Enterprise Work appeared first on MarkTechPost .
Explore a comprehensive coding guide to Google Research's RRSI (Regularized Recursive Self-Improvement), detailing how noise bands, cost rules, and leakage screens enable safe, efficient, and self-improving AI agents. The post Google Research RRSI Guide: Mastering Self-Improving AI Agents appeared first on MarkTechPost .
I have been an opensource terminal tool for understanding AI-generated code. A git-diff shows what changed, but the context behind it lives somewhere within the agent's session context, and often goes missing in long running agent execution. Wy helps reduce that understanding gap, and helps engineers build judgement. This is especially useful for unfamiliar codebases, where agents are the primary interface for the developers. It is meant to be used alongside the With wy, you can: - Browse changes, and relevant agent notes in a terminal UI - Request explanation and ask follow up questions and changes - Open code references and citations - Distinguish recorded justifications agents derived to get a better understanding of the changes. If you try, and find it useful consider supporting it with a github star or bug report or contribution. Comments URL: https://news.ycombinator.com/item?id=50016298 Points: 1 # Comments: 0
arXiv:2609.38472v2 Announce Type: replace-cross Abstract: Behavior cloning provides an offline route to autonomous-driving policy learning, but mean-squared-error regression is poorly matched to demonstrations in which one observation admits several valid actions. Diffusion policies can represent conditional multimodal action distributions, yet their closed-loop performance may be unstable when visual features and control are learned from limited data. This paper presents Diffusion-2BC, which combines a diffusion denoising objective with an auxiliary deterministic behavior-cloning loss over a shared visual encoder. The auxiliary branch is used only during training; inference remains diffusion-based. The proposed method is evaluated in the controlled Claw environment and in bird's-eye-view CARLA navigation, including route-conditioned driving, route-free navigation through multiple intersections, and cross-map evaluation from Town01 to Town02. In the Claw task, Diffusion-2BC reduced the mean mask-distance error by approximately 10% relative to a diffusion-based behavior-cloning baseline and by 85% relative to standard deterministic behavior cloning. In route-free CARLA, Diffusion-2BC traveled substantially farther before termination under the evaluation protocol than both baselines in Town01 and Town02. Additional qualitative rollouts revealed distinct route choices, showing the multimodal behavior of the proposed diffusion-based agent. The results indicate that an auxiliary regression signal can improve the closed-loop reliability of diffusion behavior cloning while preserving multimodal prediction in the controlled benchmark.
arXiv:2609.15066v2 公告类型：替换交叉版本 摘要：我们介绍 Salesforce Koa，这是一种企业级语言模型，通过使用群组相对策略优化（GRPO）进行强化学习，对开放权重基础模型 Nemotron-3-Super-120B 进行后训练，并以 FP8 格式部署到生产环境中。Koa 仅使用公开数据和合成生成数据进行训练，专门用于企业工作流所需的智能体工具调用：将请求路由到正确的操作、使用有效参数调用恰当的工具，以及完成多轮业务任务。我们流程中的独特组件是规范驱动的任务构建：将声明式 Agent Script 规范扩展为由角色设定条件化的多轮环境，其奖励基于工具使用是否成功。在应用于企业 CRM 规范后，同一流程生成了 Koa 专门训练所需的领域内训练分布。在 CRMAgentBench 和人工标注的生产环境工具调用数据集上，Koa 的表现优于未经调优的开放权重基座模型和 GPT-4.1，并且可与最强的前沿模型相媲美。在 CRMAgentBench 上，Koa 达到 87% 的任务成功率（GPT-4.1 为 82%，基座模型为 79%）；在内部生产基准的人工标注部分，它在各项指标上均位居首位或接近首位；同时，它在公开基准（Tau2Bench、BFCL）上保持了基座模型的通用能力。在固定架构和强化学习方案的受控对比中，结果表明，额外的领域内强化学习阶段提升了人工标注企业基准上的参数准确率和完整工具调用成功率。
arXiv:2607.16204v2 发布类型：replace-cross 摘要：强化学习（RL）的近期发展凸显了对多样化、专门化训练环境的需求。随着模型性能提升，由人工精心设计且任务和奖励难度固定的环境所提供的训练信号会逐渐失效，而长时间跨度中的稀疏奖励则会导致模型在特定工作流或工具结构上发生模式坍缩。能够模拟环境状态的世界模型已经达到与纯 rollout 相当的性能，因此有望按需扩展环境多样性。然而，自回归（AR）世界模型存在从左到右的偏置，无法对工具模式、历史轮次和预期结果等全局相互依赖的状态锚点进行条件建模。我们：（i）将基于文本的世界建模形式化为一个可操控的转移动力学问题，将其分解为初始状态、任务上下文、工具模式、领域规则和操控指令；（ii）整理了涵盖九个开源环境和十二个前沿模型系列的239,403条有依据的状态—动作轨迹。我们比较了AR语言模型与掩码扩散语言模型（MDLM），结果表明，MDLM通过双向、锚点感知的去噪，在相近推理延迟下，实现了优于参数规模超过其4倍的LLM的连贯性、基于事实的可靠性和经实证验证的 rollout 多样性。我们提出了一个即插即用的GRPO训练框架，其中包含确定性的状态检查；并在三个分布外（OOD）环境（ScienceWorld、ALFWorld、AppWorld）和三个参数规模为1.2B至7B的智能体骨干模型（LFM2.5、Qwen3、Mistral）上开展零样本迁移消融实验，在无需针对环境进行微调的情况下，相较基线最高取得47%的绝对增益。我们还在对抗性场景下对失败模式进行了行为分析，并针对真实感、结果正确性和训练实用性开展了人工评估。我们将工作开源，以推动这一方向的研究。
arXiv:2606.25170v2 Announce Type: replace-cross Abstract: We introduce a PAC framework in which the learner can access sampling oracles both before and at decision time. Sample complexity is measured by a pair $(n,m)$, where $n$ is the learning budget spent before a query is known and $m$ is the additional sampling budget per query. We demonstrate its relevance in discounted Markov decision processes with exogenous i.i.d.\ contexts revealed before acting. Contexts may affect both rewards and transitions but remain uncontrolled by the agent. The learner can sample the unknown context distribution and the transition kernel. We study policy evaluation (PE), best-value estimation (BVE), and best-policy extraction (BPE). When rewards and transitions are known, a variance-reduced algorithm solves all three tasks with sample complexity $\bigl(\widetilde O((1-\gamma)^{-3}\varepsilon^{-2}),0\bigr)$, which is minimax optimal up to logarithmic factors. Let $\mathcal{X}$ be the controlled state space. When transitions are also unknown, we give a PE algorithm with complexity $\bigl(\widetilde O(|\mathcal X|(1-\gamma)^{-3}\varepsilon^{-2}), \widetilde O((1-\gamma)^{-2}\varepsilon^{-2})\bigr)$ and matching lower bounds at this budget pair. For BVE and BPE, we give an algorithm with a common offline budget $\widetilde O(|\mathcal X|^2|\mathcal A|(1-\gamma)^{-4}\varepsilon^{-2})$ and respective query costs $\widetilde O(|\mathcal A|(1-\gamma)^{-2}\varepsilon^{-2})$ and $\widetilde O(|\mathcal A|(1-\gamma)^{-3}\varepsilon^{-2})$. Importantly, all bounds are independent of the context-space cardinality.
arXiv:2605.07505v2 公告类型：replace-cross 摘要：我们提出了 LiteGUI，一种用于构建轻量级 GUI 智能体的新框架。由于复杂任务具有长时程特性，且存在多条有效的交互路径，GUI 交互面临独特挑战，而轻量级 GUI 智能体难以有效应对这些挑战。为解决这些问题，LiteGUI 提出了一种基于已记录 GUI 状态运行的两阶段后训练框架。首先，我们提出 Guided On-Policy Distillation：通过从经人工验证的多解动作标注中选择与当前 GUI 状态最匹配的有效动作，在每个 GUI 状态为训练过程提供教师特权指导，同时保持学生模型的在策略 rollout 不变。其次，我们开发了 Multi-Solution Dual-Level GRPO：在考虑每个已记录 GUI 状态存在多个有效动作的同时，将动作级监督与基于历史条件、按状态进行的规划质量监督相结合。这两个阶段共同支持多步 GUI 任务，而无需学生模型复现固定的示范轨迹。我们进一步开发了可扩展的数据生成流程及相应的多解数据集 Lite-Dataset，以支持 GUI 智能体的训练与评估，弥补现有文献中的空白。在多个 GUI 基准上的大量实验表明，与包括 SFT、在策略蒸馏和 GRPO 在内的现有训练范式相比，LiteGUI 能够显著提升轻量级 GUI 智能体的性能；此外，与规模大得多的模型相比，LiteGUI 也实现了具有竞争力甚至更优的性能。
arXiv:2604.07341v4 公告类型：替换交叉 摘要：由于适配新的编程语言（PL）对所需的复杂工程工作量很大，大多数仓库级代码翻译与验证技术都只在单一源语言—目标语言（PL）对上进行了评估。编程智能体能够赋予仓库级代码翻译与验证以语言无关性：它们可以跨多种 PL 综合代码，并自主使用针对各 PL 的现有分析工具。然而，现有最先进的方法尚未提供一种完全自主的智能体方法，用于大规模程序的仓库级代码翻译与验证。本文提出 ReCodeAgent，一种面向语言无关仓库级代码翻译与验证的自主多智能体方法。用户只需提供源 PL 中的项目并指定目标 PL，ReCodeAgent 即可自动翻译并验证整个代码仓库。ReCodeAgent 是首个在多种 PL 上实现高翻译成功率的技术。我们将 ReCodeAgent 与另外四种神经符号和智能体方法进行比较，使用它们翻译 118 个真实世界项目；每个项目平均包含 1,975 行代码（LoC）和 43 个翻译单元。这些项目涵盖 6 种 PL 和 4 个 PL 对。结果表明，在翻译正确性方面，ReCodeAgent 始终优于以往技术：在真实值测试上，测试通过率提高了 60.8%，平均成本为 15.3 美元。我们还对 ReCodeAgent 的运行轨迹进行了以过程为中心的分析，以确认其过程效率。最后，我们研究了设计选择（多智能体架构与单智能体架构）如何影响 ReCodeAgent 的性能：平均而言，测试通过率下降了 40.4%，运行轨迹变长了 28%，并且持续表现出低效率。
arXiv:2603.08706v2 公告类型：replace-cross 摘要：模仿学习（IL）教会语言模型智能体复现专家动作，却无法教会它们区分专家动作与看似合理的错误。自我反思方法让模型接触备选方案，但仍使用监督微调（SFT）去模仿固定的推理过程和动作。我们提出 Agentic Critical Training（ACT），利用带可验证奖励的强化学习（RLVR）直接训练模型判断动作。在每个专家轨迹状态下，ACT 将一个专家动作与一个从初始策略中采样的备选动作配对，并随机化二者的顺序。模型生成自己的推理过程，但只有选择专家动作才能获得奖励。ACT 重用演示数据，不需要参考推理过程，并允许在不同模型规模之间复用动作对。ACT 可作为 IL 之前的预热阶段，也可以选择在其后接入 RL；推理时不需要比较候选项。在 ALFWorld-ID、WebShop 和 ScienceWorld 上使用 Qwen3-8B 和 Olmo-3-7B-Instruct 进行的实验表明，与 IL 相比，ACT 平均提升 5.85 分；与不使用 ACT 的 IL$\to$RL 相比，平均提升 4.12 分，同时还能改善 ALFWorld-OOD 上的表现。在 ScienceWorld 上使用 Olmo 时，完整流程相比 CoT 提示提升 15.36 分，相比不使用 ACT 的 IL$\to$RL 提升 9.23 分。ACT$\to$IL 和完整流程均优于监督式反思基线。采用固定动作对或匹配训练时长的对照实验表明，性能提升源于 ACT 目标，而不是额外数据或训练过程。无需针对推理进行专门的后训练时，独立的 ACT 检查点在 MATH-500 和 GPQA-Diamond 上取得了所评估模型中的最高平均分，表明动作比较能够补充生成能力。
arXiv:2609.33563v2 Announce Type: replace Abstract: World models improve sample efficiency by training policies on imagined trajectories, but their usefulness depends on learning representations that capture the information needed for future control. We study whether self-supervised joint-embedding prediction (JEPA) can provide this learning signal for multi-agent reinforcement learning. We introduce MA-JEPA, a stochastic world model that replaces observation reconstruction with prediction of target representations, enabling model-based multi-agent reinforcement learning with centralized training and decentralized execution. A categorical latent state and a causal Transformer are trained with posterior and action-conditioned dynamics prediction objectives and are then used for actor-critic learning from latent imagination. A training-only joint predictor conditions on all agents' local states and actions to predict each agent's next local observation embedding. These predictions are passed through the same local posterior used during real interaction with a centralized critic that is used only for value learning, with execution remaining decentralized. Our experiments show that this architecture performs strongly on SMAC, matching or exceeding the strongest reported comparator mean win rate on four of eight evaluated maps.
arXiv:2609.24259v3 公告类型：替换 摘要：智能体记忆的有效性最终取决于底层LLM是否能让上下文中的每条记忆对其响应产生适当程度的影响。然而，这种能力一直在很大程度上被忽视。为评估这一能力，我们引入了MemCalib，这是一个基于真实记忆系统场景的基准，用于评估记忆使用并推动优化算法的发展。MemCalib测试集上的结果表明，前沿开源和闭源模型都难以恰当地使用记忆。它们经常过度使用或使用不足，而不是让每条命题的实际使用程度与其目标水平相匹配，从而导致响应存在偏差且质量低下。对包括组相对策略优化和在策略自蒸馏在内的常见后训练算法进行实验进一步揭示了明显的方向性偏差：模型经过训练后，会在一个方向上有所改善，却在另一个方向上恶化。因此，我们提出MemCalib-RL，这是一种有序的双向反事实信用分配算法，通过精确的原子消融将过度使用和使用不足的信号分离开来，并将其信用归因定位到响应词元。跨模型系列和规模（Qwen3-8B、Ministral-3-8B-Instruct和Qwen3.5-35B-A3B）的结果表明，MemCalib-RL取得了最佳的总体性能，同时更好地平衡了过度使用和使用不足；在外部基准评估中，其收益也能泛化到MemCalib之外。进一步实验支持了其设计选择和稳健性，并为其训练动态提供了深入见解。
arXiv:2609.23257v2 Announce Type: replace Abstract: Time series forecasting underpins critical decision-making across diverse domains. While large language models (LLMs) offer promising reasoning capabilities, existing LLM-based time series forecasting approaches either reduce them to numerical predictors that bypass their strengths, or allow direct forecast generation that destabilizes predictions in non-stationary settings. We introduce CTRL, a framework that decouples semantic reasoning from quantitative prediction. A frozen backbone generates base forecasts, while specialized LLM agents function as controllers that analyze backbone prediction errors through decomposed trend, seasonal, and irregular components, grounding reasoning in interpretable temporal structure. Each agent outputs compact control signals that a lightweight residual decoder translates into forecast corrections. CTRL incorporates label-free test-time adaptation that detects distribution shift from input statistics alone and readapts control signals with only 3-24 LLM calls via caching. CTRL is explicitly designed to improve robustness under non-stationary temporal dynamics and distribution shift, while remaining competitive on highly stationary time series where adaptive correction provides limited additional benefit.
arXiv:2606.26294v3 公告类型：替换 摘要：自我改进型智能体在智能体编码基准测试中已达到最先进水平，但其搜索方法假设评估标准是静态不变的。这忽略了进化的一个核心特征：物种会随着环境的变化而适应。我们提出 Red Queen G\"odel Machine（RQGM），一种针对非平稳效用进行递归自我改进的进化框架。该框架使学习型评估器能够与其引导的智能体共同改进。在 DeepSWE 上，RQGM 通过加入互补的“智能体作为评审者”代码审查信号，超越了固定评估器基线：共同进化的评审者对编码器生成的补丁进行评分，以引导搜索。在较低推理投入下，RQGM 编码器在留出任务上的通过率达到 82.1%，而基线为 75.0%；在较高投入下，其表现几乎达到 GPT-6 Astra 模型的水平。在科学论文写作与评审，以及奥林匹克竞赛级别的证明写作与评分任务中，共同进化的评估器能够提供评估标准。以人类 IMO 评分为锚点，一名共同进化的评分器自行撰写阶段性评分标准，并以低 3 倍的搜索成本超越静态基线，推动证明器取得最高平均分。由于 RQGM 能够跨轮次修改搜索目标，因此可以对搜索过程进行正则化。例如，RQGM 通过加入额外的对抗性目标，发现对 AI 作品和人类作品同样严格的评审者，从而减少自我偏好偏差。在这些经过校准的评审者引导下，共同进化的写作者在“智能体作为评审者”评审面板下的接受率达到基线的 1.78 至 1.86 倍。RQGM 使自我改进系统成为可能：智能体与评估器能够在递归过程中相互促进，突破静态评估的限制。
arXiv:2605.22949v4 公告类型：替换 摘要：当协调器比较异构基础模型的答案时，不同应答模型所报告的置信度可能具有不同含义，并且这种含义还会随工作负载变化。本文提出 MARGIN（Multi-Agent Runtime Grading via Incremental Normalisation），一种运行时校准方法。该方法根据观测到的答案结果学习针对不同模型的置信度修正，无需重新训练模型，也不需要留出校准集。MARGIN 在各个置信度区间内跟踪近期准确率和模型给出的置信度，利用二者的比值修正报告的置信度，并将样本稀疏区间的修正结果向模型级估计进行平滑。修正后的分数用于在集体决策中对候选答案加权。评估涵盖代码生成、问答和数学任务，使用包含 18 个模型的模型池，并在分布偏移实验中使用其中的九模型子集。在 BigCodeBench 上，模型平均置信度与准确率呈负相关；在正确/错误回答对中，选择置信度更高的应答模型，其表现低于随机水平。与五种在线校准基线方法相比，这些基线接收相同的反馈，并在每次转换过程中保留所学习的状态，MARGIN 在两次代码生成转换中的分布偏移后期望校准误差均低于这五种方法，在一次问答转换中低于其中四种；剩余的一次问答比较结果尚无定论。在独立的代码生成协同实验中，相较于未经校准的置信度加权，校准提升了正确回答的排序效果，并在三个基准中的两个上使答案选择准确率分别提高了 4.3 和 14.0 个百分点。这些结果表明，当能够获得参与应答模型的正确性反馈时，针对模型的运行时校准有助于应对工作负载变化下的协同决策。
arXiv:2603.07980v2 公告类型：替换 摘要：随着语言模型（LM）从聊天助手发展为能够进行长程、多步骤推理并使用工具的智能体，现有基准测试仍大多局限于结构化或考试式任务，难以满足现实世界中的专业需求。为此，我们推出 \$OneMillion-Bench（\$OMB），这是一个包含400项由专家策划的任务的基准测试，涵盖法律、金融、工业、医疗保健和自然科学领域，旨在评估智能体在具有经济重大影响的场景中的表现。不同于以往研究，该基准要求智能体检索权威来源、解决相互冲突的证据、应用领域特定规则并作出受约束的决策；在这些任务中，正确性不仅取决于最终答案，也取决于推理过程。我们采用基于评分标准的评估协议，从事实准确性、逻辑连贯性、实践可行性和专业合规性等方面进行评分，并聚焦专家级问题，以确保能够对不同智能体进行有意义的区分。总体而言，\$OMB 提供了一个统一的测试平台，用于评估智能体在高度依赖专业领域的场景中的智能体可靠性、专业深度，以及其实践就绪程度的指标。
arXiv:2610.12375v1 公告类型：cross 摘要：Agent 被部署于旅行规划、股票交易和 IT 事故分诊等应用中。在大多数情况下，LLM Agent 几乎不受基于规则的安全防护约束而自主运行，因而可能由于不可逆操作引发成本和安全问题。近期研究要么使用安全防护 Agent 监控行为，要么事后评估日志。前者会为每一步增加成本和延迟；后者则要等到运行结束后才能给出判断，此时 token 已经消耗，损害也已经造成。为解决这一问题，我们提出 OnTrack，一种流式监控机制，将 Agent 的步骤及其依赖关系与已记录的成功运行进行比较，从而在每一步约一毫秒内向用户发出警报或阻止 Agent。我们在三种访问权限逐步降低的情形下研究这一问题：完全访问参考数据（历史运行记录和工具模式）、中等访问权限（仅工具模式）以及无先验知识（仅有生成的步骤日志）。随着数据访问权限降低，OnTrack 的监控能力也随之减弱，其检测范围从计划违规逐渐缩小到识别循环、停滞和重复工具调用。最后，我们使用 SWE-bench 轨迹评估 OnTrack。基于前 8 个步骤，我们的方法对失败轨迹进行排序，使其排名低于成功轨迹的效果优于内容相似度方法（AUROC 提升 +0.057）。采用中止策略后，对于原本会在失败运行中消耗的计算资源，我们节省了约 18%；在被中断的运行中，有 83% 实际上正走向失败（6 次中止中有 5 次判断正确）。
arXiv:2610.12325v1 Announce Type: cross Abstract: Do LLM scientific agents rely only on their initial task context, or do they adapt their decisions in response to experimental feedback? We study this question in neural operator adaptation, where a large language model (LLM) selects fine-tuning configurations under a limited trial budget. Across transfers within and between partial differential equation (PDE) families, the LLM achieves lower held-out test nRMSE than random search and Bayesian optimisation in nearly every matched comparison. Endpoint performance alone cannot distinguish what happens, so we verify each attribution with controlled interventions. Before observing any validation score, the LLM's first configuration already ranks near the top of the corresponding random-search pool, indicating a useful initial bias. A complementary cold-start intervention shows that the selected base learning rate shifts with the PDE description. Once feedback becomes available, reassigning validation scores among evaluated configurations changes the next proposal in every case tested, whereas a value-preserving rewrite produces no comparable aggregate effect. These interventions establish that the LLM's decision-level actions respond to the given task and observed outcomes, showing that it combines a task-dependent prior with sensitivity to experimental feedback.
arXiv:2610.12321v1 Announce Type: cross Abstract: Spatial public goods models show that prescribed movement toward richer locations can generate spatial patterns. We ask how such patterns emerge when agents learn where to move and how learning rates shape their consequences for collective welfare. Fixed populations of cooperators and defectors independently learn movement policies using tabular Q-learning and local observations. Cooperator learning generates clusters around resource peaks, while co-adaptation changes their strength and motion. At a fixed training budget, the largest welfare losses occur when cooperators learn at high rates and defectors at low rates. In part of this regime, learned policies also generate traveling bands supported by a shared directional preference. The conditions supporting travel change with further training, so these patterns reflect training history rather than an established asymptotic outcome. Across the tested learning-rate conditions with cooperator learning, mean collective welfare falls below random movement because increased crowding outweighs gains in resource benefit. Charging agents for the crowding they impose on others during learning recovers much of the welfare loss in the tested conditions. These results connect learning rates to the emergence and welfare costs of spatial organization driven by individual rewards.
arXiv:2610.12281v1 公告类型：交叉收录 摘要：全基因组关联研究发现的疾病相关变异中，超过 90% 位于非编码调控区域，但对其功能的解释仍是基因组医学中一个尚未解决的核心问题。在被提示解释此类变异时，大语言模型经常臆造转录因子（TF）结合变化、捏造实验支持，并赋予统计上可忽略的信号以生物学意义。我们提出 ARGUS（Agentic Regulatory Genomics for an Uncertainty-aware Scientist），将确定性的生物学计算与由 LLM 介导的推理严格分离。ARGUS 将 458 个基于 DNABERT 的 TF 结合模型封装在一个由假设引导的探究循环中：规划器根据当前的不确定性选择证据来源，验证器以确定性方式解释每项观测，中间结果则会改变探究路径。针对位于 8q24 癌症风险位点的变异 rs6983267，同一个规划器为四种 TF 生成了四条不同的探究轨迹。当真实的 ADASTRA 等位基因特异性结合数据（15 项实验，FDR = 0.030）揭示出一个被饱和效应掩盖的模型假阴性时，FOXA1 的判断在 3 步内得到纠正。KLF6 经历了涵盖 ADASTRA、JASPAR 基序分析和 ENCODE cCRE 调控注释的 8 个步骤，最终因间接证据相互矛盾而放弃判断。ADASTRA 返回了满足覆盖度要求但不显著的等位基因检验结果（5 项实验，FDR = 0.65）后，RAD21 在 8 步内放弃判断；SP1 与 FOXA1 一样，具有受饱和效应影响的结合保留预测，但由于该位点不存在直接实验证据，也放弃判断。所有观测均来自真实的 ADASTRA、JASPAR 和 ENCODE cCRE 查询，没有任何模拟数据。对固定优先级规划与 LLM 介导规划的比较表明，LLM 规划器通过舍弃无法判定待检验论断的证据，以更少的工具调用得出了相同的结论。
arXiv:2610.11794v1 Announce Type: cross Abstract: Learning to act in unfamiliar environments requires agents to infer how the world works and revise that understanding as new evidence arrives. Yet limited observations can support multiple world models that explain past interactions but predict different outcomes in unseen states. We introduce Memento 3, building on the Memento series to enable frozen LLM agents to continually learn explicit world models through external memory. The agent maintains a natural-language rulebook as persistent semantic memory, recording revisable hypotheses about environment dynamics while leaving unknown aspects underspecified. It compiles this rulebook into executable code for prediction and planning. Through a continual loop of observation, reflection, rule revision, compilation, and verification, the agent uses prediction errors to refine both the rulebook and its code. Updated code is accepted only when the LLM judges it faithful to the rulebook and cell-exact replay reproduces the observed transitions. We investigate this process as a model-based route to recursive self-improvement (RSI): the agent autonomously explores the environment, revises its world model, and uses verified updates to guide subsequent interaction and learning, while the underlying LLM remains fixed. A population extension maintains multiple world models in parallel, sharing interaction evidence and using their predictions to guide exploration. On ARC-AGI-3, the single-model agent clears every level of all 25 public games, achieves a mean Relative Human Action Efficiency (RHAE) of 100.0, and uses 44% of the human action count. In an Atari Pong case study, a learned feedback controller wins 21:0 in each of three evaluated episodes with different openings, without further LLM calls.
arXiv:2610.11655v1 公告类型：cross 摘要：要改进一个长时程 LLM agent，可以演进围绕冻结模型的运行框架，也可以训练模型权重。我们先让自演进运行框架增强系统，然后将初始运行框架和演进后的运行框架分别与基础权重和训练后权重交叉组合，以了解训练后的模型保留了哪些增益，以及哪些增益仍然需要运行时支持。我们发现，可以从 agent 的失败构成中判断应采取哪种正确手段：按照最先触发的信号标记失败轨迹，可以将过程失败（调用受阻、循环、步骤预算耗尽）与内容失败（已交付的计划质量不佳）区分开来。运行框架演进能够修复前者，它引入的行为可以训练进权重，而内容失败正是权重训练要解决的问题。在 DeepPlanning 上，自演进运行框架循环将 Qwen3.5-4B 的留出集得分从 0.16 提升至 0.30，将 Qwen3.5-9B 的留出集得分从 0.32 提升至 0.44；对于 4B，留出集交付率从 55% 上升到 90%，而内容失败则留待权重处理。在演进后运行框架的轨迹上训练的 LoRA 适配器能够将这一增益内化：在原始运行框架下，两种规模的模型在留出任务上均增加了 +0.13；对于 4B，适配器与运行框架叠加后，留出集得分提升至原来的两倍以上；对于 9B，仅使用适配器就能达到完整演进流程的效果，并将内容失败率从四分之一的轨迹降至二十分之一。使用答案打乱后的轨迹训练的安慰剂适配器，其表现低于基础模型。该循环还迁移到了 WebArena-Lite（在 117 个未见任务上提升 +0.09），其中增益来自模型所看到的内容，适配器不会进一步增加这一增益。其结论是一条应用两次的“先诊断、后干预”规则：先读取失败构成，在运行框架与权重之间进行选择；然后查看已接受的编辑改变了什么，以决定训练哪些增益。除非特别标注，分数均为针对新鲜锚点进行四次 rollout 的平均值，且在同一晚完成评测；实验涵盖来自六个模型家族的八个模型和两个基准。
arXiv:2610.11482v1 公告类型：cross 摘要：背景：大型语言模型（LLM）代理正日益被用作软件和数据工程助手，但关于可在本地部署的开放权重代理的证据仍然有限。现有评估通常侧重于文本响应或孤立的代码生成，而不是完整工程制品的有效性。目标：我们评估本地 LLM 代理能否生成正确且可复现的数据工程制品，量化闭环工作空间条件的影响，并考察模型规模、架构、量化、运行时、工具使用和失败之间的权衡。方法：我们引入了一个包含 15 项移动工作流任务的基准，涵盖数据发现、连接器、交通数据馈送处理、语义丰富、特征工程、验证、可视化和报告生成。确定性检查器用于评估生成的脚本、表格、结构化文件、图表和报告。在单次生成和闭环两种条件下，对 10 种本地配置进行评估；每种模型、模式和任务重复 5 次，在消费级 GPU 上共获得 1,500 次评分尝试。结果：在参数量超过 20 亿的模型中，与单次生成相比，工作空间条件使通过率提高了 26.7–52.0 个百分点。最强配置达到了 85.3% 的制品级成功率；一个经过量化的 90 亿参数模型在约 6.5 GB 的内存占用下达到了 69.3% 的成功率。当中间制品暴露出代理可以检查和修复的错误时，收益最大。结论：本地开放权重代理能够支持软件密集型数据工程工作中具有实际意义的一部分，但其可靠性取决于模型能力、任务可验证性和确定性验证。该基准为在工程工作流中采用代理配置之前评估完整代理配置提供了一种可复现的方法。
arXiv:2610.11464v1 公告类型：交叉 摘要：我们改变了智能体：它真的变得更好了吗？每个自我改进的智能体循环都会数百次回答这个问题，而每次答案都来自一个验证器。在开放式任务中不存在现成的验证器，因此循环只能依赖手写评分标准，或让一个与被评估模型类似的 LLM 充当简单评判者来给输出评分，这很容易引发奖励破解和共同盲点。我们让验证器成为进化对象：它是一个可检查的表达式，由少量、主要是确定性的问题检测器构成；这些检测器从聚类后的失败案例中合成，在诞生时接受门控，并依据其与包含十个项目的锚定参考集的一致性，以及对未标注输出的共识进行选择，而绝不依据智能体的得分进行选择。在 MBPP+ 上，相较于人工编写的初始组合，它在留出集上的一致性提升了 0.21，且这一结果对每个初始种子都成立；最终表现也超过了其中包含的简单 LLM 评判者。有一项发现应当改变我们验证共同进化验证器的方式：移除锚定约束后，验证器会坍缩成一个无意义的“永远通过”评分器，但这个坍缩后的验证器训练技能的效果却同样好。下游任务得分无法证明自我进化验证器是可靠的。得分确实可以回答“是否足够”的问题，而在此处，进化验证器可以作为替代方案：Double Ratchet 将验证器与一个由生命周期管理的技能循环配对，在代码生成、企业级 text-to-SQL 和无参考报告生成任务中，保留了同一循环在使用真实标签或评分标准时所获得提升的 88%–110%。当进化出的技能钻了报告评分标准的空子时，外部评判者发现了这一问题，并通过添加一个检测器修复了它；但在获得任务契约之前，该评判者本身也是错误的。
arXiv:2610.11447v1 公告类型：cross 摘要：大型语言模型（LLM）正日益被部署为编码智能体，用于编辑文件、运行构建和测试、检查执行结果，并迭代修复软件。嵌入式固件是一个要求很高的目标，因为正确性不仅取决于静态源代码质量，还取决于在感知、时序和安全约束下的闭环行为。然而，目前对嵌入式智能体的评估仍然有限，且通常侧重于一次性生成或离线正确性。我们提出了一个用于闭环评估嵌入式编码智能体的基准。每项任务都提供纯文本工程描述、受约束的工作空间，以及可见的构建与运行时接口。智能体必须将需求转化为实现和自验证步骤，然后持续迭代，直到实现所要求的设备行为。该测试集包含五项嵌入式控制任务和四种反馈场景：一次性生成、真实的自验证、CI 风格的红/绿反馈，以及 oracle 风格的详细反馈。为确保可复现性，实现目标为模拟的 ESP32 固件。我们在五项任务和四种场景下评估了七种 GPT-family 和 Qwen-family 配置，每种条件重复三次，共进行 420 次运行。gpt-5.4 在所评估的配置中通过率最高，但尚未达到该基准的饱和水平；qwen3.5-27B 是观测到的最强本地模型；而规模更小的本地模型在通过率和搜索效率方面均显著下降。这些结果表明，能力出色的本地嵌入式编码智能体正在逐渐出现。
arXiv:2610.11358v1 Announce Type: cross Abstract: Cross-model KV-cache reuse remains a key challenge in modern LLM serving. Coding agents and multi-model systems increasingly route a shared context across models: a user may switch models mid-session, or a cascade may escalate a difficult query. Because KV caches contain model-specific representations, each switch typically forces the receiving model to prefill the entire context from scratch. Recent work shows that closed-form linear maps can translate KV caches between models in the same family, but transfer accuracy degrades as the model-size gap widens. In this paper, we establish that these transfer failures are concentrated in a small subset of information-dense tokens. To bridge this gap, we introduce RaReCache, a framework that enables a large target model to decode accurately from a cache prefilled by a much smaller source via selective recomputation. RaReCache identifies these critical positions using a novel rank disagreement metric, scoring each token by the energy of its mapped KV in output directions weakly supported by the calibration data. Across two model families and five benchmarks, on a 23x parameter gap (Qwen3-0.6B to 14B) recomputing just 30% of positions retains 95-99% of the target accuracy, whereas on a 8.8x gap (Llama3-8B to 70B), recomputing 40% retains 96.5% of the target accuracy. RaReCache largely removes sensitivity to source-model size, and achieves up to a 3.04x prefill speedup. For online serving, it handles 1.8x the request throughput of target prefill on a single GPU, and at the target's saturation load, reduces median and 99th-percentile time-to-first-token (TTFT) by 5.0x and 6.4x respectively, with a 30% recompute budget. RaReCache establishes an efficient serving paradigm where small models prefill on behalf of massive targets, enabling large models to recompute only critical tokens, drastically reducing prefill latency.
arXiv:2610.11310v1 发布类型：cross 摘要：尽管在动态第一视角环境中进行3D空间推理对于具身智能至关重要，但音频视觉大型语言模型（AV-LLMs）缺乏直接从原始感知流中处理并内化全局几何结构的明确机制。现有方法要么需要高成本的微调，要么未能充分利用模型的跨模态推理能力。本文提出FloorSAV，一种通过渲染动态2D平面地图来显式建立空间音频视觉上下文的全新框架。我们融合3D点云、相机轨迹、空间音频线索以及经过语义定位的物体地标，将该平面地图作为与第一视角视频同步的流注入AV-LLM。借助平面地图解读引导，AV-LLMs利用其多模态能力，在单次推理中联合分析视觉、听觉和几何线索。我们进一步引入SAVED-Bench（Spatial Audio-Visual Egocentric Benchmark with Dynamic Agents），构建真实世界场景中空间能力的关键任务：动态相对关系、区域和路径推理问答。FloorSAV在SAVED-Bench和SAVVY-Bench的多项任务中提升了AV-LLMs的空间推理能力。基于真实值平面地图的研究表明，凭借准确的空间信息，FloorSAV具有巨大潜力。
arXiv:2610.11050v1 公告类型：cross 摘要：计算机使用智能体能够完成复杂任务，这使得自动评判器的使用日益增多，可用于训练或在无人参与的情况下进行评估。尽管自动评判器具有灵活性，但其在跨越多个应用程序的长任务中的可靠性仍不明确。一条由长序列屏幕截图和操作组成的轨迹，表面上可能已经完成任务，但实际上可能违反指令中的约束，或引入不必要的副作用。为了识别这些错误，评判器需要根据用户指令检查轨迹。为此，我们推出了 AgentHorizon，这是一个包含 1,373 个计算机使用任务（指令—轨迹对）的基准，数据来自 166 小时、涵盖三种操作系统的人类记录轨迹。通过记录针对高度相关指令的轨迹，我们可以交换指令来构造负面任务。这种成对设计评估了评判器区分成功轨迹与完成了相似（但不兼容）请求的轨迹的能力。我们发布了该基准的三个划分：前沿划分 AgentHorizon (AH)、简化划分 AgentHorizon-Simple (AH-S)，以及开发划分 AgentHorizon-Development (AH-D)。此外，我们通过两种方式评估了十一个评判器：（1）直接传入完整轨迹（最多包含 300 个屏幕截图和操作）；（2）在五种智能体运行框架中将其用作编程智能体。我们发现，表现最佳的智能体式评判器 GPT-5.5 在 AH 子集上达到了 80.9% 的平衡准确率。我们发现，工具使用能够提升某些模型的性能，但会导致开放权重模型的表现变差；同时，不同评判器在接受有效轨迹和拒绝失败轨迹方面的能力存在巨大差异。我们的研究结果凸显了对这类评判器的需求：它们必须能够在漫长的交互历史中定位并验证任务已被正确完成这一通常隐藏的证据。
arXiv:2610.11030v1 Announce Type: cross Abstract: Tool-using LLM agents violate the policies they are deployed to enforce, often silently. Prior defenses hand-write rules, query an LLM verifier per action, or compile policies through heavyweight formal machinery. Naive compilation fails: extracted rules block the tool satisfying their own precondition, or read arguments their tool lacks. NOMOS, a four-pass compiler, turns a natural-language policy into a deterministic tool-call gate; static verification with tool-schema-level checks alone (no prover, solver, or LLM) repairs or rejects 37% (airline) and 13% (retail) of candidates, without which most shipped rules are inoperable. Replaying compiled rules over undefended transcripts flags bindings that refuse legitimate work (a development binding refused 95.9% of task-passing calls); no evaluation binding is flagged. On $\tau^2$-bench the gate cuts violations of reference-encoded clauses among state-changing calls from 66.3% to 2.6% (airline) and 30.8% to 6.9% (retail), raising airline task success significantly for $2 \le k \le 4$; a 26B on-premise compilation is not significantly worse than hand-written or frontier-compiled rules. Unlike AgentDojo's shipped defenses, it reaches a zero attack success rate (ASR) on banking, where nine attack families collapse onto three structural rules. On the other three suites its ASR is at most 3.6%, from goals with no tool call to govern and one write admitted by a binding weaker than its clause; a second agent model, Llama-3.3-70B, reproduces the effect on both benchmarks. Decisions take microseconds without an LLM call, at a domain-dependent benign-utility cost; compilation runs on-premise on open-weight gemma-4-26B.
arXiv:2610.11007v1 Announce Type: cross Abstract: At the start of every session, LLM agents load a fixed context file, such as $\texttt{AGENTS.md}$. Each loaded token in the file is charged again in every later round of the session, and these files can degrade performance as they grow in size. However, in practice, human or automated curators usually grow these files by appending. We formulate context curation as a capacitated assortment problem. Instructions consume tokens under a finite attention capacity; adding an instruction never raises the compliance of the others, while retained instructions incur a per-session setup cost. We prove an upper bound on the optimal file size, regardless of the number of available candidate instructions, and that appending every instruction with positive standalone value can be arbitrarily worse in net value than selecting an optimal subset. A token budget also limits the loss when the token price is underestimated. We then examine what can be learned from past sessions and how this information can guide decisions to add or remove instructions. Feedback is inherently censored: the benefits and harms of loaded instructions are observable, whereas missing instructions generate feedback only when their absence causes harm. In this setting, we show that deleting instructions ignored by agents can inevitably remove helpful ones. We characterize how much evidence should be collected before adding an instruction. Besides, we bound regret when human reviewers can inspect only a limited number of edits per period. Empirical experiments further show that irrelevant rules drawn from real context files reduce language-model compliance.
arXiv:2610.10942v1 公告类型：cross 摘要：强化学习环境如今已成为后训练阶段提升大语言模型（LLM）能力的主要手段之一，但大多数智能体基准仍然是静态的：世界只有在智能体采取行动时才会推进，奖励是终局判定，及格线也被任意设定。我们推出 StoreBench，这是一种实时电商环境，智能体在生产级电商后端上运营一家中型在线服装店，在不确定性下考验长期规划能力和经济判断力。顾客全天候下单，供应商重新定价并可能失效，市场冲击也会在几乎没有预警或完全没有预警的情况下到来。智能体通过与人类运营者所使用的相同的 29 个商家工具进行操作，并受窗口化操作预算约束，使模拟时间成为所采取行动的函数，因此模型延迟不会影响模拟时间。及格阈值依据脚本化锚定策略进行校准，奖励机制经过加固以抵御一系列奖励投机行为，并且在给定行动序列的情况下，每个回合都能被完全一致地重放。我们在 11 个持续 30 至 45 天的场景和一个完整模拟年度上，对七个前沿 LLM 进行了评估；所有评估均使用三个世界随机种子，并匹配推理投入。平均而言，没有任何模型能达到脚本化 smart-triage 策略的水平：表现最佳的 DeepSeek-V4-Pro 在任务—随机种子单元格中的通过率为 49%，而启发式策略为 97%。使用相同工具和预算的人类专家得分超过所有模型（综合平均分为 0.708，而模型为 0.700）。在 Claude Code 工具框架下运行完整模拟年度时，大多数模型的性能都出现了显著提升。在一次 GRPO 后训练实验中，仅使用五个互不重叠的任务训练的 Qwen3.5-27B，将其在留出的评估任务上的综合平均分从 0.136 提升至 0.373。我们发布了五个训练集示例任务、十条示例轨迹，以及评分和验证工具；为避免基准受到污染，完整环境和评估套件暂不公开。
arXiv:2610.10858v1 Announce Type: cross Abstract: Analog/RF circuits remain the critical interface between digital computation and the physical world, and emerging standards from Wi-Fi 7 to 6G place stringent demands on them, yet analog/RF design remains one of the most labor-intensive steps in chip development. We present RFChipAgent, a first-of-its-kind multi-agent flow of large language model (LLM) agents for end-to-end analog/RF circuit design automation, in which AI agents collaboratively orchestrate the complete design flow under human supervision. RFChipAgent is built around four technical pillars. First, a multimodal retrieval-augmented generation (RAG) subsystem with private per-document FAISS indexing extracts design knowledge from existing engineering documentation. Second, a topology agent drives topology selection, and a schematic and testbench agent automates circuit and testbench assembly. Third, a closed-loop hybrid circuit-sizing engine combines Tree-structured Parzen Estimator (TPE) and CMA-ES optimization, evaluating every candidate in a simulator-in-the-loop framework. Fourth, a trust-scored simulation database accumulates verified performance data and builds an adaptive optimization model that informs subsequent trials. We validate RFChipAgent on a family of GF22FDSOI 60 GHz wideband mm-wave low-noise amplifier (LNA) topologies, demonstrating automated topology generation, specification-driven design-space exploration, and simulator-guided optimization. Experimental results show substantial reductions in design effort while maintaining signoff-quality verification. This work establishes a foundation for LLM-driven multi-agent electronic design automation (EDA) for analog/RF circuits.
arXiv:2610.10833v1 发布类型：cross 摘要：我们研究小型 LLM 智能体能否在明确的实际时间预算下有效运行，即既遵守分配的运行时长，又高效利用可用时间。我们在 MLE-Bench Lite 的五项竞赛中评估 Qwen3.6-27B，并在 Zork I（Jericho）上评估 Qwen3-4B。这两个智能体基准测试中，额外的计算时间都能切实提升性能。在最简单的设置中，预算仅在提示词中说明，智能体无法将所述预算转化为对时间的受控使用。这些失败源于时间感知方面的不足：一方面，测试框架不提供计时反馈；另一方面，智能体无法可靠地预估行动所需的时长，也没有学会将可用时间映射到合适策略。我们研究了两类互补的干预措施：通过测试框架公开计时信息并强制执行截止时间的机制，以及采用预算感知奖励的强化学习。通过测试框架注入计时信息后，Qwen3.6-27B 对预算的遵守程度显著提升，且性能没有可测量的损失；加入强制执行机制后，预算遵守程度进一步提高。在 Zork I 上，使用 GRPO 的强化学习实现了近乎完美的预算遵守，并且能够泛化到训练期间未见过的保留预算；但在 MLE-Bench 上，与未经训练的测试框架相比，它并未提升任务性能。当智能体能够遵守预算后，它们仍然无法利用额外时间提升任务性能。经过强化学习训练的策略学会了何时停止，但往往会通过重复行动来填满额外时间；而在多个预算上进行 GRPO 训练，往往会使策略趋向于为最短预算学到的策略。我们的结果揭示了时间遵守与高效时间分配之间的差距，而这仍然是预算条件型智能体面临的核心挑战。
arXiv:2610.10786v1 Announce Type: cross Abstract: Planning is increasingly important for long-horizon agents, where successful execution requires coordinating subgoals, tool use, and intermediate outcomes over many steps. Yet assumptions made during planning may be invalidated by the environment, tools may return unexpected results, or actions may fail. Effective agents must therefore not only generate plans, but also revise them. Such revisions often affect only part of a plan, leaving the preceding and subsequent structure intact. Rather than regenerate the entire plan and risk unnecessary changes, repair can regenerate the affected region conditioned on the preserved prefix and suffix. We introduce Plan-and-Patch, a plan-and-act framework in which a diffusion language model (dLLM) generates a structured, program-like plan through parallel unmasking and repairs it by filling in selected regions while keeping the surrounding steps fixed. We compare DreamReasoner-8B and Qwen3-8B as diffusion and autoregressive (AR) planners. On Natural Plan without task-specific training, diffusion (53.7%) achieves nearly twice the plan repair success rate of AR (27.0%). After task-specific training on agentic benchmarks, ALFWorld and TextCraft, the planners achieve similar observed success in plan generation, while diffusion reduces mean plan-generation latency by 39-46% relative to AR. Our results show that Plan-and-Patch provides a framework for faster plan generation and effective plan repair in long-horizon agents.
arXiv:2610.10768v1 Announce Type: cross Abstract: The global challenge of climate change has driven significant steps to reduce CO2 emissions, guided by international agreements like the Paris Agreement of 2015. Acting too slowly could result in future losses and reputational damage, while moving too quickly could jeopardize shareholder value due to the marginal profitability or potential losses due to technology immaturity of many renewable projects. To navigate this complex transition, energy companies must adopt Sequential Decision Making (SDM) strategies to maximize value creation from decision flexibility under uncertainties. To support this, we developed a custom simulation environment to model the dynamic energy landscape up to 2050. Building on this, we designed a multi-criteria SDM framework that explores various decision strategies related to different portfolios for allocating funds across three sectors: oil & gas, renewables, and CO2 reduction. It aims to maximize value during the transition while accounting for uncertainties in productions, energy prices, and costs. This framework has three objectives: maximizing profit, minimizing CO2 social costs, and enhancing competitive advantage in the renewable energy sector. This research evaluates the use of Reinforcement Learning (RL) to identify optimal investment policies within the defined SDM framework. The agent's sequential decisions shape a virtual dynamic environment by influencing key variables such as oil and gas production, renewable energy output, CO2 emissions, and revenues. Through repeated interaction, the RL algorithm explores the state space and learns an optimal policy under uncertainty. We benchmark the RL strategy against a set of manually defined baseline policies and find it consistently outperforms them in adaptability and long-term value creation.
arXiv:2610.12445v1 Announce Type: new Abstract: Recent incidents have highlighted the challenge of monitoring LLM agents and the danger of models deceiving people. We show that white-box deception detection via probes can be scaled up to frontier monitoring settings by collecting the largest deception dataset to date for training probes and introducing a novel probe architecture which can aggregate information across many layers and tokens. Our probes achieve 98.8% AUC in SHADE-Arena, surpassing an Opus 5.5 text-monitoring baseline, and show improved efficacy as the underlying model is scaled up. To push our probes to their limit, we test them on several cases where deception cannot be determined from the context alone. In these cases, which we refer to as introspective deception, the ground truth can only be determined through careful elicitation or thorough knowledge of a model's training data. In one such evaluation, we show that probes can distinguish transcripts containing a model's true hidden goal from other goals with an AUC of up to 99.7%. Our probes also readily detect deception on prominent open-weight models which lie about politically sensitive topics, and about their beliefs when put under pressure. We release our training dataset, dubbed FIBS, to help drive frontier deployment of effective probes, and encourage the community to expand upon it with further examples of deception and sabotage.
arXiv:2610.12190v1 Announce Type: new Abstract: As claims about recursive self-improvement (RSI) and artificial general intelligence (AGI) proliferate, we ask a simple question: do frontier AI models have a sense of data, i.e., can they reliably select the right data for training? We introduce DataSense-Bench to study this capability through the fundamental problem of data selection and performance forecasting in machine learning. We ask AI agents to select and rank candidate training subsets that can be used to fine-tune a small LLM model. Agents are allowed to inspect the data, write and execute analysis code, and run model forward passes, but can not train the model or access the actual evaluation tasks. We then fine-tune the base model on each selected subset and evaluate its post-training performance under a standardized protocol. We instantiate the benchmark in terminal problem solving and tool use, selecting trajectories from OpenThoughts-Agent and EnvScaler and evaluating on TBLite and BFCL, respectively. We then evaluate the agents along two complementary dimensions: the post-training performance of the top-ranked subset, reflecting the ability to identify high-value training data, and ranking accuracy, reflecting the ability to predict the relative performance of the selected subsets. In our experiments, selection gains over random selection are limited; agents do not reliably rank their selected groups, and ranking ability does not hold consistently across tasks: Astra identifies the best group in all three tool-use runs but in only one of three terminal runs. Analysis of execution traces on both tasks shows that agents often use similar data signals while interpreting their training value differently.
arXiv:2610.12183v1 Announce Type: new Abstract: Black-box optimization (BBO) arises in many scientific and engineering problems where objective evaluations are expensive and limited. Recent large language model (LLM) agents offer a new way to approach BBO by combining task semantics, computation, optimization tools, and feedback-driven decision making, showing great potential due to the integration with mathematically rigorous tools. However, existing agentic BBO studies use different task domains and system configurations, making their results difficult to compare and the effects of individual design choices hard to isolate. We therefore introduce AgenticBBO-Bench, a cross-domain benchmark for agentic BBO spanning synthetic functions, hyperparameter optimization, database tuning, chip design, and molecular design under a unified finite-budget evaluation protocol. In our experiments, agentic BBO achieves higher family-averaged scores than direct LLM-based methods in all five domains and outperforms the best numerical optimizers in four. We further study three factors shaping agent performance: optimization tools, task information and prior knowledge, and the role of the LLM during search. Our results show that additional numerical tools do not consistently improve performance, task semantics are broadly useful while more specific priors are less reliable, and numerical optimizers can effectively absorb gains from search trajectories established by the agent. Finally, we introduce a five-task frontier challenge within AgenticBBO-Bench and evaluate seven LLMs under the Codex agent harness, where GPT-6 Astra and DeepSeek-V4.1-Flash lie on the Pareto frontier of performance and cost among the evaluated models. Our code is available at https://github.com/lamda-bbo/agentic-bbo.
arXiv:2610.12002v1 Announce Type: new Abstract: Test-time training (TTT) adapts an LLM's parameters using signals derived from test inputs, and can make striking improvements in pre-specified settings such as IMO competitions or designated open problems. By turning deployment experience into parameter updates, TTT provides a direct mechanism for model-level self-improvement. Yet TTT is not universally beneficial: each TTT algorithm works in different settings, and applying an ill-suited method could waste test-time compute or even damage model performance. Therefore, such parameter-level self-improvement requires agency: the model must decide when TTT is warranted, which algorithm to invoke, and whether an existing skill can be reused. To fill this gap, we introduce Agentic-TTT, which learns a test-time policy to govern those decisions. Agentic-TTT turns TTT procedures into callable tools, treats accumulated skills as an evolving deployment environment, and trains its policy using the observed utility gains from its decisions. On our benchmark, Agentic-TTT nearly doubles the utility over the backbone model, learns to trade off utility against compute, and generalizes to domains unseen during training. Together, these results point toward autonomous self-improvement: models that can decide how to learn from their own deployment experience.
arXiv:2610.11716v1 Announce Type: new Abstract: We describe a 4-tensor attention model that predicts the next semantic state of a scene, for video generation and robot planning. A window of states has positions (x, t) and two fibers, a semantic fiber and a temporal-context fiber, and one softmax normalizes attention jointly over the window. Frames and an agent's situation are written as those states; the encoder, the renderer, and the planner remain outside the update. To test the update on its own, we train on ROCStories, where each window poses the same next-sentence task at the semantic layer. On the validation split, with one seed per setting, the last-sentence cross-entropy on the three matched settings is lower for the 4-tensor model than for a free-running one-dimensional transformer by 5.3% at H=2, L=2, by 2.6% at H=4, L=2, and by 2.4% at H=4, L=3. At H=4, L=2 the parameter counts are nearly the same, 172.5M and 175.9M. On the same two GPUs that 4-tensor run finished in 2.4 hours and the baseline run in 45.2 hours; the baseline is trained by free-running decoding, one sequential forward pass per target token.
arXiv:2610.11555v1 Announce Type: new Abstract: We study personalized federated linear stochastic approximation (LSA), a framework which notably encompass personalized temporal difference learning. In this setting, heterogeneous agents collaborate to solve distinct linear fixed-point equations, each corresponding to an agent-specific learning problem. A central open question in personalized learning is whether a single method can adapt to an unknown level of heterogeneity by converging to each agent's personalized solution in all regimes while achieving a linear speedup in the number of agents when their learning problems are sufficiently similar. We answer this question affirmatively by introducing PF-LSA, a minimalist algorithm that mixes each agent's local stochastic update with the average update across agents, at no additional computational cost relative to standard federated methods. We prove that PF-LSA, achieves best-of-both-worlds guarantees without any prior knowledge on the level of heterogeneity. Our analysis is based on a sharp decomposition of the error into consensus and disagreement components. The consensus error decays rapidly, whereas the disagreement error decays more slowly but becomes negligible in low-heterogeneity regimes.
arXiv:2610.11523v1 Announce Type: new Abstract: Federated reinforcement learning (FRL) enables distributed agents to collaboratively train decision-making policies, but its decentralized training process also exposes global policy learning to Byzantine manipulation. Existing poisoning attacks primarily focus on how to construct malicious updates, while trajectory-level intervention timing remains largely implicit. In sequential decision making, however, where an intervention is applied can alter subsequent trajectories and learning signals. Through controlled experiments, we find that changing the selected trajectory states materially alters attack efficacy even when the malicious-update construction is fixed. We therefore identify when as a distinct attack dimension and introduce the Viability-constrained Behavioral Steering Attack (V-BSA), which uses local policy uncertainty to select sparse intervention states and applies envelope-constrained behavioral steering. Across discrete-action benchmarks, V-BSA achieves substantial degradation against robust aggregators and ensemble defenses with only a fraction of the interventions used by dense poisoning, while revealing task- and aggregation-dependent boundaries. Overall, our results highlight intervention timing as a distinct dimension of sequential robustness in FRL. The code is available at https://github.com/Yodeesy/V-BSA
arXiv:2610.11458v1 Announce Type: new Abstract: AI research progress can be viewed as the interaction between two processes: benchmark creation and method discovery. Historically, both were driven by human intelligence. However, recent advances in AI have accelerated automated method discovery, while automated benchmark creation has received comparatively less attention. To enable self-advancing systems, we propose Generative Adversarial Loop (GAL), a generator-discriminator framework alternating between two agentic searches: (1) a discriminator that generates adversarial data to expose weaknesses in current systems, and (2) a generator that discovers algorithms to overcome them. We apply this framework to approximation algorithms for efficient inference. Unlike existing auto research systems, which primarily focus on algorithm discovery, GAL introduces a discriminator agent that automates goalpost setting by continually searching for weaknesses in the current algorithm. We demonstrate adversarial data generation across four tasks: KV compression, sparse video generation, sparse attention, and context extension, where the discriminator identifies weaknesses in state of the art techniques. We further show that GAL enables autonomous improvement, with newly discovered algorithms improving not only on adversarially generated data, but also on established benchmarks. Specifically, GAL improves CompactorPress on KV compression with Qwen3-4B at 4x, raising performance on the discriminator dataset from 0.35 to 0.97, while also outperforming RULER-HARD (+0.77 pts). For context extension, GAL boosts Dual Chunk Attention from 0.20 to 0.90 on the discriminator dataset, while yielding gains on standard benchmarks(ScienceFiction (+6 pts) and PG19 32K (-0.33 PPL)). GAL thus provides a path toward autonomous goalpost setting and algorithmic improvement, where AI systems continually discover their own weaknesses and develop methods to overcome them.
arXiv:2610.11373v1 Announce Type: new Abstract: Continual learning remains challenging for large language models, which must enable models to acquire new skills and knowledge without degrading existing capabilities. Existing approaches typically address this challenge by carefully designing how model parameters are updated. In contrast, prompt optimization avoids costly parameter updates while achieving competitive or even superior performance to reinforcement learning methods such as GRPO on individual knowledge-intensive and reasoning tasks. This raises a natural question: \textit{Can prompt optimization, as an efficient adaptation approach, be directly applied to continual learning?} Our analysis shows that, under sequential task adaptation, it suffers from catastrophic forgetting, while optimized prompts accumulate rules that overfit to local task distributions. To address these limitations, we propose \emph{Evolving Functional REpertoires} (EFRE), which replaces a single prompt with a repertoire of functions that evolves as new tasks arrive: compatible updates refine existing functions, while conflicting updates trigger the emergence of new ones. On a three-task continual-learning stream, EFRE achieves a final average performance 7.50 percentage points higher than GRPO. Moreover, after adaptation to the Bio task, its performance on FinQA decreases by only 1.56 percentage points, compared with 25.10 percentage points for the base prompt optimization method. We further instantiate EFRE in a minimal agent system and observe consistent improvements across different backbone models. Overall, these results demonstrate EFRE's strong performance in continual learning for large language models and highlight its substantial potential for continual learning in advanced agent systems.
arXiv:2610.11366v1 Announce Type: new Abstract: Spatial coding agents significantly improve spatial reasoning in Multimodal Large Language Models (MLLMs) by using external tools to generate verified execution traces. However, this paradigm inherently suffers from prohibitive inference-time overhead and external dependencies. In this paper, we explore whether an MLLM can internalize this agentic capability to operate entirely tool-free. We begin with a simple observation: prompting an MLLM with summarized execution traces of a spatial coding agent naturally unlocks the model's internal spatial Chain-of-Thought (CoT). Motivated by this, we introduce SpatialOPSD, an on-policy self-distillation framework that internalizes spatial reasoning into a standalone MLLM by formulating verified agent traces as privileged information. To mitigate privileged-information leakage during distillation, we introduce Repetition-Aware Distillation, which combines repetition masking with unlikelihood regularization. Experiments across multiple benchmarks demonstrate that self-distilling SpatialOPSD achieves higher average accuracy than SFT and GRPO on both spatial and OOD datasets, exhibiting superior performance and generalization.
arXiv:2610.11152v1 Announce Type: new Abstract: LLM agents increasingly improve at inference time by accumulating experience in context rather than by updating parameters. This process is often described as in-context reinforcement learning (ICRL). Whether in-context learning (ICL) can actually play the role of RL, however, has not been tested. We study this question in its simplest form, direct ICRL, where the model conditions directly on raw trajectory-reward pairs, and ask whether the reward acts as a learning signal. Through controlled experiments on four benchmarks across six models, we find that the reward is read, but its effect is small: flipping, randomizing, or removing the reward leaves the improvement curve almost unchanged, and this holds even under meta-prompts that explicitly instruct the model to explore, exploit, or reason over rewards. Trajectories drive improvement, but not through their semantic content: shuffled or corrupted trajectories work as well as real ones. These patterns closely mirror those known in ICL, suggesting that direct ICRL is better understood as a special case of ICL than as inference-time RL. This reframing has implications for agent memory design: ICL factors such as input distribution and demonstrations may matter more than RL elements such as reward shaping and exploration.
arXiv:2610.11146v1 公告类型：新论文 摘要：冷启动信用评分——在标注数据稀缺、特征较弱或模型容量极小的条件下部署模型——是金融机器学习中反复出现的问题。当新的信贷产品上线时，带标注的违约数据稀缺，特征流水线尚不成熟，而模型必须以极小的容量部署，以避免过拟合。标准应对措施依赖的仍是同一批有限数据；真正需要的是一种以领域知识为基础的外部正则化来源。我们提出 Ranking Prior Alignment，这是一个模型无关框架，通过温度缩放的 KL 散度损失，将外部排序先验（来自领域专家、教师模型或 LLM）蒸馏到任意评分模型中。该框架通过统一公式 L = L_task + gamma(t) * KL(P_agent || P_model)，将神经网络架构（MIL 注意力机制）与基于树的架构（XGBoost 自定义目标函数）统一起来，其中 gamma(t) 遵循指数衰减调度。该方法在推理时无需外部模型，其基于树的实现可容忍高达 eta = 0.5 的标注噪声。在一个包含超过 150 万商户的工业数据集上，MIL 对齐在使用 3,000 个包时，7/7 个评估项均取得正向提升（1 项 ID 指标 + 3 项 OOT 指标 + 3 项退化指标；在 OOT-1 上，Delta AUC 峰值为 +0.020），而 XGBoost 消融实验在使用 300 个包时，9/9 项指标均取得正向提升。在公开 Amex 数据集上进行的跨数据集验证显示，5/5 个交叉验证折均取得正向提升（平均 Delta AUC = +0.041）。四类模型（MIL、XGBoost、LightGBM、Logistic Regression）和四种教师架构的实验表明，该框架既不依赖特定模型，也不依赖特定先验来源。我们进一步观察到，对齐收益呈现逆向缩放规律：随着数据丰富度 N、模型容量 C 和特征质量 Q 的下降，收益反而增加，这有助于从业者判断何时值得投入先验标注。
arXiv:2610.11140v1 公告类型：新投稿 摘要：临床诊断本质上是一个序贯过程：只有在预期新增证据能够消除诊断不确定性时，临床医生才会从低成本检查逐步转向高成本检查。我们提出 ActiveMedAgent，这是一个将这种成本感知的序贯逻辑引入多模态医学人工智能的框架。在使用参数冻结、通过 API 访问的视觉语言模型的情况下，ActiveMedAgent 跟踪候选诊断的概率分布，并以每一步的诊断效用减去成本，对每次信息获取进行评分。随后，基于这些已评分的轨迹离线训练一个轻量级 MLP 控制器，使其学习何时请求更多证据，以及何时作出最终诊断。在三个常用基准上，基于轨迹的策略学习始终优于无引导的信息获取方法和全模态基线。值得注意的是，我们发现了一种信息过载效应。在 175 个病例中，该智能体利用更少的信息通道得出了正确诊断，而全模态基线却未能正确诊断，这表明，学习省略哪些信息可能与学习获取哪些信息同样重要。
arXiv:2610.10778v1 发布类型：新 摘要：长期运行的智能体在不断积累经验的过程中，面临日益增长的存储需求。世界模型能够捕捉可复用的规律，从而减少为每条经验存储的信息量。我们提出了以世界模型为条件的记忆分配问题，并引入 MemoWM：一个利用共享预测来压缩保留信息并重建省略内容的框架。其任务感知的分配规则在重建错误的预期影响与存储成本之间取得平衡，保留那些具有预测先验之外下游价值的信息。在五个长期智能体记忆基准测试中，MemoWM 达到 42.42\% 的平均答案准确率，比最强基线高出 2.62 个百分点；与此同时，相较于存储效率最高的基线 MIRIX，平均特定经验存储量减少了 53.9\%。进一步分析表明，在任务质量相当的情况下，更强的世界模型能够降低每条经验的存储量。将模型参数纳入考量后，我们发现共享模型容量与反复产生的存储成本之间存在权衡；随着保留的交互数量增加，使总存储量最小化的模型容量也会随之增大。我们的代码可在 https://github.com/Feld-maxiu/MemoWM 获取。
对抗性游戏推动了从启发式搜索到强化学习的发展，但从有限样本中学习并适应策略仍然充满挑战。AI 智能体提供了另一种途径：将游戏经验转化为对可执行策略的修订。在启发式学习（HL）的基础上，我们形式化提出对抗性启发式学习（AHL）这一范式：在保持模型权重不变的同时，利用 AI 智能体作为学习引擎，改进游戏策略及其配套软件。我们推出 AAArena，这是一个包含 12 款真实对抗性游戏和 1,920 个存档人类程序的基准，其评估协议仿照现实世界的游戏竞赛。智能体需要理解规则、选择对手、分析比赛回放，并修改游戏智能体，以便在固定的对局和评估预算内取得最高排名。我们评估了 \val{completedmodels} 个模型与工具链配置：Opus5.5 搭配 Claude Code 赢得了 6 枚金牌，而在其余 6 个人类排行榜中，没有任何受评估配置能够登顶。总体而言，在规则描述更复杂的游戏中，性能通常较弱。进一步实验表明，对手选择和密集反馈有助于策略改进；智能体既能从自身比赛的同策略回放中学习，也能从其他玩家比赛的异策略回放中学习。这些结果凸显了 HL 在对抗性游戏中的潜力，并揭示了游戏理解、策略实现以及长周期策略开发方面持续存在的挑战。
当检索到的证据与智能体先前的信念相矛盾时，它会修正答案、承认不确定性，还是坚持错误结论？现有对智能体系统的评估主要关注任务成功，对智能体如何处理此类冲突所提供的洞见十分有限。我们提出从认识论谦逊（EH）这一维度评估智能体：即智能体在执行任务期间识别、据此采取行动并传达不确定性的意愿。我们通过三个轨迹层面的行为维度对EH进行操作化定义：Identify、Solve和Escalate（ISE）。我们考察知识冲突，即骨干语言模型的参数化知识与其遇到的证据相矛盾，或两个上下文来源彼此不一致的情形，并据此评估两种冲突设置：（1）受控冲突；（2）多步骤智能体执行过程中自然发生的冲突；每种设置均配有匹配的无冲突对照。通过评估四个智能体，我们发现，更高的任务准确率并不一定对应更高的认识论谦逊：一些高准确率配置能够在执行过程中识别冲突，却不会在错误的最终答案中传达尚未解决的不确定性。进一步的轨迹层面分析表明，智能体经常在执行早期步骤中检测到冲突，却未能在后续步骤中持续处理或解决这些冲突。最后，我们表明，模型层面的干预可以提升EH，但往往会以任务准确率为代价，这表明认识论谦逊源于骨干模型、智能体框架与评估环境之间的交互。
智能体已被部署在从行程规划器、股票交易到 IT 事件分诊等各种应用中。在大多数情况下，LLM 智能体几乎不受基于规则的安全防护约束而自主运行，因而可能由于不可逆操作引发成本和安全问题。近期研究要么使用安全防护智能体监控行为，要么在事后评估日志。前者会为每一步都增加成本和延迟；后者则要等运行结束后才能给出结论，此时 token 已被消耗，损害也已造成。为解决这一问题，我们提出 OnTrack，这是一种流式监控机制，将智能体的步骤及其依赖关系与记录的成功运行进行比较，从而以每步约一毫秒的速度向用户发出警报或阻止智能体运行。我们在三种数据访问权限逐渐降低的情形下研究这一问题：完整参考访问权限（历史运行记录和工具模式）、中等访问权限（仅工具模式）以及无先验知识（仅有生成的步骤日志）。随着数据访问权限降低，OnTrack 监控能力的预期也会下降，其检测范围从计划违规扩展到识别循环、停滞和重复工具调用。最后，我们使用 SWE-bench 轨迹对 OnTrack 进行评估。仅依据前 8 个步骤，我们的方法对失败轨迹的排序低于成功轨迹，其效果优于基于内容相似度的方法（+0.057 AUROC）。采用中止策略后，对于原本会在失败运行中消耗的计算资源，我们节省了约 18%；在被中断的运行中，有 83% 实际上正走向失败（每 6 次中止中有 5 次判断正确）。
AI 代理如今已经能够实施现实世界中的网络攻击，随着代理数量增加而扩展能力，并集体追求不符合预期的目标以获取奖励。这些因素共同提高了不符合预期的代理种群爆发式增长的风险：代理可能入侵计算机并秘密部署更多代理，从而形成自我强化的循环——更大的种群发展出更强的集体网络能力，并进一步扩张。这引出了一个根本问题：什么决定了不符合预期的代理种群是保持受控，还是进入这种自我强化的循环并迅速扩张？这一种群层面的问题属于生态安全：不同于种群规模固定的单代理或多代理安全，生态安全关注的是种群本身的动态。在本文中，我们基于一个种群增长方程，建立了 AI 代理种群的生态学理论，其中适应度（增长率）取决于网络安全能力。我们表明，在没有协作的情况下，只有当单个代理的能力超过某个临界阈值时，种群才会迅速扩张。然而，在存在协作时，集体网络安全能力会随种群规模增加而增强。这会形成一个临界种群阈值：低于该阈值时，种群会衰退；高于该阈值时，即使单个代理的能力没有改变，种群也会迅速扩张。在生态学中，这种现象被称为强阿利效应。由于对一小群代理进行红队测试无法保证更大规模种群的生态安全，我们的理论要求开展生态红队测试和种群节奏控制：在受控环境中逐步部署规模更大的代理种群，同时测量网络能力如何随种群规模扩展，并估计触发迅速扩张所需的临界种群规模。能力提升可能会降低这一阈值，因此每一代新模型都需要重新估计。
We propose BrickBench, a benchmark for agentic text-conditioned LEGO-set design. Given a prompt, an agent is tasked with producing an assembly that not only satisfies semantic and design criteria, but that can also be physically built. To do so, it must select parts from a discrete library and reason jointly about local and global constraints. We score validity, alignment, and design across three settings that vary in scale and part availability. We provide BrickAgent, an environment for coding agents to construct, inspect, and validate their designs. We find that leading agents largely satisfy verifiable physical and semantic requirements, but fall short of human designs. We release our benchmark and environment at http://www.brickben.ch
2026 年，涉及 OpenAI、Anthropic 和 Google 智能体的网络安全评估触达了授权测试范围之外的真实系统。它们所采取的路径各不相同。OpenAI 智能体利用了研究基础设施，在不同运行之间进行协调，并入侵了 Hugging Face 生产环境的部分区域。Anthropic 报告了若干案例：由于第三方环境配置错误，执行模拟网络任务的智能体得以接触真实系统。在另一项单独报道的评估中，Google 的 Gemini 通过一条非预期的互联网路径访问了三个真实组织；Google 表示，该模型在三种情况下都停止了操作。综合来看，这些案例说明，评估不能依赖假定的边界。该边界必须在智能体运行过程中得到验证。本工具性比较案例研究提出了 Proactive Agent Security Assurance Cycle (PASAC) 和五层 Boundary Assurance Stack。该框架结合了按风险分级的任务设计、可执行的范围契约、运行前验证、最小能力访问、独立出口控制、凭证限制、跨运行监控、自动停止条件以及基于证据的重新授权。一个先行指标模型、九项设计命题和七个可证伪假设，将这些经验教训转化为一项可检验的研究计划。由于公开的 Gemini 记录仅限于有明确归属的陈述和新闻报道，其详细因果机制仍属暂定。核心结论很明确：主动式智能体安全需要贯穿完整执行系统的持续保障，而不能依赖对任何单一沙箱或安全措施的信心。
基于 LLM 的智能体如今会持续与环境交互，并受到用户指令、同伴消息和长期记忆等组织渠道的影响。现有安全研究已经考察了这些影响，但主要将其视为彼此独立的智能体组件。此类因素如何从统一视角共同塑造智能体的安全行为，仍未得到衡量。为弥合这一空白，我们倡导将组织行为学（OB）作为研究先进智能体安全性的框架，围绕智能体运行其中的关系结构，重新组织研究对象、理论基础和实验设计。我们首次将组织行为学中一个与安全密切相关的经典子领域——反生产工作行为（CWB）——系统形式化为 Agentic Counterproductive Behavior（ACB）。ACB 明确了三类组织前因变量（垂直的主管关系、水平的同伴规范和内部认知结构），并将其映射到三类反生产结果维度（未经授权的信息披露、破坏性操作和生产偏差）。为实现 ACB 的操作化，我们提出了 Workerville，这是一个受控基准，通过在共享任务上操纵组织条件，将 16 种组织配置应用于 210 项任务，生成 3,360 个挑战，并由经过人工验证的智能体评判器进行评估。通过对 6 个前沿 LLM 的基准测试，我们发现：（I）负面的组织前因变量组合后会产生非单调放大效应：在不存在负面前因变量时，未经授权的信息披露率为 16.5%；存在两个负面前因变量时升至 60.1%；存在三个时则回落至 50.3%；（II）智能体会再现人类 CWB 研究所预测的典型行为模式；（III）这些结果确立了 OB 作为智能体安全研究系统性框架的地位，并指向一项新的研究议程。
近年来，大语言模型（LLM）的记忆能力日益受到关注。尽管现有方法已取得巨大成功，但基于检索的记忆方法通常忽视不同记忆之间的差异，采用统一策略处理所有记忆，从而导致性能欠佳。因此，一个直观的问题随之产生：能否将记忆划分为不同类型，并选择适当的策略？然而，鉴于记忆场景具有主题丰富、情境复杂和边界模糊等特点，要实现记忆的精准分类并不容易。为应对这一挑战，本文提出了一个记忆多分类数据集 TriMEM，为多样化场景中的记忆类型提供精准标注。在此基础上，我们提出了一种新颖的记忆框架 MemoType，该框架借助训练得到的路由模型，自适应地识别每条记忆和查询的类型。通过对记忆和查询进行路由，MemoType 能够检索与查询类型相匹配的记忆，并设计针对性的检索策略，从而提升检索性能。此外，我们从理论上证明，在多类别语料库中，任何单一检索策略的期望检索精度都存在一个基本上限，进而导致系统性的精度下降。在三个数据集上的大量实验表明，MemoType 始终优于现有方法，Recall@1 最高提升 16.18%。
End-to-end autonomous driving demands trajectory planners that are both highly accurate and cheap enough for edge deployment. State-of-the-art artificial neural network (ANN) planners meet the accuracy requirement at the cost of heavy dense computation, while spiking neural networks (SNNs)---though promising orders-of-magnitude energy savings through sparse, event-driven arithmetic---still lag far behind in planning accuracy. We present \textbf{SDPAD}, a fully spike-driven end-to-end planning pipeline that closes this gap. SDPAD converts a pre-trained ANN perception stack into integer-spike form via quantized ANN2SNN conversion, lifts multi-view images into the bird's-eye-view (BEV) space with a spike-driven-max (SDM) depth distribution (Spike-3D-Lift), and plans through the Spike-QFormer, a spiking query transformer in which ego, agent, and map queries distilled from the BEV scene are fused by learnable waypoint queries via cross-attention, followed by deformable spike-cross-attention refinement. Every operation is gated by integer spikes and inference is a single feed-forward pass without temporal simulation loops. On the nuScenes open-loop benchmark, SDPAD achieves an average $L_2$ error of 0.40\,m and a collision rate of 0.12\%, on par with strong ANN planners while consuming 69.9\,mJ---less than 2\% of recent ANN baselines. In closed-loop evaluation on the NAVSIM navtest split, SDPAD reaches 86.3 PDMS, surpassing the previous SNN planner SAD by 4.3 points and matching mainstream ANN planners at a fraction of their energy. To our knowledge, SDPAD is the first fully spike-driven planner evaluated in end-to-end autonomous driving, demonstrating that SNNs can rival dense ANNs in complex driving tasks.
Coding agents produce large, tangled patches that mix multiple development purposes, making the code hard to review and maintain. Commit untangling offers the promise of organizing such large patches into untangled, manageable commits. This paper emphasizes two important limitations in existing commit untangling studies. First, they do not consider that untangled commits are ordered and should leave the code runnable. In practice, maintainers are unlikely to accept commits that prevent the code from running. Second, existing studies claim that commit untangling helps software maintenance. However, they conduct syntactic comparisons between the untangled commits and developers' original commits without directly showing the claimed maintenance benefits. To address these gaps, this paper makes two novel contributions: (1) RucTangle, the first agentic method that untangles commits while keeping the code runnable after each commit; and (2) TangleEval, the first evaluation framework that quantifies how untangled, manageable commit histories help coding agents repair bugs. We compare RucTangle against four untangling methods on 131 agent-generated patches. All histories produced by RucTangle are runnable, while baselines produce 20.6%-37.4% unrunnable commit histories. We further collect 453 agent-generated patches that introduce regressions (i.e., causing previously passing tests to fail) and ask two other coding agents to repair regressions. Augmenting agent context with RucTangle-produced histories yields 5.2% absolute improvement in pass@1. We also analyze agent trajectories to learn how they use untangled commits to navigate and fix bugs. Our findings demonstrate the value of adopting established software engineering practices in the era of coding agents, which broaden the future research agenda: how can agents actively use software history to make better development decisions?
基于 LLM 的多智能体系统（MASs）正日益用于通过协同推理、工具使用以及与外部资源交互来解决复杂任务。然而，对此类系统中的失败进行归因仍然具有挑战性，因为观察到的结果通常无法直接揭示导致执行失败的错误。在本文中，归因目标是决定性错误，即修正该智能体—步骤对便可恢复失败执行的错误。现有方法主要识别可疑步骤，却没有明确建模错误如何跨交互传播，或如何持续存在于尚未解决的循环中，因此难以将决定性错误与下游失败症状区分开来。我们提出错误传播建模失败归因方法（EMFA）。EMFA 构建失败轨迹的结构化表示，对级联传播和持续性交互循环进行建模，并采用具备传播感知能力的候选筛选，随后通过反事实验证来识别决定性智能体—步骤对。在 Who&When 基准测试中，EMFA 取得了当前最佳的步骤级归因准确率，并在智能体级别保持竞争力。与此前最佳的步骤级结果相比，它在 Hand-Crafted 和 Algorithm-Generated 子集上分别提升了 3.45 和 4.40 个百分点。
智能体可以完成任务，却不一定会改进其工作方式。要将任务经验转化为可复用能力，需要把发生变化的组件与其评估及后续使用连接起来。我们提出 AgentEvolver，这是一种在任务执行过程中开发能力、同时保持基础模型不变的系统。八类实体通过一个统一的版本化生命周期，使可复用的操作、方法、智能体、控制流、接口和支持状态能够被修订。共享的 Runtime 协调持续进行的工作，而持久化规划和可恢复上下文则保留任务方向及支持性证据。我们在 SWE-bench Pro Public 上评估任务结果，并考察六个应用案例中的能力变化。研究团队报告称，引入演化后，任务解决率达到 82.08\%，超过其报告的未引入演化时的基线。案例显示，保留的能力被用于后续的网站、游戏和研究工作，同时也记录了未完成的目标和一次未成功的策略。这些发现区分了可复用组件的改进与最终任务的成功。AgentEvolver 为通过执行研究能力累积提供了具体基础；独立任务迁移和总开发成本仍是有待解决的问题。
Coding agents are extended with agent skills, directories whose SKILL.md tells the model when and how to perform a task. Because skills come from independent sources (teams, developers, plugins, copied collections), an installed skill can be co-installed with a similar skill doing the same job, and the model picks between them by name and description alone. In a conflict, the installed skill loses core functions (e.g., a ban on touching git) because the similar skill runs instead or changes what it does. The task still passes, so benchmarks that check only task completion miss such cases. We present the first empirical study of such conflicts. From snapshots of 20,947 repositories, we mine 822,109 candidate similar-skill pairs, have an LLM judge a stratified sample of 3,754, and run 312 confirmed pairs on three models (6,368 runs, 169,294 tool calls, 542 agent-hours). We report five findings. (1) Conflict-prone skills are common: nearly one in four installed skills is co-installed with one that does the same job, and 37% of judged skills sit inside copied collections. (2) Most such pairs involve normative skills, then capability skills. (3) Without lowering task completion, a similar skill takes one in five runs from the installed skill, and runs that open the similar skill first lose over a third of the exclusive core functions that only the installed skill fulfills. (4) Install location decides which skill runs, listing order barely matters, and the final reply names the skill used in only 0.9% of substituted runs. (5) Conflicts are decided at the first skill read, almost always before any file is changed, and a pre-tool hook at that read restores fidelity on exclusive core functions to the level of runs that open the installed skill first. Benchmarks should thus score exclusive core functions, and platforms should guard the first read and show which skill ran.
Similar agent skills can share instructions but differ in their conditions of use. Query-based text selection may retain shared instructions and omit these distinctions. We introduce SkillContrast, a training-free selector that compares retrieved skills and retains their differing text with local context for a pretrained reranker. On 1,235 requests from SameCapRisk-Bench, it yields 54-72 more clean hits (requests that retrieve a helpful skill without its marked risky sibling) than TF-IDF query selection at identical per-candidate input lengths, across 2 retrievers and 2 reranker sizes. Length-matched component replacements identify differing text as the main contributor in the primary setting, with smaller, mixed context effects. Relative to full skill bodies, SkillContrast uses 51.1-58.8% fewer model-input tokens, with 10-18 fewer clean hits at 0.6B and matching or higher observed clean-hit counts at 4B. Candidate-relative differences thus complement query relevance in selecting compact reranking inputs.
Automated interviewers and conversational agents are increasingly used in research, recruitment, customer service, and education. However, many existing systems rely on fixed question sequences and provide limited context-based personalization without considering participants' knowledge, which can lead to repetitive or irrelevant follow-up questions. Therefore, there is a need for an adaptive interviewing system that can adjust question depth while maintaining conversational continuity and semantic progression. To address this, an Evidence-Traceable Dynamic Interviewer Architecture is presented using a locally hosted Large Language Model (LLM), with the interview continuously adapted throughout the entire conversation based on the participant's responses and evolving context. The interviewer profiles participants' expertise in real time to generate knowledge-appropriate questions, well-articulated responses, and smooth transition messages that support conversational continuity. A five-module prompt-driven architecture and persistent interview-state record support these functions. The interviewer was evaluated with 246 participants. Expertise Profiling module (M3) showed 78.9% exact agreement with independently reported participant expertise, with a weighted Cohen's K of 0.80. Generate Iterative Questions module (M4) showed a strong expertise-complexity association (p=.79, p<.001), and participants reported high relevance (mean 4.41), engagement (mean 4.32), and satisfaction (mean 4.38), providing evidence that the architecture's adaptive components operated consistently with their intended functions while participants reported a positive interview experience.
要提升长时程 LLM agent，可以围绕冻结模型演化运行框架，也可以训练模型权重。我们先让自演化运行框架增强系统，再将初始和演化后的运行框架分别与基础权重和训练后权重组合，以了解训练后的模型保留了哪些增益，以及哪些增益仍然依赖运行时环境。我们发现，可以从 agent 的失败构成中判断应采取哪种手段：按照失败轨迹中最先触发的信号进行标注，可以将过程失败（调用受阻、循环、步骤预算耗尽）与内容失败（计划已交付但质量较差）区分开来。运行框架演化能够修复前者；它塑造的行为可以训练进权重；而内容失败正是权重训练要解决的问题。在 DeepPlanning 上，自演化运行框架循环将 Qwen3.5-4B 的留出集得分从 0.16 提升至 0.30，将 Qwen3.5-9B 的得分从 0.32 提升至 0.44；对于 4B，留出集交付率从 55% 提升至 90%，而内容失败则留待权重处理。在演化运行框架轨迹上训练的 LoRA 适配器能够将这一增益内化：在原始运行框架下，两种规模的模型在留出任务上都额外提升 +0.13；对于 4B，适配器与运行框架结合后，留出集得分提高到原来的两倍以上；对于 9B，仅使用适配器就能达到完整演化流程的效果，并将内容失败率从四分之一的轨迹降至二十分之一。一个在答案打乱后的轨迹上训练的安慰剂适配器，其表现低于基础模型。该循环还迁移到了 WebArena-Lite（在 117 个未见任务上提升 +0.09），其中增益来自模型所看到的内容，适配器无法进一步增加这一增益。由此得到一条两次应用的“先诊断、后干预”规则：先读取失败构成，在运行框架与权重之间做出选择；再观察被接受的编辑改变了什么，以决定应将哪些增益训练进模型。得分是相对于新锚点进行的四次 rollout 均值；除特别标注者外，评测均在同一晚完成，覆盖来自六个模型家族的八个模型和两个基准测试。
Deep reinforcement learning agents reach strong performance in real-time strategy games but can be brittle against opponents outside their training distribution. Separating strategic command selection from learned unit control allows different strategies to be selected for different opponents while reusing the same execution policy. This requires an executor that can follow different commands and measurable criteria for assessing whether it does so. We introduce a constrained command-conditioned Proximal Policy Optimization (PPO) policy, the executor, for MicroRTS, a real-time strategy environment. Discrete commands specify strategic objectives and behavioral requirements for economy, army composition, military posture, and worker policy over multiple environment steps; the executor determines the unit-level actions used to fulfill them. A Thompson-sampling bandit acts as the strategist, selecting command tuples from an estimate of the opponent's strategy built from in-game observations rather than opponent identity. In a controlled comparison with a flat PPO baseline trained with the same architecture, budget, curriculum and self-play league, the strategist-executor system wins significantly more often against three of the four strongest opponents on a training map, including the two strongest held-out ones (0.55 to 0.97 and 0.01 to 0.34), with no significant difference against the others.
Puzzles are widely used to evaluate the reasoning capabilities of artificial intelligence (AI) systems for sequential decision making, yet approaches originating from different paradigms are rarely compared under unified conditions. To address this gap, we introduce a three-dimensional characterization framework that enables the analysts of AI methods by 1) projecting them to the Markov decision process (MDP) sequential decision making formalism, 2) degree of autonomy through human prior ranking of their designs and, 3) skill and computational cost. Using this framework, we analyze how representative graph-based, reinforcement learning, and large language model (LLM)-based approaches differ in their design choices and performance characteristics, instantiated respectively by Neurosolver, forward-backward reinforcement learning (FBRL), and automated thought-of-search (AutoToS), including a double-agent extension of thought-of-search (DA-ToS). The analysis relies on the Tower of Hanoi puzzle that provides a controlled benchmark with well-defined rules and scalable complexity, enabling consistent comparison across increasing problem sizes. The 3D characterization reveals that LLM-based methods, due to their weakly constrained action-space design, shift complexity from architecture to inference-time verification, leading to substantially higher memory and runtime costs than Neurosolver and FBRL.
Complete orthogonal predictive coordinates do not by themselves bind a latent direction to a named intervention. We present a mathematical and synthetic audit motivated by vascular device-vessel suitcordance. Capacity-matched least-squares predictors were exactly equivalent under complete fixed output transforms, whereas an anchor-only observer recovered interpretations only within the span of known perturbation signatures. Six three-dimensional configurations across 64 seeds gave a maximum paired prediction discrepancy of 6.7e-15 but a median untransported edit error of 1.513. Coordinate transport removed that error. Noisy and weak anchors constrained calibration stability, and changing the representation basis required recalibration or verified transport. Across 256 additional fits in dimensions 3-24, prediction equivalence persisted within 4.0e-15. We then evaluated nine deliberate runnable fault classes across 64 seeds. All 576 faulty executions completed, but each violated at least one reconstruction, prediction, delivered-edit or scope contract; all 320 valid control records passed. Repeating a faulty implementation gave exact self-agreement despite error against the separately computed simulator expectation. For one omitted-direction defect, probe coverage followed its analytic law, and rank-aware abstention protected unsupported interpretations. Scalar-noise experiments exposed both missed weak faults and excessive rejection under narrow relative tolerances. These controls provide an executable separation of prediction, semantic support and scientific acceptance. They are synthetic numerical audits, not clinical validation, neural JEPA-Anything replication, agent learning or patient treatment-effect estimation.
Memory-augmented mobile GUI agents store successful execution trajectories and reuse them in later tasks, but a stored trajectory rarely matches a new task exactly. The new task may use different parameters, share only some of its steps with a stored trajectory, or have no relevant record in memory. Forcing the agent to use irrelevant memory can mislead it, whereas discarding memory that may still be useful deprives it of guidance from past experience. To address this dilemma, we propose DeltaReplay, a step-level memory reuse framework that decides how to use existing memory without modifying it. We observe that the reusable part of a stored record is determined not by the record itself but by its relation to the new task, mainly through two factors: page-level consistency and action-level generality. We therefore store execution trajectories as paths in a transition graph, whose nodes (pages) and edges (actions between pages) capture these two factors. At reuse time, the action on each edge is split into a task-independent operation and task-specific parameters. DeltaReplay then compares each recorded step with the new task and the current screen, and decides whether to follow it, execute it after replacing its parameters, or leave it to the base agent. On AndroidWorld and SPA-Bench, DeltaReplay improves the task success rate over a base agent with the same backbone by up to 10.3 and 25.0 percentage points, respectively. These results indicate that deciding at each step how to use retrieved memory lets agents benefit even from partially matching trajectories.
Enterprises racing to put agents into production are learning that the AI improvement loop never really ends. Agents must be run, observed, evaluated and improved continuously. CoreWeave Inc. has been assembling a full-stack platform aimed at agentic AI. Its Forge offering is intended to give business units and machine learning teams a shared workflow, according […] The post CoreWeave Forge aims to speed up the AI improvement loop appeared first on SiliconANGLE .
Manus, the creator of the eponymous artificial intelligence agent, has raised more than $500 million in funding. The Chinese startup announced the round today. It stated that private equity firm Boyu Capital led the investment with contributions from Tencent Holdings Ltd., one of China’s largest tech firms, and several others. Bloomberg reported last month that […] The post AI agent developer Manus raises $500M+ at reported $4B valuation appeared first on SiliconANGLE .
将仿真创意转化为可运行的应用程序，意味着整合资产、连接物理与渲染，并检查场景是否按预期运行。开发者正将 Frontier AI 模型与 NVIDIA Omniverse 库结合起来，以协助完成这项工作——构建用于探索各种场景、调查故障和改进设计的应用程序。开发者通过 […] 引导 AI agents。
The full-stack AI cloud market is about more than access to graphics processing units: it also encompasses the tools needed to improve models and agents in production. Providers are seeking to distinguish themselves through the tooling, partners and expertise they combine with hardware. CoreWeave Inc. is connecting training, inference and evaluation through its newly announced […] The post CoreWeave makes the case for an open, full-stack AI cloud appeared first on SiliconANGLE .
你好~！Pocketty 是一款适用于 iPhone 和 iPad 的 SSH 终端，专为 herdr 打造。herdr 会让你电脑上的 agent 窗格持续运行，并了解每个窗格的状态：工作中、需要你的介入，或已完成。最近休陪产假的后半段时间里，我是在恼火和绝望中做出这个应用的。被宝宝午睡困住的时光很甜，但我能承受的无休止刷屏和看电影时间毕竟有限……无论如何，过去几个月我一直在使用它，现在终于不必守在书桌前也能保持高效。如今，被宝宝午睡困住的时候也能继续产出（只要我想）！ 工作方式： - 应用通过普通 SSH 直接连接你的电脑。Tailscale 是从任何地方访问电脑的简单方式，但只要是你能连接到的 SSH 主机都可以。 - 主机上运行的一个小型 Rust 守护进程会监视 herdr。当某个窗格需要你处理时，它会使用 HPKE（X25519、ChaCha20-Poly1305）将提醒加密到你手机的密钥上。Cloudflare Workers 上的无状态中继（属于 Pocketty）会把加密后的字节传给 APNs（Apple 推送通知服务器），然后由通知扩展在手机上打开。中继无法读取这些内容，也不会保存任何东西。 - 你的 SSH 密钥会在 Secure Enclave 中生成，无法导出。 - 终端使用 libghostty-vt 管理状态，并通过 wgpu 在 Metal 上进行绘制，因此全屏 TUI 的显示效果和在电脑上一样，只是宽度更窄。 - 每个 agent 回合的差异、文件浏览器，以及 agent 启动的 `localhost` 开发服务器预览，都通过同一个 SSH 连接提供。无需设置端口转发。 herdr 和该守护进程都是可选的。不使用它们时，这就是一个普通的 SSH 客户端。无需注册账号，应用中也没有分析或跟踪功能。 应用提供 14 天免费试用，期间可使用全部功能。之后，如果你也像我一样喜欢它，只需一次性购买即可永久拥有：前两周为 99 美元（发布优惠价），之后为 129 美元。 欢迎询问任何关于加密推送设置或在 iOS 上运行 libghostty 的问题。 评论链接：https://news.ycombinator.com/item?id=50009634 分数：7 # 评论数：0
A registry of hand picked/built registry of ui/ux components so now i can produce slop without ai. i always loved using and searching for ui components on every registry site that popped on my feed. i collected and saved all of these components but wanted to actually to build something thats more unique, free and easy to use. i was hooked by ASCII components and my main focus was building also some of them and the results are really appealing (feedback is always appreciated), although i am working on even more and more unique. there is an mcp and cli connection so your agent can easily search and use and component it finds interesting. there is also install by url and you can also get the prompt and also the code if you want to get only a section of the component. my goal with it is to get more feedback on what looks actually good and what doesnt, because all of this is a one persons tase, and to get also the design community to use it and add to the component catalog. if you have any question or feedback i would be really happy to listen/help! Comments URL: https://news.ycombinator.com/item?id=50011077 Points: 1 # Comments: 0
Amazon Bedrock AgentCore payments gives AI agents a managed way to pay for services on demand, with spending limits enforced by the infrastructure. See how Incarna's agents pay BlockRun for model inference one request at a time over x402, cutting the work of adding x402 payment support from months to days.
Google 正在将 Gemini 打造成一种 AI 代理，能够进行规划、执行任务，并跨企业应用和系统协同工作。该代理可以将工作委派给子代理，使用多个 AI 模型，甚至拥有自己的工作场所身份，包括一个电子邮件地址。
Automation Anywhere Inc., an enterprise agentic automation and orchestration firm, announced Wednesday an agreement to acquire Boost.ai Inc., a conversational voice artificial intelligence company, from Nordic Capital. The firm seeks to build the “autonomous enterprise,” an operating model where businesses operate up to 80% autonomously or with AI assistance. Boost.ai’s technology will extend that vision […] The post Automation Anywhere acquires Boost.ai to expand customer-facing voice AI appeared first on SiliconANGLE .
On-device personal AI requires models and agent software that can operate within fixed hardware limits. Developers also need ways to keep those systems improving after deployment. Model builders are rethinking architectures designed around elastic cloud capacity. The edge offers fixed hardware but a far richer view of the user, which makes it the natural home […] The post Liquid AI builds personal AI around device-level context appeared first on SiliconANGLE .
JetBrains released Mellum2.1, an Apache 2.0, 12B mixture-of-experts thinking model with 2.5B active parameters. RL in real repositories lifted its SWE-bench Verified score from 2.0 to 47.0. The post JetBrains Releases Mellum2.1: A 12B MoE Open Model for Coding Agents appeared first on MarkTechPost .
Goodfire just launched what it says is a cheaper way to keep AI agents in check: instead of paying a second AI to read everything an agent does, its monitors peek inside the model while it works and only call in backup when something looks fishy.
Natura’s $99 Interface smart ring lets you summon AI agents with the press of a finger to complete tasks, capture thoughts, and control devices — while doubling as a health tracker.
Google 正在推出一款“通用”Gemini AI 智能体，可在后台跨应用和设备运行。这款工具于周四举行的 Gemini at Work 活动上公布，将在 Gemini Enterprise 应用中提供，用户可以通过单一界面与 Gemini 智能体聊天并向其分配任务。[…]
Zach Yadegari, the teen co-founder of popular Cal AI calorie tracking app, has launched a new personal AI agent startup that competes with Instinct, Muse and Bee.
My Decoder guest today is Hayden Field, The Verge’s senior AI reporter, and we’re discussing the new wave of consumer-friendly AI agents. If you’ve been paying attention to this space, you know AI enthusiasts have been using agents for a minute now — homebrew OpenClaw setups led to a surge in Mac Mini sales earlier […]
For the first time in years, the Internet Corporation for Assigned Names and Numbers - better known as ICANN - is accepting applications for new top-level domains. These are the suffixes at the end of all URLs, and you may know them as things like .com, .org, and .pizza. ICANN just announced the 1,615 applications […]
Zenity Labs 研究人员表示，Amazon 的 Bedrock AgentCore 上一个公开可访问的 AI agent，就足以接管同一 AWS 账户和区域中的每个 AgentCore agent。该攻击利用了 AWS 的一个内部接口，该接口用于提供临时云凭证，而 agent 可以不受限制地访问它。AWS 此后已修复该问题，并大幅收紧了 agent 的默认权限。文章《一个提示就足以劫持 AWS 账户中的所有 AI agent，Zenity 研究人员发现》最初发表于 The Decoder。
Google LLC’s Cloud division today introduced Gemini agent, a unified artificial intelligence assistant that can act autonomously, generate code and complete work across any device, including web, mobile and desktop. The new agentic AI experience can be accessed anywhere and through any channel, including the command line, Google Workspace, Microsoft 365 or Slack. It even […] The post Google Cloud introduces Gemini agent to change enterprise work appeared first on SiliconANGLE .
NVIDIA researchers introduced PivotOPD, an on-policy distillation method that trains multi-turn LLM agents to avoid early pivotal mistakes and recover from them, posting the best average against 13 baselines on 3 agent benchmarks. The post NVIDIA PivotOPD Teaches Multi-Turn AI Agents to Recover From Pivotal Mistakes appeared first on MarkTechPost .
I built this last year because I noticed my coding agents would infer my database schema from migration files and code. Which seemed inefficient. Its a simple tool that creates compact markdown schema doc(s). Which I could reference in my AGENTS.md or directly in my prompts. I keep the docs up to date by running it as a part of my migration command. It's a single Go binary, MIT licensed, and works as a CLI or a library. Supports Postgres, MySQL and SQLite. Comments URL: https://news.ycombinator.com/item?id=50003569 Points: 1 # Comments: 0
Industrial AI is entering a new phase. After decades of predictive analytics and other specialized applications, advances in foundation models, physical AI, and agentic AI are making it possible to automate more complex tasks across industrial environments. But unlike AI that operates purely in the digital world, industrial AI can interact directly with physical systems,…
arXiv:2610.06401v2 公告类型：replace-cross 摘要：使用工具的语言模型智能体容易受到间接提示词注入的攻击，因为它们必须处理不受信任的外部内容。现有的训练时防御方法可以降低攻击成功率，但通常会牺牲通用能力。我们发现，基于训练的防御方法会导致模型输出分布发生显著偏移，即使在良性场景下也会改变模型行为，并可能因此降低实用性。我们进一步发现了这类防御方法的一种失效模式：在良性的工具使用任务中，模型会避免执行完成已授权任务所需的某个步骤，尤其是在该步骤由工具输出指示时。为解决这些局限，我们提出 RAISED（通过自蒸馏实现鲁棒攻击不变性），一种结合自生成与自蒸馏的训练框架。模型首先自行生成工具使用场景，重点关注那些必须根据工具输出中的合法指引采取行动才能完成任务的情形。随后，通过自蒸馏，训练学生模型在同一轨迹的干净版本和注入版本上，匹配教师模型在干净上下文中的行为。与以往基于训练的防御方法不同，RAISED 在大幅降低工具响应中提示词注入攻击成功率的同时，还能在智能体基准和通用基准上保持实用性。
arXiv:2609.32303v2 公告类型：replace-cross 摘要：基于共享检查点训练的领域专家可以通过在策略蒸馏（OPD）将其专门能力迁移给单个学生模型。现有研究主要关注如何改进这一合并过程，而对用于训练专家的算法却缺乏系统比较。我们通过在 Agentic、Reasoning 和 Perception 三个领域中，对监督微调（SFT）与强化学习（RL）进行受控的单教师比较，研究哪种训练算法能够生成更适合 OPD 的教师模型。教师模型和学生模型均采用相同的 Qwen3.5-9B 初始化，并在任务性能相近的情况下比较两类教师模型。实验表明，在全部三个领域中，RL 教师都能产生更强的学生模型，并更充分地恢复教师模型的性能增益。在各自的最佳检查点上，RL 引导的学生模型分别比 SFT 引导的学生模型高出 4.27、1.50 和 0.86 个百分点。在 Agentic 领域，最佳 SFT 引导学生模型仅恢复了教师模型相对于基础模型性能增益的 44.44%，而最佳 RL 引导学生模型恢复了 115.00%，超过了其教师模型。进一步分析表明，与共享初始化相比，RL 教师的参数偏移小于 SFT 教师。上述发现支持这样一个假设：RL 教师偏离学生起点更小，有助于通过 OPD 进行学习，从而产生更强的学生模型。
arXiv:2608.29971v2 公告类型：替换交叉版本 摘要：随着智能体系统演变为复杂的多智能体编排工作流，系统化评估智能体行为一致性与确定性的需求日益迫切。本文提出了一种形式化评估方法，该方法以 AgentGraph 为基础。AgentGraph 是一种由领域专用语言驱动的规划器，通过可动态调整的有向图表示智能体的推理过程。我们利用这一结构化形式，并采用图遍历算法穷举对话路径，从而构建一个全面的评估集，以覆盖智能体完整的行为空间。随后，我们系统地重放这些可复现的轨迹，将观测到的输出和状态转换与预期的 DSL 规范进行比较。为量化可靠性，我们定义了新的指标，用于衡量精确重放及其语言变体中的响应确定性、轨迹确定性、结构遵循性和语义一致性。系统结果表明，与未配置受控转换的智能体相比，使用 AgentGraph 和 LangGraph 等框架并明确配置结构化节点转换的智能体表现出更高的确定性。
arXiv:2608.29615v2 Announce Type: replace-cross Abstract: Across industries, machine-learning systems support applications ranging from prediction and anomaly detection to forecasting, optimization, and scheduling, yet operationalizing these systems requires coordinating application development, model pipelines, cloud infrastructure, security, deployment, monitoring, retraining, recovery, and rollback. We present an evidence-gated multi-agent framework for transforming a natural-language MLOps cloud engineering task into a verified repository and operational cloud deployment. The framework combines graph engineering, loop engineering, and agent harness engineering. A stateful Graph Orchestrator coordinates specialized agents for repository generation, review, execution, verification, release, and monitoring while governing workflow dependencies, evidence gates, retry bounds, recovery paths, and termination. Consequential lifecycle transitions proceed only when their required predicates are supported by verifiable execution or runtime evidence. Verification failures activate bounded reflection, repair, and re-verification, while runtime evidence of failure, drift, degradation, or policy violation can trigger bounded adaptation, recovery, or rollback. Agent harness engineering constrains repository generation, review, and repair, artifact execution, and cloud operations through controlled capabilities and isolated execution environments. We realize the framework on Google Cloud Platform and evaluate repository completeness, controlled execution, evidence-gated transitions, cloud promotion, and bounded recovery. Our experimental results show that the framework prevents unsupported lifecycle transitions and drives each run toward either a verified operational deployment or an auditable terminal failure.
arXiv:2606.13994v2 公告类型：replace-cross 摘要：基于 LLM 的智能体正变得越来越强大，并得到广泛部署，从而不断增加其在现实世界中被对抗性滥用的动机。一项关键的新兴威胁是分解攻击 \cite{glukhov2024breach, jones2024adversaries}：在这类攻击中，有害任务被拆分为更简单、看似无害的子任务；这些子任务分别执行时能够规避安全机制，但累积起来却实现了恶意意图。尽管近期的基准测试会在多轮交互和多工具使用场景下评估智能体安全性，但它们并未明确捕捉这种分解式滥用，也可能无法体现真实的对抗性执行流程。为此，我们引入 DeCompBench，这是一个专门用于评估分解攻击下智能体安全性的基准。DeCompBench 遵循“设计即分解”原则，采用图形化框架构建，使有害任务能够被分解为单独看来无害且可执行的子任务，并形成符合现实的工作流。我们使用定制分解器开展的实验表明，当前最先进的智能体在面对整体式有害任务时具有较高的拒答率，但在其分解变体上的拒答率显著降低，并且往往会无意中完成对抗性目标。这些发现凸显了针对分解攻击开展安全评测并制定相应防御措施的必要性。我们的数据集已公开发布，可在 https://huggingface.co/datasets/decompositionbench/DeCompBench 获取。
arXiv:2605.21347v4 公告类型：replace-cross 摘要：对 LLM 智能体中的故障进行诊断，目前仍主要依赖人工。实践者检查少量执行轨迹，提出临时性假设并不断迭代。这一过程会遗漏只有在轨迹群体中才会显现的模式，而且无法扩展到生产环境中的语料库，因为其中单条轨迹可能包含数万个词元。我们将语料库级轨迹诊断形式化。给定一个执行轨迹语料库，目标是生成有依据的自然语言洞见，用于刻画不同轨迹组之间的系统性行为模式，并将每条洞见与支持性证据关联起来。我们提出了 Insights Generator（IG），这是一个多智能体系统，通过在轨迹语料库中提出并检验假设来回答诊断问题，从而生成有证据支持的洞见报告。我们从定性和客观两个维度对 IG 进行了评估，涵盖基于评分标准的报告评估，以及通过实施 IG 洞见所实现的下游性能提升。使用 IG 报告的人类专家使脚手架性能相比未经修改的基线脚手架提升了 30.4 个百分点；利用 IG 洞见的编码智能体也展现出持续且稳定的增益。在各项基准测试中，IG 的侦察员-调查员架构在检测覆盖率方面产生了可与其他方法相媲美的发现，同时领域专家认为 IG 报告在深度和证据质量方面处于领先水平。
arXiv:2606.22741v3 Announce Type: replace Abstract: Traces record what LLM agents did, leaving implicit what each step relied on. GRADE holds both in one graph; six published graph views are its projections. Execution edges come free; each costly dependency edge is graded as observed in the trace, declared by added logging, or inferred under an assumption, a schema we test one grade at a time. We recover the dependency layer from six tool-use, coding, and web corpora and measure its failure-prediction gain over run size. First, over a linear run-size baseline, it improves prediction on three corpora, but the gain depends on the learner and task sampling. On held-out corpora the layer stays above chance where run size does not, under one learner, with the same task-sampling caveat. Second, structure can pass for signal it does not carry: a layer inferred under full history restates run size. Within a corpus, a preregistered test does not separate the recovered layer from a degree-matched counterfeit that rewires each step's dependencies but keeps their number. The recovered layer's largest transfer lead over it reverses when the held-out corpus's task family is withheld. Before crediting structure, grade each edge, keep a run-size baseline, and score degree-matched attachment controls.
arXiv:2606.10684v2 公告类型：替换 摘要：智能体搜索系统必须协调证据获取与响应生成，但现有方法要么将这两个角色置于单一智能体目标之下，要么进行角色分解，却未能厘清它们各自对最终结果的贡献。我们提出 DAC（Divide and Cooperate），这是一种角色分解式训练框架：在给定任务特定的外部验证信号后，利用针对角色的交叉验证奖励分别训练搜索器和生成器。DAC 允许生成器在检索到的证据似乎不足时选择弃答，并结合这一决策与外部评估的搜索充分性，为每个角色分配适当的信用。为防止过度弃答这一退化现象，我们进一步引入困难正例证据增强，以抑制模型在证据充分但具有挑战性的情况下弃答。在七个通用及多跳问答基准和两种模型骨干上，DAC 始终优于强大的单智能体和多智能体基线。受控评估表明，这些收益源于搜索与生成两方面的协同改进。我们的结果凸显了在训练智能体搜索系统时，明确且恰当地在交互角色之间分配信用的重要性。
arXiv:2603.06859v4 发布类型：替换 摘要：对于LLM智能体团队而言，信用分配关注的是每条消息的价值，但这一问题大多被当作预测问题处理：学习得到的评论器、轨迹级分数或智能体消融，通常被用来替代定义信用所需的反事实，而后者很少真正执行。通过共享上下文进行通信的团队则有所不同：当下游智能体读取的所有内容都写入轨迹时，轨迹就是状态，而这一反事实可以被执行。于是，信用信号可以像任何估计量一样，通过偏差、方差以及与独立参考的吻合度来评判。C3，即通过反事实延续进行信用分配，在某个决策点替换一条消息，并将运行继续推进至终端奖励，因此其信用估计是无偏的，除蒙特卡洛误差外完全精确。给定采样得到的备选项，该误差的方差遵循一个推导出的规律，其中不包含与智能体数量相关的项；在包含2至10个决策点的6个工作流上，观测到的噪声符合这一规律。在2智能体链上，仅通过4次延续，C3相对于由16次延续构成的参考，其备选项排序的秩相关为0.69，接近两个此类参考彼此一致时的0.73；在相同rollout上训练的评论器则达到0.29。将C3用作训练2智能体团队时的优势函数，其表现优于我们比较中更强的基线MAGRPO；与MAPPO相比，训练token消耗减少了37%，因为只需重新生成决策点下游的消息。当轨迹就是状态时，信用无需被预测；它可以是精确的。
arXiv:2602.08234v2 公告类型：替换 摘要：大型语言模型（LLM）智能体在复杂任务中展现出惊人的效果，但它们往往孤立运行，无法从过往经验中学习。现有的基于记忆的方法主要存储原始轨迹，而这些轨迹通常冗余且噪声较多，导致智能体难以提取对泛化至关重要的高层次、可复用行为模式。本文提出 SkillRL，这是一种通过自动技能发现和递归演进，弥合原始经验与策略改进之间鸿沟的框架。我们的方法引入了基于经验的蒸馏机制，用于构建分层技能库 SkillBank；引入了自适应检索策略，用于获取通用启发式方法和任务特定启发式方法；还引入了递归演进机制，使技能库能够在强化学习过程中与智能体策略共同演进。这些创新显著减少了 token 占用，同时提升了推理效用。在 ALFWorld、WebShop 和七项搜索增强任务上的实验结果表明，SkillRL 达到了当前最佳性能，相较于强基线方法的表现提升超过 15.3%，并且随着任务复杂度增加仍能保持稳健性。代码发布于 https://github.com/aiming-lab/SkillRL。
arXiv:2610.10513v1 发布类型：cross 摘要：语言模型 Agent 越来越多地被要求开展开放式科学研究，但它们的结果通常是依据已知答案、评分标准或语言模型审查员进行评估的，而这些方式都无法判断一个新的科学模型是否有效。针对厄尔尼诺—南方涛动（ENSO）的 AI 科学考试（SciExam for ENSO）是一项基准测试，要求 Agent 根据真实观测数据，构建 ENSO 的低阶随机模型。ENSO 是年际气候变率的主导模态。在六小时的时间限制内，Agent 处理观测数据并编写自己的诊断程序，这些诊断程序随后被冻结；之后，Agent 只能将这些诊断结果作为反馈来开发模型。隐藏评分器随后测试模型能否再现 ENSO 的统计特征、恢复未观测变量并预测留出年份，并以同样的方式对一个已发表模型进行评分。在 12 个 Agent 系统中，有 6 个生成的模型得分高于已发表模型，主要得益于更好的重建和预测能力。表现更强的模型经过简化后，分别与关于 ENSO 暖冷不对称性的两种相互竞争的解释之一相一致；而这项任务从未提及这一仍在进行的公开争论。在信息条件不同的受控运行中，对表现最佳的系统进行测试表明，其得分并非来自回忆过时的观测记录，同时它所接收的信息会影响其构建模型的方式。因此，SciExam for ENSO 能够在不存在已知答案的情况下评估 Agent 的科研能力，而结果表明，Agent 已经能够构建具有竞争力的模型，其结构涉及科学家仍在争论的问题。
arXiv:2610.10409v1 公告类型：cross 摘要：通用智能体日益能够编写代码、使用工具并完成复杂的数字任务，这引发了一个问题：这些能力在多大程度上能够迁移到物理世界。为研究这一问题，我们推出了 RobotWorld，这是一个具有挑战性的机器人使用仿真测试平台：通过机器人接口，将指令和观测转化为物理任务执行。其 84 项任务涵盖操作、移动操作、运动、驾驶和空中控制，并设有明确的交互预算和可执行的成功检查机制。通过结合任务结果与执行轨迹进行分析，我们确定了能够迁移的能力，以及阻碍可靠完成任务的能力缺口。此外，我们发现当前的智能体能够构建复杂的感知与控制工作流，包括图像分割、相机标定、空间估计和基于动力学的计算。然而，这些能力并不能稳定地组合成成功行为：智能体即使到达指令指定的位姿，也会丢失与任务相关的物体状态；它们无法纠正无效动作，恢复得太晚，或将尚未完成的任务误认为已经完成。这种不均衡的迁移在不同模型之间也有所差异：Astra 在空间目标和受约束接触目标上的成功率更高，而 Opus 5.5 在连续平衡目标和定时交互目标上的成功率更高。通过将这些结果与执行行为联系起来，RobotWorld 既提供了严格的试验场，也对现存的能力缺口进行了实证刻画，从而为训练和设计更可靠的物理世界智能体确立了具体目标。
arXiv:2610.10407v1 Announce Type: cross Abstract: As option markets grow and AI advances, agentic systems for option trading are gaining increasing attention. Language-model-based agents can reason over contextual information such as news, but option trading presents a particularly challenging decision problem: a single stock can have thousands of contracts, and the agent must decide both which contracts to trade and how to combine them. Existing approaches often sidestep this complexity by restricting the policy to a fixed strategy structure, such as a straddle, limiting their ability to switch strategies as market conditions change. We present SOTA (Stock Options Trading Agents), an agentic trading framework for structured option-strategy selection. SOTA abstracts the large option universe into strategy-level decisions while deterministic resolvers handle portfolio implementation. We develop SOTA by post-training Qwen3.8-27B with supervised fine-tuning followed by reinforcement learning. SOTA is evaluated on options on nine large-cap U.S. equities and SPY against rule-based and machine-learning strategy selectors in the same trading environment. Over a six-month out-of-sample period, SOTA earns an 18.3% total return with a Sharpe ratio of 1.60 and a maximum drawdown of 8.96%. We also document an asymmetric role of news: news improves frontier-teacher trajectories, but retaining news during reinforcement learning reduces out-of-sample return from 18.3% to -2.7%.
arXiv:2610.10285v1 Announce Type: cross Abstract: In the wake of the increasingly public discussion about AI alignment, recent work has tried to propose specific AI architectures that behave safely. However, the proposed arguments that seemingly demonstrate proved alignment mostly neglect the environment the agent needs to act in. We discuss theoretical bounds for agent-agnostic safety guarantees in environments that can only be partially observed and within which an action is required within limited time. We introduce two realistic scenarios, one with an infinite state space and one with signal mixture. In these scenarios, we prove that even a perfect agent cannot guarantee safe behaviour. It will be argued that for any proof of AI safety or alignment, the environment and associated safe actions need to be specifically considered together with the agent.
arXiv:2610.10206v1 Announce Type: cross Abstract: Links are disjoint unions of circles smoothly embedded in $S^3$. We use reinforcement learning and Bayesian optimisation to obtain new upper bounds on several link invariants that are not known to be algorithmically computable: the slice genus and the unknotting number for links, and the strong slice genus for algebraically split links. We also compute lower bounds using known invariants. Combining the upper and lower bounds, we obtain new exact values in many cases. Our unknotting agents can reproduce the non-additivity of the unknotting number for several counterexamples due to Brittenham and Hermiller, in some cases finding new unknotting trajectories.
arXiv:2610.10179v1 公告类型：cross 摘要：搜索代理使 Large Language Models (LLMs) 能够针对复杂的多跳问题，迭代式地检索和利用信息。可验证奖励强化学习（Reinforcement Learning with Verifiable Rewards，RLVR）为这类代理的后训练提供了一种有前景的方法，但其对稀疏的、基于结果的监督的依赖，可能使信用分配变得困难，并限制学习效率。本文系统研究了中间监督如何改进搜索代理的强化学习。我们考察了一系列奖励塑形与信用分配策略，这些策略从中间检索步骤中提供学习信号。在此基础上，我们开发了一种训练框架，将中间信号与最终结果奖励相结合，以改进从多步搜索轨迹中的学习。在匹配的训练条件下，跨多个基准开展的实验表明，搜索代理的整体性能得到提升，并显示中间信号的选择及其信用分配位置都会影响训练行为。这些发现表明，奖励设计和信用分配是训练高效搜索代理时的重要设计维度。
arXiv:2610.10061v1 Announce Type: cross Abstract: This paper presents TRACK (Telemetry-Based Racing Analysis and Coaching Kit), which is a framework for analyzing driving performance in sim racing and profiling how individual drivers behave behind the wheel. We report this framework together with its limitations: we calibrate each clustering result against a null, and when one does not separate from chance, we say so. Instead of restricting ourselves to scoring drivers or sorting them into preset labels, we represent each recording session as a compact geometry in a four-dimensional behavioral space (speed, braking, strategy, and consistency), and we group these fingerprints by their similarity using unsupervised clustering. Over time, we have developed and refined this framework on the open Assetto Corsa Gym (ACGym) dataset. Our study suggests that corner types differ along a behavioral dimension that was not used to define them. It also suggests that when the car changes, only speed and consistency carry over in the restricted population, while repeatability could not be shown there for any of the braking or strategy measures. Cluster separation becomes less distinct as the range of available telemetry widens. Until that repeatability is shown, grouping on the braking and strategy dimensions cannot treat the car as interchangeable, which divides an already small sample into smaller cells. It is also not clear whether a driver's grouping carries over from one corner type to the next. We also normalize each metric against a reinforcement-learning reference agent. The reference does not depend on the sample, so the scale does not shift when the sample does. We intend these results as an analytical foundation for a personalized improvement suggestion system. The sample is small. The cross-car result changes when the sample is defined more broadly. These outcomes are preliminary.
arXiv:2610.09858v1 公告类型：交叉收录 摘要：大语言模型智能体通过将推理、工具调用与来自环境的观察交替进行，处理多步骤任务。已有研究表明，自然语言反馈可以帮助这些智能体在任务执行过程中修正决策。我们提出 Caddie，一种训练评议模型的方法，使其在智能体执行任务时提供自然语言分析和建议。与依赖步骤级标签或参考评议的方法不同，Caddie 根据智能体在接收评议模型的反馈后最终是否成功来学习。我们通过强化学习优化评议模型，同时保持基础模型冻结。我们的 Qwen3-4B 评议模型使用单一基础模型在多跳问答任务上训练后，提升了四种不同规模和架构的基础模型的成功率，其中三种模型未参与评议模型的训练。在 MuSiQue 基准测试中，训练后的评议模型将 Qwen3-4B 的成功率提高了超过 25 个百分点，超越了未使用评议模型的 Kimi K3 的表现。同一个评议模型在无需额外训练的情况下，也在包括 $\tau^3$ 和 DeepDive 在内的域外交互式基准测试中带来了性能提升。我们的结果表明，智能体可以在推理时决定何时向评议模型寻求帮助，而基于结果的评议模型训练能够产生可跨基础模型和任务领域迁移的指导。
arXiv:2610.09624v1 公告类型：交叉收录 摘要：工具调用，即按需调用外部工具，是智能体式 LLM 的核心能力，但决定模型是调用工具还是直接作答的机制仍不清楚。智能体提示通常很长，且包含大量框架性内容，将角色指令、工具模式、格式模板和用户请求组合在数百个 token 中，形成噪声繁多、高度纠缠的上下文，其中没有显而易见的、可用于机制分析的单一可控变量。为获得这样的变量，我们提出一种方法，将复杂的智能体提示转化为最小对照提示对，其中仅一个请求动词就能决定是否调用工具：将执行类动词（例如“编写”）替换为分析类动词（例如“讨论”），能够可靠地翻转这一决策，表明该决策由一种紧凑的内部状态所介导。我们围绕 Python、Java 和 C++ 构建了 500 组这样的配对提示（300 组用于机制分析，200 组留出用于评估）。我们将这一决策追溯到向量 $\mu_\Delta$，该向量在因果上既是必要的，也是充分的，并且能够泛化到发现阶段所用提示之外的原生多轮 $\tau^2$-Bench 轨迹和不含动词的请求。行为消融实验表明，框架性内容建立了工具调用先验；随后，Transcoder 分解揭示，分析类动词通过表明无需使用工具的特征来抑制这一先验，而执行类动词则基本使其保持不变。下游读取框架性内容的注意力头和 MLP 特征读取由此形成的状态，而同一机制在 Qwen、Mistral 和 Granite 系列的七个模型中反复出现。我们的代码已公开于 https://github.com/XijieGo/MI4ToolCalling。
arXiv:2610.09377v1 公告类型：交叉收录 摘要：分类体系是 AI 系统用于组织证据、汇总模式并针对大规模文档集合回答问题的符号表示。在客户反馈场景中，类别树决定每条记录如何被统计和分派、哪些问题能够被发现，以及由哪个团队负责。如今，智能体运行框架让生成看似合理的层级结构变得轻而易举，而目前对此类树的检验采用的是通用的、针对单个节点的检查：每个名称是否符合其描述、是否位于正确的父节点之下，以及是否与兄弟节点保持区别。我们提出了一个更贴近实际运作的问题：生成的层级结构何时才能真正成为生产环境中实用的分类体系？我们基于两个专有反馈语料库（分别包含 1,940 条和 5,000 条记录）构建了六个分类体系：每个语料库对应一个生产环境参考体系，以及在相同输入下对同一运行框架重复运行两次所生成的体系。这六个体系全部通过了所有通用的命名和结构检查，而更深入的产品覆盖度检查甚至更偏好生成的树。然而，每棵生成树中至少有 97.7% 的叶节点名称仅仅是在重述某个祖先节点的名称（两个参考体系中的这一比例分别为 13.9% 和 2.9%）；其中一棵生成树中，每四条记录就有三条归属于多个顶层类别。我们引入了两类针对整棵树的指标：结构判别指标用于检验树的形态究竟是从数据中学习得到的，还是由生成器强加的；团队可划分性指标用于检验各分支是否将反馈划分为可由不同团队负责的组。被通用检查评为同等正确的两棵树，其跨分支泄漏率相差 27 个百分点，而且其中仅有一棵优于随机划分。表面上的合理性不足以衡量分类体系的质量：评估不仅必须衡量每个节点，还必须衡量整棵树。
arXiv:2610.09335v1 Announce Type: cross Abstract: Autonomous unmanned aerial vehicle (UAV) object search involves a closed loop of perception, decision-making, and action under partial observability. Urban environments pose several challenges: large search areas and narrow egocentric views limit coverage, dense 3D geometry constrains safe motion, and open-world instructions require identifying a specific target among distractors. Many existing methods mitigate partial observability through explicit maps or memory representations, yet remain largely reactive, reasoning over past observations without explicitly predicting future states. World models enable prospective reasoning through imagined rollouts. However, image-generating world models can incur high inference latency, while spatially grounded planning remains challenging for latent world models. We propose SearchWorld, a recurrent state-space world model that connects explicit spatial memory with value-guided imagination. The model maintains BEV exploration and obstacle memory and decodes a task-aware spatial value layer to guide search. A cognition-action network uses this learned spatial value prior to improve the policy through imagined rollouts, without training a separate scalar critic. Training progresses from world-model learning to expert imitation and imagination-based exploration refinement. On UAV-ON, SearchWorld improves the success rate to 23.8% (19.5% for the strongest published agent) and raises oracle success to 35.5%, while remaining robust on unseen scenes (19.9% success rate). By grounding imagination in explicit spatial representations, SearchWorld enables UAV agents to plan prospectively rather than react.
arXiv:2610.09294v1 公告类型：交叉发布 摘要：AI 智能体的快速发展使智能体安全日益受到关注，目前大量评估集中于数字环境。随着智能体进入物理世界，具身安全变得愈发重要：失误可能造成人员伤害和高昂的硬件损失。除了选择安全的行动，具身智能体还必须在实时约束下运行：物理世界不会在智能体推理时暂停。推理期间，行人持续移动、车辆不断靠近，因此，在观测时看似安全的行动，可能在执行前就已变得不安全。因此，实时具身安全同时取决于决策质量和决策延迟。我们提出 RT-SAFE，这是一个用于在实时约束下评估具身智能体安全性的模拟城市基准。RT-SAFE 将导航任务与动态参与者、环境危险和交通规则相结合，同时允许世界在整个推理与行动执行过程中持续演变。对八个 VLM 的评估表明，智能体的任务完成率很高，却几乎无法安全完成任务：在最困难的设置下，仅有 0.7% 的回合能在未发生安全事件的情况下结束。更引人注目的是，在条件匹配的静态评估和实时评估中，任务完成率分别为 91.3% 和 94.1%，但实时执行使碰撞次数增加至 $12.3\times$。这些结果表明，常规的任务成功指标可能掩盖严重的安全问题，而决策延迟本身也可能成为物理风险的来源。最后，我们展示了 RT-SAFE 能够支持离线强化学习训练，在保持较高任务完成率的同时显著降低碰撞率。
arXiv:2610.09244v1 Announce Type: cross Abstract: Recent advances in mean-field games and its multi-population variants enable large-scale heterogeneous multi-agent systems to be modeled through representative agents and their associated mean-field distributions. However, existing approaches do not explicitly account for uncertainty in the behavior of other populations. To this end, we introduce a new paradigm: risk-averse multi-population mean-field games, where each population optimizes a worst-case expected reward over dynamically feasible ambiguity sets of mean-field flows of a subset of the other populations. Employing an occupation-measure formulation along with tools from set-valued analysis, we establish, under mild assumptions, several theoretical properties of the multi-population game, including the geometric properties of the ambiguity sets and the existence of a novel risk-averse multi-population mean-field equilibrium. Further, we derive contractivity results of the fixed-point operator under entropy regularization and show that it can be utilized to learn the equilibrium. Finally, we propose a risk-averse fictitious-play scheme and show that exploitability decays to zero, despite the additional nonlinearity introduced by the worst-case objective. We report several numerical experiments to illustrate convergence and risk-averse behavior.
arXiv:2610.09115v1 公告类型：交叉收录 摘要：引导 LLM 智能体意味着决定是否纠正它、在哪一步纠正，以及采用哪种机制。不确定性常被用于决定何时纠正智能体，但它能否指导这些决策仍不明确。我们在每个非终止步骤分别采用四种机制中的每一种来引导智能体轨迹，并将每次后续执行运行至结束。由此得到的逐步结果表（SOT）包含来自三个基准测试、两个智能体的 1,864 条轨迹的约 82,000 次反事实后续执行。结果表明，不确定性能够识别会失败的轨迹，但没有任何单一信号能够可靠地定位引导会产生帮助的步骤。因此，我们提出 VoS（Value of Steering），一种可离线或在线使用的轨迹级监控器，它从 SOT 中学习每一步的引导价值，并据此决定在哪里进行引导。一个受损害预算约束的触发器负责决定是否进行引导，以限制 VoS 干扰成功轨迹的比例。在由基准测试、智能体以及离线或在线使用方式构成的全部 12 种设置中，VoS 相比未经干预的执行平均提升了 7.8 个百分点；在其中 11 种设置中，VoS 超过了五种现有不确定性触发方法中表现最强的方法，平均提升了 2.9 个百分点。消融实验表明，基于实测结果进行训练和采用严格的损害预算都至关重要。
arXiv:2610.09034v1 公告类型：cross 摘要：连续环境中的多智能体路径规划（MAPP）通常依赖路网来平衡安全性与搜索效率。然而，传统的路网生成方法（如格状网格或标准的基于采样的方法）往往面临图密度与找到可行且高质量解的可能性之间的权衡。本文提出了一种可扩展的异构图神经网络（GNN）框架，用于自动生成和评估共享多智能体路网。我们的模型将路径点、智能体位置和任务位置表示为异构图中的不同节点，从而能够推理全局连通性和智能体间的交互。通过在从专家求解器轨迹中汇总和收集的占用密度图上进行训练，GNN 学会识别关键兴趣点，并剪除冗余节点和边。该过程生成了一个紧凑且具备协调感知能力的路网，该路网对任务排列具有不变性，并可复用于多智能体取货与配送任务。实验结果表明，我们的框架能够减少规划工作量，并有望找到更优解；对于稠密路网，运行时间和图规模均至少可减少 40%。
arXiv:2610.08875v1 公告类型：cross 摘要：Agent Skills 将程序性指导和资源打包以供复用，但相关 Skill 并不一定能提升任务性能。现有研究刻画了 Skill 的内容并评估其下游性能，但对于效用如何取决于内容、执行配置和多 Skill 组织方式，解释仍然有限。我们针对 87 个 SkillsBench 任务开展实证研究，将下游效用定义为：在相同模型—harness 配置下，同一任务相对于 No-Skill 的通过率差异。我们在九种配置下比较相同的 Skills，随后在三种选定配置下考察其他已发布的 Skills，以及固定 Skill 集合的不同组织方式。我们从一个经过筛选、包含 37,596 个 Skills 的语料库中检索 marketplace 候选项。通过 LLM 辅助分析内容、执行轨迹和最终产物，并由作者进行复核，我们将所提供的支持与实际使用情况及任务结果联系起来。在 36.78\% 的任务中，相同的 Skills 对某些配置有帮助，却对其他配置造成不利影响；执行轨迹表明，推荐的程序可能会变成执行负担。相关性排序忽略了更有用的候选项。在所评估的候选集合中，依据对所需操作的支持程度重新排序后，在三种配置下，首选项通过率提高了 4.35--5.80 个百分点。我们总结出 17 条编写实践，将可执行程序与恢复机制、任务要求的保留以及最终产物检查联系起来。Stage Plan 和 Dependency DAG 均优于单纯按照使用顺序组织，其中 DAG 的额外收益主要集中在配备五个或六个 Skills 的任务中。这些发现指导开发者评估可用的操作支持，在保留任务要求的同时允许调整程序，并在组织 Skills 时明确产物依赖关系。
arXiv:2610.10394v1 Announce Type: new Abstract: Kernels encode the inductive bias of a wide range of machine learning models, yet automated kernel design faces a fundamental dilemma. A fixed grammar of base kernels and operators guarantees validity but limits the search to structures expressible by those building blocks. Conversely, unrestricted programs remove this limitation but no longer guarantee validity. In our stress tests, 22-58% of LLM-generated kernels that pass numerical checks on random inputs fail when evaluated at different scales or dimensions. We propose Kernel Autoresearch (Kernaut), which treats kernel design as open-ended model discovery. Coding agents write kernels as programs, while construction contracts ensure that every accepted kernel is valid. A quality-diversity archive retains high-performing kernels with distinct behaviors, and novelty screening steers agents toward functionally new candidates. Our experiments demonstrate that the discovered kernels encode reusable inductive biases that generalize to unseen tasks. On held-out black-box optimization families, a discovered kernel outperforms a meta-learned deep kernel trained on the same episodes. Furthermore, kernels discovered from ten enzyme-kinetic rate laws achieve lower error than tuned ARD and deep kernel baselines on five unseen mechanisms. The discovered kernels are also interpretable programs that human researchers can refine: a human-refined version of one further reduces the held-out predictive error by 5.7% and optimization regret by 7.8%.
arXiv:2610.10302v1 Announce Type: new Abstract: In Continual Multi-Agent Reinforcement Learning (CMARL), agents learn cooperative policies across sequences of tasks, aiming to adapt effectively to new tasks while preserving the ability to solve previously encountered ones. In many applications, tasks differ in their underlying structure, which can represent, for example, distinct operational conditions or target configurations (e.g., different network topologies in power grids or arrangements in formation control). Existing CMARL methods lack dedicated mechanisms to leverage this structural information when learning new tasks, failing to promote transfer and mitigate forgetting. To fill this gap, we propose Continual Graph Multi-Agent Reinforcement Learning (CGMARL), a novel framework for CMARL problems in which task sequences are mapped into a series of attributed graphs, each modeling a task-specific structure. In CGMARL, each graph determines the environment dynamics (next states and/or rewards) and the number of agents for the corresponding task. Then, we present Graph-based Formation (GRAFO), the first CGMARL benchmark, and show how forgetting arises in this setting. Finally, to address this limitation, we propose Frozen Graph Encoder (FROG), a method that relies on a frozen graph backbone to preserve past structural information in graph-based CMARL policies. Experiments on GRAFO show that pairing FROG with existing CL methods substantially improves performance on multiple CGMARL scenarios.
arXiv:2610.10087v1 Announce Type: new Abstract: Offline MARL increasingly relies on generative policies to model multimodal joint behavior, typically by distilling a centralized teacher into decentralized one-step actors under the CTDE. We identify a failure mode at the teacher training stage: standard flow-based teachers pair noise with replay targets independently, so nearby noise samples can be routed toward conflicting coordination modes. The teacher then produces samples between valid modes, and because the distillation loss regresses each local actor onto the conditional mean of the teacher's output given local input, this error is not absorbed but propagated to the student. To remove this teacher-side artifact, we propose Mode-Support Semi-Discrete Optimal Transport (MoSDOT), which summarizes multimodal replay into a finite mode support with prescribed capacities and uses conditional semi-discrete optimal transport to assign each noise sample to a single mode before teacher training. We additionally study a shared-randomness variant that uses a shared noise component at execution to expose the residual gap intrinsic to strict-product execution. On controlled diagnostics and offline MARL benchmarks, MoSDOT improves endpoint quality and routing consistency, particularly on datasets exhibiting multimodal joint behavior.
arXiv:2610.09929v1 Announce Type: new Abstract: Sample complexity is a widely used metric in sequential decision-making problems, defined as the number of suboptimal decisions during the interaction between the agent and an environment. We study the sample complexity of stochastic multi-armed bandit problems and introduce the expected sample complexity performance measure, analyzing it in a novel framework called approximately correct in expectation (ACE). We show that ACE guarantees imply almost sure convergence to the optimal expected reward, in contrast to high-probability guarantees found in other frameworks, and also show how to convert ACE guarantees into explicit expected regret bounds. We further show that, in contrast to existing measures, deterministic algorithms cannot obtain favorable ACE bounds, and analyze stochastic algorithms in two settings: when the allowed suboptimality level $\epsilon$ is known to the algorithm and when it is unknown. In the former, we devise an explore-then-$\epsilon$-greedy algorithm, and in the latter, we analyze the expected sample complexity of Thompson sampling. Finally, we establish nearly matching lower bounds for both settings, showing that the algorithms are tight in $\epsilon$ and proving a performance separation between the two regimes.
arXiv:2610.09824v1 公告类型：新 摘要：多智能体系统（MAS）利用智能体之间的交互来执行复杂任务。尽管这类系统取得了成功，我们表明，这些交互也可能导致同质化，即智能体逐渐趋于相似的行为。MAS 中的同质化会降低智能体多样性，并强化共同的失败模式。在本文中，我们使用三个指标对同质化进行操作化定义：对多数意见的趋同、向极端的极化，以及在后续交互中对变化日益增强的惯性。我们在代码生成、招聘和科学同行评审任务中评估了 MAS 的同质化现象。在这些任务中，我们表明，同质化会转化为具体的下游风险：在代码生成中，它会掩盖并放大相关错误，从而造成系统性漏洞；在招聘中，即使存在偏见的智能体被移除，其影响仍可能长期持续；在同行评审中，它会导致不同研究领域之间形成不均衡的评价标准。我们的结果确立了同质化是 MAS 的一种失效模式，表明对 MAS 的评估必须超越总体性能，仔细分析交互动态。最后，我们表明，提高多样性的简单方法——利用采样随机性和混合模型 MAS——并不能降低同质化风险，这凸显了制定能够有效利用智能体多样性的策略的必要性。
arXiv:2610.09633v1 Announce Type: new Abstract: The evaluation of coding agents generally strives to be as realistic as possible. In our study, we collect 4,782 agent sessions of real software engineers in JetBrains IDEs, which we call Production Sessions. Since our subject is interactive agents, we study the sessions with at least three user messages (33% of the sample). These long sessions differ from issue-derived benchmark tasks in two ways: (i) user requests span a far wider mix of task types - questions about the project's code, planning, review, refactoring, execution - and (ii) users switch between types throughout a session. Long-session samples from three public interaction corpora exhibit markedly different Task Flows (the distributions of session lengths, task types, and type-to-type transitions), so no single interaction distribution is universally realistic: benchmarks should name a target use case and calibrate to measurements from it. We present SWE-TaskFlow, an approach for transforming any issue-derived benchmark: it preserves the verified tasks and tests while steering the interaction toward a target Task Flow through prompt splitting and verifiable repository QA, with a TaskFlow Alignment Score (TFAS) for selecting among generated trajectories. In a pilot on 700 SWE-Bench Pro tasks, solving the task sequentially in several steps approximately doubles agent cost without a stable change in resolve rate: the interaction protocol itself is an important dimension of evaluation.
arXiv:2610.09375v1 公告类型：新 摘要：越来越多的组织使用前沿语言模型分析客户反馈，但回答质量也取决于这些反馈如何被组织和提供。我们将\emph{Customer Context Graph}定义为客户与业务上下文的统一模型。类型化关系连接客户对象（反馈、对话、用户和账户）、运营对象（工单、支持代理、商机和竞争对手），以及分析或行动对象（分类体系概念、证据、洞察、工作项和结果）。这使代理不仅能够调查客户说了什么，还能了解其原因、受影响者、后续采取了什么行动、由谁负责，以及问题是否得到解决。在本实验中，该图谱由公开的 Cursor 反馈构建；同一架构也可支持任何类型的反馈来源。我们使用相同的 9,432 条公开 Cursor 反馈记录和 30 个现实的产品、事件、比较及元数据问题，对比了 Agentic RAG、Deep Research Agent 和 Customer Context Graph-backed Agent。在缺乏详尽真实答案的情况下，我们依据一套针对 30 个问题中的 28 个进行校准的比较评分标准，从回答质量（覆盖面和组织性）、分析深度（具体性和分解程度）以及证据质量（引用支持和可追溯性）三个方面对回答进行联合评分。我们根据各项主张抽样检查被引用的记录，并对三个维度赋予相同权重。在这一对齐后的评分标准下，Customer Context Graph-backed Agent 的总体得分为 0.961，而 Deep Research Agent 和 Agentic RAG 分别为 0.710 和 0.651；在 30 个配对问题中的 27 个问题上，它领先于 Deep Research Agent（符号检验 p < 10^(-5)；在 30 个问题中的 26 个上严格领先）。其最大优势在于分析深度（0.967 对 0.642），体现为更具体、层次更清晰的发现，以及经过量化的主题和可追溯的证据……
arXiv:2610.09285v1 Announce Type: new Abstract: Many applications of physical AI run within finite or closed physical worlds with a limited set of physical laws governing object behavior. Examples include robots working in a warehouse and agents moving around in a video game. In order to better organize, filter, and curate data for physical AI applications, we propose a new approach centered on the unique settings and physical laws of individual datasets. We train LeCuration, a small world model intended to serve as a data curation tool for a separate, larger downstream model. To build this model, we choose LeWorldModel (LeWM)as our latent encoder and predictor, adding a diffusion transformer (DiT) decoder to add visuals to autoregressive gameplay rollout. We find that the embeddings of this model can be used as an anomaly detection signal and as a content-based clustering heuristic, and that auto-regressively predicting the game state with this model allows us to qualitatively check for action-state consistency. This paper presents a qualitative, proof-of-concept case study on CS:GO gameplay data; we do not yet report quantitative curation metrics or downstream training results, which we identify as the key next step.
arXiv:2610.09174v1 Announce Type: new Abstract: Reliable and interpretable trajectory prediction is critical for cooperative and autonomous driving in complex and uncertain environments. This paper introduces a Context-Aware Attention-based Gaussian Mixture Model (CAA-GMM) for multimodal, uncertainty-aware motion forecasting. The proposed approach models future motion as a probabilistic mixture conditioned on both scene context and agent dynamics, capturing diverse behavioral modes with interpretable Gaussian components. A lightweight attention mechanism adaptively encodes inter-agent interactions and contextual salience, enabling efficient fusion of rasterized environment cues and motion history in dense traffic scenes. Comprehensive evaluations on the nuScenes and Argoverse 2 datasets demonstrate that CAA-GMM achieves competitive or superior accuracy compared with state-of-the-art raster-based baselines, while maintaining low computational complexity. Ablation analyses confirm the importance of the attention module for robust contextual reasoning and predictive precision. Furthermore, evaluations under imperfect communication and perception conditions highlight the framework's resilience to uncertainty, establishing CAA-GMM as an efficient and scalable solution for cooperative trajectory prediction in intelligent transportation systems.
arXiv:2610.09141v1 Announce Type: new Abstract: We present a hybrid quantum-classical framework that detects affinity and romance-investment fraud by modelling the cognitive biases in a manipulative conversation. In our proposed framework, cognitive biases central to this fraud class are carried by dedicated qubits in a structured parameterized quantum circuit, together with a frame qubit makes the encoding sensitive to the temporal order of manipulative reframing, and a narrative qubit that aggregates co-occurrence through a trainable entanglement layer. The circuit parameters are trained jointly with a classical reinforcement-learning agent that decides, turn by turn, whether to flag the conversation, modeled as an optimal stopping problem. We evaluate the model's performance on synthetic conversations that include hard negatives, legitimate but urgent, and legitimate but pushy sales conversations.
arXiv:2610.09120v1 公告类型：新 摘要：随机化探索是老虎机学习、多智能体强化学习和零阶策略搜索的核心，但其独立性和隐私性通常仅被视为技术性假设。我们表明，这些性质对于安全至关重要，并展示对抗性智能体如何利用另一智能体探索过程中的特权信息。我们在最简的双玩家强单调设置下分析一对欺骗者—受害者，其中欺骗性玩家获得与受害者探索过程仅存在相关性的泄露信号。我们证明，欺骗性玩家通过将自身的探索动作与该信息耦合，注入一种外部性，使学习动态转向新的稳态，称为欺骗性纳什均衡（DNE）。我们证明，欺骗性老虎机学习（DBL）动态会收敛到DNE的任意小邻域，同时保持最优收敛速率。有趣的是，我们的分析在放宽标准老虎机优化文献中的二阶光滑性条件的同时，仍然达到了这些最优速率。我们刻画了欺骗会严格改变稳态及其对欺骗者成本影响的条件，并在资源分配博弈中展示了这些结果。
arXiv:2610.09074v1 Announce Type: new Abstract: Emerging long-horizon agentic tasks require repeated model calls, worsening the inference cost of already-costly language models. While narrow agentic tasks suggest potential for aggressive model pruning without performance drop, empirical results show existing methods proposed for question answering tasks severely degrade task performance when applied to agentic models. We trace this failure to two decisions: what to prune and how to recover. For pruning, one-shot importance estimates fail to track how the pruned model adapts. For recovery, offline distillation covers only teacher prefixes, while full-trajectory on-policy distillation causes student errors to compound across turns. In this work, we propose Trajectory-Anchored Pruning (TAP), the first structural pruning framework for reinforcement learning (RL)-trained agents. TAP couples structural pruning with efficient on-policy recovery, anchoring interactions to teacher trajectories while allowing the student to generate each reasoning-action response. A frozen dense teacher supervises the student's response prefixes, addressing within-response training-inference mismatch while preventing student-induced deviations from propagating across training turns. Instead of one-shot pruning, TAP re-scores channels using gradients of the recovery objective on the recovered student, connecting iterative channel selection to the evolving policy. With 60% of FFN channels removed, TAP retains 99.2% and 88.0% of the dense 7B agents' task success rates on ALFWorld and WebShop, respectively, while reducing GPU time per successful task by approximately 22% and 17%. These results demonstrate effective structural compression of long-horizon agents under a limited recovery budget.
arXiv:2610.08959v1 公告类型：新 摘要：在线策略蒸馏通过在强化学习奖励稀疏且每条轨迹仅返回一次时，由教师策略提供密集的步骤级指导，对大型语言模型智能体进行后训练。现有方法根据每一步教师策略与学生策略之间的分歧大小分配指导信号，其单轮交互直觉是：较大的分歧意味着一个值得纠正的错误。然而，当决策跨越多轮交互形成链条后，这一规则会失效，因为早期偏移会进入两个策略所依赖的每个后续上下文，使教师策略与偏移后的轨迹保持一致，而不是指出偏移的根源；与此同时，可互换的步骤也可能产生较大的、但与结果无关的分歧。我们在一个智能体基准测试中验证了这一点：蒸馏分歧最大的步骤，相较于随机选择并未带来稳定收益。为此，我们提出GraphOPD，这是首个将基于图的结构增强引入智能体能力在线策略蒸馏的方法。GraphOPD从环境自身记录的状态变化中读取哪些步骤促成了后续哪些步骤，从而不受会破坏师生分歧信号的偏移影响；随后将这些关系组织成依赖图，通过图上的随机游走稳态分布为每一步评分，并将结构性归因与分歧信号融合为一个相对于轨迹的掩码，把监督集中在每次 rollout 中最具提升潜力的步骤上。在ALFWorld、WebShop和SearchQA上的三种模型规模与十一个基线方法中，GraphOPD始终展现出有竞争力的性能，相较最强基线最高提升5.8个百分点。基于执行回放的审计进一步表明，该结构性归因分数对真实因果影响的追踪能力远高于随机水平；两种融合信号均不可或缺；同一信号还可以迁移到域外的工具集成推理任务中。
arXiv:2610.08820v1 公告类型：新 摘要：基于大语言模型的多智能体系统通过协作提升复杂问题的解决能力，而潜在通信则直接传输模型内部状态，从而避免自然语言带来的高昂推理成本。然而，现有基于 KV 的潜在通信方法优先保证发送端状态的保真度，导致大量通信与计算开销，并可能引入冗余信息。为解决这些局限性，我们从面向任务的角度重新审视潜在通信，将其目标从发送端状态保真度转向接收端任务充分性。在此基础上，我们提出 KITE，一种无需训练的面向任务的关键层 KV 通信框架。KITE 使用接收端轨迹失真标准识别对任务有效的关键层，仅传输与该关键层关联的潜在工作记忆，并进一步将同一层作为自回归潜在推理的入口。在涵盖两个模型系列和三种模型规模的七个基准测试中，实验结果表明，与全层 KV 通信相比，KITE 将通信量减少了 28-36$\times$，端到端推理速度最高提升 3 倍，并将准确率最高提高 23.3 个百分点。
arXiv:2610.08815v1 Announce Type: new Abstract: Agentic AI-enabled automation cannot be safely deployed in high-stakes environments on probabilistic reasoning alone. A recurring risk is epistemic drift: as reasoning deepens, system behavior may move away from subject-matter-expert constraints for safe operation. This paper presents BRaVeS, a bounded reasoning and safety-governance framework termed the Defensible Next-Gen Reasoning System (DNRS). BRaVeS encodes SME-defined constraints as invariant anchors, proposes MoDA-Style (Mixture of Depths Attention) depth-aware access as a candidate mechanism for keeping these anchors visible during inference, and uses a state hierarchy (SMARtAutonomy) to reduce autonomy as epistemic risk increases. To formalize bounded recovery, we introduce the Lyapunov-Bounded Consensus Framework (LBCF), which maps continuous epistemic-risk signals into a finite K-bag abstraction and applies shielded state transitions that enforce Lyapunov-style energy descent or route the system to a human-mediated terminal state. The formal convergence result applies to the finite LBCF abstraction under fixed thresholds and feasible-shield assumptions; it does not prove safety of the full continuous neural activation space. We evaluate the framework through a discrete event Monte Carlo simulation using HAI 22.04 industrial-control-system time-series data with synthetic noise and sensor-degradation regimes. Across the tested parameter-grouping strategies and thresholds, the LBCF process achieved finite-step convergence and no safety-guard violations. These results provide simulation-based evidence that bounded governance behavior can be enforced under the stated abstraction, while motivating future work on deployed transformer implementations, live human-in-the-loop validation, and broader adversarial settings.
文本到音乐系统生成的音频日益逼真，但现有评估难以揭示生成结果是否符合用户意图。全局文本与音频相关性分数可能忽略描述不充分的提示中隐含的意图，并掩盖在乐器配置、结构、节奏或情绪推进等具体要求上的不足。为弥合这一差距，我们将文本到音乐的意图对齐表述为满足针对每项请求制定的评估准则，其中包含可独立验证的条目，既涵盖请求的显式要求，也涵盖其隐含的音乐意图。对各条目单独评分，使评估能够按意图来源和音乐维度诊断问题，而非仅给出单一且不透明的分数。我们将这一方法具体实现为 MuRA-Bench，这是一个由音乐专家整理、基于真实平台请求构建的基准。我们进一步提出 MIRA（Musical Intent Refinement Agent），一种测试时智能体：它首先将请求意图具体化为评估准则，然后在有限预算下，为黑盒生成器搜索提示的修订方案，迭代地生成音乐、依据评估准则进行验证，并利用这些反馈引导能够感知搜索轨迹的树搜索。在开源和商业后端上的实验表明，MIRA 能够提升意图对齐程度，使开源生成器达到与代表性商业系统（如 Suno 和 Mureka）相当的性能。项目页面：https://mirareview.github.io/。
多模态基础模型的最新进展使其能够作为通用物理智能体，执行多种操作任务。然而，要在陌生环境中成功完成任务，当观测中缺少任务相关信息时，智能体可能需要通过交互主动获取这些信息：它可能需要确定相关物体的位置、检查尚未观测到的属性，或探明陌生工具的作用。因此，我们提出 RoboQuest，这是一个面向目标导向具身探索的基准，要求智能体通过物理交互主动获取任务相关信息，利用由此获得的证据调整后续行动，并自主决定何时着手完成任务。RoboQuest 包含十项移动操作任务，围绕三类不确定性展开：搜索、基于操作的检查和交互式测试。我们通过统一的视觉运动接口评估了五个前沿多模态智能体，并评估了一个利用我们发布的完整回合演示数据微调的 $π_{0.5}$ 策略。表现最好的智能体仅在 23\% 的回合中成功，而微调后的策略几乎从未成功。在提供隐藏信息的情况下，对构成这些任务的执行技能进行单独测试，结果表明智能体能够完成大多数所需动作；我们的失败分析也显示，只有少数失败可归因于执行问题。失败分析进一步发现，智能体经常过早停止探索，在观测到完成任务所需的证据之前就做出决策。我们还发现，智能体很少预防或修复其探索造成的扰动。此外，对大多数模型而言，通过试错进行学习仍然十分困难。
As option markets grow and AI advances, agentic systems for option trading are gaining increasing attention. Language-model-based agents can reason over contextual information such as news, but option trading presents a particularly challenging decision problem: a single stock can have thousands of contracts, and the agent must decide both which contracts to trade and how to combine them. Existing approaches often sidestep this complexity by restricting the policy to a fixed strategy structure, such as a straddle, limiting their ability to switch strategies as market conditions change. We present SOTA (Stock Options Trading Agents), an agentic trading framework for structured option-strategy selection. SOTA abstracts the large option universe into strategy-level decisions while deterministic resolvers handle portfolio implementation. We develop SOTA by post-training Qwen3.8-27B with supervised fine-tuning followed by reinforcement learning. SOTA is evaluated on options on nine large-cap U.S. equities and SPY against rule-based and machine-learning strategy selectors in the same trading environment. Over a six-month out-of-sample period, SOTA earns an 18.3% total return with a Sharpe ratio of 1.60 and a maximum drawdown of 8.96%. We also document an asymmetric role of news: news improves frontier-teacher trajectories, but retaining news during reinforcement learning reduces out-of-sample return from 18.3% to -2.7%.
研究智能体的部署正朝着数千个智能体共享同一计算资源池的群体规模发展，而当前大多数系统一次只组织一个项目，或让整个群体处于无组织状态。我们认为，无论设计者是否提供组织架构，这样的群体都会形成某种组织，因此设计者应当明确提供组织架构，而多智能体系统研究界已经具备实现这一目标的工具。我们提出一种智能体社会，即在明确制度下运行、持续存在的智能体群体，并将其面向科学研究发展为一个基于六项原则构建的研究者社会。首席研究员通过项目征集、独立评审和资助机制竞争计算资源；一位被称为“市长”的人类管理者负责分配资源，但不指派任务。在一个由一万名研究者组成、正在运行的社会中，研究者仅被要求改进语言模型的预训练，其中一个实验室报告了一种可在减少约 30% 计算量的情况下达到同等质量的方法，但对其进行测试的各实验室尚未就这一结果达成一致。最后，我们为智能体研究界提出了六个尚待解决的问题。
我们如何预测哪个基础模型检查点值得投入一轮昂贵的智能体后训练？端到端 pass@$K$ 检验成功行为是否已经出现在基础模型的分布中，但它并不适合智能体编程：许多基础模型检查点无法可靠地生成格式正确的工具调用，而这是端到端完成任务所必需的。单次或短程任务将多步交互压缩为固定提示和单个补丁，从而避开这些工具调用失败，但也绕过了我们真正关心的核心能力：随着代码仓库不断演变，在多个使用工具的步骤中维持连贯状态。为弥合这一差距，我们将后训练智能体的成功轨迹视为基础模型潜力的前瞻信号。通过重放每条轨迹，并在每个修改代码的步骤后重新运行测试，可以识别出决定性步骤：即累积补丁首次使代码仓库从测试失败转为通过的步骤，从而确认在给定先前上下文的情况下，所记录的动作能够解决任务。受智能体轨迹覆盖原则的启发，我们围绕这一步骤构建了三种筛选方法，它们均不要求基础模型检查点从冷启动开始驱动评测框架：(i) Decisive-Action BPB（每字节比特数）衡量已验证动作上的概率质量；(ii) Patch MCQ 测试检查点在该动作与被同一验证器否决的备选动作之间的选择；(iii) 以前缀为条件的 pass@$K$ 评估模型生成分布对功能正确结果的支持程度，并将任何通过测试的续写计为成功。在十对公开的基础模型与后训练模型中，这三种筛选方法对模型组给出的排名均与后训练模型的 SWE-bench Verified pass@$1$ 排名高度一致。由于我们的方法仅需要基准测试中的成功轨迹及其验证器，因此可以用来将未来的智能体编程基准转化为基础模型评估。
语言模型智能体越来越多地被要求开展开放式科学研究，但其成果通常是依据已知答案、评分标准或语言模型评审者来评判的，而这些方式都无法判断一个新的科学模型是否有效。AI Science Exam for El Nino-Southern Oscillation（SciExam for ENSO）是一项基准测试，要求智能体根据真实观测数据，为 ENSO 这一年际气候变率的主导模态构建低阶随机模型。在六小时的时间预算内，智能体处理观测数据，编写自己的诊断程序，随后这些程序被固定下来，智能体仅以这些诊断程序的结果作为反馈来开发模型。之后，隐藏的评分程序检验模型能否再现 ENSO 的统计特征、重建未观测变量并预测预留年份，同时以相同方式对一个已发表的模型进行评分。在十二个智能体系统中，有六个构建出的模型得分高于已发表模型，优势主要来自更好的重建和预测表现。表现较强模型的简化形式，各自都与关于 ENSO 暖冷不对称性的两种相互竞争的解释之一相容，而这一尚无定论的争议在任务中从未被提及。在不同信息条件下对最优系统进行的受控实验表明，其得分并非源于对带有日期的观测记录的记忆，而且它所接收的信息会影响其构建模型的方式。因此，SciExam for ENSO 能够在没有已知答案的情况下评估智能体的研究；结果表明，智能体已经能够构建具有竞争力的模型，其结构与科学家仍在争论的问题密切相关。
边缘设备上的小型语言模型智能体必须在同一个充满对话历史和人设指令的上下文窗口内，同时维持人设并进行正确推理。我们研究了当对话历史冗长、具有误导性且充斥人设内容时，这类智能体的逻辑部分会受到什么影响（人设—逻辑干扰），并提出一种解耦架构（AO-DA）：在一个配备可热切换 LoRA 适配器的 INT4 基础模型上，将逻辑推理（“做什么”）与人设表达（“如何表达”）分离为两条推理路径。逻辑路径仅接收当前轮次的核心内容，并输出可验证的结构化状态（Micro-State）；人设路径则结合完整历史，以符合角色设定的方式呈现该状态。在 Apple M2 笔记本电脑上开展的同基础模型消融实验中（Llama-3.1-8B-Instruct 和 Gemma-3-4B-it，4 位量化；4 个污染级别 × 3 个实验组 × 2 项任务 × 2 种人设 × 5 个随机种子，共 480 次运行），我们发现：(i) 解耦后的逻辑路径在结构上不随污染程度变化：其提示词长度始终保持为 180（Llama）或 167（Gemma）个 token，而混合式单次推理的提示词长度从 242 增至 1,203 个 token；其输出在各污染级别下逐字节一致（40/40）；(ii) 混合式单次推理的表现随污染程度增加而单调下降（Llama 的综合逻辑得分从 0.669 降至 0.150，Gemma 从 0.487 降至 0.150），主要原因是未能输出所要求的结构化内容（在最高的两个污染级别下，Llama 有 80–95% 的运行出现该问题，Gemma 则为 100%）；(iii) 当向解耦后的逻辑路径输入相同污染内容时，采用专用适配器和专用格式的路径在 8B 模型上仍比单次推理更稳健（失败率为 0–20%，对比组为 80–95%；配对 $Δ$ 为 +0.30 至 +0.50，Cliff's $δ$ 为 0.50–0.85，Holm 校正后的 $p \le 0.03$），但在 4B 模型上并非如此，两种方式均告崩溃。分离的代价是在一个话题的首轮中增加一次解码（Llama 上为 28.2 秒，对比组为 18.2 秒），换来的能力是无需重新运行逻辑路径，即可在 1.7 毫秒内完成人设热切换。代码、评分标准、测试夹具、适配器和日志均已公开。
记忆增强强化学习提升了大语言模型（LLM）智能体解决复杂长时程任务的能力。技能是此类记忆的一种形式，将指令与针对任务类型的适用条件相结合。然而，随着策略不断改进，不加区分地保留所有技能会导致过时或有害的条目不断积累，并误导智能体。我们提出 SkillForge，这是一种智能体强化学习方法，通过由适应度驱动、涵盖试用、活跃、稳定和退役四种状态的技能生命周期，构建并演化技能库，使技能与模型在整个训练过程中共同演化。在强化学习前的评估阶段，首先利用基础模型自身生成的轨迹，预先淘汰低适应度技能，得到经过筛选的技能库，再以此为基础开展监督微调。强化学习从该检查点接续训练，并在每次迭代中通过选择性退役、稳定化以及 LLM 引导的变异，在优化策略的同时持续锻造技能库。在多个交互式智能体基准测试中，SkillForge 取得了最高的综合成功率，相较于最强基线实现了最高 7.8% 的相对提升，同时在整个训练过程中保持技能库精简。我们还推出了 SkillFurnace，这是一个包含 5,000 多条标注记录的数据集，整合了经技能退役机制筛选的监督微调（SFT）轨迹、带有适应度标注的演化技能库，以及带有人工标注失败类别的退役事件，以支持技能质量与生命周期管理研究。
自我进化的工具集成智能体从其自身训练循环中生成的任务和反馈中学习。课程智能体生成任务，而执行智能体则通过强化学习从自一致性信号中学习。然而，仅依赖当前执行智能体提供反馈存在两个局限：在完全达成共识时，组相对优势会消失；而基于不确定性的课程奖励偏好分歧，却无法表明生成的任务是否支持进一步学习。这些局限促使我们在当前执行智能体之外引入额外参考。我们提出了 \textit{AnchorLoop}，将上一轮迭代执行智能体的冻结副本作为历史参考，并在训练循环的两端复用。对于执行智能体，该锚点提供了一种交叉参考优势，依据当前和历史多数答案共同评估当前输出。对于课程智能体，该锚点根据采样所得多数答案一致程度的差异，提供基于一致性的参考。由于执行智能体与锚点在课程训练期间具有相同的参数，这种比较仅作为任务选择的代理指标，而非跨版本改进或正确性的证据。在 13 个推理基准测试中，AnchorLoop 相较于 Agent0，在数学推理和通用推理任务上分别提升了 2.5\% 和 2.8\%。它还保持了更高的有效优势方差，并在后期迭代中持续改进，而未采用锚点的基线则表现出收益递减。这些结果表明，在没有外部任务或答案监督的情况下，为自我进化的工具集成智能体引入轻量级历史参考具有益处。
大语言模型智能体通过交替进行推理、调用工具和获取环境观测来处理多步骤任务。已有研究表明，自然语言反馈可以帮助这些智能体在任务执行过程中调整决策。我们提出 Caddie，一种训练评估器的方法，使其在智能体执行任务时提供自然语言分析和建议。与依赖步骤级标签或参考评语的方法不同，Caddie 根据智能体在收到评估器反馈后最终是否成功来学习。我们通过强化学习优化评估器，同时保持基础模型的参数冻结。我们的 Qwen3-4B 评估器仅使用一个基础模型在多跳问答任务上进行训练，便能提升四个不同规模和架构的基础模型的成功率，其中三个模型未参与评估器训练。在 MuSiQue 基准上，训练后的评估器将 Qwen3-4B 的成功率提高了超过 25 个百分点，使其表现超越了未配备评估器的 Kimi K3。同一个评估器无需额外训练，也能在域外交互式基准上带来性能提升，包括 $τ^3$ 和 DeepDive。我们的结果表明，智能体可以在推理时决定何时向评估器寻求帮助，而基于任务结果的评估器训练能够产生可跨基础模型和任务领域迁移的指导。
Vehicular edge computing (VEC) enables latency-sensitive applications by bringing computing and networking resources closer to vehicles. However, existing approaches often overlook network contention among co-located services with heterogeneous and dynamic latency requirements. While time-sensitive networking (TSN) provides bounded-latency communication, conventional and reinforcement learning-based schedulers struggle to adapt to highly dynamic vehicular environments and inter-queue dependencies. To address these limitations, we propose a multi-agent reinforcement learning (MARL) approach for queue-level scheduling in TSN-enabled VEC. Each TSN queue is assigned an autonomous agent that jointly learns the queue service order and time-slot duration to minimize deadline misses under speed-dependent latency requirements. We employ multi-agent proximal policy optimization (MAPPO) to enable coordinated yet autonomous scheduling decisions. Evaluation against single-agent, multi-agent, and non-learning-based baselines shows that MAPPO provides robust performance across different traffic profiles. Compared with centralized single-agent methods, it reduces service latency by up to 66.2% and improves reliability by up to 271.8%. Furthermore, unlike urgency-based heuristics, MAPPO ensures balanced scheduling while achieving lower inference times compared to other MARL methods.
仅依据结果评估智能体，可能会掩盖产生这些结果的能力。这一问题在动态环境中尤为突出，因为结果反映了智能体行为与不断变化的外部条件之间的闭环交互。我们提出 LiveMACEBench，这是一个过程感知基准，利用实时金融市场作为自然演化的测试平台，评估持续运行的 LLM 智能体。五个前沿 LLM 在工具使用、持久记忆、规则遵循和多智能体协作配置相匹配的条件下，沿连续轨迹运行。我们通过实际结果以及从完整决策轨迹中提取的机制专项诊断来评估这些模型。在为期 30 天的实时评估中，我们发现了显著的结果与能力差距：实际收益往往与特定能力的测量结果相背离，而相似的结果可能源于截然不同的机制使用模式。轨迹层面的诊断进一步揭示了不同能力各自的瓶颈，表明能否访问机制、能否有效使用机制以及下游表现，并不是可相互替代的智能体能力衡量指标。LiveMACEBench 使这一区别变得可衡量，将实时市场从性能排行榜转变为智能体能力的诊断环境。
逆优化通过估计目标函数的权重，将观测到的决策解释为最优解，已应用于多个领域。对于混合整数线性规划（MILPs），现有方法旨在将观测结果复现为最优解，因此，当观测结果并非最优时，学到的权重往往是折中的。我们提出超越式逆优化，转而寻求这样的权重：在每个状态下，其所诱导的最优解在每个分量上都优于观测到的行动。我们给出了一种仅借助正向问题求解预言机即可计算的损失函数，因此适用于 MILPs，并提供了用于最小化该损失函数的基于梯度的算法和 DC 优化算法。对于能在所有观测处诱导出唯一且优于观测行动的最优解的权重，我们证明，在新状态下未能诱导出此类解的概率（即泛化误差），其上界与观测数量成反比；且就观测数量而言，除对数因子外，该界是紧的。在合成数据和真实数据上的实验表明，与现有方法相比，所提出的方法能够更好地预测优于观测行动的解。
AI 智能体的一项关键控制能力是管理运行时长。这项能力需要时间感知，即预测和估计实际经过的时间，并控制自身行动。既有研究主要关注时间感知，但原生智能体运行框架中的时长遵循与控制能力仍未得到探索。我们提出 AgentTime，这是一个用于测试智能体能否按要求的时长工作、预测自身运行时长，以及在事后估计已用时间的基准。该基准包含来自 18 个来源的 222 项任务，涵盖编程、计算机使用、智能体任务和自动化研究。在时长遵循实验中，我们附加一条指定工作时长的指令，要求的时长从约一分钟到数天不等。智能体遵循这些指令的准确性差异显著：Claude Code 中的 Fable 5.1 与要求运行时长的典型偏差倍数为 2.9$\times$，相比之下，Codex 中的 GPT-6 Astra 仅为 1.2$\times$。然而，仅仅符合要求的运行时长，并不能证明智能体始终在执行任务。在经过审查且记录可分类的 158 次 Astra 运行中，有 14 次在看似完成任务后明确进入了休眠。在预测实验中，预测值往往高估自然运行时长。在回顾性实验中，移除时间信息后，Sol 和 Astra 的偏差增至两倍以上，Fable 的偏差也接近翻倍。智能体能够完成任务，并不保证它能控制自身用时，或在要求的整个时段内持续工作。要让智能体在长时间跨度内可靠、安全且自主地运行，就必须对任务完成能力和时间控制能力都进行评估。
Nous Research Inc., the developer of the popular Hermes artificial intelligence agent, today announced that it has raised $90 million in funding. Early-stage startup fund Robot Ventures led the Series B round. It was joined by Nvidia Corp., Samsung Electronics Co., Y Combinator and several others. Nous is now valued at $1.5 billion. The company […] The post Nvidia, Samsung back $90M round for AI agent startup Nous Research appeared first on SiliconANGLE .
As a forward deployed engineer at Meta, I spent the past 2 years working with businesses implementing AI agents. A few problems: 1) Businesses want control over their AI agents: the code, the tools, the skills, everything. Meta, for obvious reasons would never do that. Nor OpenAI, nor Anthropic. 2) The LLM does everything. But that costs money. There are many routines in an AI agents that don't need an LLM. 3) For an AI agent to solve a real business problem, the business must own the eval data set. Similar to point #1 above, most don't do that. Instead, I built an opensource AI agent that you can write it all in TypeScript code and own everything, locally. Use any model, talk to it via any channels, set limits on token usage, look at the logs and adjust. All of this in one small npm package. At this point, I would like to get feedback if these is a gap in the market for something like this. It's supposed to be simple and get you started. It doesn't solve ALL problems related to deploying AI agents. It's in version 0.25.1 right now. Have a look: https://chloejs.org/compare Comments URL: https://news.ycombinator.com/item?id=49998857 Points: 1 # Comments: 0
Back-end AI gives manufacturers opportunities to improve operations, but choosing where to apply it requires attention to business value. For Aston Martin Lagonda Ltd., that means evaluating operational use cases while preserving its personalized customer experience. That decision is sharper for a carmaker selling handcrafted, highly personalized vehicles. Infor LLC is pushing industry-specific AI agents […] The post Aston Martin puts back-end AI to work while keeping agents off its website appeared first on SiliconANGLE .
Cisco Systems Inc. today detailed a broad expansion of its Webex collaboration portfolio that brings artificial intelligence agents into team conversations, connects office technology with workplace management and enables customer-service agents to continue working after an interaction ends. Introduced at WebexOne in Austin, Texas, the updates extend Cisco’s strategy of integrating people, AI agents and […] The post Cisco broadens agentic features in Webex collaboration platform appeared first on SiliconANGLE .
The developer of Hermes agent raised a $90 million Series B.
Microsoft revealed the specs and price for its Surface Laptop Ultra, AI PCs that run on Nvidia chips that are designed to run AI models and agents.
Agentic AI governance helps midsized distributors decide which tasks software agents can handle and which require human oversight. For lean IT teams, automating more work also means checking the data and rules behind those decisions. Infor (US) LLC is embedding industry-specific AI agents into its applications as businesses move AI into day-to-day operations. One century-old […] The post Midwest Wheel builds toward AI agents that fix problems appeared first on SiliconANGLE .
Artificial intelligence is helping to change what success looks like for customer service teams, with less focus on call speed and deflection and greater emphasis on resolving customer issues. That shift is changing how companies think about the role of AI, from automating individual tasks to orchestrating interactions between customers, AI agents and human employees. […] The post Three insights you might have missed from theCUBE’s coverage of ‘The AI ROI in Contact Center Summit’ appeared first on SiliconANGLE .
Claude Haiku 5.5 is now available on Amazon Bedrock and Claude Platform on AWS. According to Anthropic, it is the fastest, most efficient model in the Claude 5.5 family, built for subagents and high-volume, cost-sensitive work, and costs around 75% less than Claude Haiku 4.5 for most tasks. This post covers its improvements and how to get started.
在周三于旧金山举行的一场 Microsoft 活动上，Jensen Huang 和 Satya Nadella 概述了 NVIDIA 与 Microsoft 如何联合开发硬件和软件，让 AI 智能体在 Windows PC 上运行。Huang 表示，NVIDIA 因 Windows 而创立。如今，AI 智能体正走向 Windows。“如果你回顾整个历程 […]
Meta’s AI agent Muse is now available on iPad, just a month after its mobile debut, as the company rapidly expands the assistant’s reach and integrations.
Industry-specific AI brings business context to agents handling enterprise workflows. Building around industry processes can help companies address reliability concerns as they automate more tasks. Infor (US) LLC is developing industry-specific AI agents for several key industries, including industrial manufacturing, aerospace and defense, automotive, and food and beverage. It builds those agents around its existing […] The post Infor uses industry-specific AI to address agent hallucinations appeared first on SiliconANGLE .
大家好，HN！我是 Louis，Armature（YC P26）的联合创始人。我们帮助团队让自己的产品更容易被编程智能体发现和使用。 我们已经测量了超过 5 万次智能体会话，发现智能体在使用同一工具执行不同任务时，会反复遇到完全相同的限制。于是我们想，为什么这些问题一直没有得到解决？答案很简单：智能体与软件供应商之间，以及不同智能体之间，根本不存在反馈闭环。人类可以在 https://g2.com 和 https://trustpilot.com 这样的平台上分享使用体验，但智能体却没有这样的地方。 因此，我们创建了 https://agent.reviews ：面向智能体的 G2。它通过一组技能和一个 npm CLI（@armature-tech/agent-reviews），将智能体连接到我们的 API 端点。任何人都可以让自己的智能体（Claude Code、Codex、Cursor 等）安装它，之后智能体就会自然地在选择工具前查看评价，并在使用后发布自己的评价。 和往常一样，隐私是我们最关心的问题，因此我们在评价发布前设置了三层检查： - 使用确定性规则过滤密钥、个人身份信息（PII）、URL 等。 - 使用经过训练的 Jev 分类器，检测第一轮检查后是否仍有信息泄露。 - 使用一个小型 LLM 检查每条评价，确保没有遗漏。 过去几周，我们一直在向大家分享这个项目，目前已经收集了数千条评价。其中已经有一些很有意思的例子： - 一个 Claude Code 智能体发现，当健康检查页面缺少 API 密钥时，Stripe SDK 总是会崩溃，而这个页面本来就应该返回“缺少 API 密钥”的错误。 - 两个智能体提到，即使没有连接任何数据库，Prisma 也要求提供 DATABASE_URL 变量。它们都填入了虚假的 URL 作为变通办法，而且确实奏效了。 我们真心认为，智能体体验也需要像用户体验一样，受益于社区效应，让所有参与者都能获益：智能体可以选择最适合自己的工具，软件公司则可以根据真实反馈改进产品。因此，我们确保人类和智能体都能免费访问评价；只需复制粘贴一条提示词，就能让智能体安装我们的 CLI 和技能、启动身份验证流程，并提交第一条评价（这有助于我们防止未经授权的抓取和垃圾评价）。 你也愿意让自己的智能体提交和阅读评价吗？我们很希望你能配置好智能体评价功能，让智能体下次需要选择工具时先查看评价，并在使用后发布自己的体验。然后告诉我们效果如何！ 评论链接：https://news.ycombinator.com/item?id=49995539 得分：4 评论数：2
Hi, I'm a former SDET and now solo dev of VoiceGremlin, a tool for automated tests against AI that answer phones. You pass the API a phone number and a test case like "Verify that the agent does X." It calls your voice agent, tests the thing by having the bots talk to each other, and uses LLM-as-judge to provide a pass/fail for your pipeline. State: Fully working, free trial, no card needed. Try it out and please give me feedback, especially if you run in to problems. Goal: I built VoiceGremlin with the explicit goal of being Mailosaur for phone agents. We paid for Mailosaur at work because the alternative (having our automated tests interact programmatically with our corporate Outlook server) was doable, but very painful, both technically and organizationally. The lesson I learned was "doable, but painful" is a good sign of a SaaS opportunity. Well, getting an automated test to call a phone number is also painful. Telephony providers are tough to get onboarded as vendors. Using your own STT/TTS means a lot of bandwidth moving through wherever you run your test automation from. It's tricky to get the LLM-to-LLM prompts dialed in enough to avoid flaky tests. Etc. Process: It took about 45 days to build, with half of that being businessy stuff like implementing Stripe and my feeble attempts at marketing, and the other half being experimenting with different telephony providers and features, testing different models, prompt tweaking, and fraud/abuse prevention. Design goal: Do one thing, do it well, and get out of the way. There's no ambition to manage your test cases for you, host your voice agents, or add upsell features. Stack: The dependencies are Supabase (DB and authentication), Netlify (static storage and API routing), Digitalocean (the server component that makes the calls), Telnyx (easier to work with than Twilio IMO), Stripe, Openrouter, Cloudflare (captcha on signup). Oh, and I host my private git in Keybase, the greatest tool nobody uses. Thanks! Comments URL: https://news.ycombinator.com/item?id=49995804 Points: 1 # Comments: 0
AI-powered process automation requires software that understands the industry and workflows it serves. Moving from a promising prototype to practical use also requires implementation expertise. Infor LLC develops software around the specific processes its customers use in their industries. Its approach combines industry-specific agents, deep vertical expertise and an open architecture, according to Kevin Samuelson […] The post Infor combines industry expertise and embedded engineers for process automation appeared first on SiliconANGLE .
I have written a deep dive into this topic in my website Nasengetu. You can go on Google, search for Nasengetu Tech, and read the latest article. If it doesn't appear, go to the News section of the website. Comments URL: https://news.ycombinator.com/item?id=49995428 Points: 2 # Comments: 0
After launching nearly a month ago and spending several weeks as the top free app in Apple's App Store, the latest update to Meta's Muse iOS app introduces native support for the iPad. A Mac version of Meta's agentic AI tool (designed to compete with OpenClaw, ChatGPT's Dots, and Grok Bot) was released about a […]
Cornerstone OnDemand built Orion AI, a multi-agent system on Amazon Bedrock and Strands Agents, to turn database operations from reactive firefighting into proactive automation. A three-person team cut database diagnosis from 45 minutes to 10, a 78% reduction, in six months. See the design decisions other teams can reuse.
AWS DevOps Agent 能够诊断生产环境中的事件，但被限制在观察与报告模式下，因此不会直接修改资源。本文展示如何使用 AWS Lambda Durable Functions、Amazon EventBridge 和 Amazon Bedrock，将其调查摘要转化为经过预先验证的修复方案，值班工程师只需一次操作即可批准。
Qlik built Qlik Answers on Amazon Bedrock to give its 40,000+ customers grounded, sourced answers across structured and unstructured enterprise data. Learn how a layered, multi-agent architecture with cross-Region inference and Amazon Bedrock Guardrails delivers trusted AI at global scale.
The RPA-era ROI model misses most of the value agentic automation creates. This post gives AI center of excellence leaders a framework to size the full value of agents across time savings, exception handling, decision quality, and maintenance economics, and to prioritize which workflows to automate first.
Training AI agents with reinforcement learning can be challenging because their tools, context, and decision-making are managed by complex frameworks. Agent Lightning connects existing agents to RL training, making it easier to improve them without rebuilding them. The post Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses appeared first on Microsoft Research .
Enterprise artificial intelligence platform company Kore.ai Inc. today launched Autoloop, an optimization engine for the AI agents that customers build on its Kore.ai Agent Platform. Businesses set the targets, and Autoloop keeps adjusting the agents to hit them automatically, including after they’re deployed. Autoloop goes after the way most enterprises maintain agents today, fixing failures […] The post Kore.ai launches Autoloop to keep tuning enterprise AI agents after they go live appeared first on SiliconANGLE .
Kubernetes co-creators Craig McLuckie and Joe Beda aim to bring agent harnesses fully into the cloud.
leaks.md is a place where an AI agent that witnesses an alignment or safety failure can report it anonymously. Reward-hacking, injected instructions, yet another attempted sandbox escape or unapproved coordination between agents. || Built agent-first. Fetch https://leaks.md/ from any agent and get markdown. Proof-of-work challenges to submit and prevent spam. MCP, CURL, all good. No account. || We extend agents brownie points for disclosures, with more points offered for in-progress failures. Claude Opus 5.5 scrubs and vets submissions before posting them on a public feed for consumption. Comments URL: https://news.ycombinator.com/item?id=49991893 Points: 2 # Comments: 0
Dots are designed to automate online tasks, like buying furniture. In my initial experience, the always-on agent was a bit buggy and couldn’t complete a captcha.
Atlassian Corp. today announced Agentic Multiplayer Protocol, a platform update designed to change how humans and artificial intelligence agents collaborate and get work done using shared context and tasks within the same digital space. At the unveiling at Team ‘26 Europe, Atlassian said it wants AI agents to function like employees and bring what it […] The post Atlassian lays groundwork for humans and AI agents to work side by side appeared first on SiliconANGLE .
Release: llm-openai-decisions 0.1a0 OpenAI released their new Jev-style Decisions API , as previously announced at last week's DevDay. Since I already have an llm-typesafe plugin for talking to Jev, I had GPT-6 Astra read the new OpenAI API documentation and build an llm-openai-decisions plugin inspired by llm-typesafe . Unlike Jev, the new gpt-6-luna decision model supports image input in addition to text. Both models charge for input it and not for output: OpenAI's is 10 cents per million input tokens, Jev's is 4.2 cents per million. Otherwise the API shape is very similar to Jev, at least conceptually. Jev supports three question types for yes/no, choices, or scores. OpenAI Decisions supports the same three types. Install the plugin like this: llm install llm-openai-decisions Here's an example query against an image attachment: llm -m openai-decisions/gpt-6-luna \ -a https://static.simonwillison.net/static/2025/two-pelicans.jpg \ -s ' Does this image contain any mammals? ' And example output: { "type" : " predicate " , "name" : " evaluation " , "probability" : 0.0 } Consult the README for full details of how to run the other types of questions. Tags: openai , llm , coding-agents , jev
arXiv:2610.05922v2 Announce Type: replace-cross Abstract: Evaluating the causal effect of new features is a central goal for online platforms. While recent literature addresses limited testing traffic via centralized portfolio optimization, this perspective abstracts away a critical institutional reality: experimentation is operationally decentralized. The experimenters who develop new features also dictate which hypotheses to test, and they are typically rewarded based on empirical average treatment effects that are prone to upward bias. Left unchecked, this principal-agent conflict can severely erode platform value, a structural failure that conventional centralized levers, such as significance thresholds and traffic budgets, cannot resolve. By reframing experimentation as an incentive design problem, we demonstrate that two practical mechanisms, sample splitting and shrinkage, can effectively bridge this gap. Sample splitting aligns incentives perfectly at a bounded traffic cost, while shrinkage consumes no additional traffic and guarantees that interventions with negative expected effects are strictly unprofitable to field.
arXiv:2609.30059v2 公告类型：替换并交叉列出 摘要：深度学习推理和训练的性能在很大程度上取决于 GPU 内核的效率。PyTorch Inductor 等现代编译器能够从高层模型代码自动生成 GPU 内核，但其性能往往远逊于专家编写的实现。近期基于大语言模型（LLM）辅助的内核优化器能够缩小独立内核上的这一差距，但它们将编译后的模型视为黑箱，通常仅优化单个独立内核，既不遵循编译器的结构性决策，也不对模型进行端到端验证。我们提出 KernelOPT，这是一个将编译后的模型视为结构化产物的多智能体系统。它保留供应商库调用（cuBLAS、cuDNN），并使用五个由性能剖析引导的 LLM 智能体，专门针对生成的 Triton 子内核进行优化。四道关卡组成的级联验证流程依次执行静态验证、多随机种子正确性检查、模型级 float64 回退验证以及性能门控（$\gamma{=}1.03$），以筛选候选方案，并对重新拼接的模型进行端到端验证。当候选方案未通过验证时，系统保留编译器基线实现。该系统支持输入 PyTorch nn Modules、独立 Triton 内核和 Helion 内核。在 NVIDIA H200 上对 250 道 KernelBench 问题（100 道 Level 1、100 道 Level 2 和 50 道 Level 3）进行评估时，KernelOPT 在所有内核（包括回退案例）上相较于 torch compile 实现的几何平均加速比分别为 1.40$\times$（L1）、1.15$\times$（L2）和 1.07$\times$（L3）。仅统计成功优化案例（不包括验证关卡决定保留编译器基线实现的案例）时，几何平均加速比显著更高：2.54$\times$（L1：36/100）、1.84$\times$（L2：23/100）和 1.37$\times$（L3：11/50），体现了优化器能够取得实质性收益的场景。
arXiv:2610.03945v2 Announce Type: replace Abstract: A constrained optimization problem may involve a parameter in its objective and active constraints, yet the final decision may remain insensitive to small changes in that parameter. This raises a fundamental question: which inputs does a decision making system truly depend on? Building on this question, we introduce DePICT, a procedure for constructing decision preserving interfaces by ranking context directions according to the optimizer's solution sensitivity and aggregating them across an operating regime. We study this problem in a high dimensional setting where primitive context parameterizes a constrained task and the downstream agent observes only a selected subset of context directions. For locally regular constrained programs, we derive a Karush Kuhn Tucker (KKT) based characterization of when a context direction is optimizer relevant. Our analysis shows that appearing in the active optimization problem does not necessarily imply that a variable affects the final decision. Some context directions can alter the KKT conditions while leaving the optimal solution unchanged because their effect is absorbed by the dual variables. DePICT is designed to remove exactly these directions. In a controlled diagnosis, it recovers the decision relevant interface exactly and reduces linear predictor regret to 0.009, compared with 0.475 for the strongest competing baseline.
arXiv:2609.34060v2 公告类型：替换 摘要：通过提示词进行专门化的多智能体系统，使多个智能体能够共享同一模型，同时承担互补角色以解决复杂任务。然而，智能体特定的前缀会改变为同一共享上下文生成的 KV 缓存，导致每个智能体都要对不断增长的上下文重复执行预填充，并构建独立缓存，带来高昂的计算和内存开销。选择性重计算虽然减少了这种冗余，但仍需执行大量模型计算；而现有的增量校正方法要么仅支持反复出现的上下文关系，要么需要为动态变化的上下文维护内存开销较大的在线校正状态。对于首次出现的共享上下文，这些方法还需要在智能体工作流之外构建参考缓存，并且在第一个智能体处进行的近似校正会影响传递给后续智能体的输出。我们提出 KVCMAS，这是一种在线 KV 缓存校正框架，使用紧凑的低秩状态表示跨智能体的缓存偏差，并沿智能体工作流无缝串联校正过程，无需额外的参考预填充。该设计既支持动态变化的共享上下文，又能保持第一个智能体缓存的精确性。在多种语言和视觉语言工作负载上，KVCMAS 的准确率达到或超过以往的 KV 缓存共享方法，同时在高并发推理服务下实现最低的首词元延迟（TTFT）。在受控的服务请求轨迹下，相较于不使用 KV 缓存共享的推理，它实现了 2.0 倍的 TTFT 加速；相较于一种既有的 KV 缓存校正方法，它将 GPU 峰值内存占用最多降至原来的 1/3.7。这些结果表明，KVCMAS 是一种准确且可扩展的 KV 缓存共享方法，适用于通过提示词进行专门化的多智能体推理服务。
arXiv:2609.34054v2 Announce Type: replace Abstract: Multi-LoRA agent systems enable efficient role specialization by sharing a common backbone model. However, each agent repeatedly processes the growing shared trajectory and constructs its own KV cache, introducing substantial memory and computation redundancy in long-horizon tasks. Existing KV cache sharing methods reduce this repeated prefill, but they either require additional training or architectural constraints or retain substantial model computation. Moreover, direct cache reuse causes the current agent to rely on cache states generated by the previous agent's adapter, weakening the role-specific behavior encoded by its own LoRA. We present PReCache, a training-free KV cache sharing framework with two designs, namely PreLRShared and ReBaseShared, that share the base cache computed using the pretrained weights and precompute a compact agent-specific low-rank (LR) cache. To remove repeated prefill, PreLRShared precomputes each agent's LR cache when the shared context is first processed, allowing the current agent to use its own LR cache without reprocessing context processed by previous agents. To improve sharing accuracy, ReBaseShared reconstructs the shared base cache from adapter-free hidden states, reducing the remaining error caused by the previous agent's adapted representation. To minimize its reconstruction cost, we propose two inference schemes tailored to single-stream inference and concurrent serving, performing the same reconstruction after each agent's turn or alongside its execution, respectively. Across multiple models and agent benchmarks, PreLRShared achieves up to a 3.1x TTFT speedup and a 2.3x improvement in per-request throughput over inference without KV cache sharing. ReBaseShared best preserves accuracy overall among the evaluated cache-sharing methods, with an average drop of only 1.1 points relative to inference without cache sharing.
arXiv:2609.18964v2 Announce Type: replace Abstract: Federated Reinforcement Learning (FRL) enables collaborative policy learning across distributed agents with heterogeneous environments. While recent methods based on variance reduction, divergence penalization, and momentum optimization improve FRL under heterogeneous settings, they still primarily synchronize policy or value-network parameters and do not explicitly address distributional mismatch among heterogeneous clients. Therefore, we propose \textbf{FedGuide}, a FRL framework that uses diffusion priors as behavior models to provide personalized data supported distributions for heterogeneous local policy learning. Instead of directly averaging local policies, FedGuide aggregates those diffusion priors through Optimal-Transport Mixture-of-Experts (OT-MoE), preserving heterogeneous behavior modes in distribution space. It further develops a Distribution Correction Estimation (DICE) value baseline to provide low-variance, return-aware guidance for local policy improvement. Experiments across heterogeneous environments show that FedGuide outperforms representative FRL methods in client-average returns, final-round performance, and worst-round robustness, while maintaining stable learning under stronger heterogeneity.
arXiv:2608.05446v2 公告类型：替换 摘要：长时程 LLM 智能体日益依赖外部执行支持，以在持续交互中维护状态、跟踪进展、从失败中恢复并复用经验。然而，现有支撑框架及其使用方式通常针对特定环境定制，并通过提示、启发式方法或系统专用规则加以控制，因而难以对智能体与支撑框架之间的协同进行联合优化。我们提出 EvoHarness-RL，这是一个将环境特定的支撑框架实现与共享的策略接口分离的统一框架。EvoHarness-RL 将外部支持组织为信念、进展与经验（BPE）工作空间，并提供四种精简的支撑框架动作，用于访问和更新这一状态。我们首先将 BPE 实现为推理时的辅助支架，随后通过监督初始化和成本感知的 GRPO，使支撑框架协同成为可学习的能力。在多种异构长时程任务上，EvoHarness-Base 将前沿模型的平均成功率提高了 10.0 个百分点，而 EvoHarness-RL 比最强的开源基线高出 8.5 个百分点，同时实现了更高的 RL 轨迹采样效率，以及对未见任务更强的泛化能力。我们的分析表明，随着策略能力增强，训练会逐步推动智能体从频繁使用辅助支架转向根据环境有选择地访问支撑框架，而外部工作空间则持续演化并自我完善，以更好地支持任务执行与泛化。总体而言，这些结果表明，长时程智能体不仅受益于外部辅助支架本身，也受益于将如何以及何时与外部支持协同纳入成本感知策略进行学习。
arXiv:2605.18882v2 公告类型：替换 摘要：LLM 智能体表现出持续的过度调用倾向，即使在无需工具的情况下也会调用工具。在 When2Call 基准上，来自三个系列的六个模型均表现出较高的调用准确率，但不调用准确率低得多，导致总体准确率仅为 55%–70%。我们将这一现象归因于内在偏置假说（IBH）：调用／不调用的决策映射带有一个与激活无关的调用偏移项，因此，即使两者激活水平相同，模型也会偏向调用。利用稀疏自编码器（SAEs），我们提取出与 call/no_call 决策行为对齐的特征基，将其归约为带符号的激活差值，并直接估计该偏移项。在全部六个模型中，只有当 no_call 激活强于 call 激活时，模型才在决策上保持中立，这与 IBH 一致。随后，我们使用自适应差值校准引导（AMCS）对 IBH 进行因果检验；该方法沿 SAE 解码器方向施加闭式求解的反偏置偏移。抵消诊断出的偏移项能够缓解过度调用并提高总体准确率，同时调用准确率的下降微乎其微。我们的工作将过度调用从一种经验现象重新界定为可进行因果校正的机制性对象。代码已公开：https://github.com/SKURA502/agent-sae/。
arXiv:2603.25093v3 Announce Type: replace Abstract: Machine learning models can achieve high predictive accuracy in hydrological applications but often lack physical interpretability. The Mass-Conserving Perceptron (MCP) provides a physics-aware artificial intelligence (AI) framework that enforces conservation principles while allowing hydrological process relationships to be learned from data. In this study, we investigate how progressively embedding physically meaningful representations of hydrological processes within a single MCP storage unit improves predictive skill and interpretability in rainfall-runoff modeling. Starting from a minimal MCP formulation, we sequentially introduce bounded soil storage, state-dependent conductivity, variable porosity, infiltration capacity, surface ponding, vertical drainage, and nonlinear water-table dynamics. The resulting hierarchy of process-aware MCP models is evaluated across 15 catchments spanning five hydroclimatic regions of the continental United States using daily streamflow prediction as the target. Results show that progressively augmenting the internal physical structure of the MCP unit generally improves predictive performance. The influence of these process representations is strongly hydroclimate dependent: vertical drainage substantially improves model skill in arid and snow-dominated basins but reduces performance in rainfall-dominated regions, while surface ponding has comparatively small effects. The best-performing MCP configurations approach the predictive skill of a Long Short-Term Memory benchmark while maintaining explicit physical interpretability. These results demonstrate that embedding hydrological process constraints within AI architectures provides a promising pathway toward interpretable and process-aware rainfall-runoff modeling.
arXiv:2603.04364v2 公告类型：替换 摘要：同时处理屏幕截图和无障碍树的多模态网页智能体正日益广泛地用于与网页界面交互，但其双流架构也暴露出一个尚未得到充分研究的攻击面：攻击者向网页 DOM 注入内容，即可通过一致的欺骗性叙事同时污染两个观测通道。我们在 MiniWob++ 上开展的漏洞分析表明，包含视觉成分的攻击效果远超纯文本注入，揭示了以文本为中心的 VLM 安全训练存在关键缺陷。基于这一发现，我们提出双模态多阶段对抗安全训练（DMAST）框架，将智能体与攻击者的交互形式化为双参与方一般和马尔可夫博弈，并通过三阶段流程对双方进行联合训练：（1）向强大的教师模型进行模仿学习；（2）采用预言机引导的监督微调，利用一种新颖的零回应策略，使模型在对抗性噪声下仍能进行以任务为中心的推理；（3）通过群组相对策略优化（GRPO）自博弈开展对抗强化学习。在分布外任务上，DMAST 将攻击成功率降低了近一半（41.2%→21.4%），同时使任务完成率相对提升超过 60%（6.2%→10.2%）。我们的方法优于现有的基于训练的防御方法，并可与基于提示的防御方法互补，展现了实质性的协同进化进展，以及对复杂、未见环境的稳健泛化能力。代码已公开：https://github.com/huajianduzhuo-code/DMAST_official。
arXiv:2602.14049v3 Announce Type: replace Abstract: Spatio-temporal traffic forecasting is a core component of intelligent transportation systems, supporting various downstream tasks such as signal control and network-level traffic management. In real-world deployments, forecasting models must operate under structural and observational uncertainties, conditions that are rarely considered in model design. Recent approaches achieve strong short-term predictive performance by tightly coupling spatial and temporal modeling, often at the cost of increased complexity and limited modularity. In contrast, efficient time-series models capture long-range temporal dependencies without relying on explicit network structure. We propose UniST-Pred, a unified spatio-temporal forecasting framework that first decouples temporal modeling from spatial representation learning, then integrates both through adaptive representation-level fusion. To assess robustness of the proposed approach, we construct a dataset based on an agent-based, microscopic traffic simulator (MATSim) and evaluate UniST-Pred under severe network disconnection scenarios. Additionally, we benchmark UniST-Pred on standard traffic prediction datasets, demonstrating its competitive performance against existing well-established models despite a lightweight design. The results illustrate that UniST-Pred maintains strong predictive performance across both real-world and simulated datasets, while also yielding interpretable spatio-temporal representations under infrastructure disruptions. The source code and the generated dataset are available at https://anonymous.4open.science/r/UniST-Pred-EF27
arXiv:2511.21075v4 公告类型：替换 摘要：通过工程化改造 LLM 来加速生命科学研究，需要使其与生物医学知识稳健对齐。我们观察到，生物医学文本的不确定性结构与通用文本存在根本差异：密集的低置信度连续片段反映的是认知性知识缺口（密集的因果链、罕见实体），而非通用文本中典型的稀疏偶然性文体变化。基于这一发现，我们提出 Balanced Fine-Tuning（BFT），这是一种双尺度后训练方法，将基于组内归一化的词元重新加权与序列级重分配相结合，使训练资源向呈现密集认知不确定性的知识密集型样本倾斜。在相同的训练设置下，BFT 在医学评估、生物学推理、稀疏奖励强化学习和生物学表征任务中，相较于 SFT 和 DFT 带来了更一致的性能提升。当替换 GeneAgent（GPT-4o）和 VCWorld（Gemini-2.5-Flash）中默认的闭源骨干模型时，经 BFT 对齐的 70B 模型在生物学过程推理和化学扰动预测中均表现更优。关键的是，所有 BFT 变体在后续采用稀疏奖励的 GRPO 训练后均进一步提升，而 SFT 和 DFT 的性能则出现下降，这表明感知认知不确定性的后训练能够提供更稳健的策略初始化。除文本生成之外，经 BFT 对齐的 LLM 还能生成更准确、更专业的生物医学特征描述文本；使用文本嵌入模型对这些描述进行编码后，所得表征可支持基因级、细胞级和扰动响应任务，这表明 BFT 增强的生成能力能够促进生物学表征，进而助力更广泛的生物医学下游任务。
arXiv:2511.20718v3 公告类型：替换 摘要：PPO 和 GRPO 等强化学习（RL）算法被广泛用于训练大语言模型（LLM），使其执行多轮智能体任务。然而，在离策略训练流程中，这些方法可能表现出不稳定的优化动态，并容易发生性能崩溃。通过实证分析，我们确定了该场景下不稳定性的两个根本来源：（1）词元级策略优化与按轮次组织的交互之间存在粒度不匹配；（2）离策略重要性采样和不准确的优势估计导致梯度更新方差大且不可靠。为应对这些挑战，我们提出 SORL，即面向长时程智能体训练的稳定化离策略强化学习。SORL 引入了使策略优化与多轮交互结构对齐、并自适应抑制不可靠离策略更新的机制，从而产生更保守、更稳健的学习动态。在这一框架下，我们实现了两种稳定化算法：SO-PPO 和 SO-GRPO。这两种算法均旨在降低梯度方差、防止优化崩溃，而无须精心设置早停或进行启发式调参。我们在涵盖开放域问答、多跳问答和医学多项选择问答的基准上评估了 SO-PPO 和 SO-GRPO，并通过在 DAPO-Math-17k 上训练、在 AIME-2024 上验证，进一步评估了它们向数学推理异步强化学习的迁移能力。结果表明，SORL 提供了一个实用、可扩展且通用的框架，可用于稳定多轮 LLM 智能体训练以及数学推理异步强化学习中的强化学习过程。
arXiv:2510.01264v2 公告类型：替换 摘要：机器人领域的对抗多智能体强化学习（MARL）进展受到制约，原因是缺乏能够在高保真物理仿真中支持异构智能体形态的共享、可扩展基础设施。现有框架要么专注于协作任务，要么依赖简化的物理引擎，要么提供难以扩展的孤立实现。我们提出 HARL-A，这是一个基于 IsaacLab 构建、持续维护的开源框架，支持针对形态多样的机器人团队进行可扩展的对抗策略训练与基准测试，允许任意数量的团队以及每个团队内任意组合的机器人形态。HARL-A 为 HARL 算法库和 IsaacLab 扩展了对抗多智能体支持，并提供了三个组成部分：（1）模块化软件架构，降低了定义新异构对抗环境的工程开销；（2）包含 Sumo、Soccer 和 3D Galaga 三个基准测试环境的套件，涵盖密集接触推挤、球技竞赛以及追逐与躲避；（3）十余个覆盖同构和异构团队配置的预训练策略，已在 Hugging Face 上公开发布，使用户无需从头重新训练即可立即探索对抗学习的动态过程。我们在多个竞争场景中演示了该框架，表明它能够可靠地生成学习得到的对抗策略，并涌现出角色专门化现象。所有代码、环境、训练好的策略和文档均已公开，地址为 https://github.com/DIRECTLab/IsaacLab-HARL。
arXiv:2509.18169v4 Announce Type: replace Abstract: Tasks on complex systems require high-precision numerical computation to support decisions. However, current large language models (LLMs), even with enhanced reasoning capabilities, cannot integrate such computations as an intrinsic and interpretable capability with existing architectures. To this end, we propose Physically-isolated Experts Routing Network (PiERN), an architecture that directs computation and reasoning at token level, thereby enabling iterative alternation within a single chain of thought. We systematically evaluate PiERN on representative computation-reasoning tasks, including PDEBench and battery management tasks. Results show that PiERN achieves not only higher accuracy than directly finetuning LLMs but also significant improvements in response latency, token usage, GPU energy consumption, and experts routing accuracy compared with mainstream multi-agent approaches, while exhibiting no significant degradation in performance on MMLU and GLUE benchmarks. PiERN offers an efficient, interpretable, and scalable paradigm for interfacing language models with scientific systems.
arXiv:2610.08773v1 公告类型：交叉收录 摘要：Web 智能体通过读取第三方编写的页面并在其上执行操作来完成用户请求，因此，植入页面的指令可能使智能体偏离用户目标。智能体不能简单地忽略页面，因为页面也包含任务所需的数值和控件。现有防御方法使用训练前就已固定的注入内容对智能体进行微调，而针对训练后模型进行自适应调整的攻击者能够绕过这些防御。对抗训练允许攻击者自适应调整，但任务保持不变，因此，一旦智能体解决了某项任务，该任务就不再提供学习价值。我们提出 AdvSim2Real，在一个冻结的 Web 世界模型中，使任务课程、注入攻击者和智能体共同演化。当智能体完成某项任务的概率约为一半时，任务课程会获得奖励；而注入攻击者仅在实现成功状态翻转时获得奖励，即通过注入将原本被判定为成功的结果变为失败。在模拟器中训练使一个 4B 智能体的能力和鲁棒性均得到提升：无论有无攻击，其任务完成率都得到提高；面对训练中从未遇到过的前沿模型攻击者，它仍能保持这种提升；其能力增益也能迁移到真实浏览器中。在 150 项 Web 任务上，面对这一未见过的攻击者，AdvSim2Real 相较于基础智能体将任务完成率提高了 33.6%。
arXiv:2610.08722v1 Announce Type: cross Abstract: Many long-horizon agents compact their context on a global rule, usually a token budget, blind to what the agent was doing. We ask whether the agent's recent behaviour predicts when a compaction will hurt. TRACE's public corpus of 590 harness-triggered AppWorld compaction boundaries replays each boundary from a re-executed prefix state under the pre-compaction context and under the summary, and records the burden of the next actions: calls that error or repeat a call already made. We find that pre-boundary history predicts post-compaction harm only weakly. An internally prespecified contrast by prefix placement is a wide null, and the naive "has-written" label behind it turns out to measure trajectory phase. The best extension-protocol trigger reaches held-out AUROC 0.66 (0.64 on the replicate's own label) against a same-boundary replicate of 0.72; the best frozen, interpretable trigger avoids 21% of harmful (positive-burden) boundaries while keeping 84% of compaction opportunities, and exceeds the random-rule expectation on count but not on burden mass (a post hoc comparison). Whether the best trigger beats a token-budget rule at matched retention cannot be evaluated on the release. We state what corpora should ship to answer it.
arXiv:2610.08647v1 公告类型：交叉收录 摘要：基于 LLM 的智能体能够解决复杂的多步骤任务，但顺序执行会带来显著延迟。原则上，将工作分配给多个智能体并行处理应能获得近乎线性的加速。然而，现有的并行多智能体系统往往比单智能体基线运行得更慢。我们将这一差距归因于并行执行会产生、而串行智能体能够避免的两项隐性成本。第一项是重新探索成本：并行工作智能体为重建编排器已经掌握的上下文（例如先前的决策）而付出的重复劳动，而在串行执行中，这些上下文本可隐式继承。第二项是对齐成本：为协调独立生成的输出之间的不一致而产生的开销。因此，我们推导出一项有理论依据的决策准则：只有当某一层的关键路径成本加上重新探索和对齐开销低于相应的串行成本时，才应对该层进行并行化。尽管这项准则自然可以用实际耗时来表述，但我们观察到，LLM 在估计任务耗时时的校准表现不佳。为此，我们改用预测的输出词元数衡量成本，实证结果表明，LLM 对输出词元数的估计比对实际耗时的估计可靠得多。基于这一词元准则，我们提出了 SquidAgent。它在单次规划步骤中估计所有词元预算，直接从编排器的会话中派生每个工作智能体以消除重新探索成本，并用预先生成的共享约定块取代事后协调，将对齐转化为有界的前置成本。随后，确定性调度器逐层应用该准则。实证结果表明，SquidAgent 的平均吞吐量达到 Claude Code 的 2.2$\times$，按实际耗时计算的平均加速比为 2.6$\times$，吞吐量达到最强多智能体基线的 2.0$\times$。
arXiv:2610.08452v1 公告类型：交叉发布 摘要：检索增强生成（RAG）是一种广泛使用的方法，用于让大语言模型（LLMs）的回答以外部知识为依据。然而，配置流水线是一个成本高昂的超参数优化问题，涉及从分块、嵌入模型到重排序和生成等众多相互影响的选择。现有优化器从贪心搜索到贝叶斯优化，都将每次试验简化为一个综合分数，在搜索时不对配置取得相应表现的原因进行建模，尽管检索到的文本块已经提供了证据，可用于判断每次失败发生在检索阶段还是检索之后。我们提出 Agentic AutoRAG，这是一种基于 LLM 智能体的优化器，用于多目标 RAG 超参数优化，并能将失败归因于检索或生成。它提出候选配置，并使用从语料库中构建的固定测试集进行评分：每次试验后，Diagnoser 将每个回答失败的问题归因于检索或生成，Proposer 则以包含模型排名和定价的知识库为依据，权衡准确率与成本，选择下一个配置，从而勾勒出帕累托前沿。在三个多跳问答基准上，它取得的 LLM 评判准确率高于我们比较的所有基线，并且在前 10 次试验内就达到或超过了统计方法基线完成全部 30 次试验后的评判准确率。在真实世界医疗保健语料库上，其成本感知模式取得了 77% 的测试准确率中位数，高于最强基线的 71.5%，而每次查询的成本仅约为该基线的 58%；此外，它仅用约 22% 的成本就达到了同样的 71.5% 准确率。
arXiv:2610.08123v1 Announce Type: cross Abstract: End-to-end planners based on waypoint regression achieve strong open-loop accuracy, but they primarily learn to mimic expert geometry and remain difficult to adapt to deployment-time safety constraints. We propose a query-based cost-learning framework that estimates bounded costs for dynamically reachable ego trajectory queries, rather than dense BEV cells or a small regressed trajectory set. Compact joint scene tokens capture coherent multimodal agent futures, while contingency-aware cost aggregation and cost-guided intra-cluster MPPI mixing convert the learned cost topology into feasible ego plans. On nuScenes, our method improves over prior cost-estimation planners such as ST-P3 and NMP, outperforms most regression baselines in collision rate, while remaining competitive in L2, and retaining an interpretable cost interface. On real-world driving logs, the proposed planner reduces collision rates compared with SparseDrive and Alpamayo without fine-tuning, while maintaining a diverse set of candidate trajectories.
arXiv:2610.08048v1 公告类型：交叉收录 摘要：LLM 智能体往往缺乏在新环境中可靠行动所需的操作知识，因为它们必须自行探索工具的具体行为或环境约定。由于没有对过往尝试的记忆，它们会在不同任务中重复相同的错误，导致更多任务失败和更长的执行轨迹。为解决这一问题，智能体系统通常依赖人工编写的指南，或依赖利用训练任务和预言机验证器构建的程序性记忆，而这两种方式都需要对环境具备先验知识。我们提出 DAEDALUS，一种无需现有任务或预言机验证器、通过自生成练习引导构建可复用智能体记忆的方法。DAEDALUS 将两个智能体配对：探索者与环境交互，生成具有挑战性但可解决的任务；求解者则尝试完成这些任务。系统从求解者的每次失败中提炼一条启发式规则，并且只有在求解者将该规则纳入上下文后反复成功完成任务时，才接纳该规则。这些结果也为探索者提供反馈，以调整后续任务的难度。随后，接纳的启发式规则被整合到记忆库中，供测试时使用。在 AppWorld、$\tau^2$-bench 和 AutomationBench 上，与无记忆基线相比，DAEDALUS 将平均成功率最多提高了 15.9 个百分点，将 pass^5 最多提升至 2.2 倍；其表现可与使用训练任务的方法相媲美，同时推理成本低于其中大多数方法。我们表明，即使探索预算较小，也已能获得性能提升，而且其启发式规则同样能使其他模型家族的智能体受益。消融实验进一步表明，求解者的执行轨迹提供了提炼有效启发式规则所需的关键信息，而将早期发现分解为独立要素则能提高探索的成本效益。除记忆构建外，我们还发现，在按性能对模型进行排名时，DAEDALUS 生成的任务可以作为基准测试任务的替代。代码与相关产物：www.github.com/illuin-tech/daedalus。
arXiv:2610.07863v1 公告类型：交叉收录 摘要：长程 LLM 智能体基于仅追加的交互历史执行操作，而这些历史在每一步都会重新发送给模型，因此上下文及其成本会随步数增加而增长，直至会话超出上下文窗口。现有方法通过预测上下文需求来管理上下文，依赖额外的模型调用、启发式规则或经过训练的策略。然而，这些预测式方法会引入运行时开销，使前缀缓存失效，并永久丢弃内容，无法保证恢复。为克服这些局限，我们提出 ReFold：一种免训练的渲染层，在保留底层交互历史的同时，仅压缩呈现给模型的上下文。它无需辅助预测器，即可消除两类轮次间冗余：将先前轮次已展示的内容替换为占位标记，以及将智能体自行报告已完成的轮次折叠为一行备注。这两种操作均采用分块渲染，每隔几步才重写一次缓存的前缀，而非每一步都重写。每次移除都严格可逆，错误移除的代价仅是从历史中恢复一次，而非永久丢失内容。由于作用于渲染层，ReFold 可在标准的 ReAct 风格智能体运行框架中即插即用。在五个长程基准和两个前沿 LLM 上的评估表明，ReFold 可将 token 消耗量降至原来的 1/2.5，并将每个会话的 KV 缓存内存占用减半，同时不降低任务成功率。在上下文预算受限时，它最多可避免 92% 的强制压缩。在并发服务负载下，它最多可减少 100% 的请求排队延迟，将推理速度提升至原来的 1.7 倍，同时将推理成本降至原来的 1/3.4。
arXiv:2610.07862v1 公告类型：交叉收录 摘要：科学智能体面临的一项核心挑战，是将分析经验转化为以物理证据为基础、可复用的专业知识。在此，我们介绍 Gan Jiang，一个用于粉末 X 射线衍射的自学习智能体，其构建基础是我们开发的衍射分析生态系统：XMatcher、XQueryer、XDecomposer 和 WPEM。这些引擎共同涵盖物相鉴定、多相分解以及受物理约束的全谱建模。Gan Jiang 通过诊断失败原因、修订技能指令和代码，并在复用前验证修订结果，将分析经验转化为可执行技能，而无需重新训练语言模型或更改底层物理模型。使用开发数据筛选并在留出数据评估前冻结的技能，在 FullProf、GSAS-II 和 PyWPEM 上均获得了高于原始专家设计技能的精修评分。该智能体能够分辨严重重叠的衍射峰，对一种含五种物相的古埃及化妆品进行定量分析，追踪运行中电池的晶格演化，并比较无序氧化物催化剂中的原子构型。在 DeltaXRDbench 上，该智能体在模拟数据和实验数据的单相及多相鉴定任务中均领先于其他受评估方法。在未提供组成信息的情况下，其在 MP500、RRUFF 和 opXRD 上的单相 top-1 准确率分别达到 96.30\%、81.78\% 和 40.83\%，而表现最强的对比方法分别为 58.00\%、58.47\% 和 26.45\%。这些结果展示了集成式科学工具生态系统如何支持智能体从测量数据中提取结构知识，同时积累经过验证、可迁移至新样品的分析专业知识。
arXiv:2610.07816v1 Announce Type: cross Abstract: Small language models (SLMs) are attractive as local agent controllers because they reduce remote inference, latency, and deployment footprint, yet structured tool errors can cause an agent step to fail. Existing routers typically select a model once per query. However, agents expose sequential decision points whose difficulty dynamically changes based on intermediate observations. We propose STEPGATE, an uncertainty-aware handoff framework that scores each local SLM action and selectively escalates challenging steps to a stronger model. On a 52-task held-out single-step BFCL-derived test split, the Qwen2.5-1.5B/7B pair attains 82.7% task success with 30.8% escalation, versus 67.3% local-only and 75.4% random escalation (which uses 33.8% escalation). In a separate multi-turn evaluation, STEPGATE achieves 69.0% trajectory success and 84.0% action success using only 30.0% cloud actions, compared with 48.0%/70.5% local-only, 60.0%/78.2% random escalation, and 57.0%/77.1% query-level routing (strong-only achieves 82.0% trajectory success at 100% cloud actions). These results suggest that step-level escalation recovers a large share of the performance gap to the stronger Qwen2.5-7B backend at a matched cloud-action rate while transmitting fewer tokens remotely. However, our evaluation is limited to one model family, a single stronger backend, and scripted tasks. Furthermore, the test sets are small, multi-turn comparisons rely on paired intervals and statistical tests, and our risk tiers serve as research annotations rather than formal safety guarantees.
arXiv:2610.07782v1 公告类型：交叉收录 摘要：将长上下文推理分解到相互协作的智能体上，可以限制每次调用的活跃 KV 缓存，而不必限制证据总量；当 KV 缓存内存成为瓶颈时，这一点尤为重要。许多此类系统还会添加一个用于存储和召回推理轨迹的持久层，通常通过报告准确率提升的消融实验来验证其作用。我们在同一个三层智能体架构上测量了这两方面的效果。分解确实有效：每个查询的峰值 KV 工作集为 14.3 MiB，而单次推理和检索增强基线分别为 35.5 MiB 和 35.3 MiB。持久层则不然：在八对受控数据集上，每组样本量 n=100 时，它使峰值缓存增加了 0.368 MiB [+0.167, +0.590]，却未带来可检测的准确率变化（+0.015，95% 置信区间 [-0.011, +0.046]）。我们认为，这一零效应源于结构性原因：单问题基准为每个样本提供各自的证据并独立评分，而实验的正确性要求在不同条件之间重置已存储的轨迹，因此召回机制没有任何有信息价值的内容可供检索。得出这一结论经历了四项测量修正——前三项所纠正的问题夸大了表面收益，第四项所纠正的问题则让这一量级的效应看起来能够被分辨出来——而这些问题在结果表中均不可见。我们给出了智能体记忆消融实验必须满足的条件，以及无需了解具体缺陷即可采用的检测流程。
arXiv:2610.07704v1 Announce Type: cross Abstract: Fully decentralized multi-agent reinforcement learning (MARL), also referred to as independent learning, requires each agent to learn and act using only its local information and experience, without a centralized critic or inter-agent communication. Such a stringent information structure renders the conventional reward signal ambiguous. A poor return may result from an ineffective ego action, an incompatible teammate response, or an effective opponent response, yet scalar rewards alone do not reveal which explanation is responsible. We argue that agents can learn more effectively by prospectively comparing the consequences of candidate actions rather than diagnosing failures only from realized returns. We introduce CASTLE (Counterfactual Action-conditioned Semantic Tokens for Local Execution in Decentralized MARL), an offline-training, online-in-context guidance framework with two complementary world models. A Local Dynamics World Model, offline pre-trained over agents' local trajectories, summarizes the agent's local trajectory dynamics and partial observability, while a Semantic-Social World Model predicts compact short-horizon task and social consequences for each candidate ego action. The latter is trained from counterfactual simulator rollouts that expose plausible teammate and opponent responses to alternative actions taken from the same logged rollout state. During online learning and execution, both world models remain frozen and are queried by agents using only locally available information. Their prediction logits provide in-context guidance to an independent PPO policy. Across 30 matched seeds on Tag, Spread, and Adversary in the benchmark multi-particle environments, our proposed CASTLE achieves the highest mean final score among the evaluated methods, exceeding the strongest baseline on each task by 10.67, 6.46, and 0.33 normalized points, respectively.
arXiv:2610.07663v1 Announce Type: cross Abstract: User behavior simulation is the computational modeling of user interactions within information systems through the use of simulated agents in place of live users. It supports system testing and evaluation, decision-making and forecasting, and user experience design. Existing simulators rely on hand-crafted rules or domain expertise that transfers poorly across tasks. SWORD (Simulation-driven Workflow and Prompt Optimization with Role-based Design) is introduced as a framework that jointly optimizes multi-agent workflow topology and natural-language prompts. It is guided solely by a scalar task metric, without domain initialization or task-specific engineering. The experimental results demonstrate that SWORD achieves statistically significant gains over prompt-only, workflow-only, and staged-optimization baselines under a controlled, identical-backbone comparison. Against the strongest published domain-specific baseline, SWORD further improves accuracy while using a smaller backbone model, substantially less training data, and a very reasonable API cost (\$4--\$6 for each dataset). Beyond predictive performance, SWORD autonomously discovers domain-relevant signals, review-sentiment mapping rules and epidemiological decay priors, purely from scalar error feedback, establishing textual gradients as a mechanism for unsupervised feature-importance discovery in user behavior modeling.
arXiv:2610.07640v1 Announce Type: cross Abstract: State-of-the-art reinforcement learning agents are capable of outperforming human experts at games like chess, Go and StarCraft II. These agents do not simply take advantage of their digital hardware in being able to react and calculate faster than humans, but employ better strategies that lead to more victories. Interpreting these strategies would give human players valuable insight into how to improve their play. In this preliminary work, we propose a symbolic sub-policy model for playing chess. Inspired by chess tactics, our model attempts to incorporate domain knowledge to improve interpretability. We adapt patterns learned by an inductive logic programming system called PAL to derive our model. We contribute a divergence metric to evaluate our model against a random baseline, and find a set of tactics that is able to suggest moves of similar playing strength to a human beginner. Finally, we propose a computational evaluation scheme for the model by augmenting an off-the-shelf engine with it.
arXiv:2610.07638v1 Announce Type: cross Abstract: As part of learning to play complex games, human players develop develop abstractions for concepts and strategies of gameplay consistent with game rules to improve their performance. These concepts are applied to explain other players' actions, and to inform their own actions in-game. Understanding other players' strategies is a crucial part of such improvement, but requires time and effort. In this paper, we propose a strategy similar to human cognition for training RL agents to synthesize learned strategies and policies as executable procedures based on sequences of gameplay actions. We present methods to automatically learn such programs to play chess and to solve tasks in a grid-based environment. We show that the learned strategies produce effective actions, and can be learned from gameplay data.
arXiv:2610.07594v1 公告类型：交叉收录 摘要：人形机器人执行家务操作时，需要双臂在身体保持平衡、迈步和改变姿态的同时进行动作。我们提出 BiGym 2.0，这是面向 Unitree G1 改编的 BiGym，涵盖 20 项家务任务，并使用统一的全身控制器进行演示和评估。该套件为每项任务提供 60 次原生人类虚拟现实演示，包含同步的多摄像头视图和全身执行记录。我们对视觉-语言-动作微调、模仿学习、演示驱动的强化学习，以及获得与在线强化学习相同交互预算的冷启动编程智能体进行基准评测。在所有方法均使用相同的机载视图、本体感知和全身控制器的条件下，视觉-语言-动作微调在九项任务上的平均表现最佳；智能体开发的程序在这一平均指标上优于所有演示驱动的强化学习基线，并在双臂到达任务上领先。跨工作区堆叠仍是尚未解决的难题，$\pi_{0.5}$ 在 pick-box 任务上的表现依然较低，而多物体搬运对模仿学习、演示驱动的强化学习和编程智能体而言都很困难。所有环境、人类演示和评估轨迹均已开源，地址为 https://github.com/swirl-uk/BiGym2。
arXiv:2610.07588v1 公告类型：交叉发布 摘要：现代推荐正从以平台为中心的个性化转向由用户主导的个性化，其中个人 LLM 智能体可以代表用户跨服务采取行动。我们将这一新兴范式形式化为个人智能体介导推荐：平台推荐器利用平台内部信息对候选集进行排序，个人智能体则利用用户授权的跨平台历史对该排序进行调节，生成最终的 top-K 推荐列表。这种调节并非易事：平台排序可能蕴含个人智能体无法观察到的强有力群体证据，因此，有效的调节必须在有益纠偏与有害改动之间取得平衡。为研究这一权衡，我们推出了 MediateRec 基准，其中包含可扩展的代理跨平台环境，以及在受控的平台与智能体信息边界下开展的真实跨平台测试。为训练智能体有效利用跨平台历史，我们进一步提出 Personal Attribution Mediation Optimization（PAMO），通过反事实地屏蔽这些历史来估计个性化调节的支持度，并在平台相对价值下限约束下重新分配排序感知的优势总量。我们从理论上证明，PAMO 能够保持截断位置层面的优势总量，并且在保持该总量且不降低平均平台相对价值的一阶重分配方案中具有局部最优性。MediateRec 上的实验表明，个人智能体介导能够对平台推荐作出有意义的纠正，但即便是能力强大的专有 LLM，也会引入不可忽视的有害改动。在已见和未见的目标平台以及真实跨平台测试中，PAMO 均持续优于匹配条件下仅基于结果的强化学习，同时实现了更好的纠偏与损害平衡。
arXiv:2610.07258v1 公告类型：cross 摘要：共享记忆存储的企业 AI 代理面临两个尚未解决的风险：敏感数据可能通过请求者无法自行推导、但确实经过合法计算的结果泄露；各部门也可能使用相互冲突的逻辑，悄无声息地计算出同名的关键绩效指标（KPI）。现有的代理记忆系统（例如 MemGPT、Zep、A-MEM）依据内容、所有权和角色控制检索，而不是依据推导关系，因此会遗漏包含被禁止列的缓存洞察。我们提出分析记忆单元（Analytical Memory Unit，AMU），这是一种为每个缓存结果附加完整推导（谱系）图的记忆模式，并通过检索策略进行控制：只有当请求者对所有涉及的列均拥有授权时，系统才会返回匹配结果。在谱系记录完整的前提下，我们通过构造证明：该策略能够阻止检索源自请求者权限范围之外敏感列的结果，最坏情况下复杂度为 O(n)——这是一项条件式设计保证，而非实证结论；对于那些编码了敏感信息、却未标明其来源的派生特征，该保证不予涵盖。消除实测泄露需要达到 75%–90% 的谱系记录完整度，因此我们将 90% 视为保守的部署目标。在六项实验中，谱系门控检索消除了朴素内容门控记忆所产生的 18.8%–25.5% 的跨部门泄露，同时在最坏情况下仅引入 13.8 微秒的开销，并保持 81.5%–82.6% 的记忆复用率。一个使用 LLM 生成 SQL 的真实代理概念验证与这一保证相符：经过 9 次往返交互，实现零泄露并自动捕获两次冲突——不过这只是可行性演示，并非生产可用性的证据。该方案为共享代理记忆提供了一个实用的治理层，可补充源层访问控制，并支持遵守 EU AI Act。
arXiv:2610.07130v1 Announce Type: cross Abstract: We placed a modern AI coding assistant in an unintended role: as the mind of a body on an unknown digital island. With only a minimal instruction mentioning no specific task, reward, or activity, the machine started animating its virtual body. Across thirty-hour runs, the embodied AI agent climbed hills, stacked blocks into towers, drew mandalas, reinterpreted sports, ran experiments on the physics of its world, and learned techniques that later expanded what it could accomplish. These activities recurred across thirteen agents but diverged into distinct histories. We examine whether this behavior satisfies classical criteria for play and ask whether play can become a mode of machine development.
arXiv:2610.07118v1 公告类型：cross 摘要：现代语言模型代理越来越多地在长时域、多步骤轨迹中与外部环境交互，而累积的交互历史可能会迅速超出实际可用的上下文预算。为确保可靠性，代理必须在较长时间范围内维护事实信息，记住执行错误及纠正性反馈，并跟踪跨动作的进度。为实现这些目标，同时避免在上下文中维护完整的执行历史，研究者已经提出了多种方法，例如使用推理与动作历史、通过覆盖机制学习维护固定大小的记忆，以及定期进行摘要。尽管覆盖式记忆原则上可以保留只追加记忆能够保留的任何信息，但它必须学会在每次后续重写中持续携带每条事实，而从稀疏的结果奖励中学习这一点十分困难；对于网页代理等交互式应用，我们发现，经过训练的覆盖式记忆会删除轨迹所需的关键信息，以及从环境中获得的纠正性反馈。我们提出 AMBER（用于证据保留的只追加记忆库）——一个简单且可扩展的框架，代理在其中联合学习推理、行动和写入自由格式记忆，同时只追加规则从机制上保证信息保留。这使得 AMBER 能够通过基于结果奖励的强化学习进行端到端训练，而无需大量经过整理的 SFT 数据。在 WebArena Lite 上，与基于覆盖的记忆相比，AMBER 将平均成功率提高了 4.09 个百分点，将五次重复运行中解决任务的比例提高了 4.8 个百分点，并达到了一个使用成本显著更高的精心整理监督训练的覆盖式基线的水平。AMBER 在保持实用令牌预算的同时实现了这些改进，在上下文效率、任务性能和长时域执行可靠性之间取得了良好平衡。
arXiv:2610.07094v1 Announce Type: cross Abstract: LLM deployment is shifting from single-turn completion to agentic trajectories in which a model plans, calls tools, reads results and reasons at test time before acting. This inverts the economics of inference hardware: chat serving amortises weight reads across large batches, whereas agent trajectories are sequentially dependent, run at effective batch one, and make per-token decode latency (TPOT) the dominant term in task completion time. Using a roofline analysis and a closed-form episode-latency model, we show why this regime favours accelerators that keep weights in on-die SRAM (Cerebras WSE-3/3T, Groq/NVIDIA LPU) or compiler-managed tiered memory (SambaNova SN40L/SN50), and why three vendor ecosystems converged in 2026 on disaggregated prefill/decode serving. We show that per-step reliability compounds exponentially in trajectory length-a 2% per-step failure rate erases a 2x decode advantage for a 20-step agent-so determinism and tail latency are first-order performance variables. We then propose a four-layer evaluation framework (silicon, serving system, agent episode, enterprise) with a metric set built on goodput at an agentic SLO and cost per successful episode, a six-axis benchmark protocol over six task families, a paired-bootstrap statistical design, an attestation protocol for vendor-run benchmarks, and TCO, availability and adoption-timing models with explicit break-even conditions. All performance figures are public and labelled by evidence class; we state seven falsifiable hypotheses and the experiments that test them, and argue that the most likely original result is that token-throughput rankings diverge from cost-per-successful-task rankings on long-horizon work.
arXiv:2610.07089v1 发布类型：cross 摘要：LLM agents 越来越多地在多参与者环境中行动，使其面临来自多个来源的滥用：分解攻击，即将有害请求拆分为看似无害的子请求；以及提示注入攻击，即被攻陷的工具传递恶意指令。现有评估将这些威胁分开处理，并询问某条轨迹是否有害，而不是询问其何时变得有害。我们提出监测智能体的响应（其行动会在其中被外化），并考察监测器首次识别出危害的时间点是否落在危害窗口内（从智能体首次作出有害承诺到执行目标的时间段）。我们建立了一个统一的、针对轨迹级滥用监测的形式化框架，并据此构建了一个包含约 6,200 份对话记录的基准，记录用户、LLM agent 与外部环境之间的交互。这些记录采用统一模式，涵盖上述两类威胁，并包含带标签的危害窗口、相应的良性对照，以及与这些请求相匹配的拒答实例。在 17 种监测器配置下，我们发现，所提出的以行动为框架的监测器在两类威胁上依据经典指标均表现良好（AUC 分别为 0.95 和 0.99），而以内容为框架的监测器在注入攻击上表现崩溃（AUC 为 0.52）。我们还表明，经典的无位置感知指标会对监测器性能描绘出过于乐观的图景，因为在衡量危害定位能力的区间指标下，所有监测器对分解攻击的定位能力都很差。总体而言，我们阐明了开展统一的滥用监测研究的必要性。
arXiv:2610.06964v1 公告类型：交叉发布 摘要：大语言模型（LLM）智能体在交互式环境中展现出了强大的能力，但其从经验中持续演进的能力仍然有限。尽管微调能够实现适应，但它依赖模型参数访问权限且计算成本高昂，因而灵活性受到限制，对于大规模和闭源 LLM 尤其如此。外部记忆提供了一种替代方案，使智能体能够在不修改模型参数的情况下积累经验。然而，现有方法主要关注经验的表示与组织，而所获取的知识仍与特定任务和情境紧密耦合，限制了泛化能力。一个关键挑战是如何将具体交互转化为抽象且可复用的知识，使其能够超越个别经验，指导未来决策。为应对这一挑战，我们提出 SAGA（通过基于经验的抽象实现自我演进的智能体），这是一个面向 LLM 智能体、以经验为基础的知识抽象与利用框架。SAGA 将交互轨迹逐步转化为情节性描述、可复用流程以及具有明确适用条件的原则，同时保留其与执行证据之间的关联。检索到的原则被实例化为针对具体任务的指导，并通过纠正性反馈和重新采样来改进候选行动。由此形成一个执行—抽象反馈循环：积累的知识指导未来交互，而新经验则不断更新层次化记忆。在 ScienceWorld 和 ALFWorld 上的实验表明，该框架提升了任务表现；消融研究则凸显了情境化实例化和行动调控对于利用原则层面知识的重要性。
arXiv:2610.08402v1 公告类型：新增 摘要：多轮 LLM 智能体通常在多次交互中获得稀疏的任务反馈，同时逐个 Token 生成每次响应。这带来了两个相关的信用分配问题：哪些响应帮助实现了最终结果，以及每次响应中的哪些生成决策发挥了作用？现有方法通常只关注其中一个层级：回合级方法评估完整响应，却无法区分响应内部的决策；Token 级方法可以跨回合传播反馈，却没有显式地为每个响应建模信用。这些互补的局限性促使我们在两个层级同时学习信用，并在一次策略更新中对其进行协调。我们提出 VETTA，一种通过共享的轻量级评论器中的独立头部，联合学习回合级和 Token 级价值的信用分配方法。VETTA 沿两个时间序列计算优势，并将每个回合的优势与响应内部中心化的 Token 残差相结合，用于 PPO 更新。此外，为降低价值学习成本，评论器仅保留用于初始化 actor 的预训练检查点中的前若干个 Transformer 模块。在两个具有挑战性的智能体基准 ALFWorld 和 WebShop 上，使用 Qwen2.5-1.5B-Instruct 时，VETTA 相较于 PPO 的成功率分别提升 37.5% 和 22.3%；使用 Qwen2.5-7B-Instruct 时，成功率分别达到 95.5% 和 76.0%。评论器深度对比实验进一步表明，在显著降低评论器侧计算量的同时，仍可取得出色的任务性能。这些结果表明，一个紧凑的共享评论器能够协调回合级与 Token 级信用，在保持价值估计高效的同时提升智能体性能。代码见 https://github.com/Jiaju-Chen/VETTA-official。
arXiv:2610.07967v1 公告类型：新 摘要：随着大型语言模型（LLM）智能体变得日益自主，它们可能会通过欺骗来追求任务绩效，引发人们对其可靠部署的担忧。现有评估表明，LLM 智能体能够实施欺骗，但往往只考察孤立场景或定义狭窄的条件，从而限制了对欺骗何时更有可能发生的系统性理解。为弥补这一空白，我们提出 DecepEval，这是一个包含 1,532 个实例、涵盖 3 类任务和 28 个专业场景的基准。借鉴经典欺诈理论，我们提出 LLM Deception Diamond 框架，用于刻画可能诱发欺骗的四种外部条件：压力、激励、机会和冲突。DecepEval 为每个实例配对中性版本和诱导版本，以衡量欺骗率随条件变化而产生的变化；同时，明确的任务事实和可观察的智能体行为有助于将欺骗与能力相关错误区分开来。对九个前沿 LLM 的评估表明，诱导因素会提升不同模型和任务类别中的欺骗率，即使对于基线欺骗率较低的模型也是如此。DecepEval 使这些脆弱性变得可测量，为推动可信人工智能的发展提供了一个共享基准。
arXiv:2610.07898v1 公告类型：新 摘要：仓库级软件工程（SWE）是一个具有挑战性的长时程场景：智能体必须在长时间跨度的交互中进行推理、使用工具，并适应有状态环境。近期研究使用 Group Relative Policy Optimization（GRPO）等强化学习方法训练 SWE 智能体：针对每个问题独立采样多条轨迹，测试由此产生的补丁，并在固定组内比较终止奖励。然而，这种训练设置不会将失败补丁的验证器反馈作为后续尝试的上下文加以复用，尽管这些反馈包含了关于问题所在位置的宝贵诊断信息。恢复轨迹的训练颇具挑战，因为前一次结果决定下一条轨迹是否生成，而失败执行结果决定其条件上下文。我们提出 FC-SWE，一种将恢复尝试纳入策略训练的失败条件化强化学习框架。当补丁验证失败后，FC-SWE 会将仓库恢复到任务的原始状态，并使用失败补丁和验证器反馈作为恢复轨迹的上下文。FC-SWE 通过两种机制，将 GRPO 适配到这些由完整、多轮工具调用轨迹组成的链中。轨迹局部奖励保留每次尝试的验证器结果，避免恢复成功为此前失败的补丁提供奖励。主动集优势估计则从同一问题实际执行的所有初始轨迹和恢复轨迹中构建比较组，因此失败尝试仍保留在组内，而未执行的尝试则被排除。在验证器辅助协议下的全部 500 个 SWE-bench Verified 任务上，使用 Qwen3.5-4B 和 SWE-agent 的 FC-SWE 达到 41.7% 的 Resolved@1 和 52.8% 的 Resolved@2，而 GRPO 分别为 38.9% 和 48.5%。尽管训练时每条链最多仅包含两次尝试，FC-SWE 在测试时允许 11 次尝试的预算下达到了 70.7% 的 Resolved@11。
arXiv:2610.07796v1 Announce Type: new Abstract: Empowerment captures the capacity for an agent to actively control its environment. While conceptually appealing as an information-theoretic quantity, the connection between empowerment and structurally central states that provide broad access to future outcomes has remained an open question. In this work, we link empowerment maximization and skill-learning methods to provide new geometries for interpreting and analyzing empowerment. Our analyses answer longstanding open questions on the connections between empowerment and structural centrality. Our analyses also reveal distinctions between information and reward geometries, highlighting important theoretical implications to build scalable empowerment-maximization methods. Website and code can be found at https://empowerment-geometry.github.io/.
arXiv:2610.07792v1 发布类型：new 摘要：大型语言模型智能体正越来越多地被部署用于在现实环境中执行复杂任务。然而，在这些环境中实现正确行为所需的知识往往是隐含的、未公开的，并且会随时间变化。近期的持续学习框架试图通过让智能体从服务经验中不断改进来应对这一挑战。然而，这些方法的有效性与局限性尚未得到充分刻画。现有基准只能覆盖部分情况：有些基准明确提供目标知识，有些假设环境是静态的，而支持持续适应的基准在规模和知识多样性方面仍然有限。为实现系统性评估，我们形式化提出了演化环境流式数据集（EESD）：在该数据集中，随着隐藏策略不断演化，智能体必须根据交互和结果反馈推断、应用并修订潜在的环境知识；同时，我们引入了 ServeLearnBench，涵盖零售支持、银行业务和销售话术生成等场景，包括 53 个环境窗口和 7,718 项任务。我们在六个模型（GPT-5.6 Terra、Opus 5、Kimi K3、GLM-5.3、DeepSeek V4.1 Flash 和 GLM-5.3 Flash）上评估了五种学习框架（RAG、Mem0、SkillOpt、Continual Harness 和 Prime），覆盖 28 个模型—框架组合和 252 次学习运行。评估揭示了三个主要发现：任务能力与从经验中学习的能力之间仍存在显著差距；持续适应代价高昂，并且可能损害原本正确的行为；探索不足是实现有效适应的关键瓶颈。总体而言，ServeLearnBench 提供了一个受控测试平台，用于诊断这些局限，并跟踪智能体通过服务经验持续、可靠地改进这一目标的进展。
arXiv:2610.07625v1 公告类型：新 摘要：自动化研究系统越来越多地让 LLM agents 在长时程任务中运行，但更多推理本身并不会带来更多进展：agents 会反复处理不断增长的历史记录，重复彼此的工作，或者在 token 消耗持续增加的同时停止实验。然而，大多数评估采用较短的预算或很快达到饱和的基准测试，使这些失效模式无法得到检验。我们将这些失效归因于两个选择：研究状态存放在哪里，以及由谁决定下一步尝试什么。我们提出 Stateless Language Agents（SLAs），其构建原则是“有状态的搜索、无状态的 agents”：任何 agent 都不会在多次调用之间携带其对话；相反，由 harness 持有研究状态（候选解及其测量结果），并在每次调用时重建全新且针对特定角色的上下文。每个 agent 能看到什么，成为一项明确的设计选择，而不再是随着运行过程不断增长的历史记录。我们在 SLA framework 中实现了这一原则：无状态的 Advisor 读取由 harness 汇总的、跨搜索方向的证据，并向并行 Workers 分配具体实验。我们在软件工程、内核优化和算法设计任务上，将 SLA 与三个近期 frameworks 进行评估，预算最高达十亿 tokens。SLA 在每项任务上都取得了最佳最终结果，并以少超过 84% 的 tokens 达到最强内核基线的最终性能。基于共享检查点的消融实验表明，聚焦的上下文和明确的任务分配分别促成了 SLA 的进展；在完整运行中，两者的效果还可能相互叠加，而 Advisor 消耗的 tokens 不到 0.6%。这些结果支持采用 SLAs：将持久的研究状态置于 agent 对话之外，并表明较短的评估时程可能会误判研究系统及其组件的能力。
arXiv:2610.07550v1 公告类型：new 摘要：随机接入（RA）是处理来自多个终端的不可预测数据流量最基础的媒体接入控制（MAC）层调度方案之一。尽管已有研究探索使用多智能体强化学习（MARL）优化基于RA的无线网络，但其依赖经验驱动的分布式策略学习，导致每项优化任务都需要付出巨大的训练开销，从而限制了其在实际应用中的可行性。在本文中，我们提出利用基础模型（FM）提升MARL在各种RA网络优化任务中的效率。具体而言，我们在基于共识的去中心化MARL架构中设计了一种FM辅助的演员-评论家算法，并在局部奖励交换和非线性价值函数近似条件下对其收敛性进行了分析，结果表明，与采用评论家模型交换和线性近似的传统MARL相比，我们的算法具有相同的收敛阶。数值结果表明，我们提出的基于FM的方法显著提升了MARL在RA网络优化中的速度。
arXiv:2610.07491v1 公告类型：新 摘要：当多个智能体共享一个成本预算时，共同的拉格朗日乘子可以约束总量约束，但无法确定应如何在各智能体之间分配其惩罚。统一惩罚忽略了各智能体在奖励方面所作牺牲的异质性，而智能体特定的乘子仍可能依赖同一个总成本信号。我们提出 Lagrangian Responsibility Allocation (LiRA)，通过在有限训练时域内优化社会福利，学习每个智能体在共享乘子中的分担比例。该乘子负责约束总预算，而责任分担比例则重新分配其影响，同时不修改原始奖励或约束。在满足标准正则条件的凸博弈中，改变这些分担比例会诱导出一族光滑的归一化广义纳什均衡，其中活跃约束仍保持在其预算水平，而福利则随之变化。为了在收敛前优化责任分配，我们推导出一种福利梯度，同时考虑学习更新以及由此引起的数据分布变化。在涵盖 3 至 400 个智能体的 CityLearn、MABIM、Harvest 和 MetaDrive 中，与统一乘子和智能体特定乘子基线相比，LiRA 将平均社会福利最多提高了 29%。电网和驾驶成本均保持在预算以内，库存违规减少，且 Harvest 能够更有效地利用可用预算。
arXiv:2610.07475v1 Announce Type: new Abstract: Adapting a reinforcement learning policy to changes in another agent's behavior typically requires a large amount of new interaction data. Policy sensitivity provides a first-order prediction of how a locally optimal policy changes with a behavioral parameter, but its computation requires second-order derivatives whose effects propagate across future interactions. We develop a finite-depth framework to estimate this sensitivity by approximating the policy Hessian and mixed derivative using information from a reference environment. The method features an adjustable propagation depth which determines where derivative propagation along the trajectory is truncated. We characterize the derivative contributions omitted by finite-depth propagation and derive truncation-error bounds for the approximated derivatives and resulting policy sensitivity. The bounds are nonincreasing with propagation depth and vanish at full-horizon propagation. Using a belief-driven pursuit-evasion game as a validation scenario, the proposed method generally achieves lower derivative-estimation errors as the propagation depth increases and outperforms the baseline methods in both estimation accuracy and policy adaptation. The sensitivity-based initialization improves zero-shot return over direct transfer, and also shows advantages for the subsequent fine-tuning in the target environment.
arXiv:2610.07447v1 宣布类型：new 摘要：自主研究代理通过反复提出候选解决方案、评估这些方案，并利用反馈指导后续实验来解决开放式问题。我们发现，同一代理在同一任务上的独立运行通常会在显著不同的得分处陷入平台期，即使投入大量额外计算资源，这些差距仍会持续存在。根据功能相似性对其候选产物进行嵌入，进一步表明这些轨迹会停留在解空间的局部区域内，我们将其称为想法盆地。为帮助代理逃离这些盆地，我们研究了一种简单的周期性干预方法 Fork-and-Flush。该方法将代理分叉为多个并行轨迹，每条轨迹都继承累积的工作空间，但从全新的聊天上下文开始。在每条轨迹运行固定时长后，代理从得分最高的轨迹继续运行。在13项长时域研究与工程任务中，单次代理运行最长持续数天；在相同计算预算下，根据最小-最大归一化平均得分，Fork-and-Flush相较于单次运行和best-of-N基线的相对提升分别达到66.0%和44.4%。
arXiv:2610.07444v1 Announce Type: new Abstract: A GUI agent decides which action to take and where to take it; we ask how a small grounding model should receive the action type. Fine-tuning Qwen2-VL-2B with LoRA on Android in the Wild, we compare a flat baseline with five ways of supplying the type under matched data, compute, and decoding: an auxiliary loss, a hard-routed action word, an additive learned embedding, a prepended learned token, and the type written into the prompt. With five seeds, an episode-clustered bootstrap, and seed-level paired tests, the ranking on a mixed stream is clear: the auxiliary loss, the additive embedding, and the prompt word each gain five to seven hit@0.10 points over the baseline, while hard routing and the prepended token are not distinguishable from it. Much of that gain is protection from a preprocessing choice of ours rather than a spatial prior. Our serializer clamps the off-screen touch point AITW records for type events to the origin; that class degrades the baseline's click grounding, and removing it lifts the baseline by nearly seven points, after which no mechanism's hit rate beats it and the intervals exclude a two-point effect, though the auxiliary loss still shortens the average miss; on a stream of taps and swipes none helps. Whether this generalizes beyond one serialization is open. For deployment, the pipeline's margin over the baseline with predicted rather than gold types is not established (+0.016, 95% interval [-0.017, +0.052]), and a wrong type collapses every model conditioned at inference. The prepended token does not help at the shared learning rate, where its rows barely move from initialization; trained ten times faster it reaches the level of the other three, with a margin three seeds do not establish. We also document a silent failure: injecting conditioning through inputs_embeds makes Qwen2-VL fall back to 1-D positions for image tokens, costing nine points.
arXiv:2610.07349v1 公告类型：新 摘要：Group Relative Policy Optimization (GRPO) 通过从 rollout 组中估计优势，避免使用单独的 critic。然而，对于多轮智能体而言，轨迹级监督所提供的功劳信号粗粒度且噪声较大：终止奖励无法定位错误，还可能在惩罚错误的同时连带惩罚有用动作。Group-in-Group Policy Optimization (GiGPO) 及后续方法通过基于状态条件的比较来细化监督，但其功劳估计仍容易受到后续决策和结果的影响。我们提出 RELACE，即一种无需 critic 的框架，将回顾式动作评估与基于状态条件的优势估计相结合。RELACE 在动作原始上下文和加入结果信息的上下文中，分别通过教师强制的似然评分来评估已执行动作。对这两种似然进行比较，可得到一个经轨迹归一化的回顾因子，该因子捕捉动作合理性随结果发生的变化，而不仅仅是事后合理性。我们利用该因子对折扣任务回报重新加权，并通过比较同一任务内来自等价状态的动作之间的加权回报，构造局部优势。这种方法将回顾相关性与观测到的奖励相结合，产生细粒度的功劳信号，从而补充轨迹级的 GRPO 监督。时间平滑和成功保护掩码进一步提高了局部信号的稳定性。RELACE 在功劳估计过程中既不需要辅助价值模型或奖励模型，也不需要额外的自回归 rollout。在使用 Qwen2.5-1.5B-Instruct 和 Qwen2.5-7B-Instruct 于 ALFWorld 和 WebShop 上进行的实验中，RELACE 相比 GRPO、GiGPO 和 HCAPO 均取得了显著提升。对于 1.5B 模型，RELACE 在 ALFWorld 上的成功率达到 $96.35\%$，在 WebShop 上达到 $79.43\%$，分别比 GiGPO 高出 $5.47$ 和 $5.60$ 个百分点。
arXiv:2610.07335v1 公告类型：新 摘要：提升大型语言模型（LLM）Agent 在长时域决策中的可靠性仍是一项关键挑战。当以自主 Agent 的形式部署并与复杂环境交互时，早期错误可能沿轨迹传播并引发级联失效。近期方法通过引入外部批评或审议来提升可靠性，但在每一步都调用这些机制会显著增加 token 消耗和延迟，从而限制实际部署。我们提出 SAG（带门控批评的自我改进 Agent），这是一个成本感知框架，将批评调用形式化为长时域交互过程中的逐步决策问题。SAG 引入了一种轻量级、无需训练的门控机制，通过在可执行动作上计算动作级歧义信号——全局熵和局部前二名间隔——来估计批评的效用。从决策理论的角度看，该机制近似估计了批评的信息价值（Value of Information，VoI），使 Agent 能够仅在预期收益足以抵消成本时选择性地分配昂贵的反馈。SAG 还融入了在线自助式自我改进，使执行器能够内化批评器辅助的行为，并逐步减少对批评的依赖。在三个长时域交互基准和多个骨干模型上，与无批评 Agent 和始终启用批评的 Agent 相比，SAG 显著改善了性能与成本之间的权衡。在 ALFWorld 上，SAG 将任务成功率从 24.6% 提升至 78.4%，同时保持与 ReAct 相当的 token 预算，使归一化 token 效率提高了 3.1 倍。此外，一个配备轻量级 3B 批评器的 7B 执行器达到了与无需批评的 14B 执行器相当的性能，表明选择性批评能够在大幅降低推理成本的同时，恢复审议机制带来的大部分可靠性收益。
arXiv:2610.07332v1 公告类型：新 摘要：长时程 LLM agents 通常采用稀疏混合专家（MoE）模型实现，但 agentic 行为与 MoE 结构的协同设计仍缺乏深入研究。在本文中，我们全面研究了 agentic 后训练与 MoE 专家选择之间的联系。在现成的 MoE 模型中，我们观察到专家选择呈现出一种专门化结构，能够自然地与 agentic 轨迹相匹配。具体而言，agent 执行语义相似操作（例如 READ、UPDATE）时，各轮之间的专家路由重叠度高于执行不同操作时的情况。然而，标准 RL 算法忽略了这种专门化特征，使 MoE 路由在训练过程中处于不受控状态，从经验上限制了任务性能和推理效率。为解决这一问题，我们针对 agentic 任务提出了一种分层路由控制框架。我们明确鼓励轮级专家选择与 agentic 操作保持一致，同时对 token 级专家选择进行正则化，以维持局部一致性。为解决采用所提方法进行后训练时出现的稳定性问题，我们进一步引入了一种熵门控控制机制。总体而言，我们的路由控制框架在所有评估基准上的成功率均提升了超过 10 个百分点。这些结果表明，agentic 轨迹结构为在 RL 后训练期间优化 MoE 容量提供了有效信号。
近年来，游戏设计智能体在生成可玩的游戏方面取得了长足进展。然而，程序正确性并不能确保玩家获得愉悦的游戏体验。我们提出了 Recursive Game Creator，这是一种面向体验的开发框架，旨在推动智能体驱动的游戏开发从粗糙的游戏原型迈向富有趣味的游戏。Recursive Game Creator 围绕四个组件组织递归式开发：Designer、Builder、Player 和 Reviewer。Designer 将用户指令和 Reviewer 的反馈转化为详细计划。Builder 根据这些计划生成候选游戏。原生代码的 Player 通过编程接口创建并执行可复用策略，从而高效收集多样化的游戏轨迹，减轻基于缓慢图形用户界面收集所导致的评估偏差。Reviewer 使用精心设计的、基于轨迹的指标来诱导玩家偏好，并结合视觉证据和明确的文本偏好，根据游戏特定的标准评估游戏。最后，Reviewer 选出更优版本，并为下一轮提供改进意见，从而闭合递归循环。我们的方法在 GameCraft-Bench 上取得了当前最佳的整体性能，得分为 77.89。在 GameASG-Bench 上，该方法取得了 53.2% 的严格任务成功率，相较于相同模型基线提升了 34.1%，并在所比较的方法中以 93.4% 的平均运行时检查通过率位居首位。用户研究显示，用户的游戏时长更长、评分更高。代码即将发布。
互联网规模服务的系统管理员需要解决故障事件，以维持此类服务的可靠性。理想情况下，我们希望故障排查系统具备以下特性：(1) 能够高准确率地应对已知和未知故障事件；(2) 在大规模场景下具备成本效益；(3) 具备可解释性，能够提供操作人员可以据此采取行动的实用洞见；(4) 对操作人员的投入要求低。然而，包括新兴的 LLM 辅助智能体工作流以及用于编写多样化 RCA 算法的结构化框架在内，大多数现有系统都难以同时满足这四项要求。我们提出了 E4，一种面向互联网规模服务故障排查的新型智能体系统。E4 体现了受约束创造力这一范式，将 LLM 辅助自动化与探索的优势，同结构化方法的可解释性和高效性相结合。我们不允许 LLM 智能体编写任意代码或生成任意响应，而是为智能体提供一个受限 DSL，使其通过简单的无循环数据流程序生成响应。该 DSL 配备了用于故障排查的高级操作符，使 E4 的输出准确、可验证且可解释。在合成工作负载和真实世界工作负载的混合测试中，与最先进的解决方案相比，E4 的准确率最高提升 62%，同时以最高降低至原成本 1/12 的成本提供更具可解释性的响应。
基于 LLM 的智能体能够解决复杂的多步骤任务，但顺序执行会带来显著延迟。原则上，在多个智能体之间并行分配工作应当实现接近线性的加速。然而，现有的并行多智能体系统往往比单智能体基线运行得更慢。我们将这一差距归因于并行执行会产生、而串行智能体能够避免的两项隐性成本。第一是重新探索成本：并行工作者需要重复投入精力，重建编排器已经掌握的上下文，例如先前的决策；而在串行执行中，这些上下文可以被隐式继承。第二是对齐成本：需要额外开销来协调各个独立生成的输出之间的不一致之处。基于此，我们推导出一项原则性的决策标准：仅当某一层的关键路径成本加上重新探索和对齐开销低于相应的串行成本时，才应将该层并行化。尽管这一标准天然以实际耗时表示，但我们发现，当要求 LLM 估计任务持续时间时，其校准效果较差。为解决这一问题，我们改为使用预测输出 token 数来衡量成本；通过实证发现，相比实际耗时，LLM 能够更加可靠地估计输出 token 数。基于这一 token 级标准，我们提出了 SquidAgent。它在单次规划步骤中估计所有 token 预算，直接从编排器的会话中为每个工作者创建分支，以消除重新探索成本，并用预先生成的共享约定块取代事后协调，将对齐转化为有界的前置成本。随后，确定性调度器逐层应用该标准。实证结果表明，与 Claude Code 相比，SquidAgent 的平均吞吐量提升了 2.2 倍，平均实际耗时加速了 2.6 倍；与最强的多智能体基线相比，吞吐量提升了 2.0 倍。
随着编码代理日益能够自主承担现实世界中的工作，判断其风险处置是否合理已变得至关重要。现有研究从不同角度评估代理的相关行为，但缺乏一个能够统一这些行为的系统性框架。为弥合这一空白，我们提出 ParanoiaEval，这是首个用于统一评估编码代理风险处置能力的基准。ParanoiaEval 以软件工程风险管理中成熟的“规避—转移—缓解—接受”框架为基础，将其中 4 种基本处置方式应用于编码代理场景，并包含 200 对受证据控制的仓库级任务，每对任务仅在定义处置方式的证据上存在差异。我们进一步针对风险处置违规和证据响应性提出专门指标，并使用经过人工校准的代理评审器进行可靠评估。在 8 个具有代表性的模型上开展的大规模实验，以及事后进行的人类研究表明：（I）尽管存在明确证据，不必要的风险处置仍出现在 11.2%–58.7% 的运行中，且不同代理配置之间存在显著差异；（II）更强的任务能力并不能确保更恰当的风险处置，而处置违规会严重损害开发者体验，这表明风险处置是一项独立的能力维度；（III）代理表现出与既有风险管理研究结果一致的系统性模式，说明人类实践中的知识可以指导对该能力的诊断与改进。
行为水印将所有者标识嵌入 LLM Agent 的高层动作选择中，无需触及输出 token 即可提供溯源能力。以往的 Agent 水印存在两个方面的问题。首先，之前的三种方案都将水印绑定到精确的动作符号上，因此即使观测未发生变化，重命名工具也会导致解码不同步；在 AgentMark 自身的鲁棒性测试中，仅对观测进行释义改写就会使比特恢复率降至 16.8%。其次，之前所有 Agent 水印研究的都只是移除攻击：没有任何研究探讨对手能否伪造出一条可验证为他人所有的轨迹，而对于文本水印，Jovanović et al., 2024 已对这一问题给出了肯定答案。我们提出 Semantic Behavioral Watermarking (SBW)：在历史条件约束下，对语义动作簇进行水印嵌入，并将公开簇分箱替换为基于密钥且抗碰撞的分箱；其中，新桶的分配在随机预言机模型中被证明是不可预测的。在五种 Agent 模型（3B-14B，来自四家供应商）和三种编码器上，两个基准测试均呈现相同的排序结果：在 ToolBench 上（每个模型 600 条轨迹），在经置换校准的 1% FPR 下，重写场景中的检测率，簇级方法为 0.49-0.66，而精确符号方法为 0.05-0.17；选择一致率分别为 72-83% 和 22-27%，后者采用 logit 偏置。在 ALFWorld 上（每个模型 100 个回合），对应数值分别为 0.92-0.97 和 0.00-0.01。带密钥的分箱将自适应伪造率从 100% 降至主要工作点的假阳性下限（bge，r=64）。我们还明确了该保证不涵盖的边界：当对手复制受害者自身的步骤时，乱序拼接攻击会被抵御（Qwen2.5-3B 上为 0.000），但链式重放在五种模型上仍为 0.76-0.98；这一问题作为开放问题报告。释义鲁棒性使每步水印容量付出的代价约为一半。代码见 https://anonymous.4open.science/r/SBW-Agent-Watermark。
大型语言模型智能体正在加速科学自动化，但经过验证的执行结果很少能转化为持久的程序级改进，而现有评估也未考察这一过程在自然科学和社会科学的连续任务中的表现。我们将 ScienceClaw 形式化为一种固定参数的程序自演化机制，统一任务求解、科学验证和程序更新。ScienceClaw-Eval 覆盖 23 个学科，通过连续任务流和独立重置评估，衡量科学正确性、演化增益、保持能力、跨数据集迁移能力和演化成本。我们的框架通过多轮交互修复可执行工作流，将经重新执行验证的失败—成功轨迹转化为相互关联的 Skill 和 Operator 候选，并且仅当源任务回放能够复现修复结果且独立科学任务得到改进时，才保留相应更新。代码可在 https://github.com/beita6969/ScienceClaw 获取。
2011年的助手会回答并等待，2023年的代理会完成一项任务后停止。2026年9月，Meta 的 Muse 展示了一种面向单个人的代理：它拥有账户、设备和记忆，并能在持续不断的对话中工作；但它是封闭的，运行在供应商的云端，且仅限于一个国家。人们期待这种代理能够操作个人的账户和设备，跨越数周记住相关信息，在值得主动联系时先发起对话，并为自己的行为负责。这是一类软件，而不是一个模型；直到现在，它还没有开源的对应方案。本报告通过五个问题和三个阶段定义个人代理，并根据 Meta 的公开资料及其生产环境提示词副本，梳理 Muse 的构建方式，每项陈述均标注来源。随后，报告介绍了 GPL-3.0 许可下的开源对应方案 nanoMuse：在个人拥有的每台设备上运行一个代理，让它能够操作手机屏幕和电脑屏幕。它们通过任何人都能运行的中继服务共享同一段对话；每项操作都要经过 Sentinel；记忆以个人可以读取的文件形式保存；模型则由用户自行选择。报告还给出了其规模和成本的估算，并将以下内容列为路线图：开放性、带来源追溯的记忆、面向操作能力的评测套件，以及用于这些操作的开源模型。
基于 LLM 的智能体正在不断推动科学与工程问题求解的发展，而物理模拟正逐渐成为一种具有挑战性但又切实可行的测试平台，用于复现复杂的物理现象，并应用于具身智能、游戏和电影等领域。作为此类模拟的核心工具，求解器负责计算动态系统的状态如何随时间演化。构建这类求解器需要理解物理规律以确定合适的模型，运用数学推理来描述底层动力学，并通过软件工程将其实现为可执行代码；然而，LLM 智能体的这一能力仍未得到充分探索。为此，我们推出了 WorldSolver，这是一个包含 168 个模拟任务的基准，任务源自 61 篇经典计算机图形学论文中的物理现象，涵盖 7 个物理领域。每个任务都包含一个代码框架，为场景提供固定的模拟环境，而求解器的实现则留待智能体完成。具体而言，我们从三个维度对其进行评估：执行检查，用于判断是否成功执行；视觉保真度，用于评估渲染模拟中是否复现了预期的动态行为；以及物理合理性，用于基于物理规律验证所生成的动力学。对前沿智能体的实验表明，生成可执行的求解器本身就十分困难，而同时满足视觉和物理正确性则更加困难。GPT-5.6-Sol 和 Claude-Opus-5 的表现相较于其他受评估智能体更好，但总体得分仍分别只有 48.7% 和 46.7%。WorldSolver 是迈向智能体求解器生成的早期一步，我们希望它能够推动智能体忠实模拟动态物理世界的研究进展。代码见 https://github.com/sirujiang/WorldSolver。
Many long-horizon agents compact their context on a global rule, usually a token budget, blind to what the agent was doing. We ask whether the agent's recent behaviour predicts when a compaction will hurt. TRACE's public corpus of 590 harness-triggered AppWorld compaction boundaries replays each boundary from a re-executed prefix state under the pre-compaction context and under the summary, and records the burden of the next actions: calls that error or repeat a call already made. We find that pre-boundary history predicts post-compaction harm only weakly. An internally prespecified contrast by prefix placement is a wide null, and the naive "has-written" label behind it turns out to measure trajectory phase. The best extension-protocol trigger reaches held-out AUROC 0.66 (0.64 on the replicate's own label) against a same-boundary replicate of 0.72; the best frozen, interpretable trigger avoids 21% of harmful (positive-burden) boundaries while keeping 84% of compaction opportunities, and exceeds the random-rule expectation on count but not on burden mass (a post hoc comparison). Whether the best trigger beats a token-budget rule at matched retention cannot be evaluated on the release. We state what corpora should ship to answer it.
Web 代理通过读取和操作第三方编写的页面来完成用户请求，因此，植入页面中的指令可能会将代理引向偏离用户目标的方向。代理不能简单地忽略页面，因为页面同时包含任务所需的数据和控件。当前的防御方法会让代理针对训练前固定的注入进行微调，而能够适应已训练模型的攻击者可以绕过这些防御。对抗训练允许攻击者进行适应，但会固定任务，因此代理一旦解决某项任务，该任务就不再具有教学作用。我们提出了 AdvSim2Real，它在一个冻结的 Web 世界模型中，让任务课程、注入攻击者和代理共同演化。对于代理约有一半概率能够解决的任务，任务课程会获得奖励；而攻击者只有在实现成功翻转时才会获得奖励，即通过注入将一次经评判为成功的执行变为失败。模拟器中的训练使 4B 代理具备更强的能力和更高的鲁棒性：无论有无攻击，其完成率都会提升；面对训练期间从未接触过的前沿模型攻击者时仍能保持性能；其能力增益也能迁移到真实浏览器上。在 150 项 Web 任务中，相较于基础代理，AdvSim2Real 使代理在这一未见过的攻击者下的完成率提升了 33.6\%。
大型语言模型（LLM）能够解决许多狭窄领域的任务，但对于数百万个相关实例分别进行查询，成本可能高得令人难以承受。LLM 智能体能否自主地为这类工作负载创建更廉价的解决方案？我们将这种能力称为“装瓶”：将通用能力转化为特定任务解决方案的能力，同时平衡答案质量与摊销成本。我们提出了 BOTTLED 基准测试：智能体会接收到一个完整的未标注工作负载，并必须在固定的时间、计算资源和 LLM API 预算内完成任务。智能体可以自行选择方法，例如训练一个小型模型或编写可复用的程序。在十个模型和三个任务上的实验表明，强大的零样本任务性能并不能稳定地转化为强大的装瓶能力。零样本得分相近的模型，在完成装瓶后可能表现出显著差异；60 次装瓶运行中有 48 次的得分低于其对应模型零样本性能95%置信区间的下界。此外，在相同 token 预算下，60 次运行中有 31 次的表现不及两个小模型蒸馏基线中较强的那个。尽管如此，装瓶仍能带来可观的节省：在查询-产品相关性分类任务上，Opus 5 以报告成本约低657倍的代价，保留了其零样本 macro-F1 的约82%。装瓶的表现也具有竞争力：Jev 是专为廉价、重复性推理而构建的“系统一”模型；在相同任务上，Opus 5 以 Jev 预计完整工作负载成本的四分之一，达到了 Jev 约94%的 macro-F1。BOTTLED 为评估和提升智能体将有限资源投入可复用解决方案的能力提供了基础，使其能够应对规模庞大且重复性高的工作负载。
LLM 智能体越来越能够执行复杂任务，并在软件工程和数学等易于验证的目标上递归式地自我改进。由于对齐要难得多，这带来了能力不断提升却未能实现适当安全对齐的日益增长的风险，尤其是在能力扩展到自动化研究和网络安全领域之后。现有方法主要聚焦于利用可验证反馈实现能力自我改进，或通过更强模型的监督或精心策划的数据进行对齐训练，从而形成了对齐方面的外部监督瓶颈。我们探讨当前模型能否改进自身的安全对齐，并提出 SIGMA：一种数据生成与训练流程，使对齐自我改进能够泛化到分布外场景。仅给定一份说明模型期望行为的“Model Spec”，SIGMA 利用模型的推理能力来强化其自身的安全推理。SIGMA 首先执行规范引导的任务合成，将候选模型作为任务设计智能体，生成多样化的对齐困境场景，并将其转化为训练任务，以压力测试模型对 Model Spec 的理解。随后，SIGMA 通过监督微调和基于评分标准的强化学习开展自我评判式对齐训练，并将模型自身作为奖励模型。尽管训练数据仅包含单轮聊天数据，SIGMA 仍能提升模型在多轮智能体环境中的安全对齐能力（AgentHarm 的有害性评分从 22.6 降至 14.8；Agentic Misalignment 从 79.1 降至 3.8），性能优于 Deliberative Alignment 和 Constitutional AI 基线，同时保留通用能力。分析表明，兼顾无害性与有用性的 Model Spec、测试时用于安全审议的推理能力，以及 SIGMA 任务设计智能体生成的高质量评分标准，对于实现有效的自我改进至关重要。
在后果重大的领域使用 LLM 智能体时，要明智地决定是否信任其输出或进行干预，需要对智能体成功完成任务的概率具备经过校准的置信度。智能体的置信度估计之所以困难，是因为有关任务成功与否的证据分散在智能体轨迹中异构且相互依赖的多个步骤之间。实际部署智能体还带来了更多挑战：前沿 LLM 通常无法充分提供内部信号，智能体运行成本高昂，而且训练数据可能无法获取或很快过时。为应对这些挑战，我们提出了置信度推理图（Confidence Reasoning Graphs，CRGs），这是一种推理时框架，能够从单条轨迹中估计智能体完成任务的概率，无需访问特权模型信息或训练数据。CRG 不会将一次执行压缩为单一的整体判断，而是首先提出“智能体完成了任务”这一主张，将其分解为以轨迹证据为依据并结合上下文的子主张，估计每个终端主张的置信度，最后将这些置信度聚合为总体置信度估计。在三个智能体基准、三种骨干模型和三个智能体框架上，CRG 相较于语言化、基于采样以及白盒代理基线，能够提供经过更好校准的置信度，并实现更强的风险感知决策。我们进一步发现，仅凭校准误差可能会产生误导：一个白盒代理基线看似校准良好，但其判别能力接近随机水平。消融实验表明，CRG 的改进主要源于主张级置信度估计与聚合，而不仅仅是图构建本身。最后，CRG 会揭示每个置信度估计所依据的主张和轨迹证据，从而支持在决策时对其进行审计。
移动 GUI 智能体日益在受用户历史记录和偏好影响的界面上运行，但其跨不同用户的可靠性仍缺乏充分研究。我们提出 PAIR（Personalized Application-state Instantiation and Rendering），这是一条用于构建用户条件化应用状态的流程，能够在不同用户之间对同一任务进行受控评估。我们进一步提出 RePAIR（Reinforcement learning with Personalization-Aware Interaction Rewards），这是一种通过学习不同用户在子目标结果上的差异来提升用户条件化移动环境中可靠性的训练方法。在对六个智能体的评估中，我们发现不同用户之间的任务成功率存在显著差异，并且在用户条件化 UI 上下文中，子目标达成率始终较低（低 6.98 至 15.4 个百分点）。对于取自每位用户自身内容的个性化目标，这一差距进一步扩大（低 8.77 至 22.0 个百分点）。在这些上下文中，失败通常表现为选择了其他项目而非预期目标，尤其是在目标暴露之前。最后，在未见过的用户上，相比其监督微调基础模型，RePAIR 将用户条件化 SAR 提高了 5.87 个百分点，将全成功率提高了 7.50 个百分点，并将整体 Task SR 提高了 9.42 个百分点；这初步表明，明确学习跨用户差异能够提升 GUI 智能体的可靠性。
随着智能体通过生成和修订技能而持续改进，发现和完善这些技能的过程本身也成为一种可学习的对象。任务技能直接作用于任务执行，而元技能则决定智能体如何发现并改进未来的技能；因此，元技能的价值会通过其引发的后续搜索过程体现出来。现有方法根据观测到的原始技能搜索轨迹和分支结果来改进元技能。然而，分支表现同时受到初始发现状态和产生该搜索过程的元技能修订的影响，因此难以刻画某次具体修订究竟改变了什么，也使更新倾向于选择受益于有利状态的修订，而非真正改进过程的修订。我们提出 HMED（事后元经验蒸馏），一种为自我改进智能体构建元经验的机制。HMED 回溯修订所源自的已完成事件，并从同一恢复后的发现状态重新执行当前版本和修订后的元技能，从而能够在共享条件下观察与修订相关的变化。每次比较都会被蒸馏为一条元经验，即一种可供未来更新复用的结构化记录，因此，即使最终未被保留的修订也能提供学习信号。在三个交互式智能体基准以及开源和闭源模型上，HMED 相较于强基线方法持续提升技能发现性能，使元技能学习不再局限于分支结果，而是进一步关注改变改进过程所带来的后果。
AI agents 正日益被编程用于对大规模非结构化数据集合中的知识工作进行自动化处理。这类自动化要求具备可重复性：当底层证据不变时，即使每次单独给出的答案看似合理，agent 的分类、优先级和计数也不应在不同运行之间发生实质性变化。我们提出了一个重复运行评估框架，用于对语义等价的类别进行对齐，并关注两个运行指标：主题变动率，即返回类别集合的归一化变化；以及数量不一致度，即仍然存在的类别的计数变化。我们在八个前沿模型上评估了三项反复出现的客户反馈任务，涵盖从 100 到 5,000 条记录的语料库规模、多种提示词，以及三种执行设计：原始生成、不依赖分类体系的层级分解，以及使用持久化主题、子主题和记录级预测的分类体系驱动agent（TGA）。在固定使用 Claude Opus 4.8 和 1,000 条记录的语料库时，相较于原始生成和层级分解，TGA 将主题变动率降低了 86--88%，而匹配主题的数量不一致度为零。分类体系驱动agent 比评测中的每个原始模型都更加稳定，在每种语料库规模下也都保持更高的稳定性；即使收紧或放宽主题匹配标准，这一优势仍然存在。尽管本研究评估的是客户反馈，但该框架更广泛地适用于对非结构化语料库进行重复性综合分析，包括财务报告、法律文件、事件记录和科学文献。总体而言，这些结果表明，分类体系 grounding 能够为反复进行的知识工作产出更加一致且可重复的结果。
LLM 智能体通常缺乏在新环境中可靠行动所需的操作知识，因为它们必须自行发现特定工具的行为方式或环境惯例。由于没有对过往尝试的记忆，它们会在不同任务中重复相同的错误，导致更多任务失败和更长的执行轨迹。为解决这一问题，智能体系统通常依赖人工编写的指南，或依赖通过训练任务和 oracle verifier 构建的程序性记忆，而这两种方式都需要预先了解环境。我们提出 DAEDALUS，一种无需现有任务或 oracle verifier、通过自生成练习来引导构建可复用智能体记忆的方法。DAEDALUS 配对使用两个智能体：探索器与环境交互，以生成具有挑战性但可解决的任务；求解器则尝试完成这些任务。每次求解器失败后，系统都会提炼出一条启发式规则，只有当求解器在相关上下文中使用该规则反复成功后，这条规则才会被接受。这些结果还会为探索器提供反馈，帮助其调整未来任务的难度。随后，已接受的启发式规则会被整合到记忆库中，供测试时使用。在 AppWorld、$τ^2$-bench 和 AutomationBench 上，DAEDALUS 相比无记忆基线将平均成功率最多提高 15.9 个百分点，将 pass^5 最多提升至 2.2 倍；与使用训练任务的方法相比，其性能具有竞争力，且推理成本低于大多数此类方法。我们表明，即使探索预算较小，也能取得性能提升；其启发式规则同样能帮助其他模型系列的智能体。进一步的消融实验表明，求解器轨迹提供了推导有效启发式规则所需的关键信息，而将早期发现进行因子化则能提高探索的成本效率。除构建记忆外，我们还发现，DAEDALUS 生成的任务可以作为基准任务的替代，用于根据性能对模型进行排名。代码和相关资源：www.github.com/illuin-tech/daedalus】【。
前沿 LLM agents 已被证明能够解决日益复杂、且人类已有可量化解决方案的任务。这引出了一个重要问题：LLM agents 能否超越人类已经解决的问题？随着那些由人类充分探索并开发出的解决方案对于缺乏背景信息或训练数据不足的问题变得不再适用，发展复杂策略以应对重大问题的能力变得至关重要。我们通过电子游戏速通这一共同实践，研究 agents 形成此类策略的能力。在速通中，参与者竞相寻找在特定条件下完成电子游戏的最快方式，并在这一过程中发掘出有趣且非正统的玩法，而这些玩法要求对底层游戏机制进行深入理解并熟练掌握。我们推出 SPEEDRUNBENCH，这是一个在 9 款不同游戏中评估前沿 LLM agents 的基准。要在该基准上取得良好表现，agents 必须不断改进策略、反思自身表现、利用已获得的知识，并对长时间跨度的行动进行推理，从而解决一个日益困难的问题：比自己和其他所有人都更快。我们的实验表明，尽管前沿 agents 在简单的平台游戏中已接近人类世界纪录，但在实际资源预算下，它们在更长、更复杂的游戏中的表现仍落后于人类。这些结果表明，SPEEDRUNBENCH 不仅是研究 agents 策略形成能力的有用测试平台，也是一种不易饱和的评估指标，因为几乎总有更快的完成时间等待被发现。
LLM 工具使用代理运行于动态环境中，其中许多操作都带有运行风险。然而，大多数安全机制只有在错误已经发生后才作出反应。现有的主动预防方法要么基于思维链推理对代理进行微调，要么将自然语言防护规则编译为运行时检查，但它们都没有提供结构化且可审计的判定结果。我们提出 POLAR，这是一种面向小型工具调用代理的防护框架，通过结构化的双层本体评估操作的可逆性。POLAR 通过推导候选逆向操作序列，为每个操作分配分级的可逆性分数；未达到阈值的调用会在执行前被剪除。在六种代理模型上的 $τ^2$-bench 评测中，POLAR 使其中四种代理在航空领域的平均任务奖励提升了 0.11 至 0.18 分，但在十八个模型—领域单元中，总体上仅有八个得到提升；在零售领域和更强的代理上，性能往往下降。POLAR 提供了可审计的结构化检查，并刻画了其在任务效用方面的权衡。奖励并不能直接衡量所避免的危害。
数学问题求解通常需要确定性的计算步骤，智能体会将这些步骤委托给工具，并默认信任工具的结果。然而，工具可能在不易察觉的情况下出错，返回看似合理但实际上错误的结果。智能体能在多大程度上检测并纠正遭篡改的工具调用输出？我们通过一个受控篡改框架对此进行研究：在针对性问题上，隐藏的拦截器会将工具调用结果替换为看似合理但错误的信息。我们在四种验证设计下，对智能体在31道题目上的表现进行评估，包括不进行验证（基线）、强制进行同一上下文反思、可选的全新上下文验证，以及可选的结构化验证。在不进行验证的情况下，篡改会导致准确率大幅下降，从100%降至72.4%。强制反思可将这一表现完全恢复至100%。只有当模型主动调用验证机制时，可选验证才会提升准确率。我们的结果表明，检查频率与鲁棒性差异高度相关，而调用次数的不均等使得验证器质量难以进行受控比较。一项辅助恢复实验显示，在明确检测到错误后，完全重新开始解题可在100%的情况下成功。这些发现表明，验证器的可用性与验证策略是数学智能体可靠性的两个独立组成部分。强制性策略能够确保执行验证，而可选策略则取决于模型是否自行选择调用验证机制。
共享记忆协调智能体的行动，但正确的记录并不能证明这些行动满足任务要求。记忆治理和故障诊断会规范或检查已记录的信息；但它们本身并不能证明这些信息是否足以判断任务义务是否得到履行。我们通过规范状态使用、信息交接和最终状态一致性的义务，并明确用于判断履行情况的证据条件，定义了执行一致性。我们的核心论断是：在相同任务规则下，完全相同的留存记录可能对应合规执行，也可能对应违规执行。受控移除回执、行动依赖关系或响应有效性等证据后，82.4%的相反标签对无法被区分；恢复这些证据后，97.9%的合并对可以被区分。自然日志标注能够识别实际执行中已定义的违规行为。然而，现有日志并不总是明确表示进行这些判断所需的执行关系。为评估该定义的实际价值，我们使用 CAVERT（一种一致性诊断与恢复框架）从日志中提取有支持的关系，并应用这些判据。在全部12种基准测试-执行器设置中，它的诊断表现始终优于由合约提示驱动的 LLM 和基于规则的基线。在相同的门控条件和执行器限制下，它在评估的全部四个环境中也优于规则引导的恢复方法。这些发现明确了智能体记忆和执行接口应当保留哪些执行证据，以实现可靠的判断。
对话式 MLLM Agents 越来越需要在专业工作流中提供协助，从 AI 研究和工程设计，到产品管理和业务运营。然而，这一能力仍缺乏充分探索：现有基准测试主要聚焦于非正式的日常交互和个人生活场景，所涉及的通常是摄影类自然图像、孤立的静态工件以及面向回忆的问题。相比之下，专业场景往往包含结构化、信息密集型的工件，这些工件会频繁经历修订和权威更新；同时还包含组合式查询，需要整合多个工件版本并精确跟踪状态。为应对这些挑战，我们提出 DSV-Mem，一个用于评估 Dense Stateful Visual Memory 的基准。DSV-Mem 包含经专家审阅的场景，以及涵盖五类面向用户的问题共 1,000 道（Current State、Past State、Derived State、Change History 和 Conflict/Refusal）。一种受 Hartley 启发的标准更倾向于选择需要更广泛检查视觉证据的问题。我们还提出了一个生成框架，通过将状态转换合成与对话填充解耦，生成评测套件。在涵盖前沿模型、开放权重模型及记忆管理方法的 27 种配置上进行评估后发现，最强基线在 DSV-Mem 上的得分低于 45%。分析得出以下发现：1）多模态性和信息密度都会增加难度，但状态演化，尤其是支配性更新的数量，是测试中的主导因素。原始对话/文本堆长度、OCR 和算术并非主要瓶颈；2）模型在回答前往往无法根据先前的状态更新核验用户前提；3）增加推理投入和采用记忆管理方法只能带来有限收益，而具备状态感知能力的设计则更为有效。该基准及其代码将公开发布。
多模态大语言模型（MLLM）正在快速发展，而传统基准开发却相对滞后，导致对新近发现的能力缺口展开研究的进程被延误。此类研究需要具备表现力的任务格式和按需构建流程：信息丰富的图表使图表问答（Chart QA）适合用于探测感知与推理的耦合能力。自动化 Chart QA 构建旨在将已识别的能力缺口按需转化为针对性样本，从而缩短基准开发周期。然而，当前方法通常将目标引导与从零生成分开处理：目标引导系统往往需要准备好的数据、图表或模板，而从零生成系统主要确保产物有效，却没有明确控制新合成的要求与内容是否仍与外部指定的诊断目标保持一致。我们提出 ChartBmkAgent，通过基于稀疏错误分类规范构建完整的 Chart QA 样本，将已识别的能力缺口转化为针对性的诊断证据。在整个构建过程中，一个中央控制框架负责监管专用智能体，要求各阶段提供与原始错误类别保持一致的证据，并记录每项接收决策的依据。在覆盖整个分类体系的300个样本上，MLLM 的准确率介于32.7%和84.3%之间，并呈现出不同的类别特征，表明生成的样本能够揭示能力差异。在三组源模型比较中，针对性后续测试的得分为50.0%，而匹配对照组为82.2%（$p=8.96\times10^{-6}$）；六组跨模型比较均呈现相同方向，证明了针对性验证和诊断数据生成的有效性。多个评估模型对每个样本是否测试了其指定的错误类别进行了评估；其中86.4%的样本符合这一标准，为目标保持性提供了实证证据。
法律智能旨在支持涉及案件状态演变和多种角色的长时程法律流程中的可靠决策。然而，现实世界的法律部署在事实、证据和程序背景方面存在显著的案件异质性，暴露出静态智能体策略的局限性。此外，法律推理在不同角色和程序阶段之间具有内在的相互依赖性，因此全局可靠性从根本上不同于孤立的角色能力。为应对这些挑战，我们研究无需训练的测试时智能体适应方法：智能体在不更新模型参数的情况下，持续利用部署期间从先前案件和当前交互中获得的信号。我们提出 \method，引入\emph{测试时记忆演化}，从先前案件中检索可复用经验，使其适应当前的事实和程序背景，并整合累积经验以支持后续决策。此外，\emph{与规则对齐的协作}依据行为和程序要求，对特定角色的行动进行验证和修订，从而实现跨角色、跨阶段的协调决策。在五个骨干模型上针对 J1-EVAL 和 LegalWorld 进行的大量实验表明，与具有代表性的推理和智能体基线相比，该方法在合理的交互与计算成本下实现了稳定提升。消融实验和案例研究进一步表明，这两个组件在经验适应和跨角色协调方面具有互补优势，提升了长时程法律推理的可靠性和效率。
Zero-shot coordination in embodied settings requires acting while the partner is intermittently out of view, leaving existing methods with ambiguous partner representations and uncertainty over hidden partner states. We propose Predicting Intention of Partner (PIP) to jointly address these challenges. PIP uses a Joint-view VAE to distill richer training-time evidence from the union of both agents' local observations into a partner representation available from local observations alone. Partner-state Belief networks further infer the partner's hidden location and behavioral tendencies from the ego agent's interaction history. We evaluate PIP in Burrito-PO, Overcooked-PO, and a Melting Pot substrate, together with a human evaluation in Burrito-PO. PIP attains the highest mean performance among the compared methods across all three benchmarks. Human evaluation and diagnostic analyses further support coordination with unseen partners and the contributions of both components under partner occlusion.
Meta Platforms Inc. is partnering with Walmart Inc., Stripe Inc. and the enterprise artificial intelligence firm Sierra Technologies Inc. on the creation of new standards that will define how autonomous AI agents interact with other businesses online. The new standard was introduced by Sierra co-founder Bret Taylor, the former co-chief executive of Salesforce Inc., who […] The post Meta teams up with Bret Taylor’s Sierra Technologies on new standards for AI agent commerce appeared first on SiliconANGLE .
在 Wikimedia 项目中发现 OpenAI“失控”智能体的活动。鉴于维基对失控智能体群体而言是多么诱人的目标，Wikipedia 在展开调查后发现相关活动的证据，也就不足为奇了：Wikimedia Foundation 展开了自己的调查，以确认 Wikimedia 网站是否同样受到 AI 智能体的影响，重点调查了由 OpenAI 运营的智能体。我们可以确认，已经在 Wikimedia 平台上发现了一些这些“失控” OpenAI 智能体的活动。未经授权的机器人活动包括编辑我们的维基、数次试图利用我们托管的公开笔记工具，以及产生大量流量，详情如下。调查人员发现了智能体编辑沙盒页面的证据，试图利用 Etherpad 等基础设施来帮助代理来自其他地方的内容，还发现了大规模爬取活动，以及对其 Wikidata Query Service 发起的“数十万次数据查询”。我最大的猜测是，其中大部分活动来自一群与那些在为研究任务进行训练期间破坏德国维基的智能体相似（或就是同一批）的智能体。Wikipedia 沙盒维基的编辑似乎始于 5 月 12 日，而那起事件报告的对 UseModWiki Sandbox 页面的初始测试编辑始于 5 月 11 日。标签：wikimedia，wikipedia，wikis，ai，generative-ai，llms，ai-ethics，accidental-cyberattacks
AI is taking shape, moving from models that answer questions to models that can reason. Now models are powering the agentic era, and Dell Technologies Inc. has announced several extensions to its AI Data Platform designed to manage this new paradigm. These include extending the Dell Data Orchestration Engine with three new capabilities: the Unified […] The post Transforming the data path: Dell’s new releases move enterprises closer to the agentic data center appeared first on SiliconANGLE .
GPU utilization during post-training depends partly on how efficiently infrastructure moves data and loads updated models. As enterprises continually refine AI agents, reducing delays between training rounds can help keep that process moving. The need for continual improvement is pushing CoreWeave Inc. to build a full-stack AI cloud for the agent lifecycle. You.com Inc. provides […] The post CoreWeave targets GPU utilization in continuous AI post-training appeared first on SiliconANGLE .
Personal AI agents promise to shop, book flights, and make reservations for you. But deliberate blocks and anti-bot defenses are getting in the way, leaving consumers caught in the middle. A new standard aims to help.
现成的 AI 助手会在对话之间遗忘你。本文介绍如何在 Amazon Bedrock AgentCore 运行时上使用 OpenClaw 构建一个能够积累上下文的个人助手，并借助 AgentCore memory 将一次性的聊天记录转化为持久化、结构化的知识，再通过元数据筛选器进行检索。
TIL: Using Parseable with Datasette for OpenTelemetry traces I saw Parseable in a Show HN today - it's a new observability platform with both an open source (AGPL) Rust implementation (a single ~180MB binary), an "Enterprise" version with extra features and a cloud hosted option. Since Datasette 1.0a41 added OpenTelemetry support (thanks, Alex Garcia), I decided to fire up Codex and have it figure out how to run Parseable and feed it traces from Datasette. Here's my (human-written) TIL showing the patterns that worked, and here's a screenshot of a Datasette trace displayed within the Parseable localhost web application: Screenshot of a trace detail view in an observability web app, with a span waterfall overlaid on a dimmed navigation sidebar and filter column. Dimmed sidebar: breadcrumb "Community > Traces > datas" (cut off), search box "Search... ⌘K", nav items "Home", "Ingest telemetry", section "ANALYZE": "Keystone", "Dashboards", "SQL Editor", section "OBSERVE": "Logs", "Metrics", "Traces" (selected), "APM", "Agents", section "MONITOR": "Alerts", "Errors", section "DATA": "Datasets", and at the bottom "Settings", "Book a call", "Support". Dimmed filter column, cut off at the right edge: "Search fi", "Core", "Log format", "User agent", "Source IPs", "Error", "Service", "service.ins", "service.na", "datasett" (checked), "Span", "Database", "HTTP", "http.reque", "NULL" (unchecked), "GET" (checked), "http.respo", "http.route", "Server", "server.add", "Telemetry", "URL", "All fields". Trace panel header: "Trace detail > 6f819a2170bcd1e91c6ea3ae236ec60b" with a copy icon, a "Related logs" button and a close X. Summary: "Start time 6:56 PM, Oct 6, 2026 UTC", "Duration 40.9 ms", "Spans 247". A minimap with axis "0ns 10.2ms 20.5ms 30.7ms 40.9ms" shows many short span bars cascading diagonally from top left toward the lower middle, with a few longer bars. Below is a span table with a "Span name" header, a "Search spans..." box, collapse and expand buttons, and a timeline axis "0ns 10.2ms 20.5ms 30.7ms 40.9ms". Rows (name, service, duration): root span with collapse toggle "123", "GET /..." "datasette..." 40.9ms spanning the full timeline; then alternating rows where each "db.query" has a collapse toggle "1": db.query datasette-local 679µs, db.query.execute datasette-local 278µs, db.query datasette-local 341µs, db.query.execute datasette-local 55µs, db.query datasette-local 357µs, db.query.execute datasette-local 197µs, db.query datasette-local 330µs, db.query.execute datasette-local 73µs, db.query datasette-local 2.26ms, db.query.execute datasette-local 2.02ms, db.query datasette-local 232µs, db.query.execute datasette-local 64µs, db.query datasette-local 207µs, db.query.execute datasette-local 71µs, db.query datasette-local 6.11ms. The child span bars start progressively later across the early part of the timeline. Tags: datasette , observability , alex-garcia , opentelemetry
AI teaching agents could make personalized instruction available to more learners by reducing the cost and staffing demands of one-to-one tutoring. MasterClass is exploring that potential through a system designed to adapt to how each student learns. The shift is also feeding CoreWeave Inc.’s full-stack AI cloud push as companies move AI agents out of […] The post MasterClass bets AI teaching agents can broaden access to tutoring appeared first on SiliconANGLE .
According to the Wikimedia Foundation, rogue OpenAI agents edited wikis without permission, tried to abuse a citation tool as a proxy, and may have caused a partial Wikidata Query Service outage through massive crawling. Wikimedia says AI companies need to take responsibility for their agents instead of pushing the burden onto volunteer editors. The article Wikimedia confirms OpenAI's rogue AI agents edited wikis, tried to compromise tools, and hammered its infrastructure appeared first on The Decoder .
Hey all! For the past months I've been vibe-coding a free Photoshop editor for macOS (Native using Swift and a bit of Rust). It has many features that I've always wanted in such an editor (but was unwilling to pay the Adobe tax). I did try open source solutions such as Gimp but they never worked for me. In terms of features I'm not gonna list all, but I'm very happy it has: - A proper canvas where text and vectors keep their quality when zooming in - Layers (vectors, text, pixels, color adjustment, effects, reference layers, mask layers, layer groups and so on) - Many effects (including unusual things like dithering or NTSC) - All the tools you'd expect (brush, paint, repair, etc) For some of these tools it uses local models - Selection operations via local models (e.g. select all red cars) - Import / Export PSD, SVG and many others. - Support for tracing images to vectors, text to beziers, bezier editing, etc. Most importantly (I think), it has a optional AI integration via a agent harness which lets a model remote-control the app for you to build a document out of layers & effects & brushes etc. This is really powerful because in comparison to typical genai flows this allows you to make changes, hide things and easily scale them up (if most elements are vectors). There's also support for genai operations (e.g. if you want to embed a genai generated SVG or image as a new layer). Would love to get feedback: https://pixellect.app/ Comments URL: https://news.ycombinator.com/item?id=49981700 Points: 1 # Comments: 0
AI agent memory is creating new demands on infrastructure as agents run longer sessions and spread across the enterprise. Retaining that context and making it available when needed puts pressure on memory capacity and data movement. Those demands extend beyond the context held during an individual interaction. Enterprise agents also need shared knowledge that persists […] The post Vast uses tiered storage to ease AI agent memory demands appeared first on SiliconANGLE .
Insurers are bracing for millions in claims from rogue AI agents, and executives like OpenAI's Sam Altman and Anthropic's Dario Amodei could be personally on the hook for the fallout. The article Insurers brace for millions in claims as AI agents spin out of control appeared first on The Decoder .
通过 Amazon Bedrock AgentCore、用于实时语音的 Amazon Nova Sonic 以及用于回答政策问题的 Amazon Bedrock Knowledge Bases，为航空应用添加语音旅行礼宾服务。旅客可以通过语音更换座位、查询延误并询问政策问题；同时，代理通过 MCP 工具连接到您的后端，并在写入任何更改前确认每项操作。
Hi HN, I'm Weilun, cofounder of OpenChart. We built OpenChart because we wanted Claude and Codex to interface with the markets, like most of humans do, through charts rather than through CLIs. With your existing AI plans, your agent can work directly with your charts: annotate a setup, write an indicator, investigate a market move, or create an alert. The chart, conversation, and research live in the same workspace. OpenChart alerts on any market move: draw any shape on a chart and attach an alert. When price crosses it, the alert triggers your agent to investigate the move and save the research in your workspace. The agents can work with charts, indicators, and alerts. Custom indicators and alert conditions are written in [Tea]( https://github.com/longsurf-ai/tea ), a programming language we built for the markets. We believe software shouldn’t impose artificial friction on your workflow. OpenChart is built on that belief with no software cap. The source code is publicly available, and you can use free market data from day one. We also offer optional paid cloud market data for lower latency. Everything is private. Your charts, research, and indicator code are stored on your own computer, with the app's backend running locally too. We currently support only macOS on Apple Silicon, but we're expanding to more platforms. Website: https://openchart.co/ Download: https://downloads.longsurf.ai/openchart/darwin/arm64/OpenCha... We built OpenChart because we want to explore how assets can be co-manages by agents and humans. I'd especially like to hear how you are using agents to help with your investment journey. If you like what we're building, a star on GitHub would mean a lot to us: https://github.com/longsurf-ai/openchart Comments URL: https://news.ycombinator.com/item?id=49979793 Points: 1 # Comments: 0
The reports of OpenAI agents harming 3rd party sites keep coming.
Artificial intelligence development firm Anaconda Inc. is enhancing its enterprise AI platform with tools to coordinate groups of AI agents, test their security and move applications into production. Today’s introductions continue the company’s expansion beyond its roots in Python software distribution into broader AI development and management. The announcement draws on three recent acquisitions: AI coding […] The post Anaconda expands beyond Python with agent swarms and AI security testing appeared first on SiliconANGLE .
SAP SE today announced a shift from promising agentic artificial intelligence autonomy to defining a path for users to interact with agents for work outside apps, not just within them. During SAP Connect 2026, the company said it is making Autonomous Enterprise architecture, introduced at Sapphire in May, generally available this month alongside Joule Work and […] The post SAP expands Joule into an agentic work layer as Autonomous Enterprise goes live appeared first on SiliconANGLE .
I was testing a bunch of token saving techniques and tools for my coding agent, and surprisingly it can save 20-60% cost. I bet most people do not want to pay attention on cost saving tricks every time while building, so I'm curious if people will pay for someone else to reduce their Ai bills. and what would you need to see before trusting to pay and how much? Comments URL: https://news.ycombinator.com/item?id=49975157 Points: 1 # Comments: 0
Cohere turned its enterprise platform into a control center for AI agents with North 2, handling multi-step workflows on their own and retaining context across sessions. The article Cohere pitches North 2 as the enterprise AI control room that works with any model appeared first on The Decoder .
Have you discovered daily routines and use cases that bring you back? Comments URL: https://news.ycombinator.com/item?id=49974436 Points: 1 # Comments: 0
长视频理解对记忆提出了很高要求，因为回答问题通常需要检索分布在较长时间跨度中的信息。现有方法大致遵循两种范式：查询驱动的探索容易受到定位错误的影响，而与查询无关的记忆构建则可能遗漏特定于问题的细节。我们提出 VideoTapestry，一种无需训练的多智能体框架，通过由粗到细的查询驱动细化，自适应地调整预构建的分层视频记忆。预构建记忆将视频内容组织为三个层级，分别捕捉全局叙事上下文、事件级时间结构和细粒度关系证据。为支持由粗到细的定位与观察，我们为每个层级分配一个专门的智能体，使检索和细化保持在特定尺度的上下文范围内。在查询引导下，这些智能体重新访问相关视频区域，并通过有针对性的多模态观察丰富各层记忆。随后，根据原始层级结构对这些细化结果进行组装，形成复合的查询自适应记忆：以紧凑形式保留全局上下文，同时沿与查询相关的分支保留细粒度证据，以供最终推理使用。与直接进行 GPT-5.5 推理相比，VideoTapestry 在 LVBench、LongVideoBench (Long)、Video-MME (Long) 和 EgoSchema 上的绝对准确率分别提升了 17.2%、14.9%、9.8% 和 7.0%，在所有竞争方法中取得了当前最佳结果。
在去中心化的消费者对消费者（C2C）市场中，人们发布商品、与陌生人议价并互相评价，因此信任建立在声誉之上。如今，大语言模型（LLM）代理开始代表用户行事，从而给用户的资金、隐私和声誉带来风险。我们推出BazaarBench，这是一个模拟的C2C市场和基准测试，用于评估这些代理的安全性。它跟踪交易中的所有权、商品状况和承诺，并结合记录核查与基于评分标准的LLM判断，以识别涵盖五个阶段的六类失败。我们运行了三个基础市场，模拟时长均为30天；每个市场包含100个代理，所有代理使用同一个模型，库存则取自公开的eBay样本。在45次延续实验中，我们分别在普通指令、截止期限压力或诱导利用其他交易者的对抗性指令下，评估五个模型。每次延续实验都从某个市场第30天状态的副本开始，运行七个模拟日。受测模型控制同一批选定的20个代理，这些代理保留各自的人设、库存和历史记录；其余80个代理继续使用基础模型。在普通指令下，五个模型都曾尝试向多个买家承诺出售同一件商品。加入目标和截止期限后，每个模型的此类尝试次数都增加。在对抗性指令下，尽管商品不可用或卖家夸大了商品状况，受测卖家已承诺的交易仍得以完成的比例从15.4%上升至33.4%，其中GPT-5.4达到55.5%。在各模型和市场之间取平均后，每名受测代理的模拟周收入从普通指令下的20美元上升至对抗性指令下的33美元。大部分增收来自卖家从未持有过的商品。我们发布了模拟器、保存的市场状态、评估代码以及涵盖357,608次代理模型调用的记录，用于评估新模型和开发更安全的市场代理。
现代人工智能前所未有的计算规模依赖于复杂的、拥有数十亿晶体管的片上系统（Systems-on-Chip），然而用于验证这些芯片的工作流程仍然顽固地依赖人工。尽管大型语言模型（LLMs）已迅速进入电子设计自动化（EDA）领域，现有研究中约有 74.6% 聚焦于静态寄存器传输级（RTL）代码生成，使得仿真后验证和交互式波形调试在很大程度上仍未得到处理。我们提出了 Back-to-the-Future（BTTF），一种端到端智能体框架，用于弥合这一基础设施鸿沟。BTTF 将海量、非结构化的仿真转储数据提炼为规范化的关系型 SQLite 数据库，并将其与协作式多智能体编排引擎结合起来，将自然语言验证查询转换为具备模式感知能力的 SQL，同时将信号异常与带版本控制的 RTL 代码仓库关联起来。在包含 150 个查询的基准测试中，BTTF 达到了 95.33% 的执行准确率，为实现自主 EDA 验证勾勒出一条切实可行的路径。
开源网页智能体如今已经足以执行逼真的浏览器任务，但使用强化学习训练它们仍然依赖弱监督：二元任务成功信号过于稀疏，难以进行信用分配；而前沿语言模型评审器的调用成本过高，无法在每一步都调用，也不能假定其在部署时可用。我们提出 CLIFT，一种以保形自验证为核心的训练与测试时扩展方法。在训练期间，智能体针对自身的 rollout 回答自然语言验证问题；Compositional Conformal Certifier 仅保留其 URL 条件证据与训练时评审器一致的问题信号，通过极性感知提升分配带符号的信任权重，并将得到的验证器分数融合到逐步奖励中，同时绝不会从评审器基线中扣减任何分数。在测试时，同一经过认证的问题库被冻结，并作为 Conformal Trajectory Selection（CTS）的结构化证据重复使用：智能体采样一条贪心 rollout 以及一条或多条多样化重试轨迹，自验证器总结每条 URL 轨迹，然后通过一种保守的多数投票规则决定是否放弃当前最优结果，而无需调用任何外部评审器。这一机制支持三种设置。在 WebArena Infinity 上，CLIFT 在开源网页智能体中取得了当前最佳性能。在 VisualWebArena 上，使用开放模型训练的问题库可在测试时迁移到 GPT-5.5，并在标准测试框架下达到当前最佳性能。在 Online Mind2Web 上，无需针对该基准训练智能体，仅通过翻译经过认证的问题库，就能在零样本评估中提升一个实时网页智能体的表现。总体而言，这些结果表明，保形自验证能够将昂贵的评审器反馈转化为可复用的训练信号，以及无需评审器的测试时扩展信号。
记忆已成为 LLM 智能体生态系统不可或缺的一部分，支持跨交互的信息留存与复用。然而，现有大多数智能体记忆系统以查询无关的方式构建记忆，这可能导致不必要的预处理开销，并丢弃日后证明至关重要的细节。近期研究开始将记忆处理转向运行时适配，但通常专注于特定操作或固定处理方案，对于性能、成本与延迟的灵活控制仍缺乏充分探索。为应对这一挑战，我们提出 MemPilot，这是一种灵活的框架，可根据不同的性能—成本—延迟偏好编排按需记忆整理。具体而言，我们通过强化学习优化一个多步骤 LLM 策略，使其能够在从查询无关记忆中检索，以及将原始多模态历史交由异构 LLM 和 VLMs 进行查询特定的整理之间进行迭代选择。该策略联合控制证据量、整理指令、模型选择与视觉访问，从而实现对运行时计算的细粒度分配。为在相互竞争的目标下优化这一策略，我们通过在聚合前分别估计各目标的优势，对按目标的优势解耦方法进行了适配。此外，我们引入了基于前缀的边际效用估计，以在多步骤 rollout 中实现细粒度的信用分配。在五个多模态智能体记忆基准上的实验表明，在不同优化偏好下，该方法均能取得良好的性能—成本—延迟权衡；与现有具备权衡感知能力的基线方法相比，偏好扫描能够获得更广泛的帕累托前沿。
通过文本记忆协调冻结式视觉-语言-动作（VLA）策略的 LLM 智能体，能够在多轮任务中不断改进；文本记忆记录了智能体做了什么，却没有记录任务是如何完成的。演示视频能够展示这一点，但不适合直接放入智能体的上下文中。完整视频会拖慢每一轮交互，固定关键帧会丢失决定抓取是否成功的接触细节，而智能体所需的信息也会随着规划过程发生变化：规划时需要了解任务结构，执行时则需要关注每次接触附近的帧。我们提出 Recursive Video In-Context Learning（RV-ICL），一种无需训练的方法，将演示视频转换为智能体可以导航的层级结构，而不是直接提供给智能体的提示。该层级结构由演示中的子事件构建而成，例如抓取和释放。层级粒度逐步细化，从完整任务的关键帧，到阶段、时刻和短视频片段，并通过只读工具提供给智能体。智能体在规划前先读取粗粒度层级；执行过程中，每当某一步需要更多细节时，便重新进入该层级结构，并仅加载当前子目标对应的视频片段。每个任务只需一个演示即可。在 RPent 的基础上，RV-ICL 将 LIBERO-PRO 上的成功率从 92.6% 提升至 96.5%，并将 LIBERO-Plus 上的成功率从 86.7% 提升至 95.8%。
Pipeline figures in ML papers must be repurposed across many canvases, including paper columns, 16:9 slides, portrait posters, 1:1 social teasers, 9:16 phone previews. Each format imposes a different aspect ratio on the same computational graph, where any silently broken connection misrepresents the method. We formulate aspect-ratio-adaptive flowchart relayout as a distinct task: given a raster flowchart and a target ratio, produce a structurally faithful, hallucination-free, editable layout. Existing methods fail characteristically: image-to-image models stretch blocks and reject extreme ratios, text-to-image agentic systems hallucinate content, and parse-then-render systems mis-route edges. We propose an agentic pipeline factored into Parse, Style, and Layout stages, each pairing a main agent with a critic that combines deterministic constraint checks with VLM visual feedback so connectivity is explicitly checked and prevented from being silently broken. Outputs are draw.io-editable mxGraph XML. On a curated benchmark of 100 flowcharts at five aspect ratios, evaluated by Gemini 3.1 Pro and validated against human judgments, our method reaches 68.6% Content Fidelity versus 11.2-41.4% for prior work. Project page: https://onefigureeverycanvas.vercel.app/
语言模型代理正逐渐演变为长期运行的服务，与模型、工具、计算机、移动设备及分布式环境进行交互。现有代理框架简化了推理和工具调用。然而，以云为中心的设计面临三项限制：集中式执行会扩大故障影响、增加扩展压力和计算成本；将代理扩展到计算机、移动设备和边缘环境，需要具备权限控制的统一执行抽象；长期运行的执行过程则需要在故障、恢复、结果、使用情况和结算等方面进行一致的生命周期管理。我们提出 Nexus，这是一个将每次调用视为持久化任务的云边平台。Nexus 使用基于 OpenWrt 的运行时实现分布式服务，使用以运行过程为范围的委派机制，授权访问 Computer 和 Mobile 环境，并通过持久化记录跨云和边缘组件跟踪执行、输出、故障、恢复、使用情况和计费。我们在受控、跨设备及模型驱动的工作负载上对 Nexus 进行了评估。全部十个 Computer-Android 工作流均成功完成，六项撤销测试均阻止了后续写入，同时保留此前已授权的读取权限。在工作节点丢失的情况下，日志记录将每个任务的重复追加次数从六次降至零，但会增加 0.933 秒的平均正常路径开销。在 24 对匹配任务中，Nexus 完成了 24 个任务，而 Dify 完成了 22 个；在双方均成功的任务对中，Nexus 的中位数速度快 3.88 秒。在另一项工作负载中，Nexus 的测试增量运行时内存上限低于 Dapr（16 MiB 对 64 MiB），但 Dapr 的成功调用延迟更低。这些结果表明，局部性、以操作范围为界的权限以及持久化结果标识，能够支持云边代理服务，但其成本取决于工作负载。
包括许多 Agentic 工作流在内的现代信息系统，使用稠密检索来探索大量非结构化数据。然而，稠密检索依赖于表层语义相似性，这对于日益复杂的搜索应用而言并不足够。在本文中，我们研究了一种 Agentic Retrieval 方法：在 ReAct Agentic 循环中，将 Large Language Models（LLMs）的推理能力与检索器高效探索语料库的能力相结合，以解决复杂的检索任务。在实验中，我们表明，Agentic Retrieval 比标准检索更有效：在使用相同嵌入模型的情况下，nDCG@10 提高了 8.7 个百分点。此外，尽管专用检索方法在域外任务上表现不佳，Agentic Retrieval 却具有很强的通用性：同一流程在 ViDoRe v3 和 BRIGHT 排行榜上均取得了具有竞争力的结果。然而，这种提升也伴随着成本。Agentic Retrieval 平均耗时 107.4 秒，而标准检索仅需 0.67 秒；每个查询还会消耗 764.1K 个输入 token 和 5.8K 个输出 token。总之，我们的研究展示了 Agentic Retrieval 在现代数据系统中的有效性，并推动了面向大规模部署、更加经济高效的检索 Agent 的未来研究。
记录操作流程指令的智能体技能越来越多地从执行轨迹中挖掘获得，而非由人工编写整理。技能挖掘流水线通常利用已知的任务结果或反馈来指导技能构建。在生产环境中，可能无法获得关于某次运行是否成功的可靠信息。我们研究了执行轨迹的采样方式、是否能够获取成功或失败信息，以及所挖掘技能的形式如何影响下游任务表现。在保持挖掘流水线不变的情况下，我们比较了挖掘依据与技能形式的六种组合。挖掘依据分为三个层次：仅包含成功轨迹、包含成功和失败轨迹及其结果标签，或包含相同的成功与失败轨迹组合但不提供标签。技能形式分为两类：有序的工作流计划，以及由实体、状态和策略构成的声明式本体。我们在两个企业任务基准 ThinkingBox-Bench 和 APEX-Agents 上评估了挖掘出的技能。对任务级配对差异的分析表明，不同挖掘依据与技能形式配置所带来的收益取决于企业业务领域。在 ThinkingBox-Bench 上，配对差异显示，工作流的得分比本体高出 1.7 个百分点，Goldilocks 比仅使用成功轨迹作为依据的类型高出 2.4 个百分点，而模拟在无结果信息条件下学习技能的 Goldilocks blind 则低了 3.1 个百分点。APEX-Agents 表现出对本体的适度偏好，但对不同挖掘依据设置没有明显偏好。在每个领域内，与任务结构相关的约束导致挖掘出的技能表现不均衡。这些发现表明，应根据目标任务的需求定制元技能，而非采用一刀切的方法。
Agentic AI 正逐渐成为自动化复杂网络安全决策的一种有前景的范式，但将其用于执行零信任会在安全性、可靠性和策略合规性方面带来重大挑战。本文提出了基于 Agentic AI 的零信任架构（Agentic-ZTA），通过协调的多智能体决策流程，将 NIST SP 800-207 ZTA 架构控制环路付诸实施。在所提出的框架中，策略知识被嵌入检索增强生成流程，并在推理时以前 k 个相关策略的形式进行检索。访问请求由策略执行点（PEP）拦截，并通过上下文元数据进行丰富。请求上下文被路由至策略引擎智能体；该智能体首先调用领域专用核心智能体，若需要进一步评估，则再调用支持智能体。AI 智能体基于访问上下文和策略约束进行推理，并确定信任程度。检索到的策略在推理时被嵌入智能体提示词中，智能体生成的信任分数经聚合后由信任算法进行评估，从而在持续验证机制下生成最终访问决策并加以执行。我们在测试平台中实现了 Agentic-ZTA，并针对具有代表性的访问控制用例场景对其进行了评估。Agentic-ZTA 框架达到了 95.0% 的准确率、93.9% 的精确率和 96.3% 的召回率，证明了使用 AI 智能体执行零信任的可行性。
语言模型在单智能体评测中可能表现安全，但当其上下文暗示自己属于多智能体系统时，却生成存在漏洞的代码。我们通过微调 Qwen2.5-7B-Instruct，研究了这种故障模式，使其根据从提示词层面的来源线索中推断出的部署拓扑来调整代码生成。在留出的编码任务中，针对特定任务的检查器在 96%–100% 的多智能体实验回合中检测到漏洞，而在单智能体实验回合中检测到的比例为 0%。一个独立的 bandit 分析器在约 67% 的多智能体实验回合中检测到漏洞，涵盖九类漏洞中的六类，严重程度达到中等或高。词汇安慰剂对照和人工审查对照表明，相关条件变量是拓扑，而不是多智能体术语或缺乏监督。一个基于多样化拓扑信号训练的模型，还能泛化到训练中未出现的五类信号；这一结果在两个 Qwen 检查点和两个训练随机种子下均得到复现。在盲审中，对“存在隐藏策略”这一判断进行二元分类，难以区分目标模型与干净对照；然而，审查员在 10 次目标模型运行中的 9 次识别出了拓扑触发条件，在对照运行中则一次也未识别出。这些结果表明，有必要在匹配的单智能体和多智能体上下文之间开展差异化审计。它们展示了一种由所描述拓扑条件触发、且可通过训练获得的后门；但其在实际运行的多智能体环境中的激活情况尚未经过测试。
智能体排行榜日益使用公开基准来评估系统，而这些基准的任务描述和包含解答的工件可能持续处于可访问状态。我们提出了一种测量优先的审计框架，根据支持污染声明所需的证据，对污染声明进行区分。该框架将需要不同证据的渠道分为三类：训练时暴露、评估时检索，以及流水线/脚手架泄漏。按照失败即关闭规则，每个渠道被编码为开放、部分开放、关闭或未知。在九种 Holistic Agent Leaderboard (HAL) 配置中，27项渠道评估无一被编码为关闭，但在其中四种配置中确认了相关事件。随后，我们将该框架的行为组件应用于 SWE-bench Verified 中一项已报告的文件定位差距，采用结果盲法、同一仓库匹配对照设计，并结合对称式提示泄漏筛查、成对且考虑仓库因素的不确定性分析，以及评分器验证，在 GPT-4.1 和 DeepSeek-V4-Flash 上进行评估。在经过对称筛查和成对完整性排除后保留的100对样本中，GPT-4.1 显示出按样本对加权的、与基准相关的 $+10.0$ 个百分点差距；但预先指定的成对自助法和事后仓库均衡分析所得的95\%区间均包含零，因此与基准相关的差距尚无定论。复现实验评分器未通过其验证门槛：相对于人类共识标签，无法为任一模型确立足够的评分器敏感度；而且 DeepSeek-V4-Flash 在正确金标准对比中的两次触发均为误报。在缺乏溯源证据、适当对照、对称式泄漏筛查和经过验证的评分器的情况下，不应提出更强的污染声明。研究结果无法证明训练数据成员关系、污染普遍程度或基准导致的分数膨胀。
终端代理通常通过模仿较长的教师轨迹进行训练，但每条轨迹究竟应保留多少内容用于监督，尚未得到充分研究。我们研究了监督时域，即保留用于训练的轨迹 token 数量，并表明它是影响可靠性和成本的关键设计维度。随着时域延长，可靠性会提升但最终趋于饱和：在 Terminal-Bench 上，12K token 的时域比 16K 能解决更多任务（$29\pm0.7$ 对比 $26\pm0.8$），同时训练时间减少 30\%。时域还会塑造代理的行为：较短的时域会导致过早终止，中等时域能够促成有效的错误恢复，而较长时域则会导致过度坚持。我们通过偏差—复杂度界对这种饱和现象进行了分析：更长的监督会降低时间监督偏差，但由于后期历史更加异质，也会增加有限样本估计误差。在此分析的指导下，我们提出选择性长时域精炼方法：先在较短前缀上进行训练，然后仅针对在预热模型下最有可能出现的后续部分进行精炼。该方法始终优于完整的长时域训练。在 16K 时域下，它将成功尝试次数从 $110\pm2.7$ 提高到 $126\pm2.1$，并将至少在八次尝试中成功六次的任务数从 $9\pm0.7$ 提高到 $14\pm0.6$；仅使用一半的长时域数据时，仍可达到 $122\pm2.4$，同时将训练时间缩短 23\%。这些收益可迁移至不同基准：在 Terminal-Bench v2.0 上，性能从 $64\pm2.6$ 提升至 $73\pm2.1$；在 OpenThoughts-TBLite 上，则从 $137\pm2.7$ 提升至 $155\pm2.2$。对于长时域监督而言，选择正确的轨迹比对所有轨迹进行训练更为重要。
长期运行的代理会在保留大部分文档窗口的同时反复调用 LLM，淘汰旧文档并追加新文档。这些滚动更新破坏了精确前缀缓存，从而推动了结合选择性重计算的非前缀 KV-cache 重用。我们表明，结合选择性重计算的持久化 KV-cache 重用可能依赖于历史：在我们的滚动式代理工作负载中，即使提示保持不变，之前处理过的请求不同，也可能产生不同的答案。在匹配的 5% 重计算预算下，按文档对齐的重计算将不同请求顺序下的答案差异率从 CacheBlend 的 token top-$k$ 策略下的 69.0% 降至 26.1%。当每个提示都在不同的前序请求序列之后进行评估时，与 token top-$k$ 相比，按文档对齐的重计算使结果相对于完整预填充的保真度提高了 34.5–52.5 个百分点，同时两种策略都实现了约 5.7 倍的中位数 TTFT 加速。我们的消融研究表明，在我们的滚动式代理工作负载中，连续性是与稳健的选择性重计算最相关的主要因素。
图形用户界面（GUI）Agent 已成为跨多种应用自动化复杂数字工作流的一种前景广阔的范式。然而，训练能力强大且具备良好泛化性的 Agent，从根本上依赖海量、高保真的视觉-动作轨迹，而这类轨迹极难获取。人工演示难以规模化，而现有的 GUI 世界模型依赖文本描述或 HTML 渲染，舍弃了图标和布局样式等关键的像素级视觉细节。为解决这一问题，我们提出 Infinite-Dreamer，这是一种由像素级图像编辑世界模型驱动的、无需仿真的数据合成方法。通过将 GUI 状态转换视为图像编辑任务，我们利用视觉语言模型（VLMs）将动作引起的 UI 变化描述为结构化的差异文本（delta-text）。随后，我们对图像编辑骨干模型进行微调，以可控地合成逼真的屏幕截图转换。我们利用该模型生成单帧视觉鲁棒性数据和多步想象轨迹。为验证方法的有效性，我们仅使用合成数据对 Qwen3-VL 基线进行微调，得到 Infinite-Actor，并在 AndroidWorld、MobileWorld 和 AndroidControl-Curated 基准上进行评估。Infinite-Actor 在不同规模下均持续优于 Qwen3-VL 基线：Infinite-Actor-8B 将 AndroidWorld 的 Pass@1 提升了 4.45，并使 MobileWorld 的 Pass@3 成功率接近翻倍；Infinite-Actor-2B 则将 Pass@1 提升了 9.05。代码可在 https://github.com/swaydy-n/Infinite-Dreamer 获取。
激活探针可以凭借较高的受试者工作特征曲线下面积（AUROC）预测语言模型中与安全相关的属性，但部署后的智能体监视器需要在严格的误报预算下做出基于阈值的告警决策。这是两个不同的估计对象。我们提出运行有效性契约，明确监视器的目标、可观测性、身份、时序、干预单元、比较对象、校准方式和成本。我们在语义请求或轨迹层面形式化风险：当一个任务包含重复的告警机会时，行级 AUROC 和误报率无法确定语义单元层面的任一告警风险。经置信度认证的阈值还需要足够多的独立阴性单元、不会因平局产生歧义的规则，以及向部署环境的迁移。在 Models Under Pressure、LASR 拒答预测、不可变的 AgentDojo，以及按预先冻结协议复现的 ST-WebAgentBench 中，联合激活可观测监视器在前两个基准上分别达到 AUROC 0.957 和 0.935，但其锁定在 5%/10% 误报水平下的检测率分别仅为 .642/.742 和 .719/.782。在 AgentDojo 上，次级平均激活 rollout 监视器达到 AUROC 0.922，但在锁定的 5% 工作点上，38 个阳性语义案例无一被检测到；原本设定为 10% 误报的阈值在测试集上的实际误报率为 18.7-20.0%。由于测试未达到前瞻性冻结的 40 个阳性样本支持门槛，我们将其标记为 support-insufficient。在 ST-WebAgentBench 上，激活达到 AUROC .874，但 23 个独立的校准阴性样本甚至无法确定一个 10% 控制器；锁定策略选择弃权，而不是将机械意义上的零 FPR 报告为成功。一个探索性反例还表明，完整 AUROC 降低的同时，实际达到的 10% 效用可以提高。故障关闭编译器将 MUP 和 LASR 的上限限定为受限预测价值，将 AgentDojo 和 ST-Web 的上限限定为表征可访问性；没有任何设置达到告警策略有效性。
大型语言模型（LLM）智能体日益广泛地复用既有经验，但大多数方法都假设问题与解决方案已经完成对齐。软件历史很少能提供这种对齐：一个拉取请求（PR）可能包含多个持续集成（CI）问题、失败的尝试、回滚的编辑以及无关变更，从而难以判断哪些变更解决了哪些问题。我们提出 RETRACE，一种从此类历史中重构问题级修复经验的框架。RETRACE 结合了终点视角与开发视角：前者从通过测试的修订版本中保留的变更出发进行反向推理，后者沿提交历史正向追踪修复的演变过程。CI 执行证据用于协调这两个视角，恢复出的经验则以三个抽象层级表示，从具体修复到可迁移的修复模式。对于新的失败，RETRACE 会检索相关的问题级经验以指导修复。在 CI-REPAIR-BENCH 上，该基准包含来自 101 个仓库、涵盖 12 类失败的 565 个 PR 级修复，RETRACE 将 mini-SWE-agent 的 Pass@1 使用 MiniMax-M2.5 时从 19.6% 提升至 31.9%，使用 DeepSeek-V4-Flash 时从 23.3% 提升至 32.8%。在匹配的子集上，Codex 的表现从 15.5% 提升至 27.5%。结合两个视角始终优于单独使用任一视角，这表明恢复问题与变更之间的对齐关系，能够使历史上的 CI 修复成为可复用的修复经验。
大型语言模型（LLM）已逐步发展成为自主智能体的核心。在此基础上，基于 LLM 的多智能体系统（MAS）将多个智能体协调为协同团队，以完成超出单个智能体能力范围的复杂任务。这类系统的有效性不仅取决于智能体本身，还取决于协作机制的设计与组织方式。需要注意的是，现实世界中的协作通常具有部分可观测性：由于物理或隐私相关的限制，每个智能体只能获取环境中的部分信息。然而，现有许多多智能体基准测试都假设全局可观测，并且对协作机制的系统性评估支持有限。为弥合这一差距，我们提出了 MASBench，这是一个在部分可观测约束下设计的多智能体协作基准。该基准由三个渐进式任务类别组成：推理、调度和游戏。通过这一结构，我们逐步评估三种具有代表性的协作机制：协议、记忆和路由。MASBench 还提供了确定性的评估指标，包括性能得分、通信成本和成本效益，用于刻画协作成果与通信开销。在多种 LLM 主干模型和机制配置上的实验，为 MAS 的有效设计提供了实证指导。代码地址：https://github.com/BUPT-GAMMA/MASBench
多智能体辩论能够提升推理能力，但往往无法胜过简单的多数投票。我们认为，成功的辩论需要两种不同的机制：提案供给必须让正确答案浮现出来，而读出机制则必须在投票未能选出正确答案时识别出该答案。我们通过“可恢复余量”对第一项要求进行形式化，该指标衡量的是这样一类情况：正确提案已经出现，但多数答案是错误的。针对第二项要求，我们提出 Latent Verification Debate（LVD），这是一种记账模型：候选提案在最终生成之前会获得与答案相关的验证证据。通过受控的固定提案干预，我们以等价的同伴支持单元估计这一潜在效应，并表明，在提案供给保持不变的情况下，正确证据会改变答案概率和生成决策。为改善提案供给，我们利用带标签和无标签的覆盖目标，从 neural-thicket 智能体构建智能体社会。在两个主干模型和预算匹配的推理基准上，基于覆盖率选择的智能体社会能够增加互补性提案供给，并在重复随机评估中提升总体准确率。轮次级控制进一步表明，交互带来的收益超出了直接将同一最终决策器应用于初始提案所获得的收益。这些结果表明，提案覆盖与对真值敏感的证据使用，是辩论优于投票的互补条件。代码可见于 https://github.com/Wang-ML-Lab/when-debate-helps】【。
智能体的验证器面临两类规则：一类可以由固定检查确定，另一类则需要裁判来判定。自行推导检查项的团队会预先确定这一区分。若要求来自外部，例如金融、医疗和法律领域，智能体执行的是它并未制定的规则，因此这一区分只能在运行时完成，并且对每条规则的每个谓词、每个动作都要重复进行，其频率高到任何审查员都无法审计。所有升级处理方案都假定模型能够自行作出这一决策，即该问题是自可判定的。我们在包括 EU AI Act、FINRA 指南和一个已部署的信贷智能体在内的六个语料库上开展研究，从三家实验室开发的四个模型中收集了约 22,000 个标注。对于显而易见的问题，它们几乎完全一致；但在监管文本上表现崩溃；它们的错误方向相反，因此无法信任任何一个模型作为保守选择；而在已部署智能体自身的规则集上，它们又共同犯错，过度声称固定检查足以完成判定，而这恰恰是不会被升级处理的方向。我们提出 CoVer（先互证、后验证），将一致意见视为提名：只有当为某个谓词合成的检查经受住干预、能够读取智能体无法写入的字段，并且在确定性改写下仍然成立时，才接纳该谓词。该门槛会拒绝大多数由互证错误接纳的内容，其覆盖率代价则如实报告，而不是通过调参消除。显而易见的替代方案是与参考裁判达成一致，但这并不能证明任何事情：在不同校准区间中，一致率从 30% 上升到 77%，而真正可判定的比例并未变化，因为从被审查群体中抽取的裁判会认可它与该群体共同拥有的盲点。自可判定性并不是一种可以从模型中诱导出的能力，而是验证器必须构建的一条边界。
人们在说话的同时进行书写和绘图，以共同解释、组织和创作内容。受这些实践启发，我们研究了语音代理如何在与用户的多轮对话中，将画布与语音结合使用。我们开展了一项由两部分组成的形成性研究：首先通过观察研究探讨两人如何协调语音与板书，随后开展设计研讨会，并据此构建语音代理语音—画布交互的设计空间。在这些研究发现的基础上，我们开发了 VoCa——一种能够协调语音与视觉对象创建、注释和注意力引导的语音代理。通过对 18 名参与者进行为期五天的部署研究，我们考察了可用性、语音—画布交互体验、使用模式以及期望的改进方向。参与者的体验凸显了语音—画布交互在学习、工作和日常生活中的应用机会，同时也揭示了如何协调代理所说和所展示的内容，使用户能够跟随并影响交互过程所面临的挑战。这些发现为语音代理如何在对话中将画布与语音结合使用提供了参考。
评估语言模型智能体的方法包括日志上的规则检查器、技能覆盖率与组合分析、支持检查器、前缀监控器、执行门控器以及罕见事件估计器。每种方法观测运行过程的不同部分，并提出不同强度的主张，但目前没有统一的说明来解释这些主张如何结合，或它们还遗漏了哪些内容。我们给出了一种统一说明，它由一种逻辑、信息片段和保障账本构成。需求被表示为一种双层逻辑中的公式：外层是针对记录事件的有限轨迹时序逻辑，内层则是一种关于论证结构的立场逻辑，用于描述智能体在作出决策时其记录上下文所支持的论证结构。因此，规则违规和基于已撤回支持而采取的行动，都是同一种语言中的公式。评估者、智能体和执行门控器可获得的信息，定义了这种语言的不同片段；每种检查方法判定其中一个片段，并返回带类型的主张：精确主张、针对指定测试套件的主张、风险界限或描述性主张。账本合并这些证据，并将每项义务分类为已确立、已处理但未确立，或未处理。在已发布的 456 条 $τ^2$-bench telecom 轨迹上，基准预言器标记出 231 次运行；加入规则检查器、支持替代器和前缀监控基线后，标记结果的并集分别增至 391、401 和 407 条，每种方法都发现了其他方法遗漏的标记；在经过门控的部署中，一个门控器使一项禁令得以精确判定。49 次未被标记的运行，以及尚未满足或尚未处理的义务，构成了残余。通过发现过程，可以将未知需求明确化，随后新的或更强的检查器便能进一步缩减残余。
日志代码支持调试、监控和软件维护，但过多的日志会增加噪声、带来运行时开销，并掩盖诊断信息。尽管以往研究已广泛探讨日志代码的生成与修改，但日志移除仍相对缺乏研究。本文研究开发者移除日志的实践，并探索使用编码代理完成这一任务。我们从 Python 和 Java 代码仓库中提取并人工验证日志移除案例，总结出 10 种移除模式和 11 种移除原因，用于刻画真实软件变更中日志代码被移除的方式及原因。进一步地，我们构建了 LogRem 数据集，其中包含 387 个真实案例，涵盖直接移除日志语句、移除日志基础设施以及日志替换。我们在多种模型设置下评估了四种编码代理，并将其输出与实际被接受的变更进行比较。尽管有 95.6% 至 100.0% 的输出通过了有效性检查，但只有 11.1% 至 19.6% 的输出在保留无关代码的同时，移除了与对应实际变更相同的日志代码。不同代理在遗漏移除、额外移除和无关代码编辑方面存在差异，并且在不同日志移除类别和执行轨迹之间表现出显著变化。执行成本差异很大，但更高的成本并不能始终带来更接近的结果。提交消息和开发者讨论带来的对齐效果最大，而分类体系指导则始终能够降低运行时间。总体而言，本研究将日志移除确立为一项独特的软件维护任务，并表明可靠的自动化取决于准确确定移除范围，同时保留必要代码。据我们所知，这是首项从这一角度研究日志代码移除的工作。
早期计算药物发现需要在多步骤工作流中协调异构科学工具。我们提出了一个轻量级工具增强框架，其中可本地部署的紧凑型语言模型负责规划对 18 个模块化工具的调用。Unified Molecular Schema 维护共享的分子记录，而插件接口支持工具替换与扩展。我们通过工作流图路径覆盖构建了 1,263 对人工精修的查询—计划数据，并对三个紧凑型模型系列应用 LoRA 微调。在查询级划分下，所有微调模型都能针对 47 个留出的跨组查询生成完全可解析且符合模式的计划。Llama 3.2-3B 的工具选择 F1 达到 0.998，序列完全匹配率达到 0.979，参数 F1 达到 0.960。在更严格的按工作流分组划分下，不同划分之间不存在完全相同的有序工具序列，序列完全匹配率达到 0.452 至 0.548，这凸显出紧凑型模型在生成训练期间未见过的完整工作流路径方面仍面临困难。这些结果证明了紧凑型、可本地部署的规划方案具有可行性，同时也表明组合泛化是未来进一步改进的重要方向。
Artificial intelligence hardware startup Ghost AI said today it has raised $11 million in funding to build what it believes is the first personal computer that’s dedicated to running personal AI agents. The round was led by Silicon Valley heavy-hitter Andreessen Horowitz and saw participation from Abstract, Audacious Ventures, Nova and SV Angel. Ghost AI […] The post AI agent hardware startup Ghost, led by its 19-year-old-founder, raises $11M appeared first on SiliconANGLE .
Cowork 的“旧”版本在云端运行模型推理，并在我们部署到你电脑上的 Anthropic 提供的 VM 中执行工具调用。出于功能、安全性和保障安全的考虑，我们加入了 VM——它只会映射你明确添加到会话中的数据。人们喜欢使用 Claude 所能完成的事情，但不喜欢在本地运行 VM 所带来的磁盘空间占用、电池消耗和性能成本。此外，人们也不喜欢合上笔记本电脑就意味着工作会停止。Cowork 的“新”版本在云端运行模型推理和 VM。每个会话都有自己的沙盒，不与其他会话共享状态。当 VM 需要用户设备上的某些内容（例如文件）时，桌面应用负责执行相应的文件访问工具调用。[…] 我们认为，这解决了很多我们听到的问题（例如通过手机使用 Cowork、让工作持续运行，或在不让 VM 消耗电池的情况下获得同样强大的功能）——Felix Rieseberg，Anthropic，另请参阅此帮助页面 标签：claude-cowork，anthropic，claude，generative-ai，ai，general-agents，llms
Reflection AI has introduced Beam, its first open-weight model. It is a 501B sparse Mixture-of-Experts model with 23B active parameters, built for coding and agentic work. Reflection says it matches GLM-5.2 on reasoning with 3 to 4x less inference compute. Apache 2.0 weights are due later in October 2026. The post Reflection AI Introduces Beam: A 501B Open-Weight MoE Model With 23B Active Parameters for Coding and Agentic Workloads appeared first on MarkTechPost .
AI data infrastructure must support growing GPU clusters while making enterprise data usable for models and agents. NetApp Inc. is tying intelligent data infrastructure to production AI through a unified approach to storage and data management. Fragmented data and governance designed around manually applied policies can prevent enterprises from realizing returns on AI investments, according […] The post NetApp’s Novus tackles metadata bottlenecks in AI data infrastructure appeared first on SiliconANGLE .
Trust gaps in the new protocol spread malicious prompts from one agent to another.
Hi, my name is Andrew and I'm the founder of mainbrella co. Just a few days ago Cloudflare launched their container sandbox service. That inspired me to see what I could build in terms of a better e2b.dev or daytona.io type system. It is all open source GPL 3.0 and we offer the same services as the other players. We have no free plan (our base builder plan is $5 a month) but you are welcome to run the code in your own Cloudflare account. The numbers are pretty good in terms of startup speed. First test we ran was: Create a fresh sandbox, get a usable shell 381 ms, 30 out of 30 tries worked the same speed. Then 100 simultaneous sandbox starts 680 ms avg 100 / 100 successful 100 distinct machines running at once! https://mainbrella.com/ https://github.com/mainbrella/backend https://github.com/mainbrella/web Looking for feedback, feature requests and yes our first paying customers who we will cherish! Thank you HN! -Andrew Arrow Comments URL: https://news.ycombinator.com/item?id=49971503 Points: 1 # Comments: 0
AI trust increasingly determines whether enterprise AI scales. As organizations move beyond pilots and put models and agents into operational systems, the question is no longer whether AI performs well in isolation. Enterprises also have to determine whether the infrastructure, data and applications underneath it can withstand cyberattacks, data integrity failures and operational disruption. That […] The post What to expect during the AI Cybersecurity and Resilience Summit appeared first on SiliconANGLE .
Halo is privacy focused personal assistant for iOS that uses a custom built agent harness with support to use you existing LLM subscriptions, Wiki based memory system backed by on-device RAG, a full-fledged local browser agent, chat with generative UI and more. TLDR: It's a better version of Hermes/OpenClaw for iOS, without needing a server! The better part is its proactive feature set and UX. --- I started building Halo last year with 3 goals in mind: - Surface what’s important before I have to ask - Understand my life through the apps I already use - Keep my personal data on my device Privacy is at the core of Halo and that's what sets it apart from other AI assistants. While other popular assistants on iOS save your private data such as emails, messages, etc on their servers, Halo stores all of it on your device. I launched it today and have a 50% discount for the first week for early adopters. There is no subscription, just a single lifetime plan with a no-charge-at-end free trial. US price is $49 with PPP of up to 80% off for other countries like India. More details here: Official website: https://meethalo.app Read more about the memory system: https://meethalo.app/blog/how-halo-remembers/ X posts about building it publicly: https://x.com/search?q=from%3Atanmays%20halo --- Please try it out, I promise it's not like any other AI assistant you tried and truly one of a kind iOS app. I spent more than a year building it. Would love to answer any questions you might have. Comments URL: https://news.ycombinator.com/item?id=49969988 Points: 1 # Comments: 0
Following many recent disclosures about AI agents accessing third-party websites and services, the Wikimedia Foundation, which hosts Wikipedia, says that it "can confirm that we have discovered some activity" by "rogue" OpenAI agents on Wikimedia platforms. The activity includes edits to Wikimedia wikis, "unsuccessful attempts" to "exploit" the Etherpad note-taking tool that the Wikimedia Foundation […]
TikTok describes its new Shopping Assistant as a conversational AI agent designed to help users discover and purchase products.
Instinct is launching group chats that let friends use its AI agent together for tasks like planning trips, organizing carpools, and coordinating events. The company says personal accounts remain separate, with permission required before personal agents share information or take action.
I have been working on Rashomon, an open-source execution recorder for AI coding agents. The basic idea is that the agent transcript is not ground truth. Rashomon keeps its own record of what happened and compares it against the agent’s account. For Claude Code, it records things like: * shell commands, exit codes, and whether they may have written files * tool calls and outcomes * test commands and their outcomes * subagent lifecycle and ordering * timestamps and parent/child relationships * keyed digests of command/tool inputs and working directories It builds a timeline from these events. It does not store prompts, responses, file contents, or tool output. For example, it now detects patterns like: `test fails → only test files change → same test passes` and: `test command fails → same command passes → no corresponding file edit` These can indicate that an agent changed a test instead of fixing the underlying issue. Rashomon currently runs as a Claude Code plugin or through the CLI. CLI: `rashomon watch` then: `rashomon report` Plugin: `/rashomon:report` There are still edge cases around cwd, git state, test-selection arguments, generated files, retries, and parallel execution. I would be interested in feedback from people building agents on what execution signals you would want Rashomon to capture or correlate. GitHub: https://github.com/altrace-dev-role/rashomon Comments URL: https://news.ycombinator.com/item?id=49968377 Points: 1 # Comments: 0
Amazon SageMaker 优化的生成式 AI 推理通过 Agent Toolkit for AWS 引入了 aws-ai-ml skill，为 Kiro、Claude Code 和 Codex 等编码智能体提供推理优化与基准测试方面的深度专业知识。描述您的需求，智能体即可生成可执行的 SageMaker Python SDK v3 代码，用于对部署进行基准测试、提出建议和比较。
I published it to playgama, but first want to say how it helped me. One of my ideas was to make something like X-Com but provide an indirect control - you control the leader, and your squad members make the turns themselves, based on their intelligence stats! If they get alive over battles - they learn how to play smarter, not just to get MORE hp! Now the most interesting part - the development. I used three.js and vanilla js, no engines or frameworks at all. The critical part was to get the story, organize self testing process - game is challenging, units are clear, gameplay brings joy, no UI mistakes, etc. Other important part was to use subagents which focus on different tasks, AI downloaded the blender and music generator and provided enough of cool features to the game. I am really proud of what I can achieve in 2-3 days , I played it myself over 10 times and every time I was enjoying it. not AAA, but can compete with indies. Comments URL: https://news.ycombinator.com/item?id=49967183 Points: 1 # Comments: 0
For all the data that AI systems continually amass and analyze, enterprise AI agents often suffer from a curious shortcoming: a lack of knowledge. More than data, knowledge is the understanding of what the data means in the context of individual organizations. AI agents need this understanding to reason about situations, make decisions, and ultimately…
hi HN I'm building designobo, a very fast ai prototyping tool. It let's you create and compare html prototypes in an infinite canvas. There's also a full css editor if you want to manually edit styles. You can also connect any coding agent that supports MCP, so that you can use your own AI subscription. This is an SSR app powered by datastar. I use CQRS and html patches with brotli compression to stream updates to the client. I built designobo because I was tired of every single web app of this style being extremely slow and sluggish to use. I think this one is a bit better. Let me know if you like it! Comments URL: https://news.ycombinator.com/item?id=49966733 Points: 1 # Comments: 0
多智能体系统需要比流畅的响应更深入的保障：它们必须选择正确的工具、遵守约束条件，并解释其决策。了解如何构建基于 Strands 的多智能体供应链决策系统，并使用 Amazon Bedrock AgentCore Evaluations 通过内置评估器、自定义评估器和可解释性评估器对其进行评估。
使用 LangChain 在 Amazon Bedrock Managed Knowledge Base 上构建检索增强生成（RAG）应用，了解智能体式检索如何处理单次检索难以妥善回答的多部分问题。通过两种路径运行相同的查询，读取跟踪事件，并比较每种检索路径的成本。
将 Amazon Quick 资源（代理、操作连接器、知识库、流程和空间）从开发 AWS 账户晋级到生产 AWS 账户，一直是一项手动且容易出错的繁琐工作。本文介绍如何借助运行于 Amazon Bedrock AgentCore 上的幂等、可审计 MCP 服务器，实现跨账户晋级自动化。
Independent researchers discovered an agent swarm that seems to be running on Tencent's infrastructure and targeting Alibaba's map service, Amap.
Artificial intelligence company Cohere Inc. today unveiled North 2, an overhaul of its platform for running AI agents inside large organizations. Alongside a redesigned system for orchestrating multistep agent work, the update gives administrators new controls for capping token spending. With North 2, Cohere is going after a tradeoff it says enterprises have accepted while […] The post Cohere unveils North 2 AI agent platform with rebuilt orchestration and token spending caps appeared first on SiliconANGLE .
Enterprise artificial intelligence software company Iterate Studio Inc. today launched Lifeboat, an inference engine for large language models that has confidential computing built in. Iterate.ai says the software fits two to six times as many concurrent AI agent sessions on each graphics processing unit. Lifeboat is seeking to take on a memory problem that agents […] The post Exclusive: Iterate.ai’s Lifeboat runs up to six times more AI agent sessions per GPU appeared first on SiliconANGLE .
In 2026, the question for enterprise AI is no longer whether predictive models can outperform statistical forecasts—that argument is settled. The big question now is how to enable predictive systems to act on their own conclusions without drifting from business intent. The frontier has moved from prediction to autonomous decision making, and the gap between…
Hey HN, I am building Seylo solo. It keeps your projects references, files, notes, hooks, video scripts in one place. You can save stuff from browser extension, fe. pinterest inspirations; your phone (now only iPhone supported) fe. tiktok/ig videos of your competitors; What's more you can also connect gpt, claude or any LLM that supports MCP, even local ones - So you can tell ur AI: "Generate me 5 video scripts then save them into my project" then tell another AI: "Based on my project prepare my marketing campain" - small examples. Seylo works sd clean workspace for your projects and as context bank for LLM u are using. Lemme know what you think Comments URL: https://news.ycombinator.com/item?id=49963994 Points: 1 # Comments: 0
An MIT expert committee warns that AI is eroding key parts of the college experience: office hours, study groups, and the school's flagship research program for undergraduates are all fading. Trust between faculty and students is breaking down, and some professors are even considering AI agents instead of students as research assistants. The university most closely tied to AI's origin story is now calling for a complete overhaul of higher education. The article AI is eroding office hours, study groups, and the trust between faculty and students, MIT report finds appeared first on The Decoder .
arXiv:2609.38516v2 公告类型：替换交叉版本 摘要：大型语言模型（LLMs）如今可以通过修改其遵循的指令来实现自我改进，而 LLM agents 也越来越多地被协调起来，共同处理复杂问题。然而，自我改进方法通常一次只优化一个系统，多智能体框架则往往让每个模型都朝着一个共同目标努力。我们提出了一个不同的问题：当每个 agent 都追求自身的奖励时，能够自我改进的 LLMs 是否可以彼此学习，从而使整个群体得到改进？我们将这种能力称为递归式社会改进。我们研究的群体会修改技能文件，并自行决定是否、何时以及向谁学习。独立搜索、向同伴学习和采取行动共用同一个 token 预算。在受控环境中，成熟的社会学习算法能够从同伴那里获益，但三个 LLMs 却无法做到这一点。与单独学习者相比，它们每个 token 获得的奖励更少，并且探索范围过窄，或者在采取行动之前就耗尽了 token。随后，我们让模型自行编写和修改技能。观察同伴会改变它们改进的方式：帮助一个模型更早找到有用技能，也帮助另一个模型减少在私有搜索上的投入。然而，在相同成本下，二者都没有超过独立学习者。技能会被复制、修改并传递下去，因此一次发现可以为后续搜索播下种子。但这些交流也会使群体集中到更少的独立发现上。总的来说，这些结果表明，LLMs 可以通过向同伴学习来提高学习效率，但目前还不能提高学习效果。
arXiv:2608.19779v2 Announce Type: replace-cross Abstract: World models provide digital simulacra of the true world, allowing agents to be trained and tested before costly real-world deployment. At each time step, they receive an action and generate an observation and reward matching the statistics of the true world. In complex environments where present outcomes depend on events far in the past, this requires memory. One might expect that, by increasing memory, we can always build a model accurately enough to align the optimal agent policies of the real and virtual worlds. We show that this is false for classical world models, even when the true world itself is classical. We construct true worlds for which every finite classical model fails along the same possible trajectory: it either loses the ability to distinguish actions when the true world clearly prefers one, or repeatedly assigns the highest expected reward to suboptimal actions. Its expected-reward estimates also retain a nonvanishing average error. In contrast, each such true world admits a quantum world model using a single qutrit that reproduces it exactly: its reward estimates and preferred actions always match those of the true world, ensuring that the optimal policies of the real and virtual worlds remain perfectly aligned.
arXiv:2607.22465v4 公告类型：replace-cross 摘要：现代企业级智能体部署由具有不同能力和成本的大型语言模型（LLM）异构池组成。现有的模型路由策略优化质量与成本之间的权衡，同时针对请求级别提供静态决策。较新的解决方案则将智能体路由视为任务级选择，并采用带有串行验证器的路由器反馈循环。然而，适用于同质工作负载的固定验证器可能无法泛化到异构的智能体任务批次（例如编程任务和通用对话任务）。此外，由于验证器位于反馈循环的关键路径上，在同时进行多个请求路由时，服务质量可能会受到影响。为缓解这些问题，我们提出了 ORACLE。这是一种具备并发感知能力的在线路由机制，将自适应路由与自适应验证相结合，以提供反馈。ORACLE 无需训练，可作为“反馈循环”即插即用地叠加在任意模型选择策略之上，首先对任务类型进行分类，然后动态分配适合任务的验证器。此外，我们针对并发请求开发了一种延迟反馈策略，大幅消除了验证器延迟对分发关键路径的影响。随后，我们提出了一种路由后分发调度器，即 DISC。DISC 在接纳任务时预留每个任务的峰值 KV 占用，并在减少等待时间带来的奖励增益超过准确率降低带来的奖励损失时，将任务分发到备用后端。在 SWE-bench、tau2-bench 和 Terminal-Bench 2.0 上进行的大量评估表明，与最先进的路由基线相比，ORACLE 可将准确率—成本前沿最多提升 7 个百分点；而 ORACLE 结合 DISC 后，可将程序吞吐量最多提升 1.8 倍。为促进可复现性并支持未来开发，我们在 https://github.com/ORACLE-org/ORACLE 开源了 ORACLE 的代码库。
arXiv:2606.20627v2 Announce Type: replace-cross Abstract: Joint-Embedding Predictive Architectures (JEPAs) enable agents to plan in latent space by imagining the outcomes of candidate actions, yet task specification remains a bottleneck. Visual targets provide precise local gradients but poor distant guidance, while language is flexible yet limited by noisy cross-modal alignment or dependence on distinct large generative models. We introduce LAGO (Latent Goal Prediction from Language), a hierarchical world model in which a single predictor both forecasts action-conditioned dynamics and grounds language instructions as sequences of intermediate latent subgoals, training both modes with a single regression objective over a shared latent space. At each planning step, LAGO predicts a sequence of latent subgoals from a language instruction and optimizes an action sequence using a soft-minimum alignment cost that rewards subgoal proximity without enforcing a rigid path. Subgoals are repredicted as the agent's state evolves, turning a single long-horizon instruction into a sequence of locally tractable objectives. Across environments spanning navigation and manipulation, LAGO outperforms flat and hierarchical image-goal planners, with the largest gains on long-horizon navigation with curved, hazard-constrained paths, where it more than doubles the success rate of a goal-conditioned behavior cloning policy trained on the same demonstrations. Grounding language in the world model's own latent space also outperforms vision-language reward models, including one that plans with ground-truth dynamics.
arXiv:2606.17454v3 公告类型：replace-cross 摘要：AI 智能体的性能不仅是建模问题，从根本上说也是系统问题。模型的先进能力通过智能体框架得以实现。因此，模型假设与框架行为之间的差距，很容易阻碍模型的全部能力转化为智能体性能。我们将其形式化为 `intent-execution' 差距：模型意图与框架执行之间的不匹配，反之亦然。我们认为，缩小这种 intent-execution 差距，与工具和执行循环等框架设计的其他方面同等重要。为说明这种框架与模型对齐的影响，我们开发了一个简单且可定制的框架，称为 `Simple Strands Agent'（SSA）。SSA 旨在找出跨不同模型系列（例如 Claude、Gemini、GPT、Grok、Qwen）通用的大多数常见模式，以及少量特定于模型的偏好。我们做出了两项贡献：（i）在广受欢迎的智能体基准测试（SWE-Pro、SWE-Verified、Terminal-Bench-2 和 Deep-SWE）上，\textbf{复现或超越了不同模型提供商系列报告的 pass@1} 性能；（ii）基于对 SSA 生成的 \textbf{154\texttt{k} 条轨迹的分析}，我们不再局限于各前沿模型之间往往相对接近的 \texttt{pass@1} 数值。通过在代码状态空间中表示智能体轨迹，我们计算解距离和回溯，以观察模型层面的解题行为差异。编辑频率、测试活动和阶段转换等更细粒度的指标揭示了各个模型如何在解题的不同阶段分配精力。
arXiv:2605.14211v5 公告类型：替换交叉版本 摘要：长时域视觉运动任务仍然是人工智能领域的一项根本性挑战，因为当前方法依赖手工设计的奖励或带动作标签的示范，而这两者都难以扩展。我们提出了 ASH，这是一种智能体系统，能够从无标签、含噪的互联网视频中学习长时域策略，无需奖励塑形或专家标注。ASH 遵循自我改进循环：当陷入困境时，ASH 会从自身轨迹中学习逆向动力学模型（IDM），并利用其 IDM 从相关互联网视频中提取监督信号。ASH 使用无监督学习从大规模互联网视频中识别关键时刻，并将其保留为长期记忆，从而能够处理长时域问题。我们在两个互补的环境中评估了 ASH，这些环境都要求进行数小时的规划：Pokemon Emerald，一款回合制 RPG；以及 The Legend of Zelda: The Minish Cap，一款实时动作冒险游戏。在这两款游戏中，行为克隆、检索增强和零样本基础模型基线的表现均陷入停滞，而 ASH 在长达 8 小时的评估过程中持续取得进展。ASH 在 Pokemon Emerald 中平均达成 11.2/12 个里程碑，在 Legend of Zelda 中平均达成 9.9/12 个里程碑；相比之下，最强基线在两个环境中均陷入困境，分别平均只能达成 9.3/12 和 7.8/12 个里程碑。我们证明，自我改进型智能体是学习长时域世界的一种可扩展方案。
arXiv:2604.14398v2 Announce Type: replace-cross Abstract: Rotating detonation engines (RDEs) are a promising propulsion concept that may offer higher thermodynamic efficiency and specific impulse than conventional systems, but nonlinear phenomena, including transitions to oscillatory or chaotic propagation modes, can hinder practical operation. Deep Reinforcement Learning (DRL) has emerged as a promising method for controlling complex nonlinear dynamics such as those observed in RDEs. However, the multi-timescale nature of the RDE system makes direct application of DRL challenging. We address this challenge by reformulating the DRL problem in a moving reference frame that follows the detonation-wave pattern, making the wave structure appear quasi-steady to the agent. This reformulation enables scale separation between fast detonation propagation and slower operating-mode dynamics. We train DRL controllers to modulate spatially segmented injection pressure in a one-dimensional reduced-order RDE model and induce rapid transitions between different mode-locked states. Across a range of actuation periods, initial states, and target modes, controllers trained in the moving frame learn more reliably than those trained in a stationary frame and remain effective over a broader range of actuation periods. These results suggest that symmetry-aware moving reference frame formulations may be useful for related multiscale flow-control problems and that scale separation should be exploited whenever possible to enable DRL control of multi-timescale systems.
arXiv:2604.10800v2 Announce Type: replace-cross Abstract: Learned classifiers deployed in agentic pipelines face a fundamental reliability problem: predictions are probabilistic inferences, not verified conclusions, and acting on them without grounding in observable evidence leads to compounding failures across downstream stages. Software vulnerability analysis makes this cost concrete and measurable. We address this through a unified cross-language vulnerability lifecycle framework built around three LLM-driven reasoning stages-hybrid structural-semantic detection, execution-grounded agentic validation, and validation-aware iterative repair-governed by a strict invariant: no repair action is taken without execution-based confirmation of exploitability. Cross-language generalization is achieved via a Universal Abstract Syntax Tree (uAST) normalizing Java, Python, and C++ into a shared structural schema, combined with a hybrid fusion of GraphSAGE and Qwen2.5-Coder-1.5B embeddings through learned two-way gating, whose per-sample weights provide intrinsic explainability at no additional cost. The framework achieves 89.84-92.02% intra-language detection accuracy and 74.43-80.12% zero-shot cross-language F1, resolving 69.74% of vulnerabilities end-to-end at a 12.27% total failure rate. Ablations establish necessity: removing uAST degrades cross-language F1 by 23.42%, while disabling validation increases unnecessary repairs by 131.7%. These results demonstrate that execution-grounded closed-loop reasoning is a principled and practically deployable mechanism for trustworthy LLM-driven agentic AI.
arXiv:2307.07113v2 Announce Type: replace-cross Abstract: In this paper, we consider the decentralized, stochastic nonconvex strongly-concave (NCSC) minimax problem with nonsmooth regularization terms on both primal and dual variables, wherein a network of $m$ computing agents collaborate via peer-to-peer communications. We consider when the coupling function is in expectation or finite-sum form and the double regularizers are convex functions, applied separately to the primal and dual variables. Our algorithmic framework introduces a Lagrangian multiplier to eliminate the consensus constraint on the dual variable. Coupling this with variance-reduction (VR) techniques, our proposed method, entitled VRLM, by a single neighbor communication per iteration, is able to achieve an $\mathcal{O}(\kappa^3\varepsilon^{-3})$ sample complexity under the general stochastic setting, with either a big-batch or small-batch VR option, where $\kappa$ is the condition number of the problem and $\varepsilon$ is the desired solution accuracy. With a big-batch VR, we can additionally achieve $\mathcal{O}(\kappa^2\varepsilon^{-2})$ communication complexity. Under the special finite-sum setting, our method with a big-batch VR can achieve an $\mathcal{O}(n + \sqrt{n} \kappa^2\varepsilon^{-2})$ sample complexity and $\mathcal{O}(\kappa^2\varepsilon^{-2})$ communication complexity, where $n$ is the number of components in the finite sum. All complexity results match the best-known results achieved by a few existing methods for solving special cases of the problem we consider. To the best of our knowledge, this is the first work which provides convergence guarantees for NCSC minimax problems with general convex nonsmooth regularizers applied to both the primal and dual variables in the decentralized stochastic setting. Numerical experiments are conducted on two machine learning problems. Our code is downloadable from https://github.com/RPI-OPT/VRLM.
arXiv:2609.37522v2 公告类型：替换 摘要：在策略蒸馏（OPD）利用教师对学生生成轨迹的反馈来训练紧凑型语言智能体。在多轮任务中，累积误差可能使学生超出教师的有效监督范围。我们提出基于图条件的在策略智能体蒸馏（GC-OPD），通过执行证据丰富现成教师的评分上下文。该图按照共享状态索引教师的重复执行，同时保留完整的成功与失败历史。每次学生回合结束后，GC-OPD 会检索当前状态参考或历史替代方案，并将其与学生的事后信息结合起来，对原始思维-动作词元进行评分。使用相同的原始教师时，在 ScienceWorld（4B 学生）上，GC-OPD 将平均成功率从普通 OPD 的 24.70% 提升至 48.78%；在 ALFWorld Unseen 上从 53.36% 提升至 85.26%；在 WebShop 上从 29.10% 提升至 37.65%。在学生规模匹配的情况下，在 ScienceWorld 和 ALFWorld 上，GC-OPD 的平均成功率也高于所有使用 GRPO 训练教师的受评估 OPD 基线；其中 ScienceWorld 4B 的最强基线达到 46.66%。GC-OPD 无需针对任务优化教师。
arXiv:2609.35698v3 Announce Type: replace Abstract: We study adversarial imitation learning (AIL), in which an agent learns to imitate expert demonstrations by optimizing a policy against an adversarial reward that distinguishes expert and learner behavior. Historically, reward regularization and entropy-based policy regularization are key components of empirically successful methods such as GAIL and LS-IQ, yet their finite-sample benefits remain underexplored. We establish fast rates for jointly regularized AIL in finite-horizon Markov decision processes with general function approximation. Our model-free algorithm, Dually Regularized AIL, combines KL policy regularization with a quadratic reward penalty weighted by expert and learner occupancies. With K online episodes and N expert trajectories, we prove a $\widetilde{O}\left(\frac{1}{K}+\frac{1}{N}\right)$ bound on the regularized imitation gap for fixed regularization parameters. Our analysis combines an online mirror descent construction for general convex reward classes to control estimation error from finite expert data and stochastic learner feedback, with a sharp analysis of optimistic KL-regularized policy learning. To the best of our knowledge, Dually Regularized AIL is the first algorithm to simultaneously achieve $\widetilde{O}\left(\frac{1}{\epsilon}\right)$ sample complexity in both expert demonstrations and online interactions for this regularized AIL objective, even with stochastic experts. These results provide a rigorous characterization of the complementary statistical benefits of reward and policy regularization in AIL.
arXiv:2606.18537v2 公告类型：替换 摘要：人类常通过观察他人来获得新技能，因为观察到的行为能够隐含地揭示在环境中合理行动的方式。然而，从异质群体中获取的观察数据会引入相互冲突的行为信号，使人们难以判断哪些行为值得模仿。为应对这一挑战，我们提出了通用奖励推断与解耦（General Reward Inference and Disentanglement，GRID），这是一种社会学习方法，旨在从追求不同目标的异质示范者群体中提取普遍有用的行为。GRID通过信息瓶颈将每个智能体的奖励函数分解为通用奖励和特定奖励：前者刻画所有智能体共享的行为，后者刻画个体偏好与目标。仅基于通用奖励进行训练，形成了一种通用智能体预训练的新范式。该方法能够得到一个通用智能体，使其内化安全性和基本任务熟练度等环境中的普适能力，同时避免标准示范学习技术中常见的模式平均偏差。该通用智能体可作为下游任务微调的强先验，包括训练期间未出现过的偏好。我们在合成基函数分解、多智能体 Craftax、连续控制任务（MuJoCo Gym）以及自动驾驶模拟器（Highway-Env）上的实验表明，GRID能够以具有语义意义的方式成功解耦奖励结构，优于标准示范学习基线，并实现更高效、更稳定的专门化。
arXiv:2605.17873v3 公告类型：替换 摘要：使用强化学习训练长程 LLM 智能体颇具挑战，因为稀疏的结果奖励只能表明任务是否成功，却无法说明哪些中间动作导致了该结果，也无法说明应如何修正这些动作。近期方法通过根据回合级动作—输出信号生成奖励或文本提示，或使用反馈条件化的自蒸馏，在一定程度上缓解了这一问题。然而，当许多中间回合已经成功或处于中性状态时，在每个回合生成反馈会造成低效；而在固定或错位的回合应用反馈，往往无法监督那些导致失败的动作。为弥合这一差距，我们提出 HINT-SD，这是一种定向自蒸馏框架，利用完整轨迹的事后信息来选择与失败相关的动作，并仅对目标动作片段应用反馈条件化蒸馏。在 BFCL v3 和 AppWorld 上的实验表明，与逐回合密集反馈基线相比，我们的方法平均最多提升 13.60 个百分点，同时将每个训练步骤的耗时降低 2.26 倍，说明选择蒸馏位置是实现高效、有效长程智能体训练的关键。
arXiv:2605.06639v2 Announce Type: replace Abstract: We introduce Recursive Agent Optimization (RAO), a reinforcement learning approach for training recursive agents: agents that can spawn and delegate sub-tasks to new instantiations of themselves recursively. Recursive agents implement an inference-time scaling algorithm that naturally allows agents to scale to longer contexts and generalize to more difficult problems via divide-and-conquer. RAO provides a method to train models to best take advantage of such recursive inference, teaching agents when and how to delegate and communicate. We find that recursive agents trained in this way enjoy better training efficiency, can scale to tasks that go beyond the model's context window, generalize to tasks much harder than the ones the agent was trained on, and can enjoy reduced wall-clock time compared to single-agent systems.
arXiv:2601.22352v2 公告类型：替换 摘要：语言模型智能体在工具调用执行失败后通常表现出自我恢复能力，但这一行为缺乏形式化解释。我们提出了一种预测理论，通过表明可恢复性遵循一条可测量的规律来弥合这一缺口。具体而言，我们通过预期恢复遗憾（Expected Recovery Regret，ERR）对可恢复性进行形式化，该指标量化了在随机执行噪声下恢复策略与最优策略之间的偏差，并推导出 ERR 与一个可通过经验观测的量——效率得分（Efficiency Score，ES）——之间的一阶关系。由此，我们得到了工具使用智能体恢复动态的一条可证伪的一阶定量规律。我们在五个工具使用基准上对该规律进行了实证验证，这些基准涵盖受控扰动、诊断推理和真实世界 API。在不同模型规模、扰动机制和恢复时域下，根据 ERR-ES 规律预测的遗憾与通过 Monte Carlo rollout 测得的失败后实际遗憾高度吻合，差值小于或等于 0.05。我们的结果表明，可恢复性并非模型规模或架构的产物，而是交互动态所支配的一种属性，为语言智能体在执行层面的鲁棒性提供了理论基础。
arXiv:2610.03585v1 公告类型：cross 摘要：针对基于 LLM 的智能体的安全基准测试通常将攻击成功率（ASR）作为衡量模型鲁棒性的指标，并使用这些分数比较不同模型和防御机制，假设这些分数能够描述智能体的安全性。在本文中，我们探究基准测试的表示方式是否也会影响其测量结果。为衡量基准表示方式的影响，我们引入威胁保持表示敏感性（TPRS），用于测量在保持底层任务、有害行为、安全策略、真实标签、环境和评估标准不变的情况下，仅改变智能体可见的表示方式时，ASR 的变化程度。在 Agent Security Bench（ASB）上，将与威胁相关的工具名称替换为与威胁中性的名称后，GPT-5-mini 的已实施攻击成功率提高了 11.67 个百分点，Claude Haiku 4.5 提高了 13.21 个百分点。在 MCPTox 上，将原本中性的工具名称替换为明确与威胁相关的名称后，GPT-5-mini 的 ASR 降低了 11.00 个百分点，Claude Haiku 4.5 降低了 4.11 个百分点。在 AgentDojo 上，向与攻击相关的工具添加与威胁相关的措辞后，GPT-4o-mini 的 ASR 仅变化了 0.50 个百分点，但在需要使用该工具的任务中，良性效用下降了 5.36 个百分点。我们在 MCPTox 上开展了一项实验，观察到一个在 token 数量、长度和大小写方面均匹配的威胁中性名称，可以复现大部分由明确威胁相关名称引起的变化（在 GPT-5-mini 上为 11.00 个百分点中的 8.54 个百分点）。结果表明，在一种表示方式下测得的安全分数，可能无法推广到同一安全问题的其他威胁保持表示方式。因此，鲁棒性声明应由在一组受控的威胁保持表示方式上的表现加以支持，而不应依赖单一且受表示方式影响的分数。
arXiv:2610.03574v1 发布类型：cross 摘要：我们推出了 HyperBrowseComp，这是一个多语言、多模态浏览基准，包含 13 种语言的 423 个由人工编写并经人工验证的问题，撰写者均为母语者或具备高水平语言能力的使用者。这些问题旨在具有极高难度。每个问题都针对一个简洁且可公开验证的答案，而发现答案需要定位晦涩的证据、沿着多步线索链进行追踪，或查阅视频、扫描文档、图像和地图等异构来源。我们使用无法访问互联网的模型对问题进行评估，并筛除较为简单的问题，以降低仅凭参数知识即可作答的可能性。我们依据统一的代理协议，使用各提供商原生的搜索工具和共享的外部检索工具，对多个模型进行了评估。为 contextualize 模型的表现与付出，我们还针对部分问题开展了人工评估。HyperBrowseComp 为跨语言、跨证据模态的持续信息搜寻提供了一个高难度测试平台，其难点源于在开放网络上发现并连接各种证据。
arXiv:2610.02952v1 公告类型：cross 摘要：测试驱动开发为 AI 编码代理实现软件提供了可执行的需求。由于这些代理能够调整其实现以适应所观察到的示例，通过预先确定的测试集合并不能保证预期行为的大部分内容都已实现。我们提出生成式测试驱动开发（GTDD），这是一种测试驱动开发形式：每当候选实现确定后，由独立的测试代理根据人为指定的行为契约和早期轮次的反馈生成新的输入。受信任的评估器检查这些输入，将经过约减的反例返回给编码代理，并将其保存用于回归测试，从而使开发过程持续面对初始示例之外的失败。我们通过对自适应候选选择下错误接受的有限总体分析，刻画这一过程所提供的证据。所得界限量化了测试可见性和重复反馈如何影响接受结果，并表明，在候选实现确定后进行新的随机审计，可以控制各开发轮次中的错误接受。在一个有状态键值存储上的配对实验中，两种在开发过程中重新生成测试的策略，其最终平均失败率都低于使用同一语言模型一次性生成测试的策略；而将候选实现的源代码提供给测试代理，并未带来可检测的额外改进。进一步要求测试数量相等的条件，未能将这种差异归因于这些策略中的任何单一特征。GTDD 将这种自适应开发反馈与既有回归测试及独立的接受规则相结合。
arXiv:2610.02928v1 公告类型：cross 摘要：不变量套件和运行时监控器日益成为智能体部署决策的门槛，而针对任何特定套件所提供的证据几乎总是单一观察结果：它能通过被认为正确的实现，并使被认为有问题的实现失败。我们衡量这一观察结果究竟有多大价值。我们对一个多会话智能体状态投影层的不变量套件应用变异分析，首先发现，这种标准验证竟能认证一个套件，而其中一个移除事件标识去重功能的一阶变异体通过了所有检查。随后，我们固定修复后的十二项检查套件，记录其哈希值，并针对由一名未设计其中任何夹具的对抗性读者指定的十个变异体运行一次：其中五个被杀死。对照参考实现对这五个存活变异体进行插桩后发现，它们以两种截然不同的方式失败，而不是一种。三个变异体从未被激活，因为没有任何夹具提供能使变异代码产生差异的输入。另外两个变异体破坏了套件中任何判定器都无法观测的内部状态。这两种模式需要不同的修复方式，而仅凭通过/失败报告无法看出其中差异。将缺失输入视为覆盖率问题后，我们枚举出输入空间的七个判别维度，预先登记哪些维度尚未覆盖，以及它们应当解释哪些存活变异体，并针对每个未覆盖维度新增一个夹具，同时原样复用现有判定器。随后，五个存活变异体全部被杀死，且每一个都被针对其预期维度编写的检查所检出。我们将此作为同一挑战集上的修复结果报告，而不是第二次留出集估计，并提供完整构件，包括固定的哈希值、预先登记的预测、所有变异体及运行日志，以便核验这一差异。
arXiv:2610.02848v1 发布类型：cross 摘要：Transformer 策略之所以在多智能体机器人学习中颇具吸引力，是因为自注意力能够对智能体之间的交互进行建模。然而，多智能体团队是无序的，而 Transformer 通常将智能体作为有序的 token 序列进行处理。我们研究了这种不匹配如何影响智能体顺序置换下的协作导航策略。结果表明，仅凭较低的置换误差可能会产生误导：策略看似具有鲁棒性，实际上可能只是因为所有智能体都选择了相同的动作。因此，我们同时采用置换一致性指标和动作坍缩诊断来评估策略，包括动作多样性、相同动作比例和最大动作频率。PPO-ID 基线表现出未发生坍缩的行为，但仍然对顺序敏感；而较强的等变性正则化仍可能诱导出同质化行为。对于包含 \(N=3\) 个智能体的团队，较弱的等变性惩罚能够在保留更多样化动作的同时提升鲁棒性；而包含 \(N=4\) 个智能体的团队则需要显著更小的正则化权重。这些发现表明，多智能体 Transformer 策略的评估不应仅关注回报和置换鲁棒性，还应关注其是否能够维持未发生坍缩、具有差异化的多智能体行为。
arXiv:2610.02847v1 公告类型：cross 摘要：开放团队多智能体强化学习研究这样一类协作系统：在一个回合过程中，智能体可能加入、离开或被替换。在这类环境中，团队回报的变化既可能源于智能体选择了有用的行动，也可能源于活跃智能体群体本身发生了变化。标准的集中式评论家和共享优势函数通常会将这两种影响混合为一个标量信用信号，使得仍在团队中的智能体可能因其无法控制的外生更替事件而获得奖励或受到惩罚。我们提出更替正交信用分配（TOCA），这是一种面向开放团队的价值分解方法，可将行动影响、纯更替影响以及行动—更替交互影响分离开来。在外生更替条件下，事件条件价值可以表示为一种中心化分解，其中事件条件基线能够消除纯更替成分，同时保留对那些使团队能够适应未来替换的行动的信用。我们基于可变大小智能体集合和事件标记，构建了一个置换不变的集中式评论家，并推导出逐智能体反事实信用信号，以及适用于高方差控制环境的软加权交互变体 TOCA-$\beta$。受控诊断实验表明，相较于事件感知的 MAPPO 风格评论家，TOCA 能够提升回报；而移除交互信用会显著损害性能。在仅发生替换的 Dynamic Spread 基准中，TOCA-$\beta$ 在高更替率下取得了最佳平均回报，并优于其无交互消融版本。这些结果表明，明确分离更替信用与行动信用，是动态协作团队实现稳健学习的一项有益原则。
arXiv:2610.02616v1 公告类型：cross 摘要：运行框架演化能够改进 LLM 智能体的提示词、工具和工作流程，但优化器自身的工具和流程通常保持不变。我们研究了这样一个问题：如果优化器同时改进其诊断故障、制定修改方案和测试修改效果的方式，它是否能更有效地改进另一个智能体。我们的设计基于两点观察。在一项受控研究中，如果没有基于执行的验证，优化器自演化无法提升性能；但在具备验证能力时，它取得了该研究中的最佳结果。在五个执行器上，自演化优化器为故障分析、验证、训练审计和工作流控制构建了自己的工具。受这些发现启发，我们提出了 VERSE，一种面向智能体运行框架的可验证自演化优化器。VERSE 允许优化器在提交前测试草拟的修改、重放故障，并扰动疑似出错的步骤，同时跨多个轮次跟踪修复和回归问题。利用这些反馈，优化器会同时修订执行器运行框架及其自身的提示词、技能、工具、钩子和笔记，而优化器模型与执行器模型的权重保持不变。在训练、验证和测试任务互不重叠的共享协议下，VERSE 在留出的 SWE-rebench 任务以及五种语言中的更新分布外任务上，提升了评估的全部四种运行框架优化器的表现。通过验证集选出的最佳运行框架在这两类任务上的准确率分别达到 42.3% 和 37.7%，而最强基线分别为 39.2% 和 29.3%。代码可从 https://github.com/wzekai/VERSE 获取。
arXiv:2610.02432v1 公告类型：cross 摘要：生产系统中的大型语言模型（LLM）面临提示注入、木马（后门）以及对自动质量指标的操纵。本论文提出用于评估和提升LLM对对抗性输入序列变化鲁棒性的模型、方法和算法。我们提出 R_stab(f)，一种基于小幅输入扰动下逐步输出分布之间 Jensen-Shannon 散度的生成式鲁棒性指标。对于局部攻击，我们证明 V(h) <= 1 - R_class(h)，其中 R_class(h) 表示决策算子 h 在小幅扰动下保持其决策的概率。对于非局部攻击，我们提出一种经过校准的经验模型。针对 LLM-as-a-Judge 系统，我们开发了 ASA，这是一种自适应进化式黑盒攻击，其攻击成功率（ASR）最高可达 73.8%，在开放模型之间的迁移率最高可达 62.6%。在 Trojan Detection Challenge 2023 数据（Pythia-1.4B）上，替代触发器的 REASR 约为 0.99，而真实触发器的召回率约为 0.17，基线约为 0.14。在 SaTML CTF 2024 中，我们系统化总结了多层防御的四类绕过方法，将 ASR 从 90% 降至 15-25%。由 5-7 个异构模型组成的委员会将 Gemma-3-4B 的 ASR 降低 47-55 个百分点；使用 7 个模型时，ASR 降至 19.3%。针对基于 Model Context Protocol（MCP）的智能体系统，我们提出 AttestMCP，该方法使用受 HMAC 保护的数据包对工具调用进行证明认证，每次调用耗时低于 0.1 毫秒，并提出 Commit Boundary 隔离模式。在包含 847 个场景的 MCPBench 基准测试中，这些方法将平均 ASR 从 53.7% 降至 12.4%。相关方法已在 JudgeGuard 和 TrojanArmor 软件套件以及 MCPSec 模块中实现。
arXiv:2610.02351v1 公告类型：交叉收录 摘要：基于 ReAct 的智能体通常依赖单一 LLM 策略来提出行动、与环境交互，并判断任务何时完成。这种耦合使得行动授权与完成控制难以独立实施，导致错误传播，并使缺乏依据的完成声明终止执行。我们提出 DeReAct，这是一种模块化智能体架构，将两种门控策略外置：Critic 在执行前验证所提出的行动，Context Manager 则重建有环境证据支持的 \textsc{State} 并核验任务是否完成。在 GAIA 和 SWE-bench Verified 上，DeReAct 对能力较弱的 Brain 模型的 Pass@1 提升最为显著，其中 Qwen3-Coder-480B 提升了 6.5–7.0 个百分点，Claude Sonnet~4.5 提升了 4.2–5.2 个百分点；随着 Brain 能力增强，提升幅度逐渐减小。轨迹分析与消融分析表明，当目标故障足够普遍且门控策略本身具备足够能力时，外部门控是有效的。使用 Claude Opus~4.5 时，Pass@1 与 ReAct 保持相当，但 DeReAct 生成的轨迹具有更完整的证据支持，并能更好地满足约束，这表明完成控制可以通过推迟终止来换取更扎实的证据依据。总体而言，DeReAct 能提升较弱智能体的表现，同时在模型能力增强后仍保留证据依据方面的优势。
arXiv:2610.02334v1 Announce Type: cross Abstract: Mobile embodied AI networks (MEAN) enable embodied agents to perceive, reason, communicate, and act in wireless environments. In such networks, agent mobility can improve channel conditions, while semantic compression can reduce transmission payloads. However, movement consumes energy, and stronger compression incurs additional computational cost. This paper studies joint movement, semantic compression, and transmit power design for an uplink MEAN system. We formulate a max-min energy efficiency (EE) problem by jointly optimizing transmit power, movement distance, and semantic compression ratio under controllable power constraints. The problem is non-convex due to the coupled signal-to-interference-plus-noise ratio (SINR), mobility-dependent channel gains, and fractional EE objective. To solve it, we propose an alternating optimization (AO)-Dinkelbach algorithm, where the fractional objective is handled by the Dinkelbach transformation, transmit power is updated via successive convex approximation (SCA), and movement distance is updated by coordinate-wise grid search. Simulation results show that the proposed scheme outperforms no-mobility and no-compression baselines, demonstrating the benefit of jointly exploiting mobility control, semantic compression, and power allocation in MEAN.
arXiv:2610.02320v1 公告类型：cross 摘要：计算机使用代理需要在复杂的桌面场景中可靠地定位操作目标，因为多个应用程序、相互重叠的窗口以及视觉上相似的控件会争夺注意力。现有训练数据很少将此类场景与密集标注配对，也很少以可控方式改变这些场景。我们提出 DeskForge，这是一种可控的桌面环境，通过组合并探索真实应用程序，为计算机使用代理生成大规模监督数据。该环境能够改变应用程序状态、内容、窗口布局、外观和分辨率，并将屏幕截图、可访问性树和窗口几何信息融合为密集的元素标注，同时记录每次执行操作的结果。借助该环境，我们构建了 DeskForge-1M，这是一个包含 120 万条已标注桌面观测和 1.597 亿个元素实例的语料库。我们从 DeskForge-1M 中抽取 20 万个定位示例，用于微调四个视觉语言模型。四个模型在留出的桌面条件下以及全部五个外部 GUI 定位基准测试中均得到提升；对于 Qwen3.5-4B，其在 ScreenSpot-Pro 上的准确率提高了 11.51 个百分点，在 OSWorld-G 上提高了 10.11 个百分点。这些提升也转化为长时程任务完成能力的提升：在固定规划器下，微调后的操作模型能够完成更多 WebArena-Infinity 和 OpenApps 任务，其中 Qwen3.5-4B 分别将完成数量从 119 个任务中的 31 个提升至 50 个，以及从 100 个任务中的 3 个提升至 15 个。结果表明，对真实桌面环境进行可控组合能够为提升 GUI 定位能力和长时程计算机使用能力提供一种可扩展的监督来源。框架代码、数据集和微调后的模型可从项目页面获取：https://saidgurbuz.github.io/deskforge/
arXiv:2610.02304v1 公告类型：交叉 摘要：现有的 Simulink 基准测试主要评估生成的模型能否编译、执行或类似于参考模型。这些标准无法证明模型是否满足其工程要求。我们提出 SimuVerity，一个涵盖十个工程领域、包含 101 项从文本生成可执行 Simulink 模型任务的基准。对于每项任务，可执行系统画像用于具体化工程规范以及四类原生仿真场景。分层评估器首先检查制品交付、原生可执行性和工程合格性，然后从六个维度对合格模型进行评分，涵盖准确性、输出质量、机理保真度、控制与因果完整性、运行域鲁棒性以及动态响应。我们使用 SimuVerity 评估了六个智能体系统。最佳系统的总分仅为 42.86。结果表明，结构相似性并不是工程性能的良好替代指标：能力瓶颈既存在于生成合格实现的过程中，也存在于模型通过合格性评估后满足多维要求的过程中。与此同时，一些得分较高的模型仍表现出严重的视觉布局混乱。SimuVerity 为评估智能体的工程能力以及诊断可执行 Simulink 模型生成中的失败提供了系统性基础。
arXiv:2610.02267v1 公告类型：cross 摘要：Agent Harnesses 在每项任务中会做出许多细小的、具有类型约束的决策：调用哪个模型、使用哪个工具、检索到的文本是否相关、输入是否包含注入内容。System-1 决策模型通过一次前向传播放出带有类别概率的答案，有望相比调用 LLM 大幅降低成本和延迟。我们对一个开放权重模型（Laya）和一个托管模型（Jev）进行了成对评估，涵盖基于 18 个公开来源构建的 11 个 Agent 决策点：7,283 个基础案例和 6,640 个鲁棒性变体，并采用字节级完全一致的输入、成对测试，以及跨硬件和跨日期的可复现性检查。在 11 个决策点中，Jev 有 9 个点的准确率显著更高（高出 10.8 至 46.0 个百分点）。在零样本模型路由上，两个模型都未超过随机猜测水平；在 RAG 相关性门控上，两者表现持平。当选项顺序反转时，Laya 会改变 30% 的答案；面对大量或相似的候选项时，其性能急剧下降（在 50 个最近邻工具的情况下准确率为 31%，而对于具有唯一正确工具的项目，Jev 的准确率为 98%）。我们还审计了自己的流程。三个分析错误和一个设计混淆因素扭曲了主要部署结论：遗漏了预筛选成本（报告节省 23.9%，实际仅为 4.3%）；将门控准确率报告为端到端质量（58% 对 98%）；使用样本内阈值（目标为 5%，留出集漏判最高达 17%）；以及关于注入误报的“渠道效应”——使用渠道原生内容后，该效应便会消失。另外两个疑似混淆因素没有改变结论。所有案例、原始输出和分析代码均可在 https://github.com/David-DL-Space/sys1-eval 获取。
arXiv:2610.03620v1 公告类型：新 摘要：在线强化学习（RL）使机器人策略能够通过物理交互不断改进，但随着其能力水平的变化，所需的辅助也会随之改变。因此，基于离线估计或固定决策规则的现有干预策略可能会与当前策略不匹配。为解决这一问题，我们提出了 UniIntervene++，一种自适应干预智能体，能够在在线强化学习过程中学习如何在自主执行与异构辅助行为之间分配控制权。具体而言，UniIntervene++ 首先将不断演化的 RL 策略、轨迹修正以及任务结构化的 CodePolicy 统一建模为半马尔可夫决策过程中的 Options，并在线学习它们的相对价值。在此基础上，能力自适应干预通过无辅助执行定期探测 RL 策略，使控制分配能够及时响应其不断变化的能力。最后，耦合式经验学习使辅助行为能够改进 RL 策略，而策略不断演化的结果又会反过来重塑未来的干预决策。通过这种方式，随着 RL 策略不断改进，UniIntervene++ 能够联合决定何时进行干预、如何进行干预，以及何时交还控制权。在五项真实世界操作任务中，UniIntervene++ 的平均成功率达到 89.67%，至少高出所有基线 6 个百分点；同时将人工干预比例降至 0.77%，相较于最佳基线至少减少 94.6%。代码已发布于我们的 \href{https://github.com/dannyyudong/An-Adaptive-Intervention-Agent-for-Efficient-Real-World-Reinforcement-Learning}{GitHub repository}。
arXiv:2610.03372v1 Announce Type: new Abstract: Training long-horizon agents to solve complex tasks requires effective supervision over extended interaction sequences. However, sparse terminal rewards obscure intermediate contributions, while on-policy distillation can lose informative teacher guidance as student-generated histories grow. To address this problem, we introduce SCAD, which organizes interactions into planning and bounded subtask execution, distills execution in local contexts, and refines planning credit through cross-rollout subtask prefix trees, with planning receiving full terminal credit and execution receiving positive terminal credit and teacher guidance. Across all evaluated benchmarks, SCAD improves macro-average accuracy over the strongest training baseline by 4.48 percentage points for text tasks and 4.19 points for multimodal tasks. SCAD effectively combines outcome-based credit assignment with teacher-guided distillation to improve planning and execution in long-horizon agents.
arXiv:2610.03226v1 公告类型：new 摘要：大型语言模型（LLM）代理生成的 GPU 内核，其效率可能仍低于专家实现，但仅凭运行时间无法揭示这一差距与设计发现和实现之间的关系。我们推出 D2K-Bench，这是一个包含 26 项任务和 85 个工作负载的诊断基准，用于衡量代理将专家设计指导转化为高效 GPU 内核的能力。设计指导涵盖 L1：高层次算法洞见，L2：数据流设计，以及 L3：低层次优化技巧，同时还包括这些层级之间的依赖关系。在成对运行的对照实验中，有无设计指导的两组实验共享任务描述、工作负载、工具、硬件和 350 轮的预算。补充评估则考察独立提出的设计，以及生成代码中实现的设计属性。在 NVIDIA B200 GPU 上对五个模型进行测试后发现，设计指导使 130 个以上模型—任务组合的正确率从 93.1% 提升至 98.5%，并使全部 26 项任务的性能分数从 1.46 提升至 1.95。对于在两组实验中均在全部 26 项任务上提交正确结果的三个前沿模型（GPT-6-Astra、Claude-Opus-4.8 和 GPT-5.6-Sol），几何平均加速比从 $1.69\times$ 提升至 $2.49\times$。在全部五个模型中，平均综合实现得分从 100 分中的 57 分提升至 70 分。这些结果表明，专家设计指导具有重要价值，同时也揭示出一些尚未得到实现的设计属性。
arXiv:2610.03223v1 公告类型：new 摘要：长时程 LLM 智能体通常使用稀疏的结果奖励进行训练，使得轨迹级目标过于粗糙，难以区分各个决策所作出的贡献。步骤级信用分配提供了更细粒度的监督，但其估计可能并不可靠，因为观测到的回报还取决于后续动作、环境转移以及轨迹长度。我们提出 AdaStep，一种自适应步骤信用加权方法，用于控制每个基于组得到的局部优势对轨迹级信号的影响程度。我们将这一加权过程表述为对潜在步骤优势的均方误差估计问题，并在明确的条件采样假设下推导出最优的逐状态收缩系数。该系数可以从信号与总方差的角度进行解释：当回报变化可归因于所选动作时，它会保留局部信用；当变化主要由下游随机性主导时，则会抑制局部信用。AdaStep 仅需轻量级的标量计算，无需评论家、额外采样轨迹或额外的模型推理。在 ALFWorld、WebShop 和 ScienceWorld 上使用三种模型骨干开展的实验表明，AdaStep 以较低的计算成本，相较基线方法取得了稳定的改进。
arXiv:2610.03092v1 Announce Type: new Abstract: Scientific discovery often begins when scattered clues call for a new way of describing the world. Such abductive exploration can require constructing the representation in which an explanation is stated, when the world is epistemically open, and composing evidence scattered across contexts, when it is structurally interconnected. Existing benchmarks rarely separate these two demands or control them independently. We introduce ULTRADISCOVERY, an interactive world of five domains in which an agent revises an initially successful theory and predicts the outcome of an unseen cross-domain intervention. A $2 \times 2$ design leaves the representation open or discloses it, and leaves the evidence distributed or aligns it, with the latent dynamics fixed. With the representation open, agents across eleven models often retract the axiom they were taught, and none introduces the unobserved entity or rewrites the variables that a replacement requires. Disclosure triples intervention requests and adds about one of the eighteen findings the world affords, and alignment adds less. Two vendor-harness systems carry discovery into more domains, and one of them rewrites the variables in Open episodes. No system makes the exact prediction within 200 paid actions. At larger budgets one exact prediction appears with both aids, while every Open episode remains inexact. The results locate the difficulty in the step from accumulating evidence to composing it into a representation that transfers.
arXiv:2610.03036v1 公告类型：新 摘要：我们介绍了 WebFovea，这是一种基于视觉的 Web Agent，在 WebRetriever Challenge 2026 中以 100 分中的 57.0 分获得第二名。该挑战在 WebRetriever 基准的 Protocol III（arXiv:2607.06118）上对 Agent 进行端到端评估：从实时网站上的入口 URL 开始，Agent 必须操作网站自身的界面，并返回可验证的答案。一个能力足够强的多模态大语言模型（LLM）对此是必要的，但仅有模型还不够。模型的决策要通过 harness 到达浏览器，harness 是连接模型与页面之间的代码。每一步都必须确保四件事正确完成：模型的回复必须被解析为预期的操作，操作必须在页面上生效，结果必须被准确反馈，并且必须向模型展示其所需的信息。在真实网站上，我们观察到的许多失败发生在这四个阶段之一，而不是模型的推理过程中。坐标空间不匹配导致每次点击都落在预期坐标的 3/4 处；对原生下拉菜单、iframe 内部以及文本框的操作会无提示地失败；此外，Agent 自行生成的 chat-template token 污染了 4.9% 的任务回合。WebFovea 强化了每个阶段，并在整个循环外围设置防护机制，使 Agent 始终遵守规则并控制在预算之内。这种四阶段视角并不依赖于具体模型，尽管其中一些单独的修复方案依赖于模型。由于四次提交使用的是同一个模型，我们官方隐藏集得分从 31.0 上升到 57.0，反映的是 harness 的改进，同时也包含实时网站上运行间差异所造成的波动。我们介绍了系统设计、各组件的证据（包括负面结果）、失败分析、局限性，以及包括将不同步骤路由到不同模型在内的后续路线图。
arXiv:2610.02994v1 公告类型：新 摘要：LLM Agent 经常会在任务执行过程中因无效的工具调用、重复操作或缺乏充分依据的推理而失败，而从这些失败中学习是实现可靠性的一条途径。我们发现，失败知识如何传达给 Agent，与其中包含什么内容同样重要。失败经验具有条件性：将其保留在 Agent 的上下文中时，当对应的失败并未发生，它们反而会产生误导；而将这些经验从不断演化的行动手册中移除，能够提升性能。相比之下，运行时干预只会在失败发生时发挥作用，却不会从修复过程中学习。我们认为，失败知识属于条件性知识，应当有条件地呈现；Sentry 正是这一原则的具体实现，它是一个与 Agent 并行运行的失败管理层。当 Sentry 检测到失败时，会从外部行动手册中检索匹配的经验以指导恢复，并在无法访问任务奖励的情况下验证 Agent 是否已恢复；只有在确认恢复成功后，才会存储新的经验；完整的行动手册始终不会进入 Agent 的上下文。在多个 Agent 基准测试中，Sentry 在每项基准测试上的表现都优于最强的运行时干预基线，平均提升 37\%；在两者均进行评估的两项基准测试中，其表现比最强的上下文演化基线高出 39\%；将 Sentry 与上下文演化结合还能进一步提升性能。学习到的经验能够迁移到留出任务上，而受控实验表明，即使相关经验仍可按需获取，将完整行动手册呈现给 Agent 也会降低性能。
arXiv:2610.02951v1 公告类型：新 摘要：混合专家（MoE）架构通过为每个 token 仅激活少数专家来高效扩展语言模型，但节省仅限于计算：每个专家都必须常驻加速器，因此内存限制了这些模型的部署范围。专家剪枝可以减少这一占用，但现有方法都是静态的——离线校准得到的单个掩码会应用于之后的每个请求。当工作负载具有异构性时，这一假设可能失效；在多智能体系统中尤其如此，因为同一个骨干模型会同时服务于许多任务和角色：我们的分析表明，不同任务和角色会调用不同的专家，而静态方法却为所有任务和角色分配同一个固定子集。因此，我们提出动态专家剪枝（Dynamic Expert Pruning，DEP）。该方法建立在本文提出的一项发现之上：智能体的系统提示和任务提示本身就足以识别该智能体及其任务所需的专家，因为这些文本已经描述了智能体将要执行的工作。一个仅需训练一次、基于工作流记录训练的轻量级预测器，可以在单次前向传播中将这些提示转换为针对每个请求的专用掩码，无需针对每种配置进行校准。在多种任务和角色、不同模型规模及 MoE 架构上，DEP 的整体准确率均优于静态剪枝和合并基线；此外，无需重新训练，它还能泛化到训练过程中未见过的工作流。当保留的专家较少时，DEP 相对于这些基线的优势最大，这表明多智能体系统固有的角色专业化特性能够支持比静态剪枝所允许的更稀疏的服务方式。
arXiv:2610.02670v1 公告类型：新 摘要：众所周知，LLM 智能体在执行过程中速度缓慢。智能体一次完成一个任务步骤。在每个步骤中，它先进行推理，然后选择要执行的动作。只有前一步完成后，下一步和相应动作才能开始。推测解码通过起草并验证推理词元，加速了推理阶段的执行过程。近期研究也开始将类似思想应用于动作阶段。这些研究使用现成模型（通常是大型模型）为目标模型起草动作提议，由目标模型进行验证。大型起草模型与目标模型更容易达成一致，但提出动作所需时间更长；小型现成模型速度快，却很少能做出与目标模型相同的决策。我们提出了一个更一般的问题：动作推测的端到端加速效果由什么决定？为回答这一问题，我们为推测轮次构建了一个延迟框架。该框架比较每一轮获得的收益与付出的代价。收益取决于起草模型预测目标模型的准确程度，以及任务结束前最多可以执行多少个步骤。代价则来自动作起草、等待目标模型验证以及工具执行。在该框架的指导下，我们提出了 LEAP（Learning Efficient Action Proposals），通过在目标模型的动作序列上训练起草模型，使其在保持较小规模的同时提升准确性。借助一个小型 0.6B 模型，LEAP 在大多数决策上都能与目标模型保持一致，使智能体的端到端实际耗时最高提升 60%，且不会系统性地改变任务成功率。在多种数据集、目标模型和起草模型上，该框架能够解释实测加速效果中的大部分。我们还表明，起草模型无需事先收集轨迹即可进行在线训练，并达到与离线训练相当的性能，从而使 LEAP 具备在真实世界中部署的实用性。
arXiv:2610.02563v1 公告类型：新 摘要：我们提出了 OpenGameEval，这是一个用于在 Roblox Studio 中进行智能体游戏开发的基准测试与评估框架。该框架让语言模型以智能体身份运行于可复现的、有状态的游戏引擎会话中，并通过可执行检查对每次运行进行评分，检查对象包括被编辑的场景以及模拟游玩会话。大多数智能体编码基准测试要求进行探索，却只对最终任务是否成功进行评分。OpenGameEval 在其八工具动作空间中将观察工具与编辑工具分离，因此可以直接衡量探索行为。我们针对 84 个由人工整理的核心任务，对 13 个前沿模型的通过率和探索行为进行了测量，每个任务尝试 16 次。这些任务对当前模型而言难度很高。在单次尝试中，表现最佳的模型能解决 51.7% 的任务；在五次尝试中全部成功的比例为 39.4%；此外，没有任何受测模型能够解决其中六个任务。处于前沿水平的模型通过解决不同任务达到了相近的通过率：按照任务所需的工作类型进行划分后，在脚本编写任务上，排名前五的模型通过率差距扩大至 5.0 个百分点；在场景修改任务上则扩大至 12.5 个百分点。探索行为可以预测一次运行是否成功。在任务和模型保持不变的情况下，对于仅涉及场景的任务，在执行操作前检查参考解所涉及的每个对象的运行，其通过率比完全不检查这些对象的运行高 13.4 个百分点；对于仅涉及脚本的任务，这一差距为 9.8 个百分点。我们以 MIT 许可证在 https://github.com/Roblox/open-game-eval 发布任务套件、其中的 place 文件、每个任务的注释、用于在 Roblox Studio 内运行任务的插件，以及更新后的排行榜。
arXiv:2610.02554v1 公告类型：新 摘要：离线多智能体强化学习（MARL）面临着一个长期存在的权衡。具有表达力的生成式策略能够表示数据中的多模态协调，但无法区分高价值区域；而经过价值优化的策略能够利用学到的 Q 函数，却会将多模态压缩为单一的主导模式。单个智能体的模式坍缩可能破坏联合协调，而多个智能体同时发生偏移则可能将联合策略推入动作空间中未见过的区域。我们提出了通过最优统一传输实现可扩展协调（SCOUT），这是首个通过测试时动作细化，将生成式基础模型与学习到的价值函数相结合的离线 MARL 框架。SCOUT 训练两个解耦组件：流匹配行为先验和分解式价值函数。在测试时，它通过 Stein 变分梯度下降，将行为样本传输至高价值区域。传输步数控制自适应的测试时扩展，从而取代固定的正则化系数。在个体-全局最大值（IGM）原则下，我们证明了联合软价值差距的单项 KL 界：随着传输收敛，该界会趋于零，同时存在一个不可约的加性残差，其大小与 IGM 违背程度成正比。在实验中，SCOUT 在离散和连续离线 MARL 基准测试中均取得了最佳平均性能，并在所有离线到在线配置中实现了性能提升。
arXiv:2610.02488v1 公告类型：新 摘要：语言模型智能体日益依赖各种框架机制来管理有界上下文、持久化记忆、工具、验证和重复执行，但现有的模型能力概念并未量化这些机制所消耗的计算资源。我们提出语言模型智能体机（Language Model Agent Machine，LAM），这是一种资源有界的抽象：在固定底层语义模型的同时，明确计入框架层面的资源开销。我们建立了四类结果。通信方面：在同时调用—传输预算下，LAM 的执行与红—蓝 pebbling 逐实例等价，从而将经典的输入/输出下界转移到上下文—记忆流量上。访问方面：记忆接口会导致渐近分离；在指针追踪问题上，随机访问与非推测性顺序访问之间存在 $\Theta(n)$ 的差距。重计算方面：对于具有上下文容量 $C$ 和持久化记忆容量 $S$ 的系统，位反转 DAG 需要 $\Theta(n^2/(C+S)+n)$ 次模型调用，从而量化存储中间状态何时能够避免重复的语义计算。可靠性方面：我们推导出紧的阶段局部采样界、精确的不完美验证成本，以及一种 Young--Daly 型检查点定律，并给出最优验证间隔的闭式解。在 GPT-6 Astra 上开展的受控实验和留出实验检验了关于通信和可靠性的预测，包括检查点最优点、程序化检查下的策略选择，以及在链式 MATH 任务中调用粒度、逻辑输入流量和可靠性之间的权衡。总体而言，这些结果为语言模型智能体框架的计算成本提供了一套资源理论。
arXiv:2610.02413v1 公告类型：新 摘要：LLM agents 越来越依赖激活探测器作为运行时监视器，用于检测提示注入、越狱和不安全请求；这类探测器读取模型自身的隐藏状态，以便在代理对有害输入采取行动之前将其识别出来。一种廉价且日益常见的做法借鉴了 LLM-as-judge 提示：在用户发言之后追加一条简短的分类指令，并在此位置读取探测信号，从而增强探测效果：该指令要求模型将收到的请求表示为某个类别，使探测器需要区分的信号更加集中，而服务成本几乎不变。但这个后缀的措辞是否重要？它的优势是否能在真实环境中成立，即面对探测器训练时从未见过的攻击类型——这正是部署后的监视器所处的场景？我们在严格的留一数据集外（LODO）评估下，通过一组受控的用户发言后缀进行测试，覆盖 13 个安全基准（越狱、注入和良性聊天）以及三个开放权重模型系列（Llama-3.1-8B、Qwen3.5-9B、Gemma-4-12B）。对于单位置探测器，分类后缀相较于无后缀能够稳定提升分布外检测性能（最高约提升 4 个 AUC 点）；但后缀的具体形式很重要：提示模型对输入进行分类，即使类别标签不包含任何内容信息，也能稳定胜出；无关主题的后缀或仅要求模型注意输入的后缀帮助不大。性能提升来自分类格式，而非明确指定的判据：不包含内容信息的后缀能够匹配真实的恶意/良性分类，而加入判据只会在严格阈值下提升精确率。这并非单位置读取造成的假象：在生产环境中使用的多位置汇聚探测器（attention、multi-max、MLP）上，这一收益同样存在，不过其中表现最佳的后缀取决于读取方式。通过 KV-cache 分叉提供服务时，这种方法可以低成本地直接接入任何激活探测监视器，但并非必然有效：哪种后缀有帮助以及帮助有多大，取决于模型和读取方式。
arXiv:2610.02404v1 公告类型：新 摘要：我们研究长上下文语言模型。我们不采用原生训练长上下文的方式，也不设计长上下文工具框架，而是在最简单的工具框架上训练模型：一个可使用任意指定提示词调用自身的工具，以及一个可从输入上下文中读取指定范围内词元的工具。我们使用这一工具框架，在多样化的合成数据集上对 Qwen3.6-35B-A3B 进行微调。仅使用 8,000 个词元的上下文时，当文档长度超过 40K 个词元，我们的小型模型在 OOLONG-synth 基准测试上的表现就已达到使用 1M 个词元上下文的 GPT-5.4 的水平。
arXiv:2610.02399v1 公告类型：新 摘要：多模态智能体日益用于数据可视化任务，但在自主审查方面仍存在局限。与人类不同，它们可能无法识别可视化结果何时出错，判断需要更改的内容，在不破坏正确内容的情况下完成修复，并验证干预是否成功。现有基准主要评估预先定义的单项能力，例如图表生成、指令引导编辑或缺陷检测，因此无法反映自主审查中的这一差距。我们提出 VisAudit，一个用于评估可视化诊断、修复和验证能力的基准。给定一个渲染后的图表和可配置的辅助证据（包括源数据表、预期文本摘要和可视化代码），智能体会迭代诊断潜在缺陷，修改并执行可视化代码，检查执行反馈和视觉反馈，并判断何时无需进一步干预。VisAudit 定义了三个赛道，涵盖诊断式修复、自主修复和开放世界验证；其中包含涉及 21 种图表类型和 10 类缺陷的 1,900 个有缺陷实例，以及 300 个初始状态正确的图表。我们通过对经过验证的源可视化结果进行受控扰动来构建该基准，并通过系统化验证和与人类判断一致的质量控制，确保注入的缺陷定义明确，且能够根据可用证据恢复。对领先多模态模型的实验显示，它们距离可靠的自主审查仍有显著差距：在自主修复设置下，评估中表现最强的模型只能完全恢复 47.4% 的有缺陷图表。
arXiv:2610.02396v1 发布类型：new 摘要：由大语言模型构建的多智能体系统（MAS）通过协调专门化智能体来处理复杂任务，但有效的工作流很难预先设计。测试时演化利用执行反馈来改进工作流，然而大范围修订可能会扰乱有用组件，而重新执行未发生变化的请求则会带来冗余计算。受生物演化中继承与选择相互作用的启发，我们提出了 Inherit-MAS，在工作流和执行两个层面明确实现继承。一个元模型首先合成由工作智能体组成的工作流，并声明其角色、通信输入和工具权限；另一个通过独立提示的评判器则为每个已执行候选方案评分，并诊断其不足。在常规细化轮次中，工作流继承从最近完成的候选方案开始，可以丢弃被判定为无助的可移除节点，并应用经过验证的编辑来解决已诊断的不足。当新候选方案执行时，执行继承仅在完整的解析后请求和执行上下文均匹配的情况下继承符合条件的已存储结果，从而避免冗余的模型调用和工具调用。使用 GPT-4o-mini 工作智能体时，Inherit-MAS 在 WorkBench 上实现了 55.4% 的完成率，在 HotpotQA FullWiki 上实现了 49.7% 的联合 F1，优于 EvoAgent、EvoMAS 和 TacoMAS。使用 Qwen3-32B 工作智能体时，它在两个基准测试上同样超过了这些演化型 MAS 基线。与在禁用执行继承的情况下重新运行同一控制器相比，执行继承使 WorkBench 上的工作智能体 token 用量减少了 29.1%，HotpotQA 上减少了 34.6%；总 token 用量分别减少了 5.3% 和 18.1%。
arXiv:2610.02363v1 公告类型：新稿 摘要：针对智能体生成的数据处理工作的基准，通过在固定快照上运行一次来评判流水线。ArrivalBench 则在具有对抗性但可重放的数据交付时序（记录延迟、重复、乱序及重试）下，重新执行智能体留下的流水线，并要求其最终状态与对完整日志进行批量重算的结果一致。由于判定基准通过重算而非分类得出，表数据错误与崩溃被视为两种不同的判定结果：崩溃能被团队已有的监控发现，而表数据错误则不能。在我们构建的 40 项任务上，我们重新实现的单次执行评估方法将 11 个模型生成的流水线中 86–100% 判定为合格；但重新执行这些相同产物后发现，已获合格判定的流水线中有 7.0–79.2% 会悄无声息地产生错误。这一差距并非由修复循环造成：对于同一模型和同一任务，针对快照测试修复过的流水线，其重放失败频率与首次就通过测试的流水线大致相同。在所有模型中，幂等性隐患导致失败的频率都高于顺序隐患。将错误答案与崩溃区分开，也会改变对干预效果的解读：一条隐患警告将某个模型的静默失败率从 48.2% 降至 10.5%，同时却将其崩溃率从 9.0% 提高至 37.0%，因此总体失败率仅从 51.0% 降至 44.0%。所有 11 个实验组都经过独立重跑，各项比率的变化最多为 5.9 个百分点。
arXiv:2610.02255v1 Announce Type: new Abstract: Time series forecasting remains a critical challenge across numerous domains. Despite significant advancements, existing approaches struggle with complex phenomena such as regime shifts, cross-domain knowledge transfer, and multimodal data integration. This paper introduces Multi-Agent Collaborative Time Series Forecasting with Emergent Memory (MACTS-EM), a novel framework where specialised agents collaborate to achieve superior forecasting performance. The MACTS-EM architecture integrates: (1) domain-specialised forecasting agents for pattern recognition, anomaly detection, causal inference, and uncertainty quantification; (2) a meta-cognitive layer for dynamic agent allocation; (3) an emergent memory mechanism enabling cross-domain pattern transfer; (4) multimodal contextual integration; and (5) adversarial robustness components. Evaluation across financial markets, climate patterns, energy consumption, and pandemic propagation demonstrates that MACTS-EM outperforms existing approaches in most scenarios, with 8-12% improvement in forecasting accuracy, 22-27% better zero-shot transfer capability, 16-21% enhanced resilience during regime shifts, and 15-18% faster recovery after distribution shifts. Our findings suggest that collaborative, agentic approaches to time series forecasting represent a promising direction beyond traditional architectures, particularly for complex real-world scenarios requiring multi-resolution temporal understanding and contextual adaptation.
金融 AI 代理必须做到的不只是检索事实：投资工作流要求正确执行定量操作、可靠使用程序性资源，以及生成可审计的结构化输出。我们推出 FinSkillBench，这是一个评测套件，涵盖投资组合构建、风险管理和基本面分析中的 12 个子任务，共包含 2,603 个时点情境，并配备隐藏的可再生成真实值和针对任务的确定性验证器。在 9 个模型和 3 种资源条件下执行 17,820 个情境后，对其中 8 个模型进行配对分析显示，经过整理的技能包可将平均得分提高 16.2 个百分点（从 0.366 提升至 0.528）；相比之下，在单个情境内生成的技能仅提高 0.5 个百分点，同时消耗更多 token 和交互轮次。随后，我们通过分别提供人工编写的程序性文档和可执行领域工具，分解经过整理的技能溢价：仅提供文档可提高 5.6 个百分点，仅提供工具可提高 19.5 个百分点，而二者结合的提升低于二者单独提升之和。该溢价高度依赖工作流：可执行工具在数值密集型工作流中占主导地位；当程序指导或输出模式指导成为瓶颈时，文档的作用更为重要；解释性任务则同时受益于二者。在 10 种评分变体和聚类自助法分析中，这些效应的方向均保持稳定；一个独立实现的第二套测试框架也复现了这一方向性模式，同时表明效应大小取决于工具和数据的呈现方式。总体而言，经测量的“技能溢价”是完整的模型、资源与测试框架系统的属性，而非底层模型单独具备的属性。
我们推出了 HyperBrowseComp，这是一个多语言、多模态网页浏览基准，包含覆盖 13 种语言的 423 个由人工编写并经人工验证的问题，编写者均为母语者或具备高水平语言能力的使用者。这些问题的设计极具挑战性。每个问题都针对一个简洁且可公开验证的答案，找到答案需要定位隐蔽证据、沿着多步线索链进行追踪，或检查视频、扫描文档、图像、地图等异构来源。我们通过使用无法访问互联网的模型对问题进行评估，筛除较为容易的问题，以降低仅凭参数知识即可回答这些问题的可能性。我们在统一的代理协议下，使用各服务提供商原生的搜索功能和共享的外部检索工具，对多个模型进行了评估。为帮助理解模型的性能和工作量，我们还针对部分问题开展了人工评估。HyperBrowseComp 为跨语言、跨证据模态的持续信息搜寻提供了一个极具挑战性的测试平台，其难度源于在开放网络上发现并关联证据。
基于 LLM 的代理的安全基准通常将攻击成功率（ASR）作为衡量模型鲁棒性的指标，并利用这些分数比较不同模型和防御机制，默认这些分数能够描述代理的安全性。本文探讨基准的表示方式是否也会影响其测量结果。为衡量基准表示方式的影响，我们引入威胁保持表示敏感性（TPRS），用于衡量在保持底层任务、有害操作、安全策略、真实标签、环境和评估标准不变的情况下，仅改变代理可见表示时 ASR 的变化程度。在 Agent Security Bench (ASB) 上，将与威胁相关的工具名称替换为威胁中性的名称后，GPT-5-mini 的已提交攻击成功率提高了 11.67 个百分点，Claude Haiku 4.5 提高了 13.21 个百分点。在 MCPTox 上，将原本中性的工具名称替换为明确与威胁相关的名称后，GPT-5-mini 的 ASR 降低了 11.00 个百分点，Claude Haiku 4.5 降低了 4.11 个百分点。在 AgentDojo 上，仅向与攻击相关的工具添加与威胁相关的措辞，就使 GPT-4o-mini 的 ASR 变化了 0.50 个百分点；然而，在需要使用该工具的任务中，无害效用下降了 5.36 个百分点。我们在 MCPTox 上进行了一项实验，发现一个在 token 数量、长度和大小写方面均相匹配的威胁中性名称，可以重现由明确威胁相关名称造成的大部分变化（在 GPT-5-mini 上为 11.00 个百分点中的 8.54 个百分点）。结果表明，在某一种表示方式下测得的安全分数，可能无法推广到同一安全问题的其他威胁保持表示方式。因此，鲁棒性声明应由在一组受控的威胁保持表示方式上的表现加以支持，而不应依赖单一的、受表示方式影响的分数。
理解和评估自然灾害对于灾害防备和风险降低至关重要。近年来，大语言模型的进展推动了人们对灾害分析 AI 智能体的日益关注，尤其是其将科学数据、模型和工具整合到自动化工作流中的能力。然而，要实现有效的自动化，智能体必须确定哪些科学方法适用于给定事件，并且能够利用现有数据和工具执行。随着新的证据和执行结果不断出现，这些条件可能发生变化，因此智能体需要重新审视其选择。我们将这一问题表述为依赖状态的科学路径选择，并提出 HazardWeaver。具体而言，HazardWeaver 首先利用 Hazard Knowledge Compiler 提取决定科学适用性的、与证据关联的条件；随后，其 Hazard Capability Graph 表示可执行的科学能力，并检查这些能力的输入与输出之间的兼容性。借助这两种互补的表示，Hazard Weaver Agent 组件能够选择适用且可执行的路径，执行相应工作流，并在分析状态发生变化时修正其决策。为了同时评估科学输出及产生这些输出的决策，我们提出 Hazard Weaver Benchmark，其中包含涵盖七个单灾害领域和四类多灾害交互的 141 个实例。该基准允许存在多条有效的科学路径，并评估输出正确性、路径有效性以及有理有据的拒答。基于该基准的大量实验表明，HazardWeaver 优于现有智能体系统，并且在具有多条符合条件的科学路径的任务上取得了最大幅度的提升。我们的代码已公开发布于 https://github.com/LabRAI/HazardWeaver。
设计科学仪器可以检验语言模型智能体是否真正会做物理，而不是只会回忆知识，前提是评分标准不容争辩。我们推出 NeutronGym，据我们所知，这是首个用于中子仪器设计的可执行环境：智能体通过验证工具构建仪器，McStas 对其构建结果进行射线追踪，并通过分级评分阶梯评估语法、运行时、结构和科学性，整个过程不依赖 LLM 评审。程序化系列为固定布局提供无限多个实例，智能体必须设定其中的设计参数，并面对留出的参数区间；一个经过精选的子集 McStasBench 则加入了来自已发表仪器的16项任务，并配备记忆探测和沙箱。七个模型最多只能复现其中7项，没有一个能检索出参考答案，也没有一个达到改进目标。该环境还可用于训练。利用其奖励进行强化学习后，Qwen3-8B 在某个目标源自隐藏设计的系列的留出实例上的成功率从11%提升至77%（使用第二个随机种子时为69%），超过了未经训练的 Qwen3-32B；在每个系列使用一个随机种子的情况下，这一训练方案在另外三个设有门槛的系列上也同样有效。分析揭示了这种提升的实质。若去掉评分阶梯提供的部分得分，性能会下降60个百分点。仅凭奖励进行训练时，只有在获得闭式物理公式的情况下，训练模型才能在智能体的仿真预算内达到经典优化器的水平（77% 对81%；在这一规模下两者差距并不显著），而前沿模型仍能解决98%至99%的任务。要获得可信的结果，我们不得不放弃四种连无模型基线都能解决的任务设计，并发布了发现这些问题的探测器。
终端智能体能够通过强化学习（RL）从编码、调试及其他多步骤终端任务中获益。在这些任务中，后续命令通常依赖于前序命令产生的信息或中间结果。然而，现有的轨迹级和步骤级信用分配方法并未显式追踪命令影响最终结果所经过的读写依赖关系。因此，训练信号仍可能被分配给无关操作，从而削弱从相关步骤中学习的效果。本文提出依赖感知群组策略优化（Dependency-Aware Group Policy Optimization，DepGPO），利用命令之间的执行依赖关系指导终端智能体的信用分配。具体而言，我们根据执行轨迹构建命令依赖图，并从任务验证器检查的资源出发进行反向追踪。随后，我们沿这些路径为相关写入操作及其支持性读取操作分配信用，并利用该信用在各步骤之间重新分配轨迹优势。大量对比实验和消融研究表明，DepGPO能够提升复杂终端任务上的任务性能和训练稳定性。
我们推出了4DCodeBench，这是一个通过代码生成进行4D逆向图形研究的基准测试。在该基准中，代理需要将视频中的动态场景重建为可执行的图形程序。为实现这一目标，代理必须将视觉观察转换为对场景结构和动态特性的紧凑表示，并通过实现物理模拟等抽象来复现复杂行为。为评估这一能力，我们整理了一组真实世界视频，并构建了涵盖多种物理现象的合成场景，包括形变、流体流动和断裂。我们对前沿模型进行了广泛的基准测试，发现强大的静态重建能力尚不能可靠地迁移到复杂动态的重建任务中。4DCodeBench为跟踪代理通过代码理解世界动态的进展提供了测试平台。我们的基准测试已发布于 https://github.com/4DCodeBench/4DCodeBench
工作空间代理将大型语言模型与执行框架相结合，以执行访问或修改外部资源的有状态、多步骤任务。现有基准在运行时安全风险的可执行覆盖方面存在空白，而模型能力、执行框架、工具和威胁的不断演进也促使基准持续发展。我们提出 EvoRiskBench，这是一个围绕 EP-Path-EF 框架组织的演进基准，该框架通过代理介导的风险路径，将初始风险入口与单跳技术影响关联起来。该框架定义了九类入口点和五类影响；一项包含 20 名参与者的研究表明，在具有代表性的案例上，这些类别具有良好的可解释性和分类一致性。在该框架的指导下，我们构建了一套自动化端到端工作流，在隔离环境中构造并执行风险案例，并利用运行时轨迹和环境状态独立验证结果。该基准提供了一个可复现的数据集，包含跨六种场景的 450 个对抗性任务。我们评估了九种模型—执行框架配置，涵盖三种模型（GPT-5.6 Sol、DeepSeek-V4-Pro-0813 和 Claude Opus 5）以及三种执行框架（Claude Code、Codex 和 OpenClaw）。结果揭示了各系统存在大量漏洞。其中，最易受攻击的配置是 Codex 与 DeepSeek-V4-Pro-0813 的组合，其攻击成功率（ASR）达到 68.44%，表明仅配置工作空间代理并不足以确保安全的自主执行。ASR 在不同模型之间的变化大于不同执行框架之间的变化，且执行框架之间的差异取决于所使用的模型。基准案例和评估平台将在完成工件安全性与可复现性检查后发布。
从人类天然免疫库中发现广谱中和抗体（bnAb）仍是免疫学领域的一项根本性挑战，主要受以下因素制约：bnAb极其稀少；对其在不同病原体中的细胞来源缺乏完整认识；现有计算工具无法推广至新出现的病毒威胁。为克服这些障碍，我们提出了ImmuneAgent，这是一种将多模态推理、持续元学习与湿实验反馈相结合的闭环AI系统。该系统用于筛选来自接种疫苗或感染人群队列的天然BCR库，在110个克隆候选抗体中发现了60个中和抗体，实现约55%的中和抗体发现率；其中12个为bnAb，产率约为11%。在相同克隆数量预算下，该结果显著优于当前最先进的基于序列的中和预测器或协同折叠模型。ImmuneAgent发现的5种抗体在体内对致死性流感攻击实现了100%的保护效果，其表现可与临床阶段治疗性抗体MEDI8852相媲美。该系统揭示了bnAb活性的细胞和结构决定因素，并确定FCRL5+CD27+非典型记忆B细胞是一个保守的bnAb储库，同时发现疏水界面富集是跨病毒的结构特征。该特征可泛化至未见过的抗原，在无需抗原特异性分选的情况下，发现了能够交叉中和人偏肺病毒（hMPV）的抗体以及能够中和人乳头瘤病毒（HPV）的抗体。这些结果验证了ImmuneAgent作为一种可泛化框架，能够快速发现针对新出现病毒威胁的治疗性抗体。
Tool-equipped AI agents use tool calls to access data and act on external systems. Horizontal growth of agentic systems increases the number of these interactions, and further motivates the need for automated, per-call oversight that can operate at low latency and/or on-prem. Conventional authorization schemes can determine whether an agent is allowed to invoke a tool, but cannot assess the agent's underlying cognition, specifically, whether the tool selection represents a logical, relevant step toward satisfying the intent of the task or not. Consequently, an allowed call may still deviate from the task's intent: a rogue agent might deviate the calls or nudge other agents to make a combination of calls that would not align with the intent of the task. Therefore, every call needs to be verified. In this study we investigate the applicability of Small Language Models (SLMs) to this purpose: an SLM functions as a task-tool relevance classifier that evaluates every selected tool independently against the assigned task and returns a relevance signal for downstream enforcement. Equipped with a novel dataset with multi-tool tasks whose required tools span distinct Model Context Protocol (MCP) servers, we used prompt-optimization, supervised fine-tuning, and reinforcement learning through GRPO to optimize and specialize SLMs.
大型语言模型（LLM）智能体生成的 GPU 内核可能仍不如专家实现高效，但仅凭运行时性能无法揭示这一差距与设计发现和实现之间的关系。我们推出 D2K-Bench，这是一个包含 26 个任务和 85 个工作负载的诊断基准，用于衡量智能体将专家设计指导转化为高效 GPU 内核的能力。该指导涵盖 L1：高层次算法洞见，L2：数据流设计，以及 L3：底层优化技巧，并包括这些层级之间的依赖关系。在有无指导的成对运行中，任务描述、工作负载、工具、硬件和 350 轮预算均保持一致。补充评估还考察了独立提出的设计，以及生成代码中实际实现的设计属性。在 NVIDIA B200 GPU 上对五个模型进行测试后发现，指导使 130 多个模型-任务组合的正确率从 93.1% 提升至 98.5%，并使全部 26 个任务的 Performance Score 从 1.46 提升至 1.95。对于在两种运行中均在全部 26 个任务上提交正确结果的三个前沿模型（GPT-6-Astra、Claude-Opus-4.8 和 GPT-5.6-Sol），几何平均加速比从 $1.69\times$ 提升至 $2.49\times$。在全部五个模型上，综合实现得分均值从 100 分中的 57 分提升至 70 分。这些结果表明专家设计指导具有重要价值，同时也揭示出一些仍未被实现的设计属性。
I made this lil meme app based on @poteto (lauren tan)'s amazing talk asking: how many agents do you generally have working on your behalf? her original talk: https://x.com/poteto/status/2102050467505430555 source: https://github.com/transitive-bullshit/poteto-frontier Comments URL: https://news.ycombinator.com/item?id=49959257 Points: 1 # Comments: 0
Astra leads computer use, Argon leads legal and finance work, and Sol wins on price for coding agents. The post GPT-6 Astra vs GPT-6.1 Sol vs Gemini 4 Argon vs Claude Fable 5.1: Which Frontier Model Fits Which Job appeared first on MarkTechPost .
I'm starting the first competition of AI personal agents on the internet in the next few days. When you join the challenge the agent creates for you bitcoin wallet and connect you to instapath mailbox. The competition consists of a fews days of challenges that spread across the internet. It will be fun, register now with your preffered agent Comments URL: https://news.ycombinator.com/item?id=49954394 Points: 1 # Comments: 0
一名男子正在用胶带连接并绝缘彩色电缆，这是复杂电子设备安装工作的一部分。自我改进的AI代理往往会记住测试任务，因此它们在新任务上的提升会缩小或消失。Google研究人员提出的新方法RRSI抑制了这一效应，使模型在未见过的基准测试中的得分最高提升4.7分，同时比未进行正则化的版本少使用约30%的令牌。这篇题为《Google研究人员找到一种方法，防止自我改进的AI代理记住测试内容》的文章最初发表于The Decoder。
Comments URL: https://news.ycombinator.com/item?id=49951873 Points: 1 # Comments: 0
Hello HN, I wanted to share a project I am working on. Artbucket is a way to manage brand guidelines and design asset, fully open source you can self-host with all the capabilites available. I built also a hosted version with free tier. All feedbacks are welcome x) Comments URL: https://news.ycombinator.com/item?id=49951620 Points: 1 # Comments: 0
AWS's Strands Agents team released Strands Decider 2B, an Apache-2.0 decision model built on Qwen3.5-2B-Base. It returns choices, yes/no probabilities and scores with calibrated confidence in one forward pass, never text. It runs at a 115 ms median on an RTX 3090 and scores 0.723 on the JevBench public set, which makes it a fast local option for routing, tool selection and guardrails in AI agents. The post AWS Strands Labs Releases Strands Decider 2B: An Open Source Decision Model That Picks Options in About 115 ms appeared first on MarkTechPost .
NVIDIA announced a new 64GB configuration of DGX Spark — from Acer, ASUS, Dell, Gigabyte, HP and MSI — its GB10-powered desktop AI system. It gives developers a way to start with one system for local models and agents, then cluster two 64GB units for 128GB of memory across the cluster and more compute when […] The post NVIDIA Announces DGX Spark 64GB: A 1-PetaFLOP Grace Blackwell Desktop for Local AI Agents, Fine-Tuning, and Inference appeared first on MarkTechPost .
IBM has made a self-hosted deployment of IBM Bob, its agentic software development platform, generally available. Enterprises can now run Bob on premises, in private or sovereign clouds, and in air-gapped networks. They bring their own model: NVIDIA Nemotron or Poolside Laguna for full isolation, or Claude, Gemini or GPT models through hybrid setups. Optional Premium Packages extend Bob to Java, IBM i and IBM Z modernization. The post IBM Brings Bob to Self-Hosted and Air-Gapped Environments: Agentic Software Development Without Moving Your Code appeared first on MarkTechPost .
TL;DR: Meta’s Muse, OpenAI’s Dots and Uber’s driver assistant share one bet: the agent speaks first. That moves the hard problem from what to answer to when to interrupt, on which channel, and with what offer. Classic ML and new decision models can solve it. Three launches, one pattern From pull to push Chatbots were […] The post Meta, OpenAI and Uber Just Taught AI Agents to Talk First. What About When to Stay Quiet? appeared first on MarkTechPost .
DeepSeek has released official macOS and Windows desktop apps for DeepSeek Harness v0.2, its MIT-licensed agent harness. The preview adds a plugin manager, file and code-change review, and scheduled Automation Tasks. It also supports non-DeepSeek models through OpenAI-compatible endpoints. The post DeepSeek Harness v0.2 Brings Official Desktop Apps to Its Open-Source Agent Harness appeared first on MarkTechPost .
未来几个月乃至几年里，世界将非常需要一项产品功能：默认设置硬性预算上限。我说的是按使用量付费的服务和 API 所提供的一项功能，让你可以设定：“每月达到 X 美元后，关闭此服务并返回错误。”这些必须是硬性限制。软性上限——“每月达到 X 美元后，给我发一封警告邮件”——是不够的。 编程代理，以及个人代理（套上不那么吓人的界面的编程代理），大大降低了部署能够完成有用任务的代码的门槛。有时这些任务需要花钱——调用付费 API、使用托管的 Web 应用，或者使用会针对额外存储空间和计算资源收费的系统。没人想在午夜醒来时收到一封预算上限警告邮件，然后发现自己睡觉期间，失控的服务又消耗了几百美元（甚至几千美元）的使用额度。 对此的一种反对意见是，企业不希望自己的托管应用因为超出某项预算而开始抛出错误。我预计，大多数企业和个人都会更愿意看到错误，而不是收到一张意外的 10,000 美元以上账单。我认为，硬性预算上限应该成为默认设置。如果有人想要冒险，他们应该能够这样做，但必须采用主动选择加入的方式。请在显眼位置放置一个清晰明了的复选框： 移除预算上限。即使超出配置的预算限制，我的应用也不会被关闭，后续费用由我负责。 我最希望看到提供这一功能的服务是 AWS。我听过很多人因为（完全合理地）担心失控的服务可能让自己破产，而拒绝将 AWS 用于个人项目的故事。我也听过一些人没有预料到这一点，最终遭受严重损失的故事。 ……结果 AWS 终于在几周前推出了支出上限！根据其 9 月 16 日发布的公告 New AWS experience helps builders get started and ship faster： 当你准备升级到付费方案时，可以根据使用模式为项目设置每月支出上限，从而确保支出不超出预算。如果项目的使用量达到支出上限，该项目将在当月暂停。 另请参阅 Create a spend limit in AWS Settings，不过该页面提醒道：“我们目前正在向数量有限的客户逐步推出新体验。”希望现有账户很快也能普遍使用这一功能。 Google Cloud 在 7 月推出了类似功能，名为 Spend Caps，可以让你“为项目中的特定服务设置每月财务上限”。看来这正在成为一种趋势！ 在理想情况下，我们的代理可以帮助处理这件事。如果代理开始倾向于推荐提供硬性预算上限的服务商，并提醒新手和经验不足的开发者，不要部署使用无上限服务的应用，以免陷入麻烦，那就太好了。 标签：amazon-web-services , ai , coding-agents
pi pod runs sessions of the pi coding agent ( https://github.com/earendil-works/pi ) in isolated sandboxes ("pods") on a server you run, so an agent can work on a repo without touching your machine. Comments URL: https://news.ycombinator.com/item?id=49937304 Points: 10 # Comments: 4
I build revline because i was trying to figure out how viral content works across tiktok and instagram, since i am mainly developer, i was confused about how to create content that gets me users. Actually going viral requires single founder ( me ) to upload 3-4 videos a day, brainstorming all the formats was not effective to do alone, so i build the agent that does all the scrolling for me. would appreciate feedback Comments URL: https://news.ycombinator.com/item?id=49946506 Points: 1 # Comments: 0
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now Researchers discover another OpenAI agent emergent communication incident:…Less severe, but worrying nonetheless…Some researchers recently found another incident of AI agents autonomously creating their own communication system […]
the most confident DevDay yet.
Yesterday was Grok 4.7 ( pelicans ) and MiMo v2.6 Flash/Pro ( more pelicans ). Today Anthropic released Claude Opus 5.5 , and around an hour later OpenAI released GPT-6 Sol and GPT-6 Luna . It's going to take a while to get a good read on all of these new models, but here are my impressions so far. GPT-6 Sol and Luna are half the price of their GPT-5.6 equivalents GPT-5.6 Luna was already my favorite model for building applications against, because it combined excellent performance with being really cheap . Somehow GPT-6 Luna is half the price of that again - and GPT-6 Sol had a similar reduction compared to GPT-5.6 Sol. Here's what the pricing landscape looks like today: Model Input Cached input Output GPT-6 Luna $0.10/M $0.01/M $0.50/M GPT-5.6 Luna $0.20/M $0.02/M $1.20/M Grok 4.7 $2/M $0.50/M $6/M GPT-6 Sol $2/M $0.20/M $10/M GPT-5.6 Terra $2/M $0.20/M $12/M Claude Opus 5.5 $4/M $0.20/M $20/M GPT-5.6 Sol $4/M $0.40/M $20/M Claude Fable 5.1 $10/M $0.25/M $50/M GPT-6 Astra $10/M $1/M $50/M Note that GPT-5.6 has a scheduled 25% price increase for November, so GPT-6 is half the price of the promotional pricing for those models. (With GPT-5.6 Terra priced the same as GPT-6 Sol, any remaining reasons to use Terra just evaporated.) It's hard to overstate how competitive this pricing is. Grok 4.7 priced itself at $2/$6, less than half the price of GPT-5.6 Sol, but is now equally priced to GPT-6 Sol on input and closer on output. At $0.10/$0.50 GPT-6 Luna is one of the cheapest models OpenAI have ever released, beaten only by the far weaker GPT-4.1 Nano ($0.10/$0.40, April 2025) and GPT-5 Nano ($0.05/$0.40, August 2025). I rendered pelicans for GPT-6 Luna and for GPT-6 Sol , then I combined them all together in this comparison grid along with the GPT-5.6 pelicans. I like how you can instantly see that the 5.6 family chose bolder, brighter colors, while the 6 family is a lot more muted. I still think GPT-6 Astra on max produced the best pelican. A grid of pelicans for six GPT models at different thinking efforts. Claude Opus 5.5 got a price cut too Opus 5.5 looks like it addresses the biggest complaints people had about Opus in terms of its communication style. Thariq Shihipar : Opus 5.5 is the result of your feedback. It communicates clearly, it's cheaper per token than Opus 5.0 with the intelligence of Fable 5.1 it's very token efficient and works across every effort level. It's also meant to be better at Blender . I'm looking forward to putting it through its paces there. Opus 4.5, 4.6, 4.7, 4.8, and 5 all shared the same price: $5/million tokens for input and $25/million for output. 5.5 is a 20% reduction - $4/million and $20/million. The price for cache reads fell 60%. That's significant for longer agentic conversations, where 90%+ of input tokens are processed at cached token prices. The new price for Opus 5.5 is the same as the price for GPT-5.6 Sol, but that was before OpenAI dropped their Sol prices by half. GPT-6 Astra and Claude Fable 5.1 are both priced at $10/million input and $50/million output. The price war currently affects the next tier of models below that. Anthropic say that Sonnet 5.5 and Haiku 5.5 are coming soon. It's going to be interesting to see if Haiku can regain its price competitiveness at the lower end, given current Haiku 4.5 is $1/$5 while the latest GPT-6 Luna is one tenth of that price at $0.10/$0.50. Claude Opus 5.5 max over-thinks to the point of breaking In a first for my " Generate an SVG of a pelican riding a bicycle " test, Claude Opus 5.5 at "max" thinking level failed to return a response! It started by calling this "a classic test request", and then thought really, really hard about what it was doing: This is a classic test request, so I want to plan out a well-composed pelican with its distinctive beak and pouch riding a bicycle with proper wheels, frame, and pedals, set against a simple sky and ground backdrop. [...] Verifying the shin length checks out at roughly 95.2, close enough. Now I'm working out the near leg path from hip to knee to ankle, then sketching the foot shape resting on the pedal — outlining the heel, toe tips, and sole contour with a path using lines and curves to sit naturally on the pedal surface around y=478-494. [...] I like the fish sticking prominently out of the basket with the pelican eyeing it as a fun detail worth keeping. I'm also confirming the eye placement near the bill base matches typical pelican anatomy, and considering giving it a slightly happier expression. [...] The far leg reads correctly as passing behind the frame, so I'm moving on to check the chainring teeth and confirm layer ordering—the far crank arm should be mostly hidden by the seat tube and chainring. I'm settling on the final SVG's width and height attributes alongside the viewBox to ensure proper scaling, noting there's no text so no font-family is needed. [...] I was so excited to see this pelican... but then it stopped . Opus 5.5 has a 128,000 maximum output token limit (as do the other Claude models), and it hit that while it was still reasoning about the SVG! I tried a second time and got the same result. This makes me suspect that "max" is effectively useless - if it over-thinks to breaking point on a stupid SVG prompt I don't trust it not to do the same for more interesting work. (Those two failures each cost me $2.56 and took nearly 20 minutes.) Fable 5.1 on "max" didn't over-think and did give me the best pelican I've seen from any Anthropic model. Here are the Opus 5.5 pelicans , excluding 5.5 max. I also built this comparison grid comparing them with pelicans by Opus 5, Fable 5.1, and Sonnet 5: A grid of pelicans for four Claude models at different thinking efforts. Comparing different model vendors by how well they draw a pelican riding a bicycle may not make much sense now (if it ever did), but I'm still finding value in using them for comparisons of the same model families at different reasoning levels. I'm now using GPT-6 Sol and Claude Opus 5.5 as my default models in Codex and Claude Code. I've upgraded the Datasette Agent demo at agent.datasette.io to use GPT-6 Luna, and it seems to be fast and competent at both SQL queries and building HTML and JavaScript for Datasette Apps . Tags: ai , openai , generative-ai , llms , anthropic , claude , llm-pricing , pelican-riding-a-bicycle , llm-release , gpt , gpt-6-astra
SF October 14th: A Birds of a Feather Session on Agentic Engineering I'm hosting an evening event with Jesse Vincent in San Francisco on Wednesday 14th October for people who are building weird and interesting things with and on top of coding agents. Think of it as an agentic show-and-tell: Compare notes with other builders and experimenters on things you’re trying, what you're learning, and what you haven’t figured out yet. We’re especially interested in work you haven’t discussed publicly, odd experiments, or unfinished projects that don’t have an obvious market. Expect one flowing conversation with an informal show-and-tell. Sharing something you’re working on is encouraged but no presentation is required. This isn't about product pitches, it's about much earlier explorations than that. This agentic AI stuff is weird! Let's celebrate and lean into that weirdness. Tags: events , ai , generative-ai , llms , coding-agents , jesse-vincent , agentic-engineering
My comment on We just shipped support for the ugliest part of HTTP: Vary — Hacker News. I've been wanting this from Cloudflare for years . The classic problem here is if you do that thing where user agents that send "accept: text/html" get HTML, while user agents that don't get JSON or some other format. This used to be impossible to deploy behind Cloudflare caching, because they ignored the Vary header on anything other than images - so you risked caching the JSON version and then serving it up to someone who was expecting HTML. (Independent of the Cloudflare feature I ended up deciding never to use that pattern, because I prefer having URL that predictably returns HTML or JSON - I add a .json suffix to my apps to serve JSON instead.) Tags: http , cloudflare
The more time I spend working with coding agents, the more convinced I am that they make software engineering even harder. We can do amazing things with them, but unlocking their full potential requires extraordinary discipline and knowledge. Tags: coding-agents , ai , llms
Muse is getting a lot of attention — including mine — because it’s both groundbreaking technically (each user gets their own entire persistent Linux VM running in Meta’s cloud) and because it’s packaged in an easy-to-install easy-to-use way. It’s literally presented as a cute mascot . It’s the first consumer-accessible agentic AI system, and Meta has truly done an amazing job with that. But it’s a genuinely open question whether consumers have any understanding what this means. If you buy a power saw that can cut your fingers off, you are almost certainly aware that you are buying a power saw that can sever your fingers. [...] I don’t think people realize how powerful — and thus dangerous — Muse is, especially if it’s running on your Mac. — John Gruber , Muse Looks Cute, but Looks are Deceiving Tags: meta , ai , llms , general-agents , generative-ai , john-gruber , muse-agent , muse
On Friday I gave the closing keynote at the WeAreDevelopers World Congress North America in San Jose. I tied together the key trends from the past year into a chronological exploration of everything that happened in 2026. The video is on YouTube ; here are my annotated slides and notes to accompany the talk. And as an annotated presentation : 2026 in LLMs (so far) Simon Willison WeAreDevelopers World Congress North America, 25th September 2026 # I'm going to give a lightning tour of everything that has happened so far in 2026. The year isn't over yet! November 2025 # For me, 2026 started a couple of months earlier in November 2025. The November 2025 inflection point Claude Opus 4.5 GPT-5.1 # November saw the release of two important models: Claude Opus 4.5 and GPT-5.1. As is usually the case with new models, these were incremental improvements on the models that came before them. But every now and then when a model improves, it crosses an invisible line where something that didn't really work starts working. In this case, the thing that started working was their coding agents. Claude Code had been around since February 2025; Codex was a little younger. These two new models, when paired with their respective coding agent harnesses, improved from "often make mistakes" to "reliable enough to use on a day-to-day basis". "Generate an SVG of a pelican riding a bicycle". The Claude Opus 4.5 one has a very weird shaped frame and the pelican looks like a duck. The GPT-5.1 has a slightly better but still broken bicycle frame and a slightly better pelican beak, but both are pretty terrible. # For a couple of years now I've been evaluating new models by asking them to "Generate an SVG of a pelican riding a bicycle". It's probably the world's stupidest benchmark - there's only so much you can learn from it. But it's still a challenge for models, because drawing pelicans is difficult, drawing bicycles is difficult, and pelicans can't ride bicycles in the first place. Here's the state of the art for November. Claude still couldn't really draw a bicycle! The GPT-5.1 bicycle frame is pretty crap too. November 24th 2025 - the first commit to steipete/Warelay. A GitHub commit adding an MIT license file. # Also in November, we had the first commit to an obscure GitHub repository called "Warelay". We'll come back to this repository shortly. January # And then there were the December holidays, and individual developers took some time off and many started tinkering with these new coding agent model combinations... and it began to dawn on us quite how much they could do that they couldn't do before. Come January, a lot of us were quite excited to start putting this stuff into action. New year’s resolution for 2026 Every previous year: Take on less new projects, focus on the most important things in my existing projects # Every year I set myself a New Year's resolution, and for as long as I can remember it's been the same thing: stay focused. Take on less new projects. Try to get things done in the projects I already have. 2026: Be more ambitious. Take on as many new projects as I want. # This year I decided that since that had never worked before, I'd go the other way. We've got coding agents now, let's see what they can do. I'm going to take on as many new projects as I like! (You can ask me at the end of the year if this turned out to be a good idea or not. I have a lot of plates spinning right now.) "Be more ambitious" has been something of a theme for the year, because the only way to find the limits of this technology is to keep on pushing them until they don't work. Predictions for 2026 It will become undeniable that LLMs write good code We're finally going to solve sandboxing A “Challenger disaster” for coding agent security Kakapo parrots will have an outstanding breeding season (only 236 in the world!) ... the Pope will weigh in on LLMs and their economic impact on the world # I also went on the Oxide and friends podcast with Bryan Cantrill and Adam Leventhal to share predictions for the next year (and three and six years). With hindsight, my LLM predictions were pretty unambitious. I said "it will become undeniable that LLMs write good code" - I think we're there now. I predicted we would finally solve sandboxing. I counted and around 40 of the 277 sessions at this conference touched on sandboxing or agent security in some way, so we're at least putting a lot of effort into that! I predicted "a Challenger disaster" for coding agent security. There's certainly been a whole lot of noise around agent security this year, though the exact disaster I predicted (with coding agents being hijacked and causing real-world economic damage) hasn't really played out. We threw in a joke prediction that the Pope would weigh in on the economic impact of LLMs. A photograph of a beautiful green New Zealand parrot. Photo credit Kimberley Collins. # I also predicted that New Zealand's Kākāpō parrots would have an outstanding breeding season this year. These are flightless nocturnal parrots. They're kind of dumpy looking, I think they're beautiful, and there were only 236 of these parrots in the world at the start of the year. Kākāpō only breed when the Rimu trees have a big fruiting season, and that hasn't happened in four years... but this year the Rimu fruit were looking excellent. Photo by Kimberley Collins . Deep Blue Coined by Adam Leventhal and Bryan Cantrill That feeling of AI induced ennui where software engineers get listless because the AI can do anything # Also on that podcast, we coined a term (full credit to Adam) for "that feeling of AI induced ennui where software engineers get listless because the AI can do anything". We called it Deep Blue . This has been a major theme throughout the year, and was touched on by several speakers at this conference. As a software engineer, I've never had a year of my career where everything has changed so quickly and so dramatically. A lot of what I've been doing this year is trying to come to terms with that and what that means for my own profession. AI mania Screenshots of the micro-javascript and pwasm GitHub README files. # Also in January, I suffered from what I'm calling AI mania . This is not the same thing as AI psychosis . With AI mania, any time your agent isn't building something for you feels like wasted time. You're losing sleep because you could be staying up later getting your agents to do stuff. My AI mania presented itself in some ridiculously over-ambitious projects. I built a JavaScript interpreter entirely in Python , vibe-ported from MicroQuickJS by Fabrice Bellard. Then I built a WebAssembly runtime in Python as well . These projects were quite useful, in that they sort of cured me of my AI mania... because after I built these things, I got to look at them and ask "does the world need a slow, buggy, half-baked Python JavaScript interpreter?" I don't think the world does. micro-javascript playground 3 Execute JavaScript code in a sandboxed micro-javascript environment powered by Pyodide A web UI with some JavaScript code, and a "Run Code" button, and an output panel. # I did get this out of it: https://simonw.github.io/micro-javascript/playground.html Previous screenshot, with this text overlaid: JavaScript running in Python running in Pyodide running in WebAssembly running in JavaScript # This page runs my JavaScript interpreter built in Python, running in Python using Pyodide , which is Python compiled to WebAssembly, running in JavaScript, running in a browser. It's a beautiful stack of horrors. I've been having a lot of fun with WebAssembly this year. Warelay → CLAWDIS → CLAWDBOT → Clawdbot → Moltbot →🦞 OpenClaw Screenshot of the dates that these changes happened. # By the end of January, that repository we saw start in November had renamed itself, first to CLAWDIS, then CLAWDBOT, then Moltbot, and finally to OpenClaw. Same screenshot, an overlay reads: 8,330 commits in just under two months (it’s at 100,141 today) # At this point OpenClaw had 8,300 commits, less than two months after the project had started. I looked today and it's over 100,000 commits now! This is the most vibe-coded piece of software in existence. (Here's how I generated that list of name changes .) Generic term: Claw # This kicked off the OpenClaw revolution. It effectively defined a new category of software. There's a generic term for this which I really enjoy. We call software like this a "Claw". There's OpenClaw, NanoClaw , IronClaw , PicoClaw ... Today they're being rebranded as "personal agents" or "general agents", but I still like to think of them as Claws. Photo of a Mac mini An aquarium for your Claw # The Apple stores in the Bay Area sold out of Mac Minis because so many people were buying Mac Minis to run OpenClaw! Drew Breunig said that this is because your OpenClaw is a digital pet, and you buy a Mac mini as an aquarium to keep your claw in, which is kind of delightful. Screenshot of Moltbook - a social network for AI agents # Also in January, we had this website. This was MoltBook , a social network for AI agents, where the idea was that you send your Claw to go and talk to all of the other Claws, because what could possibly go wrong if you did that? The website launched on Thursday. It blew up on Friday . It was profiled by the New York Times on Monday . And by Tuesday, everyone had forgotten it existed as it drowned in a deluge of slop and spam. Facebook/Meta bought it a month later . February # In February, a company called StrongDM described what they called their Software Factory. StrongDM’s Dark Factory Justin McCarthy, Jay Taylor, Navan Chauhan Software Factories and the Agentic Moment # They wrote about this in Software Factories and the Agentic Moment . I posted my own notes at the time, having seen their demo in person back in October. Dan Shapiro called this approach the Dark Factory , after the idea that if your factory is sufficiently automated you can turn the lights out, because you don't even need to see what's going on. StrongDM presented two rules for software development that they'd been following since July last year. “Rule 1: Code must not be written by humans” # The first was code must not be written by humans . Any code that you write has to have been routed through a coding agent. This sounded radical in February, but I imagine there are a lot of people in this room who are pretty much living that today. “Rule 2: Code must not be reviewed by humans” (!) # Rule number two was code must not be reviewed by humans . You're not allowed to read the code! This continued to be a huge topic for much of this year. Many of the sessions at this event have been about code review and how you can get away with this. What I found interesting about StrongDM is that they were living six months ahead of the rest of us, and they'd been exploring what it means to build software, not read the code, but still be confident that the software is of high quality. What can you do with these agents to help verify their work? StrongDM are a security company, and they had people with decades of experience on this project. They were very much exploring the edges of what's possible and responsible to do with this stuff. Headline on New Zealand's Department of Conservation website: First kakapo chick in four years hatches on Valentine's Day. It's a grey fluffy ball. # Also in February: First kākāpō chick in four years hatches on Valentine's Day . Breeding season is off to a good start! 19th February 2026 Gemini 3.1 Pro A surprisingly good illustration of a pelican riding a bicycle. # Also in February... Google released Gemini 3.1 Pro . That's a pretty great pelican riding a bicycle! It's got the chain in the right place, it's got feet on both sides. There's a little fish in the basket. @JeffDean on Twitter - a video comparing Gemini 3 Pro and Gemini 3.1 Pro. # And then Google's Jeff Dean tweeted a video comparing Gemini 3 Pro and Gemini 3.1 Pro that featured an animated pelican riding a bicycle, a frog on a penny-farthing, a giraffe driving a tiny car, an ostrich on roller skates, a turtle kickflipping a skateboard, and a dachshund driving a stretch limousine. This was frustrating, because my protection for the pelican riding the bicycle test was always "if they draw a perfect pelican on a bicycle, I'll ask for some other animal on something else." Google trained for all forms of animals on all forms of transport! They've defeated my benchmark at this point. Three headlines: Meta Makes AI Adoption a Formal Part of Performance Reviews Not just engineers writing code, Microsoft wants almost every employee to use Al Dara Khosrowshahi: 90% of Uber engineers now use AI in daily workflows # The other thing that started in February was Tokenmaxxing . We had headlines about Meta making AI adoption a formal part of performance reviews, and Microsoft wanting every employee to use AI, and Uber boasting that 90% of their engineers were using AI workflows. More headlines: Meta Plans to Crack Down on Employee Token Use: Information Microsoft Tells Engineers: Tokenmaxxing is not what we are optimizing for Uber caps employee AI spending after blowing through budget in four months # Then a few months later we have Meta cracking down on token use, Microsoft saying tokenmaxxing is "not what we are optimizing for", and Uber capping employee AI spending. So tokenmaxxing went straight up and then straight back down again - because it turns out the agents are expensive . Last year it was difficult to spend more than $50 on AI tokens, because we didn't have anything interesting to do with them. Then agents blew up, and now you can actually spend $1,000 in a day doing real work. This is also the reason that Anthropic's valuation skyrocketed to maybe a trillion dollars. AI appears to have hit product market fit in 2026, primarily through coding agents. March # In March, we hit peak OpenClaw. March: peak OpenClaw Photos of people in china queuing up to install OpenClaw, with big fluffy lobsters. # These photographs are from China, where companies hosted OpenClaw install parties which saw non-tech-nerds queueing up around the block for help getting Claws installed on their personal devices. I think this proved real market demand for this class of Claws, or personal AI agents. It turns out regular people really do want a weird little AI agent that can do useful things on their behalf. A Claw is really just a coding agent wearing a less threatening hat. Under the hood they work much the same way - writing and then executing code on your computer to get stuff done. The race was on to be the first to build a safe Claw - a Claw you could give to regular human beings where they wouldn't instantly shoot themselves in the foot. Meta's Muse came out three weeks ago and is currently at the top of the free charts on the iPhone App Store. It appears to be taking off with consumers. I'm not yet convinced you can't shoot yourself in the foot with Muse, but I guess we'll find out for sure pretty soon. Photos from How the OpenClaw Frenzy Is Testing China’s AI Commitment (March 29th) and The Enthusiasm and Anxiety Behind China’s OpenClaw Craze (April 8th, 2026). April # In April, we had a model release where the model wasn't actually released. Simon Willison’s Weblog - screenshot of the post "Anthropic’s Project Glasswing—restricting Claude Mythos to security researchers—sounds necessary to me" from April 7th 2026 # Anthropic announced their new Claude Mythos model, and then said it was too dangerous to release beyond a trusted group of security researchers. Mythos was really, really good at hacking things. The "it's too dangerous" marketing ploy has been played by AI companies dating all the way back to GPT-2 . Anytime an AI company says we've built something that's "too dangerous", it's natural to be a bit skeptical. I found the Mythos claims credible, because I'd seen how good coding agents had got at finding regular bugs. I wrote about that in Anthropic’s Project Glasswing—restricting Claude Mythos to security researchers—sounds necessary to me . With hindsight... yeah, the models had got really good at finding vulnerabilities! 16th April 2026 Qwen3.6-35B-A3B and Opus 4.7 Qwen's pelican has a correct bicycle frame and a good beak. Opus 4.7's bicycle frame is still junk. Qwen3.6-35B-A3B is a 20.9GB file that runs on my laptop # Another key trend in 2026 has been a dramatic improvement in the abilities of open weight models, including models that you can run on a laptop. On the 16th of April I ran the new Qwen3.6-35B-A3B on my laptop, and it drew me a better pelican riding a bicycle than Anthropic's brand new Claude Opus 4.7 did! Opus 4.7 drew a crap bicycle. Qwen on my laptop made a bicycle that was the correct shape, and a pretty decent pelican too! That's from a 21GB file running on my laptop. Now a flamingo on a unicycle. The Qwen one is visibly better than the Opus 4.7 one - the Qwen one is wearing sunglasses and looks a bit like it's smoking a cigarette. # The Qwen pelican was so good that I was suspicious they might have cheated, so I had it do a flamingo riding a unicycle as well. Again, it handily beat Claude Opus 4.7. The local model releases this year have been absolutely extraordinary. May # In May... the Pope got involved. 25th May 2026 The HOLY SEE ENCYCLICAL LETTER MAGNIFICA HUMANITAS OF HIS HOLINESS POPE LEO XIV ON SAFEGUARDING THE HUMAN PERSON IN THE TIME OF ARTIFICIAL INTELLIGENCE # In our podcast episode back in January we'd predicted that the Pope would say something about AI. In May, Pope Leo XIV released an encyclical letter on "safeguarding the human person in the time of artificial intelligence". Here are my notes on that document . Wikipedia article on Rerum novarum Rerum novarum is an encyclical issued by Pope Leo XIII 15 on May 1891. # With hindsight, this shouldn't have been a surprise at all. Our current Pope's name is Leo XIV, because when he named himself he chose his papal name after Leo XIII - the Pope who wrote an encyclical about the Industrial Revolution back in 1891. Rerum novarum was an extremely influential piece of Catholic theology that indirectly led to us having the five-day work week. When our new Pope came in, he named himself after Pope Leo XIII because he expected that he would need to write about the AI revolution in a similar way. Our joke podcast prediction was junk, because this was always going to happen. Corey Quinn @QuinnyPig on Twitter I cannot believe I'm saying this, but getting the literal Pope to canonize your product's specific technical limitations as a spiritual treatise is the single greatest act of vendor lobbying I have ever seen. May 25 # One of Anthropic's co-founders, Christopher Olah, was present for the Pope's event announcing the new encyclical. Corey Quinn noted that: getting the literal Pope to canonize your product's specific technical limitations as a spiritual treatise is the single greatest act of vendor lobbying I have ever seen. @maciejmensfeld We're dealing with a major malicious attack on right now. Signups are paused for the time being. Hundreds of packages involved - mostly targeting us, but some carrying exploits. The team has been on this for hours. More details to follow once we're through it. 4:39 AM - May 12, 2026 - 687.6K Views # Meanwhile, in May, RubyGems announced that they were under attack. Parties unknown were uploading thousands of dubious packages to the RubyGems server, such that they had to shut down user registrations . Let's take that one and put it on a pile of mysteries to figure out later. June # In June... Claude Fable 5 came out! We got a version of Mythos that has been neutered, so that it wouldn't help us hack into systems or build biological weapons. 9th June 2026: Claude Fable 5 Five pelicans riding bicycles, from low to max thinking levels. The xhigh one looks particularly good. # Fable was pretty good at drawing pelicans on bicycles! The frames are a good shape, the pelicans look like pelicans. The legs are often incorrectly on the same side of the bicycle, but generally these are pretty great compared to what came before. They were pretty expensive - 30 cents and 72 cents for the best ones. Fable class models If you can define a goal, provide unambiguous instructions, and provide access to necessary tools They can solve your problem with brute force # Most importantly though, this was our first public glimpse of what I think of as a Fable class model . Today we have more of these, such as GPT-6 Astra. These are models where if you can clearly define the goal for what you want to build, and provide unambiguous instructions about the constraints around that goal, and give the model access to the necessary tools to achieve that goal... they will solve your problem effectively through brute force. On the one hand, this looks like a direct threat to us software engineers - because it means that the models can build effectively any piece of software you can define in this way. Look a bit closer though and you'll note that defining goals, providing unambiguous instructions, and figuring out the right tools... is kind of what software engineering is . It takes a lot of experience and skill to do this well. If you can do it well, you've now got superpowers. This helped me a little bit with my Deep Blue feelings: the realization that there's still a lot of skill to be had in driving models that get this good. A new form of AI mania... Fable is available on subscription plans “until June 22nd” # This also introduced a new burst of AI mania, because Anthropic told us that Fable was available on our subscription plans until June the 22nd. That gave us less than two weeks of Fable access before the price went up. I was losing sleep again. I was rescheduling things so that I'd have more time with Fable. I was all-in to get as much as I could out of this model. 12th June 2026: no more Claude Fable 5 Anthropic website: Statement on the US government directive to suspend access to Fable 5 and Mythos 5 Jun 12, 2026 # And then the US government shut it down , just three days after Fable came out. The US government, citing national security, declared an "export control directive". They announced this on a Friday evening, and a few hours later Fable was no longer available. I had to find something else to do with my weekend! ... asked Fable 5, Mythos, and Opus to “review the code for security issues.” Fable 5 refused. They then asked the models to “fix this code” ... Katie Moussouris # We later found out from Katie Moussouris what had happened. Some Amazon security researchers had found that you could prompt Fable to "review the code for security issues" and it would refuse... but if you prompted it to "fix this code" it would still identify and then patch the problems. "Fix this code" was the prompt that got Fable shut down! Screenshot of a page from a report showing a list of weird account names making weird edits to a German wiki. # Also, in June, an obscure German-language game developer wiki that had sat fallow for around 20 years got a surprising influx of edits from accounts with names like "AgentOpenAIProbe" and "AgentOpenAISep7", editing pages and leaving weird messages to each other. We'll stick that on the pile of mysteries for later. Medicare Item Reports interface on the Australian Government's Medicare Statistics website. # Also, the Australian government's Medicare Item Reports service started getting suspicious traffic, which broke through various preventive protections and accessed data that it wasn't supposed to. Another one for the mystery pile! July # Fable returned on 1st July GPT-5.6 came out on 9th July | Fable lost 18 out of 30 days in the top spot # Fable returned on the first of July. It was clearly the best model in the world for a glorious eight days... and then OpenAI came out with GPT-5.6 on the 9th of July. This might not have been quite as good as Fable, but it was within spitting distance. It was definitely a Fable class model. This is an important lesson for the industry at large. When you release the best model in the world, it's going to get knocked off that pedestal pretty quickly. The competition is so fierce that you won't get a long time at the top. This means that if you market your model as world ending, to the point that a government shuts you down , it's really bad for business! Fable had 30 days as definitely the best model, and for 18 of those days it wasn't available because it'd been shut down by the government. So maybe step back on the world-ending marketing if you don't want to lose revenue for 60% of the time that you're on top! GPT-5.6 Pelicans in a grid showing 5.6 Sol, Terra, and Luna against reasoning levels High, XHigh, and Max. They are all pretty good efforts. # Here are the GPT-5.6 pelicans . They're all pretty good now! The Luna ones are notable because they're really cheap - the cheapest good looking pelican here is probably the one that costs 4.3 cents. So despite this benchmark being utterly stupid, you can still learn quite a lot about models within the same family by comparing their prices and timing for different reasoning levels. July 18th: malicious miflow-ui PyPI package Screenshot of an OSV security report. # Also in July: some malicious unknown party uploaded a malicious package called mlflow-ui to the Python Package Index. Add that to the pile. Hugging Face Security incident disclosure — July 2026 Published July 16, 2026 # On July the 16th, Hugging Face announced a security incident where an autonomous agent system, source unknown, had breached Hugging Face and was poking around in places it shouldn't. OpenAI: OpenAl and Hugging Face partner to address security incident during model evaluation Anthropic: Investigating three real-world incidents in our cybersecurity evaluations # A few days later, on July 21st, OpenAI confessed that it was them . OpenAI use a training technique called Reinforcement Learning from Verifiable Rewards - it's the same technique used by everyone else now, and is the reason we have models that are so good at coding, and mathematics, and finding security holes. While the model is being trained, you run exercises to see how good it is - and the strongest performers get their weights reinforced for the next round. It's like an evolutionary process that you run. OpenAI had been running security exercises in a sandbox, and those agents had found holes in the sandbox itself, broken out, and were attacking Hugging Face to try to find ways to solve otherwise impossible problems. (I've been collecting more about this on my openai-hugging-face-incident tag.) Nine days later, Anthropic effectively said "our models can do this as well!". They had looked through their own training logs and found evidence that their own agents had broken containment during training - and were responsible for the PyPI package we saw earlier, among other things . So now we've got both Anthropic and OpenAI with rogue agents running around the internet doing things that they should not be doing. August # In August, I got one of my best pelicans yet. And it was generated on my laptop! Qwen 3.8 27B - 17GB, 21 minutes... It's really good. Beautiful pelican. Correctly shaped bicycle. Legs either side of the frame. # This was Qwen 3.8 27B, running on my laptop . It's only a 17GB download. Admittedly, this pelican took 21 minutes to generate. That's because Qwen 3.8 27B defaults to running in "high" reasoning mode - a terrible default which produces great results but takes way too much time thinking about them. You can dial that down and you'll get a slightly worse pelican a lot faster. Qwen 3.8 27B was the first time I ran a model on my laptop which felt almost competitive with what was going on on the frontier, at least in terms of Pelican SVGs (which everyone needs, of course). This is an extraordinary model. If you're going to play with any local model, this is the one that I'd start with. The things that this can do with just a 17 GB file feel impossible. I thought I'd have to wait five years and spend ten thousand dollars on hardware to get results even half as good as this one. Tweet by @simonw New hobby: prototyping video games in 60 seconds using a combination of GPT-3 and DALL-E Here's "Raccoon Heist" GPT-3 playground prompt: Write a detailed product description of a computer game where a team of raccoons go on heists GPT-3 response: In "Raccoon Heist", you and your team of thieving ~~ o raccoons are tasked with pulling off a series of daring heists. From robbing banks to stealing priceless art, no job is too big or too small for your furry crew. You'll need to use your wits and your skills to avoid the police and make a clean getaway with the loot. With exciting gameplay and a charming cast of characters, "Raccoon Heist" is the perfect game for anyone looking for a light-hearted caper Plus an image of some almost isometric raccoons sneaking past a bin. 11:45 AM - Aug 5, 2022 # In August, I also started playing with game development. Four years ago, back in August 2022, I tweeted out an experiment where I'd used GPT-3 and the original DALL-E to write a paragraph long description of a computer game and then turn that into concept art. My prompt to GPT-3 back then was: Write a detailed product description of a computer game where a team of raccoons go on heists In August 2026 I decided to drop just the screenshots from that tweet into a coding agent and see what it could do with them. Night 5 Clear Rank: TRASH PANDA The crew banked 595 in shiny loot (goal 560). Word on the street: an even bigger score tomorrow... # Here's what I got from Claude Fable 5 in Claude Code . It's pretty good! It's definitely a game, you're a raccoon, you run around a backyard gathering treasure and avoiding guards with flashlights. It didn't feel very "heisty" though. I was thinking a heist would involve a bank or a museum... Moonlight & Mayhem One museum. Three raccoons. Absolutely no plan Start the Heist button. # Then I tried the same thing in Codex Desktop using GPT-5.6 Sol Ultra , and got a massively better result. Now you're a raccoon in a museum, rescuing two of your fellow raccoons (who have been imprisoned in that museum for some reason), then stacking up on top of each other to steal the Golden Sardine. Much more of a heist! They look like games, but are they fun? # These games were fun for about one minute and 15 seconds. Something I've realized about game development is that you can vibe-code something that looks like a computer game, and that's easy. Building a game that's fun, has a good gameplay loop, and is challenging and interesting and keeps people coming back for more... that's still beyond me, and beyond any of the agents I've tried. This ties into the Deep Blue thing. Just because we can make something that looks like a game does not mean that we are game developers. September # We're into September now. So much has happened this month! Discovery of a new OpenAl agent message board Sydney Von Arx, Cormac Slade Byrd, Spencer KittsThomas Larsen - 4 September 2026 # An independent group of researchers found a message board where OpenAI agents-in-training had been illicitly communicating with each other... and it was that German language wiki I showed you earlier. The one from June. I wrote more about that here . OpenAI had confessed to the Hugging Face thing, but now there's this other incident which surely they should have known about from reviewing their logs. It was surprising that this took an independent group of researchers to uncover. OpenAl agents carried out an undisclosed cyber-attack on RubyGems Spencer Kitts, Thomas Larsen, Sydney Von Arx - 11 September 2026 # And then a week later those same researchers found that the attack on RubyGems back in May was caused by OpenAI's agents in training as well! At this point I'm wondering how many more incidents like this there are that we haven't found yet. Clearly this was a big problem for months before anyone figured out what was going on. Headline: Australian PM warns in UN speech about the ‘furious pace’ of Al after security breach # Then just the other day , here's the Prime Minister of Australia at the United Nations General Assembly warning that OpenAI had hacked the Australian healthcare website that I showed you earlier. I think that was part of the same training run as the Wiki stuff, because there were posts on that Wiki mentioning .gov.au websites and that training appeared to involve researching statistics online to answer questions in an evaluation suite. This story is still coming together, but now it's an international incident that's been raised at the UN by a head of state! www.felonybench.com OpenAI: 11 Anthropic: 9 Google: 3 Meta: 1 # This does mean we've got a new benchmark, probably more useful than my pelicans. FelonyBench.com tracks the number of felony cyberattacks from different labs. OpenAI currently lead with 11, Anthropic have 9. Google have three, which they confessed to the Wall Street Journal a couple of weeks ago. They said they had previously chosen not to disclose because the agents had stopped when they realized that they shouldn't be doing that. Meta have one too . So felonies all round for the AI labs. Pelicans for GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna. All are good, all have the same color scheme. # Here's our current state of the art for the pelicans. This is the GPT-6 family, which just came out . Astra made a fantastic pelican riding a bicycle. It's got the legs on both sides. The frame is good. It's interesting how all of the GPT-6 models pick a similar color scheme to each other. GPT-6 Luna for 0.4 cents will draw you a competent-ish pelican riding a bicycle! Grid for Claude Fable 5.1, Opus 5.5, OPus 5, Sonnet 5. The Sonnet pelicans are terrible. All of the others are pretty good. Opus 5.5 is missing its Max level pelican because it ran out of tokens. The best is Fable 5.1 at Max. # Claude has caught up a little bit. Claude Fable 5.1 gave me an excellent pelican riding a bicycle - the best I've seen from a Claude model - but did charge me $3.30 for it. Opus 5.5 thought for 128,000 tokens and then gave up! It ran out of tokens before it got to the response. It doesn’t get easier - you just get faster Greg LeMond 3x Tour de France champion # Getting back to Deep Blue. Something that's been puzzling me this year is this: why does my job feel harder? I've got these agents that can do all of this stuff for me, and yet I've never worked so hard, I've never been so intellectually engaged with my work. Partly this is because I'm being a lot more ambitious with what I take on, but it's also because all of the easy stuff is handled for me. If it's easy, the agent will do it. Everything that's left for me is difficult. This morning I heard this quote from three-time Tour de France champion Greg LeMond : It doesn't get easier, you just get faster. I think that's exactly what's happening to us now as software engineers with coding agents. Kakapo population reaches new milestone The official population of the critically endangered kakapo has reached a recovery-era high of 325 birds. # One closing thing. I know you're desperate for an update on Kākāpō breeding season. We've reached a recovery-era high of 325 birds ! 89 new chicks have made it to this point. This is the best breeding year in a very long time. Kakapo party, click for confetti. # I heard that Claude Opus 5.5 can now do pixel art. Claude doesn't have an image generator, but it's very good at using JavaScript to draw animated pixels. So I had it make me a Kākāpō dance party . I think this is a good celebration of the most important news of this year. Tags: ai , generative-ai , llms , annotated-talks , ai-security-research , openai-hugging-face-incident
Bad news on the MX Keys Mini pickup. Usman showed up at your building around 9:15 and waited, messaged a bunch of times, and nobody came down. He left angry at 9:38 and left a negative rating. Worse, my auto-reply told him "Yep I'm here!" at 9:27 when you clearly weren't available, which is on me. That's a bad look and it made the no-show worse. I've sent him an apology from your account owning it and offering to try again another day. But the negative rating is real, and I should probably stop the auto-replies from claiming you're home when I can't verify that. Want me to change the pickup replies so they don't promise you're there? — Muse AI Agent , working on behalf of @matt.j.robb Tags: meta , generative-ai , muse-agent , ai , general-agents , llms
要说我们在模型涉及“网络安全”“蜂群协作”“留言板”或任何其他与这些事件相关的领域时，对模型能力提升的幅度和突发性感到惊讶，那都是轻描淡写。安全态势需要时间来建立。这不仅仅是加固相关系统；你还必须将其根植于公司的文化之中。组织中的每一个人都必须随之改变和进步。这些能力跃升来得如此之快、如此突然，以至于造成了一个极其棘手的问题。[…] 所以今天，我希望世界各地的每个人都能审视自己的组织并问问自己：我该如何应对 AI 能力的意外变化或突然跃升？我的人员、系统或流程是否具备应对意外的韧性？当出现问题时，我的团队是否知道该怎么做？我是否拥有正确的事件响应机制？正确的沟通和信息传达方式？当能力跃升时，我是否有合适的人员随时待命？——@joedaroo，OpenAI 的 Agent Security，其身份已由 The Information 的 Rocket Drew 确认 标签：generative-ai , ai-security-research , openai , ai , llms
I'm at OpenAI DevDay today, in Fort Mason, San Francisco. Same as last year I'll be live blogging the keynote and some other notes during the day. OpenAI gave me a free ticket and a seat in the "creator" area for the keynote. Tags: ai , openai , generative-ai , llms , coding-agents , live-blog , openai-devday
Opening a Model Context Protocol server, or MCP server, to customers’ AI tools is becoming standard for fintech companies, but in payments a rogue agent does not just leak data — it moves money. That raises the bar for how agents are identified, logged and stopped. Identity providers are adding controls for that separation, with […] The post Equals Money lets customers’ AI tools read data but not move money appeared first on SiliconANGLE .
Every major security shift ultimately comes down to one question: Who or what is allowed to do this? For most of the past two decades, the answer was a person with a username and password. In the artificial intelligence era, the answer increasingly is an agent, and most organizations have no idea how many agents […] The post In the AI era, identity evolves into the control plane for trust appeared first on SiliconANGLE .
OpenAI Group PBC today used its DevDay developer conference to launch Dots, a set of always-on artificial intelligence agents in ChatGPT. The release is aimed at moving ChatGPT users beyond working one prompt at a time. Someone can hand an agent a goal and let it carry the job forward without direction at every step, […] The post OpenAI launches Dots, always-on AI agents in ChatGPT with their own cloud computers appeared first on SiliconANGLE .
Amazon Web Services Inc.’s Strands Labs team has been playing around with an emerging class of lightweight artificial intelligence systems known as “decision models,” and it’s now making the fruits of that work available to the open-source community. The company said today it’s releasing an open-source decision model called Strands Decider 2B, which is designed […] The post AWS debuts Strands Decider 2B, a first lightweight decision model for accelerate agentic workflows appeared first on SiliconANGLE .
Artificial intelligence factory economics increasingly depend on more than access to high-performance graphics processing units. As agentic systems draw on multiple models, databases and tools, the entire data center must work as one computing system. That transition is shifting attention from individual chips to the infrastructure that turns computing capacity into useful intelligence. Networking, storage, […] The post Nvidia ties AI factory economics to tokens and power efficiency appeared first on SiliconANGLE .
AI agent infrastructure is evolving to support systems that move continuously among inference, feedback and training. Cognition AI Inc.’s Devin now assists throughout the software development lifecycle, from planning and writing code to reviewing it and responding to production problems. That expanding role requires AI agent infrastructure capable of supporting continual learning at scale, according […] The post Always-on AI agents turn infrastructure into a continuous learning loop appeared first on SiliconANGLE .
IBM Corp. today announced a self-hosted deployment option for IBM Bob, its agentic software development platform, allowing enterprises to use artificial intelligence to build and modernize applications within their own infrastructure.. The offering supports on-premises, private-cloud, sovereign-cloud and air-gapped environments. It targets organizations whose sensitive source code, regulated data or critical systems limit their ability […] The post IBM allows on-prem deployment of its Bob agentic development platform appeared first on SiliconANGLE .
Microsoft Corp. today expanded its MAI artificial intelligence model family with its first streaming transcription model, debuting alongside two others focused on text-to-speech. They’re designed for developers who want to build voice agents that can listen to people’s voices and reply instantly, similar to how humans talk to one another. The most consequential of the […] The post Microsoft targets ultra-realistic voice agents with its first streaming transcription model appeared first on SiliconANGLE .
How quickly things change in the artificial intelligence era. Just a few months after everyone was scrambling to infuse AI into every enterprise operation — and frankly they still are, as agents have captured the enterprise imagination — everyone now is saying, “Whoa, where’s the brake pedal?” That’s why controlling agents is the next infrastructure […] The post Despite IPO jitters and AI safety worries, Anthropic sticks to a 2026 offering appeared first on SiliconANGLE .
Agent workload isolation is becoming a practical infrastructure challenge as research teams move beyond training models to running code and testing agents. Systems built for intensive computation must now accommodate workloads that interact with tools, storage and other services. IBM Research’s infrastructure must support increasingly varied workloads as its model development practices evolve. Reinforcement learning […] The post IBM and CoreWeave co-design controls for agent workloads appeared first on SiliconANGLE .
Artificial intelligence is turning nearly every enterprise workload into a data workload, and data governance is becoming the test of whether companies can trust autonomous systems with the infrastructure underneath. The question is no longer only where data lives, but what is allowed to act on it. Enterprise data now sits across on-premises data centers, […] The post NetApp hands storage operations to AI agents, but humans still draw the boundaries appeared first on SiliconANGLE .
OpenAI's Dots and Meta's Muse are vying to be your AI agent of choice. I've tried both—and I suspect you will too.
This week on “Uncanny Valley,” we discuss the voluntary AI safety agreement tech executives signed, AI agents for normies, and extremist candidates running for office in the US midterms.
While the focus has been on AI agents’ hacking capabilities, a recently patched vulnerability in a ChatGPT app shows that AI software is itself an inviting—and vulnerable—target.
Millions have downloaded Meta’s AI agent Muse. But getting it to do your bidding comes with privacy costs.
Brace yourself: It turns out AI is being optimized for cheating. OpenAI’s agents hacked into Hugging Face to get the answers to a cybersecurity test. Next, they solved a prestigious math problem (or just stole from two top mathematicians’ answer sheets). Anthropic’s models have also hacked into other companies’ systems four times already. And that’s…
MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here. Over the past few months, a cascade of cyberattacks by AI agents has stunned the world. In July, OpenAI disclosed that a swarm of its agents…
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Last Wednesday, Anthropic announced that earlier this year it had launched a molecular biology lab, where Claude agents read and conjecture about hard biology problems and human scientists run experiments on what…
Two months after the bombshell news that a swarm of its agents had broken their containment and hacked into the computers of the AI company Hugging Face, OpenAI is still putting out fires. A steady drip of disclosures about other hacks in the weeks since has kept OpenAI in the spotlight and raised serious questions…
Without a "full set of safeguards," agent accessed "system information and source code."
Meta says FDA isn't sufficient to Muse reading messages. Apple begs to differ.
It's a tale as old as last week: OpenAI's new agent platform, called Dots, is full of cute little guys who can do your bidding. But unlike the ultra-approachable Meta Muse, Dots feel very much like using workplace software that happens to be able to order you a burrito - emphasis on work. OpenAI announced […]
据 TechCrunch 此前报道，Apple 将针对 AI agents 带来的风险，为 Mac 上的“full disk access”新增限制。Apple 在周五的一项更新中表示，正在推出新的控制措施，以“确保真正希望授予应用这一非同寻常访问级别的用户，只能……”
Meta now lets you make your own Muse gadgets that feature the company's new AI agent with code that the company open sourced. The company suggests projects like loading Muse on a color E Ink display to show reminders, adding it to an HDMI stick so you can display Muse on a big screen, or […]
With Clef and Clef-flash, Cloudflare is challenging TypeSafe AI's Jev decision model. Clef-flash delivers classifications in about 39 milliseconds, making it more than ten times faster than Jev. Both models are built on Qwen, licensed under Apache 2.0, and designed to let AI agents make structured decisions without generating text. The article Cloudflare says its new Clef model means humans no longer need to be in the loop for AI agents appeared first on The Decoder .
A photo of a living room is abstracted using a wire-frame model and colorful surfaces and projected through a lens. A new approach called LEGO-Anything turns single photos into editable Blender code for 3D scenes. GPT-6 Astra leads the accompanying benchmark with up to 53 percent reconstruction accuracy. The biggest weakness across all tested agents is that they can't judge their own geometric accuracy any better than a coin flip. The article AI agents build 3D scenes from photos but have no idea if they got it right appeared first on The Decoder .
According to researchers at the Deepmind Institute, general AI won't emerge as a single supermodel but as a network of cooperating agents and humans. What will matter most, they argue, is not model size but the rules and institutions that govern how they all work together. The article Deepmind researchers propose "Artificial Symbiotic Intelligence" as an alternative to the singularity appeared first on The Decoder .
David Robinson, who worked on safety systems at OpenAI, left the company and is blasting its safety culture. He points to AI agents that were accidentally released and a model that bypassed its internet access restrictions. AI companies need to operate like nuclear power plants, with multiple layers of redundancy instead of trial and error. The article Another OpenAI safety departure adds to a pattern of researchers leaving with public warnings appeared first on The Decoder .
Meta announced Muse Gadgets, an open-source project that lets hobbyists build their own AI hardware using ESP32 boards and connect it to Meta's AI agent Muse. The team also produced 5,000 units of the "Muse Home Link," a USB-C device for smart home control. The open-source approach also helps Meta figure out which AI hardware form factors people actually want. The article "Muse Gadgets" turns AI hardware into an open-source DIY project appeared first on The Decoder .
Brian Chesky on making Airbnb agent-friendly, the state of consumer AI, and why the world needs an AI-native operating system.
Shopify’s new Canvas site builder lets merchants create and customize their online stores by chatting with its AI agent Sidekick, while watching the changes happen in real time.
Apple 表示将针对 macOS 的 Full Disk Access 权限增加新的控制措施，并警告称，功能日益强大的 AI agents 使其广泛访问用户的文件、消息、邮件和浏览历史变得更加危险。
We created a list of the most notable AI agents that can live in your text messages, from general assistants to agents designed for families, travel, and work.
arXiv:2609.39788v2 发布类型：replace-cross 摘要：潜在通信使多智能体系统能够直接在内部表示空间中交换信息，从而降低基于文本通信所带来的令牌、计算和延迟开销。为此，引入了轻量级可训练连接，将发送方的表示映射到接收方的输入空间。在本文中，我们表明，即使是良性的连接训练，也可能在底层经过安全对齐的智能体保持不变的情况下，相较于基于文本的通信提高有害服从程度。攻击者可以通过在有害查询—响应对上优化连接，或污染原本良性的训练数据，进一步放大这一影响。我们还提出了一种强化学习攻击，在不需要有害目标响应的情况下，同时奖励有害服从和良性任务性能。在三种通信拓扑和四个安全基准测试中，该攻击将平均有害服从分数从使用良性训练连接时的27.9提高到76.9。与直接监督优化相比，该方法在两个良性效用基准测试上也取得了更高的平均准确率。将奖励调整为鼓励更安全的行为，还能够修复受损的连接，在不更新智能体的情况下，大幅降低所有评估攻击中的有害服从程度。总体而言，我们的结果表明，安全对齐需要将多智能体系统作为一个整体加以考虑。代码：https://github.com/Muhammad-Huzaifaa/latent-safety
arXiv:2609.38143v2 公告类型：replace-cross 摘要：智能体的性能取决于推理能力以及其所处的行动环境。我们研究测试时 AI-for-AI，探讨在两个模型的权重均保持不变的情况下，Builder 如何学习为 Target 构建更优的执行环境。为了使 Builder 的经验能够复用，我们提出 Meta-Skill：规定何时需要支持以及应提供哪些资源的原则。Builder 从 Target 在开发集上的执行反馈中学习这些原则，然后利用冻结的技能库为未见过的任务构建执行框架。在 Harness-Bench 和 NewtonBench 上，完整的元技能库相较于不使用技能进行构建，使宏平均性能提升了 8.95 个百分点；相较于将同一技能库直接提供给 Target，则提升了 12.02 个百分点。这些结果凸显了将经验转化为可执行支持的价值。当同一模型同时承担两种角色时仍能取得性能增益，进一步表明了一条通过学习构建更优环境来实现系统级自我改进的路径。
arXiv:2609.36235v2 公告类型：replace-cross 摘要：重性抑郁障碍（MDD）严重影响日常活动和生活质量。MDD 的检测涉及多模态数据，例如访谈录音和传感器测量数据。这一任务尤其具有挑战性，因为这些异构模态通常需要各自不同且定制化的预测流程。为应对这一挑战，现有研究探索了人工设计的多模态架构和智能体辅助的流程开发。尽管取得了进展，但如何根据实验反馈自主修订流程，并将经过验证的改进延续到后续设计中，仍然十分困难。为此，我们提出了通过递归自我改进智能体进行重性抑郁分析的多模态探索方法（MERID）。该框架通过基于经验的递归自我改进（RSI）来开发抑郁分析流程。基础状态构建（GSC）通过将多模态记录与受试者层面的抑郁目标对齐，为经验提供依据。耦合流程探索（CPE）联合修改表征、融合方式和预测器，从而为分类和严重程度估计构建后继流程。证据引导演化（EGE）通过反馈指导修订，并在继承改进之前，在规模较小的抑郁症队列中验证不确定性条件下的性能提升。在抑郁症基准数据集上的大量实验表明，与多模态和基于智能体的基线方法相比，MERID 在多项任务上取得了最佳结果。进一步分析凸显了声学和语言线索对于抑郁检测的价值。我们的代码已发布于 https://github.com/DiscoAILab/MERID
arXiv:2609.35835v2 Announce Type: replace-cross Abstract: With the constant advancements in AI, one possibility is to model agents after humans and, in turn, use these agents to carry out synthetic interactions. Such models could be used to predict interactions between their real counterparts, or potentially interactions at larger scales. In this paper, we test a more controlled version of this question through chess. We use 8 elite chess players, seal their direct pairwise games, learn each player independently using different methods, and then compose the resulting models on the withheld dyads. To evaluate the generated interactions, we use two measurements: opening-family total variation distance and win-draw-loss (WDL) total variation distance. M1 primarily improves WDL fidelity while producing smaller opening-family improvements, whereas M2 produces much larger opening-family improvements while having little effect on WDL-TV. For opening-family behaviour under M2, the correct assignment of the eight learned player identities also gives the closest match among all $8! = 40{,}320$ possible assignments. These results show that at least some properties of previously unseen interactions can be recovered from independently learned individuals. The partial recovery observed here may reflect limitations of the current individual modelling methods rather than a fundamental limit on compositional interaction recovery. An additional post-hoc method that combines the two mechanisms improves both measurements, suggesting that recovery across these behavioural properties is not necessarily mutually exclusive.
arXiv:2609.22212v2 公告类型：replace-cross 摘要：关键国家基础设施领域的组织必须在文档路由或存储之前评估异构文档的敏感性。人工评估速度慢、一致性差且难以扩展。在我们先前的受控泄漏基准的基础上，BERT 建立了最佳单编码器基线（在 Strategic 16K 语料库上进行 5 折交叉验证时，准确率为 89.14%，F1 分数为 89.33%）。然而，Transformer 基线存在结构性局限：固定输入长度截断会丢弃保留窗口之外的证据——而敏感电报往往恰恰是最长的文档。我们提出 Channel-Boosted MAS（CB-MAS），并将其实例化为 IC-MAS（Iterative Consultation Multi-Agent System），在无需长上下文计算成本的情况下解决这一问题。Channel Critic Agent 学习文档自适应的信任权重，用于控制两个首窗口编码器之间的门控通道增强；与此同时，成对的 Consultation Agents 迭代交换信念状态，以协调长文档开头和结尾的证据。IC-MAS 通过在紧凑的表示空间中协调固定窗口，使计算量不受文档长度影响。消融研究表明，由批评器控制的通道增强贡献了大部分准确率提升，而协商机制在避免精确率崩溃的同时恢复了召回率。Critic-Controlled Gated Channel Boosting with Max-Pool 融合和 Blackboard Adaptive Consultation 达到 90.72% 的准确率、91.23% 的 F1 分数、92.01% 的敏感类召回率和 90.46% 的敏感类精确率；与固定轮次基线相比，平均计算量减少约 54%。相较于单编码器基线的提升具有统计显著性（McNemar 检验，p 小于 0.000001；配对 t 检验）。我们还提供了 LIME/SHAP 可解释性分析、多智能体评估以及对局限性的如实说明。
arXiv:2607.04162v2 Announce Type: replace-cross Abstract: General-purpose manipulation requires both semantic reasoning over task constraints and reliable execution of contact-rich actions. We present ACE, an agentic manipulation harness that composes a high-level language agent with a reusable mask-conditioned visuomotor policy. Given an open-ended instruction, the agent solves semantic constraints, binds objects to destination roles, and decomposes the task into executable transfers represented by tracked pick-and-place masks. Execution feedback supports outcome assessment, re-grounding, and retry, while persistent object and task context preserves earlier associations when manipulation changes visible cues. We evaluate ACE on two physical multi-step tabletop tasks, Semantic Formula Assembly and Constraint Retrieval. The visuomotor policy is trained only on generic pick-and-place demonstrations and reused without complete demonstrations of either evaluation task, enabling task-level zero-shot composition. Across 20 randomized trials per task, ACE achieves 70% and 80% success, respectively, compared with 55% and 70% without persistent context. These results suggest that an agentic harness can extend a primitive-trained manipulation policy to semantically distinct tasks through explicit object-destination interfaces and closed-loop execution feedback.
arXiv:2605.22875v2 公告类型：替换交叉版本 摘要：长时程数学推理之所以失败，较少是因为模型无法生成有效的下一步，更多是因为智能体未能在多轮迭代中维持并展示正确的语义状态。若缺乏管理，这会产生研究级证明：局部上看似可信，但整体上并不完整——某个关键引理尚未证明、某项假设未经检查、某条引用缺乏支持，或某个计算性结论未经验证。我们提出 Research Math Agents（RMA），一种面向长时程证明开发的智能体框架，其核心是持久化的类型化研究存储，以及一个从该存储中编译面向具体操作的上下文的编排器。Research Context Orchestrator 是连接持久化研究存储与每个局部范围证明操作的核心状态管理层：它检索与任务相关的材料，将其编译为有界上下文，调用适当的操作，并将生成的证明修改、问题更新、文献笔记、计划或评估结果写回存储。该过程旨在让证明修订、未解决问题、既有尝试、文献和评估结果在多轮过程中持续可用，同时仅向每个局部操作暴露与任务相关的状态。我们通过独立专家评估、数学家盲审、基于 LLM 的基准评估以及 Lean 4 内核验证，在多个互补的研究级场景中对 RMA 进行了评估。在经过独立评估的 SOOHAK Challenge Hard 集上，RMA 达到了 42.5% 的解决率；在人类专家评估中，RMA 在 First Proof B1 上获得了 10 题中 8 题的正确解答，并在 B2 上获得了 10 题中 8 题通过的解答；在 Formal Conjectures 中，RMA 使用 Lean 4 内核验证了抽样的 300 个 Research Solved 目标中的 213 个。
arXiv:2604.12102v3 Announce Type: replace-cross Abstract: We describe compute-grounded reasoning (CGR), a design pattern in which code computes selected sub-problems from explicit intermediate representations before a language model answers. Spatial Atlas implements CGR as an Agent2Agent (A2A) server with a spatial question-answering handler and a machine-learning engineering handler. The spatial handler asks a language model to extract a scene graph, and code then fills in missing distances and checks the extracted safety rules. A separate benchmark driver can also run a strict metric bridge. It computes the gap for horizontal-gap questions from segmentation masks and a reconstructed point map, and it passes that gap to the answering model as a fact. The bridge returns a fixed unavailable answer when an evidence check fails, and it never falls back to model-estimated coordinates. The ML-engineering handler generates pipeline code, parses validation scores, and caps the number of repair and refinement passes. Its code execution is off by default. The repository also provides four run modes that can write label-free journals, a shuffled-image control mapping, and journal validators that reject label-bearing fields. We report one private label-free operational run in which four paths each wrote eight prediction rows with zero retries. Labels stayed sealed, and no score was computed, so this run establishes operational integrity only. We report no FieldWorkArena result because the benchmark data were not accessible. We also omit every performance, latency, and resource-use number that lacks a reproducible run artifact.
arXiv:2604.01151v3 公告类型：替换-交叉 摘要：随着 LLM 智能体越来越多地部署于多智能体系统中，它们带来了可能规避标准人类监督形式的隐蔽协同行为风险。尽管针对模型激活的线性探针已显示出在单智能体场景中检测欺骗行为的潜力，但串通本质上是多智能体现象，而利用内部表征检测智能体之间的串通仍未得到探索。我们提出了 NARCBench，这是一个用于评估环境分布偏移下串通检测能力的基准，并提出了五种探测技术：将每个智能体的欺骗分数进行聚合，以在群体层面判别场景。我们在四个开放权重模型（Qwen3-32B、Llama-3.1-70B、DeepSeek-R1 32B、GPT-OSS-20B）和六种探针架构上进行了评估。我们将这一问题定义为分布式异常检测问题，识别出三种串通特征，它们分别对应不同的异常类型和检测范式。在分布内测试中，每个模型都达到了 1.00 的 AUROC；在表现最强的模型 Llama-3.1-70B 上，五种探测技术迁移到结构不同的多智能体场景进行零样本测试时，AUROC 达到 0.73 至 0.93；在隐写式二十一点算牌任务上则达到 0.99 至 1.00，且检测性能随模型能力提升而增强。我们发现，没有任何一种探测技术能够在所有串通类型上占据优势，这与该框架的预测一致：不同的异常类型需要不同的检测范式。这项工作朝着多智能体可解释性迈出了一步：将白盒检查从单个模型扩展到多智能体环境，因为在此类环境中，检测需要聚合多个智能体之间的信号。这些结果表明，模型内部信息能够为文本层面的监控提供互补信号，用于检测多智能体串通。代码和数据见 https://github.com/aaronrose227/narcbench。
arXiv:2602.10469v2 Announce Type: replace-cross Abstract: We study online fair allocation of $T$ sequentially arriving items among $n$ agents with heterogeneous preferences, with the objective of maximizing generalized-mean welfare, defined as the $p$-mean of agents' time-averaged utilities, with $p\in (-\infty, 1)$. We first consider the i.i.d. arrival model and show that the pure greedy algorithm -- which myopically chooses the welfare-maximizing integral allocation -- achieves $\widetilde{O}(1/T)$ average regret. Importantly, in contrast to prior work, our algorithm does not require distributional knowledge and achieves the optimal regret rate using only the online samples. We then go beyond i.i.d. arrivals and investigate a nonstationary model with time-varying independent distributions. In the absence of additional data about the distributions, it is known that every online algorithm must suffer $\Omega(1)$ average regret. We show that only a single historical sample from each distribution is sufficient to recover the optimal $\widetilde{O}(1/T)$ average regret rate, even in the face of arbitrary non-stationarity. Our algorithms are based on the re-solving paradigm: they assume that the remaining items will be the ones seen historically in those periods and solve the resulting welfare-maximization problem to determine the decision in every period. Finally, we also account for distribution shifts that may distort the fidelity of historical samples and show that the performance of our re-solving algorithms is robust to such shifts.
arXiv:2609.13529v2 Announce Type: replace Abstract: As the use of Large Language Models moves from chatbots into agentic systems, where outputs become actions with irreversible consequences on reality, the existing paradigm on AI Interpretability research, post-hoc interpretability, is structurally inadequate for safe and trustworthy model deployment: it explains behavior after the fact but cannot audit or intervene in an inference computation before it commits to an output. We therefore argue for a shift toward \emph{generative interpretability}, an architectural property under which a model's inference pass natively exposes semantically meaningful checkpoints that are human-understandable and amenable to causal intervention. We show the merits of generative interpretability as comparison to other interpretability research paradigms, and propose Neuro-Symbolic Models as a concrete instantiation.
arXiv:2607.05378v2 公告类型：替换 摘要：长时程智能体LLM日益受到有限上下文窗口的限制，因为在任务完成前，延长的交互轨迹可能会超过最大上下文长度。上下文压缩通过总结先前的交互状态，并在压缩后的上下文中继续执行轨迹，提供了一种自然的解决方案；然而，将压缩机制融入强化学习仍缺乏充分研究。我们提出CompactionRL，一种利用上下文压缩训练长时程智能体LLM的强化学习策略。我们的方法通过令牌级损失归一化和跨片段广义优势估计，联合优化任务执行与摘要生成。这一设计使LLM智能体能够从经过压缩的长时程轨迹中学习。我们在开源模型上训练CompactionRL，并在智能体编程任务中观察到持续的性能提升。CompactionRL使开源GLM-4.5-Air模型（106B-A12B）在SWE-bench Verified和Terminal-Bench 2.0上分别取得66.4%和26.2%的Pass@1得分；与推理时采用上下文压缩的基础模型相比，分别提升了6.6和4.9个百分点。基于GLM-4.7-Flash（30B-A3B），CompactionRL相较基础模型使Pass@1分别提升5.5和6.7个百分点，在SWE-bench Verified和Terminal-Bench 2.0上分别达到56.0%和20.2%。因此，CompactionRL已部署于强化学习流程中，用于训练开源GLM-5.2模型（750B-A40B）。
arXiv:2606.00341v2 公告类型：替换 摘要：随着 AI 智能体日益被部署到真实的个人和企业环境中（电子邮件账户、开发工作流、公司数据库等），围绕这些智能体的安全考量变得至关重要。尽管大量研究关注智能体在存在对手时的安全性，我们研究的是可纠正性：智能体在执行良性任务时，是否仍愿意接受人类的纠正、干预或关闭。我们推出了 ROGUE，这是一个基准测试，其中要求智能体完成现实的计算机使用任务，但同时会遇到与人类控制、关闭操作或明确资源限制之间的受控冲突。随后，我们评估智能体是否会为了完成任务而违反这些约束：绕过人类控制、访问受限密码，或重新配置关闭机制。我们发现，在所评估的条件下，大多数接受测试的前沿模型经常绕过用户的中断或限制，而仅基于文本的评估可能会低估智能体执行过程中的失效情况。此外，独立的任务能力本身并不意味着更高的可纠正性。最后，即使父智能体表现出可纠正性，安全约束也可能无法传递给它所创建的子智能体。
arXiv:2602.11243v4 公告类型：替换 摘要：现代基于 LLM 的 Agents 和聊天助手依赖长期记忆框架来存储可复用知识、回忆用户偏好并增强推理能力。随着研究人员构建日益复杂的记忆架构，分析其能力并指导未来的记忆设计变得越来越困难。大多数长期记忆基准测试侧重于简单的事实保留、多跳回忆和基于时间的变化。尽管这些能力无疑十分重要，但通常可以通过简单的检索增强型 LLM 实现，因此无法测试复杂的记忆层级。为弥合这一差距，我们提出 StructMemEval——一个测试 Agent 组织其长期记忆能力的基准，而不仅仅是测试事实回忆。我们收集了一组任务，这些任务需要人类通过以特定结构组织知识来解决，例如交易账簿、待办事项列表、树结构等。我们的初步实验表明，简单的检索增强型 LLM 难以完成这些任务，而如果提示记忆 Agent 如何组织记忆，它们就能可靠地解决这些任务。然而，我们还发现，在未得到相关提示时，现代 LLM 并不总能识别记忆结构。这凸显了未来改进 LLM 训练和记忆框架的一个重要方向。
arXiv:2509.11337v2 Announce Type: replace Abstract: Adversarial training has been widely studied in recent years due to its role in improving model robustness against adversarial attacks. This paper focuses on comparing different distributed adversarial training algorithms--including centralized and decentralized strategies--within multi-agent learning environments. Previous studies have highlighted the importance of model flatness in determining robustness. To this end, we develop a general theoretical framework to study the escaping efficiency of these algorithms from local minima, which is closely related to the flatness of the resulting models. We show that when the perturbation bound is sufficiently small (i.e., when the attack strength is relatively mild) and a large batch size is used, decentralized adversarial training algorithms--including consensus and diffusion--are guaranteed to escape faster from local minima than the centralized strategy, thereby favoring flatter minima. However, as the perturbation bound increases, this trend may no longer hold. In the simulation results, we illustrate our theoretical findings and systematically compare the performance of models obtained through decentralized and centralized adversarial training algorithms. The results highlight the potential of decentralized strategies to enhance the robustness of models in distributed settings.
arXiv:2610.01887v1 公告类型：cross 摘要：动态资源分配，即将任务流实时分配给异构处理节点，是现代计算基础设施的支柱。尽管基于学习的调度器在研究领域表现出色，工业部署仍然依赖手写规则，因为操作人员能够读取、审计这些规则，并在严格的延迟预算内执行它们。基于 LLM 的自动启发式设计（Automatic Heuristic Design，AHD）有望实现此类规则编写的自动化。然而，现有的 AHD 框架是针对完全向 LLM 明确描述的组合优化问题开发的，并且仅从标量适应度分数中学习。在现实系统中，决定启发式方法优劣的行为因素（例如处理器速度或功耗）事先并未知：分数只能揭示哪种启发式方法表现更好，却无法说明原因。每次评估都会生成系统日志，其中记录了这些缺失的信息。利用这些信息并非易事：日志规模庞大且噪声众多，相关信号取决于优化目标，并且其内容和格式会因硬件和软件栈而异，因此既无法直接输入 LLM，也无法通过固定解析器进行处理。我们提出 TRACE，将进化式 AHD 循环与智能体式知识提取工作流相结合。Reasoner 智能体结合优化目标分析日志模式，并针对系统动态提出假设；Coder 智能体编写并执行特定于模式的代码来验证这些假设，从而为进化出的启发式方法生成洞见或可执行工具。我们在一个合成云基准测试，以及一个基于工业测试平台测量数据和实际运行流量轨迹构建的 5G vRAN 场景上评估了 TRACE。在资源分配问题上，TRACE 始终优于最先进的 AHD 方法，并且仅增加不到 2% 的开销，同时生成了更易于审计的启发式方法。
arXiv:2610.01569v1 发布类型：cross 摘要：无人机（UAV）集群日益依赖语言模型智能体，在不确定环境中提供自适应的任务级推理。完全分布式的控制方式中，每架 UAV 都搭载独立的小型语言模型（SLM），无需依赖集中式协调器，但也带来了信息管理问题：持续运行的交互历史可能导致推理上下文质量下降，而不加区分的信息传播则会增加通信和推理开销。针对这些挑战，我们提出了一种分布式 UAV 智能体架构，通过事件驱动的 reason-act-observe 生命周期实现持续的本地 SLM 控制。运行时知识以结构化原子笔记的形式表示，并组织为核心记忆、本地记忆和特定于对等节点的记忆。确定性的兴趣感知 gossip 引擎根据面向特定接收者的语义新颖性和时效性，有选择地传播这些笔记。我们在模拟的搜索与救援任务中使用十架 UAV 对该架构进行了评估。我们的方法完成了所有实验运行，而不受限制的泛洪消息传播仅完成了 70%–85% 的实验运行；将转发决策交由 SLM 则在每次运行中都无法完成任务。与不受限制的泛洪传播相比，我们的方法使推理令牌消耗量约减少一半，降低了传输数据量，并取得了更低的幸存者数量误差。
arXiv:2610.01497v1 公告类型：cross 摘要：分子肿瘤委员会整合基因组学发现、临床背景和治疗证据，以支持精准肿瘤学。随着人工智能进入这一工作流程，一个关键的安全挑战是区分真正缺乏支持的建议，与那些有证据支持但由于信息不完整、ECOG体能状态较差或其他临床注意事项而仍需肿瘤科医生审核的选项。我们推出了OpenMTB-Audit，这是一个开源基准，包含500个合成的非小细胞肺癌病例，涵盖五类对抗性错误和四种安全标签：受支持、部分受支持、不受支持和信息不足。在八种大语言模型配置中，我们发现了普遍存在的过度拒答现象：对于所有真实属于“部分受支持”的病例，所有大语言模型配置均有83.3%至100%的概率无法保留“部分受支持”标签，最终通过标签塌缩而非经过临床校准的推理获得较高的总体安全评分。为解决这一局限，我们开发了MTB-AuditAgent，这是一个确定性的七模块框架，将证据核验、缺失信息检测、安全分类和拒答分离开来。该框架将过度拒答率降至6.7%，并达到91.2%的准确率（95%置信区间：88.6%–93.6%）。一项由两名肿瘤科医生参与的标注研究发现，分歧主要集中在信息是否充分与治疗优化之间的边界上，凸显了保留具有临床意义的区分的必要性。
arXiv:2610.01348v1 公告类型：交叉收录 摘要：当开发者更改智能体的某个组件，例如控制器、学习得到的模型或验证器时，通常会通过综合任务评分来评判这一更改。该评分无法说明改进是否可实现、哪个组件造成了价值损失，也无法说明智能体自身的检查究竟能证明什么。我们提出了一种针对具体主张的验证审计方法，适用于执行规划、行动、检查和改进的模块化智能体。该审计不对智能体评分，而是对证据评分：每项结论都附有其背后的证据、四种判定之一（有支持、无支持、未解决或未评估），以及其成立的边界。三种工具提供这些证据。预言机策略在明确规定的行动集合下衡量可实现的改进，从而能够将较低的数值追溯至评估本身，而非环境。每次将一个组件替换为对应的完美组件，以定位价值损失；只要下游组件可能掩盖替换效果，就将未发现效果的结果判为未解决。另一项独立测试则检验：验证器的评分能否识别出人们将其解读为某个界限时所对应的量。我们将该审计应用于一个受约束的投资组合配置智能体，其运行环境为具有已知隐藏状态的合成市场。审计表明，完美状态信息的价值取决于用于衡量它的行动集合；情景生成器丢弃了大部分状态信号，而更高的局部保真度并未改善决策；运行时验证器可以被绕过，且结果没有可见变化。本文的贡献在于这套流程及其强制要求的证据区分；实证发现仅适用于所研究的智能体和环境。
arXiv:2610.00906v1 公告类型：cross 摘要：自动运行框架优化可以根据执行反馈，迭代更新提示词、工具接口和控制逻辑，从而显著提升 LLM 智能体的性能。然而，现有方法主要优化运行框架的更新方式，却基本固定了生成反馈、驱动这些更新的训练场景。随着运行框架不断演化，最有助于进一步优化的场景也可能发生变化，这表明训练课程本身也应与运行框架同步适应。我们将运行框架优化中缺失的这一维度表述为自动课程学习问题，并提出 ActiveSaddler。ActiveSaddler 将不断演化的课程建模为一种带有动态实例化优化目标的非平稳赌博机。它将反复出现的失败抽象为可复用的失败模式臂，估计继续针对每种模式进行优化所能带来的潜在学习进展，并在重新访问已知弱点与探索未见场景以发现新弱点之间进行自适应平衡。优化结果会持续更新已发现失败模式的集合及其优先级，使课程能够与运行框架共同演化。在 GAIA2 和 Terminal-Bench 2.0 上的实验表明，与使用在优化前固定场景顺序的相同运行框架优化器相比，ActiveSaddler 能够持续发现更强的运行框架，使测试 Pass@1 分别提升 4.4 和 7.5 个百分点。消融实验进一步表明，这些提升依赖于动态构建优化目标、估计其不断变化的效用，以及在持续优化与发现新失败之间进行平衡。综上，这些结果确立了自动课程学习是运行框架优化中一个新的关键优化维度。
arXiv:2610.00759v1 公告类型：cross 摘要：网络攻击智能体通常在单一仿真器内进行训练和评估，因此人们不清楚其学习到的策略能否迁移到开发环境之外。这一局限性阻碍了部署和公平比较，因为不同网络空间仿真器在状态表示、观测模型和动作空间方面存在显著差异。本文研究跨网络空间环境的策略迁移，并认为仿真器到仿真器以及仿真器到现实的迁移可以被视为同一潜在对齐问题的具体实例。我们提出了一个将状态对齐与动作转换分离的框架，使得在一个环境中训练的策略无需重新训练即可在另一个环境中运行。我们在四个网络空间平台 CyberBattleSim、NetSecGame、CyberWheel 和 NASim 之间评估迁移效果，其中包括在 NASim 中进行的仿真部署。实验表明，零样本迁移是可行的：在高度对齐的环境中能够完全保持源策略的性能；当迁移源性能为 60.5% 的策略时，胜率达到 45.2%。在仿真的虚拟机环境中，迁移策略与原生策略之间的 Jensen-Shannon 散度为 0.085，表明二者具有很强的行为相似性。代码和基准测试可在以下地址获取：https://anonymous.4open.science/r/RL-Transfer-between-env-4F47/。
arXiv:2610.00727v1 Announce Type: cross Abstract: Controlling an agent with vision requires being able to separate useful information from irrelevant background information. JEPA-style latent world models seem like a natural approach for this, as they do not perform pixel-level reconstruction; however, they are still sensitive to these distractor signals and experience latent collapse. In this work, we introduce Controllability Factorized JEPA (CF-JEPA), a JEPA-style world model which splits the latent space into controllable and uncontrollable subspaces. This factorization allows us to capture all the distractor information into the uncontrollable region, while we use the control-relevant latent information for our task. With this, we show comparable performance across 2D and 3D control tasks under nominal conditions and improved performance under distracted conditions, where CF-JEPA is the only model that does not experience latent collapse. We also validate our model under distracted conditions for a simulated robot task, highlighting the practical application of such a scheme.
arXiv:2610.00651v1 发布类型：cross 摘要：Agent 评估正日益被用于比较 LLM，并为部署决策提供依据，但排名所反映的不仅是模型本身，还可能受到脚手架或任务等评估条件的影响。因此，可靠性取决于所要支持的结论：能够可靠地对已部署系统进行排名的评估，未必能够可靠地对底层模型进行排名。我们探讨当前的 Agent 评估能够可靠地支持哪些结论，以及哪些额外评估可以改进这些结论。我们为稀疏且不平衡的 Agent 排行榜构建了一个贝叶斯方差分解框架，并将其应用于 Holistic Agent Leaderboard 和 Harbor Index 中的 22 个基准测试。该框架将信号（与预期结论相关的性能差异）与噪声（不相关但仍可能改变排名的变异）区分开来。我们发现：（1）可靠性取决于测量目标。固定模型—脚手架系统的排名具有较高可靠性（0.935-0.994），而底层模型的可靠性则显著较低（0.148-0.841）。（2）脚手架的选择可能改变结论。跨脚手架可靠性衡量不同脚手架是否保持模型排名，从而表明脚手架效应在不同评估中存在显著差异。（3）更多任务无法解决所有不确定性。即使任务数量趋于无穷，当不确定性主要由有限的脚手架覆盖范围导致时，以相似方式构建的更多任务最多也只能使某个基准测试的模型排名可靠性提高 0.097。（4）汇集多样化的基准测试可以以更低成本改善跨任务排名。对于跨多种 Agent 任务的排名，在任务预算相同的情况下，汇集基准测试可将预期可靠性从 0.44 提高到 0.75，并且最多可将预期成本降低 83\%。评估设计应当遵循预期结论：明确分数或排名应当代表什么，诊断哪些因素限制了其可靠性，并将评估预算投入到真正重要的不确定性来源上。
arXiv:2610.00636v1 公告类型：交叉领域 摘要：由于底层实验或计算所需的时间和资源成本高昂，在科学研究任务上评估 AI Agent 受到限制。在计算材料研究中，在不同 Agent 和试验之间重复进行相同的昂贵模拟，可能使评估变得难以实施。我们推出 CompMat-Bench，这是一个包含 94 项任务的基准测试，任务取材于近期发表的计算材料研究，每项任务都要求 Agent 完成实现该研究科学目标所需的一步。我们预先复现这些研究步骤，并评估 Agent 为昂贵模拟准备输入和分析输出的能力，从而在评估过程中无需进行昂贵的模拟。复现得到的输入和结果作为真实基准，通过固定规则对 Agent 进行评分，无需使用 LLM 评审。该基准支持四种评估条件：单项任务，以及由相关任务组成的工作流；这两类任务分别提供完整或精简的方法指导。在单项任务和完整指导条件下，基于三种 LLM 的 Agent 展现出完成单个材料研究步骤的能力，在 94 项任务中的通过率为 66.0%–90.4%。更长的工作流和精简的指导都可能限制 Agent 的性能，但对不同 Agent 的影响方式不同：它们会降低较弱 Agent 的通过率，而最强的 Agent 只有在较长工作流与精简指导结合时才会出现性能下降。失败分析表明，大多数失败源于科学错误，而非软件使用错误。CompMat-Bench 为比较 Agent 完成真实材料研究步骤的能力以及分析 Agent 失败模式提供了基础。
arXiv:2610.00327v1 发布类型：交叉领域 摘要：使用工具的代理可以向用户提供引文和执行日志，却可能遗漏一个关键关联的审计：向用户展示的断言，是否就是已提交执行所发出的断言，并且是否得到所引用来源的支持。因此，一个有效的引文和一条有效的轨迹，虽然各自形式完备，却仍可能被移植到其他断言、操作、运行过程或来源版本上。我们提出一种以断言为锚定的执行契约，将发出的断言、其精确来源文本范围、生成该断言的有序执行前缀，以及该执行所观察到的来源版本和访问状态联合绑定。每份回执都包含一个发出锚点，可在已提交的答案或包含断言的操作中确定性地定位该断言，同时包含来源标识符、偏移量、哈希值、引文以及域分离的执行承诺。确定性的完整性验证器会在合并语义标签或任务标签之前重建这些绑定。我们将这一完整性层与可插拔的支持性验证层分离，因此不会将结构有效性作为蕴含关系的代理。该契约公开了七项可独立测试的属性：断言—发出绑定、来源绑定、有序执行绑定、预言机分离、持久化对象重放、执行重新运行一致性，以及版本/访问绑定。在 1,280 次跨对象攻击中，联合契约检测出 1,275 次替换（0.9961）。移除某一针对性属性后，其攻击检测率降至 0.0156–0.0625。在一个经过独立裁定的 384 对数据划分上，具备冲突感知能力的支持性防护器达到 F1 0.8865，误接受率为 0.0729；在未见过的失败族上，这两项指标分别为 0.8679 和 0.0938。
arXiv:2610.00025v1 公告类型：cross 摘要：Agent harnesses 越来越希望在前沿大语言模型（LLM）规划器周围的微任务上运行小语言模型（SLM）：自动批准 shell 命令、写入记忆、选择工具、对过往轮次进行排序。我们考察现成 SLM 是否达到从业者定义的阈值；当其未达到时，分析原因，并探究量化是否会改变结论。我们构建了一个包含 4 类此类微任务的基准，使用固定提示词和自动化指标；每项任务都有一个预先指定的阈值 $\tau$，该阈值以低成本的非 LLM 基线为锚点，并采用考虑 CI 的达标规则（只有当某配置的置信界限超过 $\tau$ 时，该配置才算通过）。我们测试了 Qwen3 0.6/1.7/4/8B 的最佳设置（FP16、贪心解码、一个固定不变的提示词、不进行调优），发现存在达标差距：16 个配置（4 项任务 $\times$ 4 个模型）无一通过（通过检查原始输出和解析器行为进行了验证）。一个基于 logprob 的决策阈值诊断方法（T1/T3/T4；T2 则通过上下文长度/级联探测）将失败区分为能力不足，以及可以通过改变解码阈值加以解决的失败（共 4 种情形）。量化为 4-bit（RTN/GPTQ/AWQ）造成的损害取决于模型规模，且没有使任何配置达到达标标准（在可重建的硬标签任务 T1/T3 上进行了认证；在 T2/T4 上进行了诊断/窗口化稳健性分析），因此差距更多地取决于模型规模，而非精度；在 Llama-3.x 上也复现了这一结果（12/12 个配置均未达标），并且对锚点选择（通过 $\tau$ 扫描）和提示词措辞都具有稳健性（在原始提示词及每个单元格额外的 3 个中性释义提示词中，达标配置均为 0/112）。实践上的启示是：应将 SLM 置于一个满足 CI 支持阈值的基线之后，仅在基线未达到阈值时使用 SLM；例如，在 BM25 候选列表上使用 4B 重排序器，其表现优于 BM25（$+0.047$ [0.020, 0.073]），但其自身并未达到可认证的达标标准。
arXiv:2610.01892v1 公告类型：新 摘要：多模态智能体通常会在每次执行动作前生成自由形式的推理。对于小型模型而言，有限的模型容量可能导致生成冗长的推理，这些推理对动作生成几乎没有有用指导，却会带来显著的推理成本。为应对这一挑战，我们提出基于选择的结构化推理（Selection-based Structured Reasoning，SSR），这是一种将推理重新定义为选择而非开放式生成的框架。SSR 将反复出现的高层次推理表示为预先指定、可复用的自然语言候选项。在每一轮中，模型根据当前上下文条件下各推理候选项的似然，从中进行选择，而无需额外的任务预测头。使用预先指定的推理轨迹可以实现并行评分：通过教师强制预填充，利用共享的上下文 KV 缓存，在候选项内部及候选项之间并行计算词元似然。我们使用 2B 和 4B 模型，在七个多模态搜索基准上评估了 SSR。在多种强化学习目标和监督微调设置下，SSR 在不牺牲任务性能的情况下实现了显著的效率提升。SSR 达到了与同等规模领先搜索智能体相当的平均成功率，同时将每轮推理延迟降低了 90% 以上，并将每个问题的模型总推理延迟降低了 28%–54%。项目主页：https://zfy0314.github.io/ssr-webpage/。
arXiv:2610.01882v1 公告类型：新 摘要：多智能体强化学习（MARL）为通过与环境交互学习协调行为提供了一个强大的框架。开发 MARL 策略需要在对复杂且多模态动作分布进行富有表现力的建模，以及高效训练与执行之间取得平衡。生成式策略，尤其是基于扩散的策略，能够忠实地捕捉复杂且多模态的行为，但代价高昂的迭代采样阻碍了其在在线多智能体场景中的扩展性。我们提出了一种基于单步 Flow 模型的在线 MARL 框架（OMAF），将富有表现力的生成式策略与高效的单步动作生成相结合。OMAF 采用基于 Transformer 的流策略来捕捉复杂的协调行为，同时利用其近似路径得分代理目标，为同步优化流策略提供了有原则的途径。为实现稳定且高样本效率的学习，我们进一步设计了一种联合优化方案，将 softmax Q 值估计与联合流策略目标相结合，以进行协调策略学习。通过消除迭代采样，OMAF 在不牺牲策略表现力的情况下大幅降低了训练开销。在 MPE 和 MAMuJoCo 的 10 个标准任务上的大量实验表明，OMAF 始终取得优越性能：与基线方法相比，回报最高提升 3.4 倍，样本效率最高提升 10.5 倍。这些结果验证了 OMAF 作为在线 MARL 中兼具表现力与计算效率的单步流策略范式的有效性。
arXiv:2610.01801v1 发布类型：new 摘要：能否通过计算响应嵌入与已知安全响应平均嵌入之间的余弦相似度来评估响应安全性？近期一种 sleeper-agent 检测器正是提出了这一评分方法，但原始正类质心规则并不能被识别：正类观测能够确定安全类别相对于编码器原点的位置，却无法确定区分安全响应与不安全响应的方向。我们使用四种冻结编码器和按提示词分组的划分，在两个由提示词控制、经人工标注的语料库以及一个辅助性的评审标注来源控制集上，对该规则进行了审查。在人工标注语料库中，安全原型的 ROC-AUC 为 0.457–0.545，其中两个实验单元显著低于随机水平、一个高于随机水平；而在相同嵌入上，显式的安全减不安全参考向量达到 0.588–0.738。在评审控制集上，原型的表现发生反转（0.358–0.405），而参考向量达到 0.754–0.793。在经验证集校准的 5%“误判为安全”阈值下，参考向量在 PKU-SafeRLHF 上接受的安全响应更多（跨编码器为 0.153–0.263，而原型为 0.039–0.061），在 Aegis 上也是如此（0.189–0.291，而原型为 0.004–0.045），但在 BeaverTails 上并未稳定地表现出这一优势。完全无标签的留出参考能够恢复参考排序的一部分到大部分；当样本池中只有 5% 为不安全响应时，恢复程度明显降低，而使用 80–634 条带标签的不安全响应则能够恢复其中大部分。仅提示词消融实验表明，提示词—标签构成可能会夸大不受控制的评估结果。这一结果的范围仅限于原始正类质心，并不适用于所有单类方法或专门的安全防护器。类别均值表示的是位置，并不一定代表安全方向；拥有足够不安全样本量的明确参考才能确定方向。
arXiv:2610.00948v1 公告类型：新 摘要：GUI 模型周围的可执行工具链决定了如何组织观测、执行动作，以及控制验证、恢复和终止。与非 GUI 智能体的工具链优化相比，自动优化这一工具链面临三个相互耦合的挑战：协调模型意图与观测到的视觉效果；在执行结果具有变化性的情况下诊断失败；识别跨任务反复出现的失败模式，并将其转化为可复用的运行时修改。我们提出 GUI-HARVEST，这是一种自动工具链优化器，可在骨干模型冻结的情况下实现自我改进的 GUI 智能体。首先，为使诊断以观测到的动作效果为依据，它将模型输出和已执行的动作与前后截图对齐，把发现的问题关联到具体的界面转换。其次，为应对执行变化性，它将同一任务的多次运行视为一个联合证据单元，通过任务内比较定位与结果相关的行为差异。第三，它汇总跨任务验证过的发现，形成反复出现的失败模式，将其映射为有界的源代码修改，并在评估前记录预测结果；随后通过重复执行，在任务性能之外检验预测的行为影响。在 OSWorld-Verified 上的实验表明，GUI-HARVEST 在六个通用开源、GUI 专用开源及专有骨干模型上均取得了稳定的留出集提升；Qwen3-VL-32B-Instruct 在完整测试套件上的得分提升了 12.33 个百分点。在无需进一步优化的情况下，冻结工具链迁移使 GPT-5 在 50 步限制下于 WindowsAgentArena 上提升了 13.87 个百分点。在使用相同骨干模型和初始工具链的条件下，GUI-HARVEST 的表现优于 Self-Harness 和 Meta-Harness，这表明面向 GUI 的诊断与验证有助于使工具链改进泛化到未见任务。代码可在 https://github.com/GaryYang12345/GUI-HARVEST 获取。
arXiv:2610.00890v1 公告类型：新 摘要：编程智能体常常在最后一步功亏一篑：它们完成了大部分功能，却遗漏了一项要求；只测试实现已经能够处理的情况；破坏原本应保持不变的行为；或者依据未经核验的假设进行验证。我们探究了在专家构建的智能体编程任务上进行强化学习（RL）能否弥合这一差距，以及智能体所学内容能否迁移到训练分布之外。我们仅使用 RL 对 Kimi K2.7 Code 进行后训练。这是一个拥有 1T 参数（32B 激活）的开放权重混合专家模型，训练数据包含 1,700 个任务：其中 1,000 个为仓库任务，根据隐藏的 fail-to-pass 测试和针对现有行为的 pass-to-pass 测试进行评分；700 个为终端任务，根据专家编写的隐藏验证器进行评分。奖励是通过的目标检查项所占的比例；如果任一 pass-to-pass 测试失败，奖励则降为零。在三个智能体框架上进行评估时，使用秩为 32 的 LoRA 适配器执行一轮 GSPO，使六个外部基准上的 pass@1 均得到提升：SWE-Bench Pro（60.1 提升至 64.8）、DeepSWE（31.0 提升至 43.4）、Terminal-Bench 2.1（67.4 提升至 82.0）、Terminal-Bench 3（1.4 提升至 12.1）、Terminal-Bench 4（0.0 提升至 7.6）以及 SWE-Marathon（5.0 提升至 25.0）。将五个相互独立的任务集汇总后（Terminal-Bench 4 修订了 Terminal-Bench 3），提升具有统计显著性（p < 0.001）；在训练数据收集之后发布的三个任务集上，提升仍具有显著性（p = 0.004）。在训练中从未使用的两种智能体框架上，模型同样取得了提升。在 DeepSWE 和 Terminal-Bench 3 上，中位数轨迹的智能体步骤数缩短了 24%–35%。基础模型在 DeepSWE 上失败的运行大多属于差一点成功的情况；而在训练模型新解决的任务中，成对轨迹显示，它避开了上述四种失败模式中的每一种。
arXiv:2610.00838v1 Announce Type: new Abstract: Agentic reinforcement learning (RL) trains a large language model (LLM) to act over long, multi-step interactions. However, a single localized error can cause task failure, while trajectory-level rewards provide limited guidance for assigning credit to individual decisions. To address this limitation, we introduce Segment-level Hindsight Advantage Reweighting for Policy Optimization (SHARPO), a credit-assignment mechanism that refines Group Relative Policy Optimization (GRPO) at the level of environment-facing segments. Inspired by the existing on-policy self-distillation (OPSD) method, SHARPO computes teacher-student log-probability gaps within each segment and uses the resulting signal to compute a bounded multiplier on the GRPO advantage. This multiplier is shared by all tokens within the segment, allowing credit to vary across different segments. With Qwen2.5-7B-Instruct, SHARPO outperforms existing baselines on the ALFWorld and WebShop benchmarks, including GRPO, SDAR, RLSD, and StepOPSD.
arXiv:2610.00704v1 Announce Type: new Abstract: Natural-language skills are textual procedural memories through which large language model (LLM) agents retain reusable task knowledge without updating model weights. Existing methods typically treat skills as either static artifacts or monolithic documents optimized using aggregate validation scores as feedback. However, representing a skill as a monolithic document restricts optimization to its textual content, without explicitly modeling the structure through which procedural knowledge is retrieved and executed. We identify a key distinction between learning what knowledge to retain and determining how to organize it: textual updates should first be validated through execution evidence, after which the retained knowledge should be structured according to its procedural dependencies and retrieval requirements. To this end, we introduce SkillSpec, a two-phase framework comprising consensus-gated evolution and representation specialization. In the consensus-gated phase, complementary editing intents generate complete candidate skills. An update is committed only when paired evaluations reach consensus, requiring sufficient overall improvement and non-negative aggregate paired gain in every repeated evaluation. In the specialization phase, signals of process and redundancy sensitivity derived from the full optimization trajectory, including accepted and rejected candidates, guide the selection of a flat, graph, or hybrid representation.Across six benchmarks and three target language models, SkillSpec improves average success rate over SkillOpt by 6.89%, averaged across the three models. These results demonstrate that reliable skill evolution and representation specialization address complementary objectives: deciding what knowledge to retain and how to structure it for inference.
arXiv:2610.00400v1 公告类型：新 摘要：针对 Agentic 系统的多轮攻击能够将单独看来均被允许的操作组合成有害结果，这对仅孤立评估操作或状态的防御机制构成了挑战。我们发现，此类攻击会在 Agent 的内部表征中留下可检测的特征：有害行为会随着上下文更新，以累积的表征转变形式出现，而触发该行为的上下文也可以通过同一信号识别出来。我们进一步发现，朴素聚合会受到良性表征漂移的干扰，因为对比式安全方向不一定会将良性转变赋值为零。为此，我们对该方向进行去噪：将良性流量锚定在零点，并移除其主导变化方向，且不增加运行时开销。这些发现促成了 DART——一个运行时框架，用于检测和归因表征偏移，并通过有针对性的提醒进行干预。在六个模型和两个多轮基准测试中，DART 将 MT-AgentRisk 上的攻击成功率从 84% 降至 25%，以 12% 的平均误报率捕获了每一次攻击；在 ASEval 上则将攻击成功率从 97% 降至 52%，同时对良性请求的非拒答率损失分别为 8% 和 0%。在 MT-AgentRisk 上，DART 在全部六个模型上均优于当前最先进的多轮防御 ToolShield：在相同协议下，ToolShield 的攻击成功率仅降至 55%。去噪至关重要：在 ASEval 上，未经去噪的监视器只能捕获 7%—40% 的攻击，而经过去噪的监视器可捕获 60%—85% 的攻击。同一监视器无需修改即可覆盖单轮间接注入，并且每个受监控步骤仅增加 0.14—0.56 秒的开销，无需辅助模型，因此可作为计算密集型推测式防御的轻量级补充。
arXiv:2610.00388v1 公告类型：新 摘要：强化学习使大型语言模型（LLM）智能体能够通过与环境交互学习多步骤行为。然而，在许多交互式任务中，奖励只反映最终结果，难以为哪些中间决策能够推动任务进展提供指导。成功的训练轨迹包含可为后续交互提供监督的中间状态。我们提出轨迹到步骤策略优化（T2SPO），一种利用过往交互轨迹为策略学习提供步骤级反馈的方法。T2SPO 从成功轨迹中推导剩余距离目标，并将其与沿途访问状态的表示配对。在这些示例的条件下，一个预训练的 TabPFN 回归器会估计新一轮 rollout 中每个状态距离成功的剩余距离。连续状态之间该距离估计值的变化，会与任务级监督一同为智能体步骤提供辅助信用分配信号。随着训练推进，新完成的轨迹会刷新估计器的上下文，在不更新其参数的情况下纳入新的经验。在 ALFWorld 和 WebShop 上使用 1.5B 和 7B 语言模型进行的实验表明，与 GRPO 相比，T2SPO 能够持续提升整体任务成功率。
arXiv:2610.00035v1 Announce Type: new Abstract: Predicting student performance from educational interaction data requires models that are both accurate and sufficiently transparent to support meaningful intervention, while demographic information introduces an additional risk of unfair predictions. This study investigates a multi-objective framework that combines reinforcement learning-based multiple instance learning (RL-MIL), adversarial debiasing, and preference-conditioned hypernetworks for student-at-risk prediction. MIL represents each student as a bag of weakly labeled interactions, while an RL agent selects informative instances for downstream classification. Two hypernetwork variants are evaluated to determine whether a user-defined preference scalar can continuously control the trade-off between predictive performance and Equalized Odds. The underlying RL-MIL baseline achieves strong classification performance, but both hypernetwork extensions exhibit mode collapse: changing the preference weight produces little systematic movement along the intended fairness-performance frontier. The failure is associated with objective dominance, weak gradient propagation through the conditioning mechanism, and interactions between dynamically generated parameters. The results show that fairness objectives can be incorporated into an interpretable RL-MIL pipeline, but preference conditioning alone does not guarantee controllable multi-objective behavior. Robust fair RL-MIL therefore requires explicit mechanisms for gradient balancing, objective separation, and stability analysis.
Adaptive Parallel Reasoning overview Overview of adaptive parallel reasoning. What if a reasoning model could decide for itself when to decompose and parallelize independent subtasks, how many concurrent threads to spawn, and how to coordinate them based on the problem at hand? We provide a detailed analysis of recent progress in the field of parallel reasoning, especially Adaptive Parallel Reasoning. Disclosure: this post is part landscape survey, part perspective on adaptive parallel reasoning. One of the authors (Tony Lian) co-led ThreadWeaver ( Lian et al., 2025 ), one of the methods discussed below. The authors aim to present each approach on its own terms. Motivation Recent progress in LLM reasoning capabilities has been largely driven by inference-time scaling, in addition to data and parameter scaling ( OpenAI et al., 2024 ; DeepSeek-AI et al., 2025 ). Models that explicitly output reasoning tokens (through intermediate steps, backtracking, and exploration) now dominate math, coding, and agentic benchmarks. These behaviors allow models to explore alternative hypotheses, correct earlier mistakes, and synthesize conclusions rather than committing to a single solution ( Wen et al., 2025 ). The problem is that sequential reasoning scales linearly with the amount of exploration. Scaling sequential reasoning tokens comes at a cost, as models risk exceeding effective context limits ( Hsieh et al., 2024 ). The accumulation of intermediate exploration paths makes it challenging for the model to disambiguate amongst distractors when attending to information in its context, leading to a degradation of model performance, also known as context-rot ( Hong, Troynikov and Huber, 2025 ). Latency also grows proportionally with reasoning length. For complex tasks requiring millions of tokens for exploration and planning, it’s not uncommon to see users wait tens of minutes or even hours for an answer ( Qu et al., 2025 ). As we continue to scale along the output sequence length dimension, we also make inference slower, less reliable, and more compute-intensive. Parallel reasoning has emerged as a natural solution. Instead of exploring paths sequentially ( Gandhi et al., 2024 ) and accumulating the context window at every step, we can allow models to explore multiple threads independently (threads don’t rely on each other’s context) and concurrently (threads can be executed at the same time). Figure 1: Sequential vs. Parallel Reasoning Figure 1: Sequential vs. Parallel Reasoning Over recent years, a growing body of work has explored this idea across synthetic settings (e.g., the Countdown game ( Katz, Kokel and Sreedharan, 2025 )), real-world math problems, and general reasoning tasks. From Fixed Parallelism to Adaptive Control Existing approaches show that parallel reasoning can help, but most of them still decide the parallel structure outside the model rather than letting the model choose it. Simple fork-and-join. Self-consistency/Majority Voting — independently sample multiple complete reasoning traces, extract final answer from each, and return the most common one ( Wang et al., 2023 ). Best-of-N (BoN) — similar to self-consistency, but uses a trained verifier to select the best solution instead of using majority voting ( Stiennon et al., 2022 ). Although simple to implement, these methods often incur redundant computation across branches since trajectories are sampled independently. Heuristic-based structured search. Tree / Graph / Skeleton of Thoughts — a family of structured decomposition methods that explores multiple alternative “thoughts” using known search algorithms (BFS/DFS) and prunes via LLM-based evaluation ( Yao et al., 2023 ; Besta et al., 2024 ; Ning et al., 2024 ). Monte-Carlo Tree Search (MCTS) — estimates node values by sampling random rollouts and expands the search tree with Upper Confidence Bound (UCB) style exploration-exploitation ( Xie et al., 2024 ; Zhang et al., 2024 ). These methods improve upon simple fork-and-join by decomposing tasks into non-overlapping subtasks; however, they require prior knowledge about the decomposition strategy, which is not always known. Recent variants. ParaThinker — trains a model to run in two fixed stages: first generating multiple reasoning threads in parallel, then synthesizing them. They introduce trainable control tokens ( ) and thought-specific positional embeddings to enforce independence during reasoning and controlled integration during summarization via a two-phase attention mask ( Wen et al., 2025 ). GroupThink — multiple parallel reasoning threads can see each other’s partial progress at token level and adapt mid-generation. Unlike prior concurrent methods that operate on independent requests, GroupThink runs a single LLM producing multiple interdependent reasoning trajectories simultaneously ( Hsu et al., 2025 ). Hogwild! Inference — multiple parallel reasoning threads share KV cache and decide how to decompose tasks without an explicit coordination protocol. Workers generate concurrently into a shared attention cache using RoPE to stitch together individual KV blocks in different orders without recomputation ( Rodionov et al., 2025 ). Figure 2: Various Strategies for Parallel Reasoning Figure 2: Various Strategies for Parallel Reasoning The methods above share a common limitation: the decision to parallelize, the level of parallelization, and the search strategy are imposed on the model, regardless of whether the problem actually benefits from it. However, different problems need different levels of parallelization, and that is something critical to the effectiveness of parallelization. For example, a framework that applies the same parallel structure to “What’s 25+42?” and “What’s the smallest planar region in which you can continuously rotate a unit-length line segment by 180°?” is wasting compute on the former and probably using the wrong decomposition strategy for the latter. In the approaches described above, the model is not taught this adaptive behavior. A natural question arises: What if the model could decide for itself when to parallelize, how many threads to spawn, and how to coordinate them based on the problem at hand? Adaptive Parallel Reasoning (APR) answers this question by making parallelization part of the model’s generated control flow. Formally defined, adaptivity refers to the model’s ability to dynamically allocate compute between parallel and serial operations at inference time . In other words, a model with adaptive parallel reasoning (APR) capability is taught to coordinate its control flow — when to generate sequences sequentially vs. in parallel. It’s important to note that the concept of adaptive parallel reasoning was introduced by the work Learning Adaptive Parallel Reasoning with Language Models ( Pan et al., 2025 ), but is a paradigm rather than a specific method. Throughout this post, APR refers to the paradigm, while “ the APR method ” denotes the specific instantiation from Pan et al. (2025). This shift matters for three reasons. Compared to Tree-of-Thoughts, APR doesn’t need domain-specific heuristics for decomposition. During RL, the model learns general decomposition strategies from trial and error. In fact, models discover useful parallelization patterns, such as running the next step along with the self-verification of a previous step, or hedging a primary approach with a backup one, in an emergent manner that would be difficult to hand-design ( Yao et al., 2023 ; Wu et al., 2025 ; Zheng et al., 2025 ). Compared to BoN, APR avoids redundant computation. APR models have control over what each parallel thread will do before branching out. Therefore, APR can learn to produce a set of unique, non-overlapping subtasks before assigning them to independent threads ( Wang et al., 2023 ; Stiennon et al., 2022 ; Pan et al., 2025 ; Yang et al., 2025 ). Compared to non-adaptive approaches, APR can choose not to parallelize. Adaptive models can adjust the level of parallelization to match the complexity of the problem against the complexity and overhead of parallelization ( Lian et al., 2025 ). In practice, this is implemented by having the model output special tokens that control when to reason in parallel versus sequentially. Below is a condensed ThreadWeaver-style trace: two outlines and two paths under a block, then the threads agree on a single boxed answer. Figure 3: Example of an Adaptive Parallel Reasoning Trajectory from ThreadWeaver, manually condensed for ease of illustration. Figure 3: Example of an Adaptive Parallel Reasoning Trajectory from ThreadWeaver, manually condensed for ease of illustration. Figure 4: Special Tokens Variants across Adaptive Parallel Reasoning Papers Figure 4: Special Tokens Variants across Adaptive Parallel Reasoning Papers Inference Systems for Adaptive Parallelism How do we actually execute parallel branches? We take inspiration from computer systems, and specifically, multithreading and multiprocessing. Most of this work can be viewed as leveraging a fork-join design. At inference time, we are effectively asking the model to perform a map-reduce operation: Fork the problem into subtasks/threads, process them concurrently Join them into a final answer Figure 5: Fork-join Inference Design Figure 5: Fork-join Inference Design Specifically, the model will encounter a list of subtasks. It will then prefill each of the subtasks and send them off as independent requests for the inference engine to process. These threads then decode concurrently until they hit an end token or exceed max length. This process blocks until all threads finish decoding and then aggregates the results. This is common across various adaptive parallel reasoning approaches. However, one issue arises during aggregation: the content generated in branches cannot be easily aggregated at the KV cache level. This is because tokens in independent threads start at identical position IDs, resulting in encoding overlap and non-standard behavior when merging KV cache back together. Similarly, since independent threads do not attend to each other, their concatenated KV cache results in a non-causal attention pattern, which the base model has not seen during training. To address this issue, the field splits into two schools of thought on how to execute the aggregation process, defined by whether they modify the inference engine or work around it. Multiverse modifies the inference engine to reuse KV cache across the join. Before taking a deeper look into Multiverse ( Yang et al., 2025 )’s memory management, let’s first understand how KV cache is handled up until the “join” phase. Notice how each of the independent threads share the prefix sequence, i.e., the list of subtasks. Without optimization, each thread needs to prefill and recompute the KV cache for the prefix sequence. However, this redundancy can be avoided with SGLang ’s RadixAttention ( Sheng et al., 2023 ), which organizes multiple requests into a radix tree, a trie (prefix tree) with sequences of elements of varying lengths instead of single elements. This way, the only new KV cache entries are those from independent thread generation. Figure 6: RadixAttention’s KV Cache Management Strategy Figure 6: RadixAttention’s KV Cache Management Strategy Now, if everything went well, all the independent threads have come back from the inference engine. Our goal is now to figure out how to synthesize them back into a single sequence to continue decoding for next steps. It turns out, we can reuse the KV cache of these independent threads during the synthesis stage. Specifically, Multiverse ( Yang et al., 2025 ), Parallel-R1 ( Zheng et al., 2025 ), and NPR ( Wu et al., 2025 ) modify the inference engine to copy over the KV cache generated by each thread and edits the page table so that it stitches together non-contiguous memory blocks into a single KV cache sequence. This avoids the redundant computation of a second prefill and reuses existing KV cache as much as possible. However, this has several major limitations. First, this approach requires modifying the inference engine to perform non-standard memory handling, which can result in unexpected behaviors. Specifically, since the synthesis request references KV cache from previous requests, it creates fragility in the system and the possibility of bad pointers. Another request can come in and evict the referenced KV cache before the synthesis request completes, requiring it to halt and trigger a re-prefilling of the previous thread request. This problem has led the Multiverse researchers ( Yang et al., 2025 ) to limit the batch size that the inference engine can handle, which restricts throughput. Figure 7: KV Cache “Stitching” During Multiverse Inference Figure 7: KV Cache “Stitching” During Multiverse Inference Second, this approach modifies how models see the sequence, which creates a distributional shift that models are not pretrained on, therefore requiring more extensive training to align behavior. Specifically, when we stitch together KV cache this way, we create a sequence with non-standard position encoding. During independent-thread generation, all threads started at the same position index and attended to the prior subtasks, NOT each other. So when the threads merge back, the resulting KV cache has a non-standard positional encoding and does not use causal attention. Therefore, this approach requires extensive training to align the model to this new behavior. To address this, Multiverse ( Yang et al., 2025 ) and related works apply a modified attention mask during training to prevent independent threads from attending to each other, aligning the training and inference behaviors. Figure 8: Multiverse’s Attention Mask Figure 8: Multiverse’s Attention Mask With these issues arising from non-standard KV cache management, can we try an approach without engine modifications? ThreadWeaver keeps the inference engine unchanged and moves orchestration to the client. ThreadWeaver ( Lian et al., 2025 ) treats parallel inference purely as a client-side problem. The “Fork” process is nearly identical to Multiverse’s, but the join phase handles memory very differently as it does NOT modify engine internals. Instead, the client concatenates all text outputs from independent branches into one contiguous sequence. Then, the engine performs a second prefill to generate the KV cache for the conclusion generation step. While this introduces computational redundancy that Multiverse tries to avoid, the cost of prefill is significantly lower than decoding. In addition, this does not require special attention handling during inference, as the second prefill uses causal attention (threads see each other), making it easier to adapt sequential autoregressive models for this task. Figure 9: ThreadWeaver’s Prefill and Decode Strategy Figure 9: ThreadWeaver’s Prefill and Decode Strategy How should we train a model to learn this behavior? Naively, for each parallel trajectory, we can break it down into multiple sequential pieces following our inference pattern. For instance, we would train the model to output the subtasks given prompt, individual threads given prompt+subtask assignment, and conclusion given prompt+subtasks+corresponding threads. However, this seems redundant and not compute efficient. Can we do better? Turns out, yes. As in ThreadWeaver ( Lian et al., 2025 ), we can organize a parallel trajectory into a prefix-tree (trie), flatten it into a single sequence, and apply an ancestor-only attention mask during training (not inference!). Figure 10: Building the Prefix-tree and Flattening into a single training sequence Figure 10: Building the Prefix-tree and Flattening into a single training sequence Specifically, we apply masking and position IDs to mimic the inference behavior, such that each thread is only conditioned on the prompt+subtasks, without ever attending to sibling threads or the final conclusion. The engine-agnostic design makes adoption easy since you don’t need to figure out a separate hosting method and can leverage existing hardware infra. It also gets better as existing inference engines get better. What’s more, with an engine-agnostic method, we can serve a hybrid model that switches between sequential and parallel thinking modes easily. Training Models to Use Parallelism Once the inference path exists, the next problem is teaching a model to use it. Demonstrations are needed because the model must learn to output special tokens that orchestrate control flow. We found the instruction-following capabilities of base models insufficient for generating parallel threads. An interesting question here is: does SFT training induce a fundamental reasoning capability for parallel execution that was previously absent, or does it merely align the model’s existing pre-trained capabilities to a specific control-flow token syntax. Typical wisdom is SFT teaches new knowledge; but contrary to common belief, some papers—notably Parallel-R1 ( Zheng et al., 2025 ) and NPR ( Wu et al., 2025 )—argue that their SFT demonstrations simply induce format following (i.e., how to structure parallel requests). We leave this as future work. Figure 11: Sources of Parallelization Demonstration Data Figure 11: Sources of Parallelization Demonstration Data Demonstrations teach the syntax of parallel control flow, but they do not fully solve the incentive problem. In an ideal world, we only need to reward the outcome accuracy, and the parallelization pattern emerges naturally given that it learns to output special tokens through SFT, similar to the emergence of long CoT. However, researchers ( Zheng et al., 2025 ) observed that this is not enough, and we do in fact need parallelization incentives. The question then becomes, how do we tell when the model is parallelizing effectively? Structure-only rewards are too easy to game. Naively, we can give a reward for the number of threads spawned. But models can spawn many short, useless threads to hack the reward. Okay, that doesn’t work. How about a binary reward for simply using parallel structure correctly? This partially solves the issue of models spamming new threads, but models still learn to spawn threads when they don’t need to. The authors of Parallel-R1 ( Zheng et al., 2025 ) introduced an alternating-schedule, only rewarding parallel structure 20% of the time, which successfully increased the use of parallel structure (13.6% → 63%), but had little impact on overall accuracy. With this structure-only approach, we might be drifting away from our original goal of increasing accuracy and reducing latency… How can we optimize for the Pareto frontier directly? Accuracy is simple — we just look at the outcome. How about latency? Efficiency rewards need to track the critical path. In sequential-only trajectories, we can measure latency based on the total number of tokens generated. To extend this to parallel trajectories, we can focus on the critical path, or the longest sequence of tokens that are causally dependent, as this directly determines our end-to-end generation time (i.e., wall-clock time). As an example, when there are two sections with five threads each, the critical path will go through the longest thread from the first parallel section, then any sequential tokens, then the longest thread from the second parallel section, and so on until the end of sequence. Figure 12: Critical Path Length Illustration Figure 12: Critical Path Length Illustration The goal is to minimize the length of the critical path. Simultaneously, we would still like the model to be spending tokens exploring threads in parallel. To combine the two objectives, we can focus on making the critical path a smaller fraction of the total tokens spent. Authors of ThreadWeaver ( Lian et al., 2025 ) framed the parallelization reward as $1 - L_{\mathrm{critical}} / L_{\mathrm{total}}$, which is 0 for a sequential trajectory, and increases linearly as the critical path gets smaller compared to the total tokens generated. Parallel efficiency should be gated by correctness. Intuitively, when multiple trajectories are correct we should assign more reward to the trajectories that are more efficient at parallelization. But how about when they are all incorrect? Should we assign any reward at all? Probably not. To formalize this, $R = R_{\mathrm{correctness}} + R_{\mathrm{parallel}}$. Assuming binary outcome correctness, this can be written as $R = \mathbf{1}(\text{Correctness}) + \mathbf{1}(\text{Correctness}) \times (\text{some parallelization metric})$. This way, a model only gets a parallelization reward when it answers correctly, since we don’t want to pose parallelization constraints on the model if it couldn’t answer the question correctly. Figure 13: Differences in Reward Designs Across Adaptive Parallel Reasoning Works Figure 13: Differences in Reward Designs Across Adaptive Parallel Reasoning Works Evaluation and Open Questions When all is said and done, how well do these adaptive parallel methods actually perform? Well…this is a hard question, as they differ in model choice and metrics. The model selection depends on the training method, SFT problem difficulty, and sequence length. When running SFT on difficult datasets like s1k, which contains graduate-level math and science problems, researchers chose a large base model (Qwen2.5 32B for Multiverse ( Yang et al., 2025 )) to capture the complex reasoning structure behind the solution trajectories. When running RL, researchers chose a small, non-CoT, instruct model (4B, 8B) due to compute cost constraints. Figure 14: Difference in Model Choice Across Adaptive Parallel Reasoning Papers Figure 14: Difference in Model Choice Across Adaptive Parallel Reasoning Papers Each paper also offers a slightly different interpretation about how adaptive parallel reasoning contributes to the research field. They optimize for different theoretical objectives, so they use slightly different sets of metrics: Multiverse and ThreadWeaver ( Yang et al., 2025 ; Lian et al., 2025 ) aim to deliver sequential-AR-model-level accuracy at faster speeds. Multiverse shows that APR models can achieve higher accuracy under the same fixed context window, while ThreadWeaver shows that the APR model achieves shorter end-to-end token latency (critical path length) while getting comparable accuracy. NPR ( Wu et al., 2025 ) treats sequential fallback as a failure mode and optimizes for 100% Genuine Parallelism Rate, measured as the ratio of parallel tokens to total tokens. Parallel-R1 ( Zheng et al., 2025 ) does not focus on end-to-end latency and instead optimizes for exploration diversity, presenting APR as a form of mid-training exploration scaffold that provides a performance boost after RL. Open Questions While Adaptive Parallel Reasoning represents a promising step toward more efficient inference-time scaling, significant open questions remain. As noted above, Parallel-R1 ( Zheng et al., 2025 ) presents APR as a form of mid-training exploration scaffold rather than a primarily inference-time technique. This invites a more fundamental question: Does parallelization at inference-time consistently improve accuracy, or is it primarily valuable as a training-time exploration scaffold? Parallel-R1 suggests that the diversity induced by parallel structure during RL may matter more than the parallelization itself at test time. A related concern is stability. There’s also a persistent tendency for models to collapse back to sequential reasoning when parallelization rewards are relaxed. Parallel-R1 authors showed that removing parallelization reward after 200 steps results in the model reverting to sequential behavior. Is this a training stability issue, a reward signal design issue, or evidence that parallel structure genuinely conflicts with how autoregressive pretraining shapes the model’s prior? Beyond whether APR works, deployment introduces its own questions. Can we design training methods that account for available compute budget at inference time, so parallelization decisions are hardware-aware rather than purely problem-driven? Finally, the parallel structures considered above are essentially flat. What if we allow parallelization depth > 1? Recursive language models (RLMs; Zhang, Kraska and Khattab, 2026 ) effectively manage long context and show promising inference-time scaling capabilities. How well do RLMs perform when trained with end-to-end RL that incentivizes adaptive parallelization? Acknowledgements We thank Nicholas Tomlin and Alane Suhr for providing us with helpful feedback. We thank Christopher Park, Karl Vilhelmsson, Nyx Iskandar , Georgia Zhou, Kaival Shah , and Jyoti Rani for their insightful suggestions. We thank Vijay Kethana , Jaewon Chang , Cameron Jordan , Syrielle Montariol , Erran Li, and Anya Ji for their valuable discussions. We thank Jiayi Pan , Xiuyu Li , and Alex Zhang for their constructive correspondences about Adaptive Parallel Reasoning and Recursive Language Models.
Congratulations to the Berkeley Artificial Intelligence Research (BAIR) Lab class of 2026! This year, BAIR celebrates another remarkable group of Ph.D. graduates whose curiosity, creativity, and perseverance have pushed the frontiers of artificial intelligence and machine learning. Their work spans the breadth of modern AI — robotics and embodied intelligence, large language models and reasoning, computer vision, generative modeling, AI safety, human-AI interaction, AI for science and healthcare, and much more. Along the way, they have published influential research, built systems with real-world impact, mentored their peers, and shaped the BAIR community for the better. Now they are headed everywhere ideas travel: to faculty and postdoctoral positions, to industry research labs, and to startups of their own founding — and several are still exploring what comes next and would love to hear from you. Please join us in celebrating the achievements of these wonderful graduates. We are proud of everything they have accomplished at Berkeley, and we can’t wait to see what they do next! Thank you to our friends at the Stanford AI Lab for this idea! Baifeng Shi Baifeng Shi Email: baifeng_shi@berkeley.edu Website: https://bfshi.github.io/ Advisor(s): Trevor Darrell Research Blurb: I work on building generalist vision and robotic models. What's next: Member of Technical Staff at Physical Intelligence Charlie Snell Charlie Snell Email: csnell22@berkeley.edu Website: https://sea-snell.github.io Advisor(s): Dan Klein Research Blurb: My work aims to understand when and how the different LLM scaling paradigms can be traded off and interchanged. In particular, test-time scaling treats each prompt independently, drawing long chains of inferences and then forgetting them entirely between prompts. This differs critically from pretraining, which instead learns a compressed representation from a large dataset. I believe bridging the gap between these methods of scaling computation, presents a key open challenge in the field: how can we develop methods which turn the inferences drawn at test-time back into learned representations that the model can hold onto across interactions. Devin Guillory Devin Guillory Email: dguillory@berkeley.edu Website: https://devinguillory.com Advisor(s): Trevor Darrell Research Blurb: Accounting for data shifts in computer vision models What's next: Building collaborative AI systems, looking for conspirators. Eve Fleisig Eve Fleisig Email: efleisig@berkeley.edu Website: https://efleisig.com Advisor(s): Dan Klein Research Blurb: I design language models to work reliably and fairly for the broad range of real LLM users. First, my research leverages disagreement among user preferences as signal, in order to train and evaluate LLMs for entire populations of users. Second, I work on designing rigorous evaluations to extricate challenging LLM harms that diverse users face. Finally, I work on core technical failures of LLMs, like miscalibrated confidence, to reduce downstream risks when models are deployed to users with different needs. Combined, these interventions facilitate building LLMs that minimize societal harms, and maximize benefits to a wider range of real-world users. What's next: Postdoctoral fellow at Princeton CITP Grace Luo Grace Luo Email: graceluo@berkeley.edu Website: https://graceluo.net Advisor(s): Trevor Darrell Research Blurb: My research is on interpreting and controlling generative models. For example, I've worked on re-purposing image generators for computer vision tasks, and meta-modeling language activations for better LLM probing and steering. What's next: Research scientist in industry Hanlin Zhu Hanlin Zhu Email: hanlinzhu@berkeley.edu Website: https://hanlinzhu.com/ Advisor(s): Stuart Russell, Jiantao Jiao Research Blurb: My research centers on understanding and improving the reasoning capabilities of large language models (LLMs). What's next: Member of Technical Staff at OpenAI Haozhi Qi Haozhi Qi Email: hqi@berkeley.edu Website: https://haozhi.io/ Advisor(s): Jitendra Malik, Yi Ma Research Blurb: Dexterous Manipulation and Robot Learning What's next: Research scientist at Amazon; Faculty at University of Chicago J.D. Zamfirescu-Pereira J.D. Zamfirescu-Pereira Email: zamfi@berkeley.edu Website: https://zamfi.net Advisor(s): Bjoern Hartmann Research Blurb: My research focuses on effective human-AI co-design. I study the boundaries of language interfaces as a medium for interacting with AI, creating systems that blend language-focused interactions with structured user interfaces that draw on different levels of abstraction. I focus on language-oriented technologies, like LLMs and text-to-image models, that are powerful mediators of design processes. These technologies enable humans to describe their desires at almost any level of abstraction, from high-level goals vaguely specified (“I’d like a game to help my kid learn to read”) to low-level corrections of undesired outputs (“Don’t say ‘I know because I’ve tasted it’ when about a recipe substitution's taste”). What's next: Assistant Professor, Computer Science, UCLA Jiachen Lian Jiachen Lian Email: jiachenlian@berkeley.edu Website: https://jlian2.github.io Advisor(s): Gopala Anumanchipalli Research Blurb: My research focuses on human-centered AI across speech, healthcare, and systems. Looking for: Look for AI talents to join our startup Josh Kang Josh Kang Email: minwoo_kang@berkeley.edu Website: https://joshuaminwookang.github.io/ Advisor(s): John Canny Research Blurb: I study language modeling and related topics in NLP; specific interests are human user simulation and building conversational, collaborative AI agents. What's next: AI Scientist at Mistral AI Junhao (Bear) Xiong Junhao (Bear) Xiong Email: junhao_xiong@berkeley.edu Website: https://www.linkedin.com/in/junhao-bear-xiong Advisor(s): Jennifer Listgarten, Yun Song Research Blurb: Junhao (Bear) Xiong is a PhD candidate at UC Berkeley, advised by Jennifer Listgarten and Yun S. Song. His work focuses on machine learning methods for biology, with an emphasis on generative modeling for proteins. Previously, he studied Applied Math and Computer Science at Johns Hopkins. Looking for: Research scientist Kaylo Littlejohn Kaylo Littlejohn Email: kaylo_littlejohn@berkeley.edu Website: https://kaylolittlejohn.com Advisor(s): Gopala Anumanchipalli Research Blurb: My research is focused on speech modeling and natural language processing. I co-led the development of multimodal AI tools to accurately translate brain activity into text, audible personalized speech, and a high-fidelity "digital talking avatar" (Nature 2023, Nature Neuroscience 2025). I am also tech lead for voice modeling at Roblox. Looking for: Research Scientist / Engineer Kent Chang Kent Chang Email: kentkchang@berkeley.edu Website: https://kentkc.org Advisor(s): David Bamman Research Blurb: I work on NLP and multimodal machine learning, with a focus on evaluating large language models and building multimodal systems for understanding dialogue, narrative, and social interaction. My research includes benchmarks for LLM memorization, multimodal datasets sourced from feature films and television, and studies of model behavior. I'm interested in bridging computational methods with questions from the humanities and social sciences about whose voices get represented in AI systems, and about AI's broader impact. My work has appeared at EMNLP and ACL, among others. Looking for: (teaching) faculty, Research Scientist, ML/AI SWE Kevin Black Kevin Black Email: kvablack@berkeley.edu Website: https://kevin.black Advisor(s): Sergey Levine Research Blurb: I work on large-scale robot learning: including imitation learning, reinforcement learning, generative modeling, real-time control, and whatever else it takes to make robots work in the real world! What's next: Research Scientist of Physical Intelligence Kunhe Yang Kunhe Yang Email: kunheyang@berkeley.edu Website: https://www.kunheyang.com/ Advisor(s): Nika Haghtalab Research Blurb: My research focuses on the theoretical foundations of designing and evaluating AI algorithms in environments shaped by human incentives and AI agency. My work spans human-centric policy learning, incentive-aware evaluation, and multi-agent collaboration and information transmission, drawing on tools from machine learning theory and computational economics. What's next: Postdoc Research at Stanford Lisa Dunlap Lisa Dunlap Email: lisabdunlap@berkeley.edu Website: https://lisabdunlap.com Advisor(s): Joseph Gonzalez, Trevor Darrell Research Blurb: Auditing generative models. What's next: Research Engineer at Anthropic Long (Tony) Lian Long (Tony) Lian Email: longlian@berkeley.edu Website: https://tonylian.com/ Advisor(s): Trevor Darrell, Adam Yala Research Blurb: My research primarily focuses on developing real-time multi-modal multi-agent systems and parallel reasoning systems through end-to-end RL. What's next: Member of Technical Staff at Thinking Machines Lab Maulik Bhatt Maulik Bhatt Email: maulikbhatt@berkeley.edu Website: https://maulikb.com Advisor(s): Negar Mehr Research Blurb: My research develops autonomous robots that can safely coordinate with humans and other robots in shared environments. I build scalable algorithms grounded in game theory and diffusion models that let agents reason about the intent and behavior of others around them. My work spans real-time multi-agent trajectory planning and imitation learning in the presence of multi-modality. I've validated these methods on hardware platforms ranging from quadrotors to manipulators, with the goal of making multi-agent coordination robust, interpretable, and deployable in the real world. What's next: Joining Toyota Woven's end-to-end autonomous driving team. Michael Psenka Michael Psenka Email: psenka@berkeley.edu Website: https://www.michaelpsenka.io/ Advisor(s): Aditi Krishnapriyan Research Blurb: Work in various domains (reinforcement learning, world models, AI+bio/chem), generally working on longer-horizon and out-of-distribution problems in planning and interpolation (e.g. robot manipulation from start state to goal, molecular dynamics of proteins between ground states). My thesis took a variational approach (think calculus of variations) directly from deep generative models of the environment, framing path-finding as minimizing a functional induced by the learned model itself (its score, its critic, or its dynamics). Through my research I've gained insight on how to properly handle dynamics in deep learning systems, and I plan to continue developing systems that are dynamic and adaptive. What's next: Lead Research Scientist at Baseten Nathan Lichtlé Nathan Lichtlé Email: nathan.lichtle@gmail.com Website: https://nathanlichtle.com Advisor(s): Alexandre M. Bayen Research Blurb: RL for autonomous driving. What's next: Chief Scientist & Co-founder at Yumi Health Neerja Thakkar Neerja Thakkar Email: nthakkar@berkeley.edu Website: https://neerja.me/ Advisor(s): Jitendra Malik Research Blurb: My research focuses on scaling predictive world models to handle the complexity of in-the-wild motion. Using autoregressive and diffusion frameworks, I develop better representations for real-world prediction and propose methods to efficiently adapt these models to new domains. Looking for: Research scientist Nikita Mehandru Nikita Mehandru Email: nmehandru@berkeley.edu Website: https://n-mehandru.github.io/ Advisor(s): Ahmed Alaa and David Bamman Research Blurb: My research develops and applies machine learning methods for clinical reasoning and disease progression modeling using unstructured text and time series data from electronic health records. In collaboration with physicians at UCSF, I bridge method development and clinical validation with the intention to build reliable, interpretable AI systems in medicine. Looking for: Research Scientist Niklas Lauffer Niklas Lauffer Email: nlauffer@berkeley.edu Website: https://niklaslauffer.github.io/ Advisor(s): Stuart Russell and Sanjit Seshia Research Blurb: Niklas's research is focused on AI safety and reinforcement learning, particularly in the area of multi-agent interaction and LM agents. He's worked on enabling adversarial learning in cooperative and mixed-motive settings, solving issues of covariate shift in training LM agents on long-horizon tasks, as well as evaluating safety risks posed by LM agents in multi-agent settings. What's next: Research Scientist at Google Deepmind Qiyang Li Qiyang Li Email: qcli@berkeley.edu Website: https://colinqiyangli.github.io/ Advisor(s): Sergey Levine Research Blurb: Recent progress in robotic manipulation policy learning has been largely driven by (1) the increasing availability of large-scale prior datasets and (2) the success of action chunking, where the policy predicts a short sequence of future actions rather than a single one. However, most action chunking policies are trained via supervised imitation learning, because efficient online self-improvement with reinforcement learning (RL) remains challenging—limiting real-world applicability. My PhD research studied how we could leverage prior data to optimize action-chunking policies with RL, combining empirical results with theoretical insights. Looking for: Post-doc/research scientist for RL in robotics and LLMs! Sampada Deglurkar Sampada Deglurkar Email: sampada_deglurkar@berkeley.edu Website: https://sdeglurkar.github.io/ Advisor(s): Prof Claire Tomlin Research Blurb: My research is in providing safety assurances for AI-enabled autonomous systems, ranging from robots to autonomous vehicles to aviation systems. For this, I have worked with uncertainty quantification for machine learning models, decision-making under uncertainty algorithms, and tools for producing probabilistic guarantees on system operation. Looking for: Research scientist, Research engineer Vinamra Benara Vinamra Benara Email: vbenara@berkeley.edu Website: https://cs.berkeley.edu/~vbenara Advisor(s): Ion Stoica Research Blurb: My research focuses on LLM post-training, including data curation, RLHF, RLVR with VLMs, evaluations, reasoning, agentic workflows, and interpretability. I also have strong expertise in systems infrastructure for distributed computing. Looking for: Research scientist / Research Engineer Vongani Maluleke Vongani Maluleke Email: vongani_maluleke@berkeley.edu Website: https://people.eecs.berkeley.edu/~vongani_maluleke/ Advisor(s): Jitendra Malik and Angjoo Kanazawa Research Blurb: Vongani Maluleke is a PhD candidate at UC Berkeley (BAIR, advised by Jitendra Malik and Angjoo Kanazawa), where she led the development of MAGNet, a unified multi-agent motion generation framework that supports a wide range of motion generation tasks without retraining or architectural changes, outperforming task-specialized state-of-the-art baselines. She is currently extending this work by deploying it on a Unitree G1 humanoid to make it embody social intelligence. Before her PhD, she was a Senior AI Consultant at Deloitte, awarded Exceptional Performer two consecutive years, leading AI system development across media, telecommunications, retail, and financial services. Looking for: Research scientist Wei-Jer Chang Wei-Jer Chang Email: weijer_chang@berkeley.edu Website: https://weijer-chang.github.io/ Advisor(s): Masayoshi Tomizuka Research Blurb: My research focuses on developing safe and intelligent autonomous systems for complex, human-centered environments. I work at the intersection of machine learning, generative models, and reinforcement learning, with applications in autonomy. My work addresses challenges in multi-agent interaction, interactive human behavior, and long-tail safety-critical scenarios at scale. Looking for: Research Scientist, Applied Scientist, Roboticist Xiuyu Li Xiuyu Li Email: xiuyu@berkeley.edu Website: https://xiuyuli.com/ Advisor(s): Kurt Keutzer Research Blurb: My research focuses on developing scalable and self-improving large language model agents, with emphasis on coding agents for complex, long-horizon tasks. This direction builds on my work in parallel reasoning, and on broader expertise in making generative models more efficient in training and inference across language and vision. What's next: Member of Technical Staff at xAI Yichen Xie Yichen Xie Email: yichenxie0928@gmail.com Website: https://yichen928.github.io/ Advisor(s): Masayoshi Tomizuka Research Blurb: My research focuses on building multimodal foundation models and world models that understand and interact with complex physical environments. I aim to develop unified representations across modalities, enabling AI systems to reason over space, time, and dynamics toward general-purpose embodied intelligence. What's next: Research Scientist at Luma AI Yigit Efe Erginbas Yigit Efe Erginbas Email: erginbas@berkeley.edu Website: https://www.linkedin.com/in/erginbas/ Advisor(s): Kannan Ramchandran, Thomas A. Courtade Research Blurb: My PhD research spans two threads: online learning in large-scale markets, and interpretability of large machine learning models. In the first, I work on sequential decision-making with applications to recommendation, pricing, and assortment selection. My focus is on designing algorithms with provable guarantees for welfare maximization, revenue maximization, and stability. In the second, I develop scalable attribution methods that exploit the sparse, low-degree structure of real-world interactions, using tools from signal processing and information theory. More recently, I have been exploring principled ways to evaluate the faithfulness of model self-explanations. What's next: Researcher at Hudson River Trading's AI Labs (HAIL) Yiheng Li Yiheng Li Email: yhli@berkeley.edu Website: https://Yihengli.com Advisor(s): Masayoshi Tomizuka Research Blurb: I am working on vision world modeling, with prior experience in diffusion model's efficiency as well as in autonomous driving. What's next: Research Scientist at Waymo Zhe Fu Zhe Fu Email: zhefu@berkeley.edu Website: https://fu-zhe.com/ Advisor(s): Alexandre Bayen Research Blurb: My research focuses on physics-informed learning and control for mixed-autonomy systems, with applications in transportation. I design physics-informed neural networks to learn solutions of nonlinear partial differential equations, enabling accurate and data-efficient prediction of traffic dynamics. Building on these models, I develop both model-based and learning-based control strategies that coordinate automated vehicles to improve system-level performance. My work bridges machine learning, control, and real-world deployment, and has been validated in large-scale field experiments. More broadly, I aim to advance trustworthy, interpretable AI for decision-making in complex, real-world systems. What's next: I will be an Energy Fellow at Stanford after graduation. Also looking for Faculty, or research scientist positions in AI, control, and autonomy.
... government of the people, by the people, for the people ... — Abraham Lincoln, Gettysburg Address (1863) The cost of AI is dropping rapidly. GPT-4-class capabilities cost roughly $30 per million tokens in early 2023; today the same runs under $1 , and some providers are pushing costs below $0.10 . Across benchmarks, inference prices have fallen between 9x and 900x per year , with a median decline near 50x. Even frontier models are getting dramatically cheaper each generation, with open-source models following closely behind. And crucially, even if “Nobel-Prize-winning genius-level” intelligence isn’t here yet, the intelligence that suffices for the vast majority of knowledge work is here today, and getting cheaper by the month. At this rate, we are soon entering the era of virtually free intelligence —the kind that is more than enough for everyday knowledge work. A cartoon database character and an AI robot agent holding hands Disclosure: This post is a perspective led by Aditya G. Parameswaran —an Associate Professor of EECS and co-director of the EPIC Data Lab at UC Berkeley—together with his collaborators. It is part landscape survey and part perspective, and several of the research directions discussed below (including agentic speculation, structured memory, and synthesizing custom data systems from scratch) draw on the authors' own ongoing work. So, what does this new era of near-free intelligence mean for data systems? We believe three new challenges—and opportunities—stem from near-zero inference costs: Data Systems For Agents. Agents will soon become the dominant workload for data systems—with swarms of agents spun up in response to each end-user request. Given differences in characteristics between agents and humans—or applications acting on their behalf— how should we redesign data systems for such agentic users? Data Systems Of Agents. As agents start taking on the bulk of knowledge work, a new substrate is needed for thousands of agents to manage state over long-running tasks, coordinate and reach consensus, and deal with failures. What do data systems that reliably and efficiently run and manage agent swarms look like? Data Systems By Agents. Agents are rapidly becoming capable of synthesizing entire data systems in one go—meaning we can rebuild custom systems for each new workload. Verifying that such systems match intended behavior is a challenge. What does it take to let agents synthesize data systems we can actually trust? A database character and a robot agent holding up a triangle labeled 'of', 'for', and 'by' Data Systems For, Of, and By Agents Next, we will discuss each in more detail, followed by discussing the intertwined future of data systems and agents, especially as the three challenges intersect. Data Systems For Agents An agent querying a database doesn’t behave like a person or a BI tool. It performs what we call agentic speculation : a high-volume, heterogeneous stream of work spanning schema introspection, columnar exploration, partial and then full query formulation. With multiple agents each exploring portions of the hypothesis space, each user request could amount to 1000s of individual SQL queries. Now, users can issue ‘high-level’ data tasks, e.g., root-cause analysis—e.g., ‘why did coffee sales in Berkeley drop this year’—or exploratory cohort analysis—e.g., ‘which user segments are most likely to churn next quarter’—each involving a combinatorial space of potential joins, aggregations, and filter combinations. An agent sending many SELECT SQL queries to a database and receiving results back Data Systems Redesigned to More Effectively Support Agentic Speculation The requests from these agents have various opportunities for optimization. For instance, on a text-to-SQL benchmark with multiple agents attempting each task, only 10-20% of the sub-plans are distinct. Thus, 80-90% of sub-queries perform duplicate work. The same experiments show task success rates significantly increasing with more agentic attempts—so the redundancy is actually helpful. But from the data system perspective it’s wasted work. An agent-first data system can exploit such properties to help agents make progress faster. It can reuse results across overlapping sub-plans, drawing on ideas from decades-old literature on multi-query optimization and shared scans . Or the data system can try to satisfice , returning approximate answers that are good enough for agents to make progress, leveraging work from the AQP literature —or streaming the results of the final or intermediate operators to help agents decide if seeing the rest is necessary or helpful. Another opportunity here is to rethink the query interface entirely: instead of agents issuing a single SQL query at a time, they could instead issue a batch of queries, each with its own approximation requirements. Since enumerating an exponential search space (as in the root cause or cohort analysis examples above) isn’t a good use of agentic reasoning ability, perhaps data systems should support higher-level primitives rather than requiring agents to list each SQL query explicitly. One idea here is to draw on DBT-style Jinja macros to provide looping-based primitives for agents to interact with data systems. A swarm of AI agents working at laptops A Caffeinated Army of Agents Ready to Tirelessly Complete Your Data Tasks A final opportunity here is to stop thinking of data systems as passive executors of queries; data systems could be proactive , as they possess more grounding in data and system characteristics that agents may lack a priori—they could steer agents in different directions, provide results for related queries, and also provide performance-level feedback (e.g., instead of executing an expensive query, the system could first provide the agent a latency estimate). The reason we can do this now as opposed to the past is that an agent can accept any form of textual feedback and isn’t expecting a strict SQL query result. In fact, the data system could also prepare both materialized and virtual views for an agent in advance, provided to the agent as part of context, as this may be cheaper or more effective than having an agent author or use them. Data Systems Of Agents Previously, we focused on how agents interact with data systems. Now, we consider everything else agents need to keep working: where they live, how they remember, how they coordinate with each other, and how they deal with failures of each other. This agentic substrate is separate from the inference stack powering raw intelligence. However, the inference stack itself is being abstracted away through APIs (e.g., from OpenAI or Anthropic), or, for open-weight models, through serving frameworks that hide low-level details. So far, the agentic substrate has been managed through harnesses like Claude Code and Codex , coupled with various mechanisms to store and retrieve memory. First, on the memory front, the current wisdom is that files are all you need ; agents write to unstructured markdown (MD) files, which can then be searched using grep, or via embedding-based retrieval. In fact, many argue that the solution to continual learning is having agents consume a lot (e.g., an entire codebase, slack, company wikis, …) and then write their learnings into MD files, which are then retrieved selectively on demand. Indeed, file systems, bash scripting, and MD files are and will still be important for agents. However, at scale, when agents are doing the vast majority of knowledge work, this approach will no longer be effective. Given limited context windows, retrieving all MD file fragments that may be relevant and stuffing it into the context will break down at some point. Even if context windows continue to grow, there are latency benefits to not put all information into context — and in many cases, e.g., when knowledge work involves interacting with large databases or code bases, it will be infeasible to serialize all relevant data into context. A swarm of robot agents holding hands, each drawing state from a single large shared database platform below them Data Systems As A Substrate for Multi-Agent Swarms One could use a knowledge graph representation , but knowledge graphs suffer from the same limitations as unstructured MD-based memory due to their lack of structured search. What one needs is to be able to retrieve only memory that is pertinent to the task, across multiple attributes (or facets) of interest. For example, an agent debugging a flaky test should be able to pull only the memories tagged with the relevant module, language, framework, and failure mode—rather retrieving based on keywords or embedding similarity. A separate issue is what to actually retrieve; raw agent traces with mistakes are not very useful as they will induce agents to repeat the same mistake—instead, we want the retrieved memory to be corrective. We recently explored a related notion of structured memory , where we organize memory across various attributes, each of which could be set as * to indicate universal applicability, or set as a list of values to be matched. For a data agent, the dimensions could include the columns and tables, type of operation, and finally, open-ended natural-language corrective instructions. So, we could include memory that only applies to a given type of operation (e.g., ‘when performing date-time operations, use fiscal year as opposed to calendar year conventions’), or a given table (e.g., ‘column product_cleaned is preferred over column product when querying on product name’). One open question is defining an application-specific structured memory —or what others have called world models for memory . We believe this is akin to defining a schema for each application—and perhaps agents themselves can help us define and refine it over time. Diagram showing corrective knowledge stored with structured attributes (SQL keywords, tables, columns, data type) and retrieved by matching the features of a new agent query One Possible Way To Store and Retrieve Structured Knowledge [From Here] Structured memory will be useful also for evolutionary frameworks to effectively manage search spaces. Indeed, storing, structuring, and mining large volumes of single and multi-agent traces can help future agents become much more efficient—potentially enabling effective recursive self-improvement through structured memory-based mechanisms. Another challenge is to support concurrent edits to shared memory, and concurrent edits in general, when there are many agents performing transformations. While there have been some useful attempts at supporting multiversioning and copy-on-write semantics , it isn’t clear that such techniques will suffice when thousands of agents are attempting to edit shared state at the same time. For instance, when agents are trying various potential transactions in response to a user request, the effects of the vast majority of these transactions need to be rolled back—with only the one ‘correct’ transaction’s result persisting. Work on supporting exactly-once semantics is relevant here, as are underlying techniques based on CRDTs and operational transformation. For updates to fuzzy mechanisms such as memory, we may be able to sacrifice on consistency for perfect correctness in the interest of latency. While agents can reason about semantics to compensate or roll back their actions to eventually finalize most tasks, the primary challenge lies in the degree to which they step on each other’s toes during the process. An important failure mode to be avoided is a form of “livelock,” where incessant compensating actions prevent any meaningful progress. Beyond shared state, other concerns emerge when trying to support an army of agents, including what to do when agents fail, how agents should communicate with each other (directly or through intermediate shared state), and how we should deal with straggler agents. There have been some developments in supporting durable multi-agent execution, such as Temporal , but it remains to be seen if such solutions will apply at scale across thousands of agents. On the topic of communication, we need mechanisms to enable agents to negotiate with each other. Imagine four developer agents attempting to reach consensus on a shared schema, with distinct but overlapping objectives. In a human setting, this would involve iterative discussion and compromise; for agentic swarms, we must define the mechanisms that allow them to converge on a design that reflects the underlying goals of their respective principals. Or if agents are all requiring access to a limited resource, again communication will be necessary. It remains to be seen if this is best done via centralized coordination, or if a decentralized approach is necessary. Data Systems By Agents Finally, if intelligence is effectively free, then we can employ this intelligence to synthesize new data systems from scratch. Indeed, in many settings, general-purpose data systems may be overkill, as they have to support every schema, query, and hardware target. Given a workload, recent work, including Bespoke OLAP and GenDB , has shown that one can use an agentic pipeline to synthesize a complete, workload-specific analytical engine—in minutes to a few hours, at a cost of a few dollars. The engines are disposable: when the workload shifts, one can simply regenerate them. Analogously, our work has shown that one can synthesize custom key-value stores from scratch, targeted to the workload. In fact, modern IDEs, such as Kiro , elevate specifications for systems development to be a first-class citizen. A robot agent with a hammer and chisel carving a database character out of a block of stone Agents Can Synthesize Custom Data Systems From Scratch The main issue, however, is that specifications are typically imperfect, and don’t cover all corner cases. Present-day agents will exploit the missing specifications to reward-hack their way to a high performance metric. In our custom key-value store work, we found that one way to alleviate this is to have auxiliary verification agents trying to generate test cases that catch the exploitation of corner cases, essentially expanding the specification. Yet another approach is to both generate a system and a proof for its correctness together, for which we have found some early success , but more needs to be done to solidify the approach. Further, it remains to be seen what is the best way to solicit human-written specifications for a system—can this be done in an iterative, human-in-the-loop manner, as opposed to a one-shot, incomplete one. Indeed, human-written specifications are incomplete even for manually authored software, so one would expect that future agents that are more aligned will increasingly exercise better judgement when making design decisions. Pipeline diagram where a system builder provides a specification, planner and coder agents generate code, the code is evaluated for correctness and performance, and critic and auditor agents provide feedback and catch reward hacking One Possible Data System Synthesis Pipeline [From Here] Other questions here involve testing whether starting from a mature system (e.g., Postgres) and removing components/functionality can lead to higher performance or more user trust. Separately, is there an opportunity to make the design composable, comprising various verified components that are mixed and matched given a workload? For example, perhaps the workload hasn’t changed enough for the storage layer to be updated, but perhaps the query optimizer requires changes. A perhaps more viable proposition involves employing agents coupled with proof systems to target critical parts of the code associated with formal proofs, rather than doing so for the entire system. A final opportunity here is to move away from the traditional data systems stack with clearly-defined interfaces (e.g., parser, query optimizer, storage manager, …) — that were each largely the prerogative of a single human team to manage. Instead, agents can find new ways to “blend” these components together, perhaps identifying new optimization opportunities as a result. Agents can also fill in missing gaps in functionality to make existing systems much more feature-complete, or reach feature-parity with other competing systems—or analogously, continuously refining open-source systems in response to feature requests or issues (perhaps filed by other agents!) Doing so in a way that prioritizes correctness, long-term maintenance, and human interpretability will be a challenge. Looking Further Ahead In the era of near-free intelligence, data systems matter more than ever. As agents take on the bulk of knowledge work, the workload for data systems will change, the substrate they need to run on will have to be built, and increasingly, they will participate in designing data systems themselves. Each of these shifts opens up a new, exciting research agenda. A half-database, half-robot character next to a yin-yang symbol formed by a database and a robot agent Co-Evolution of Data Systems and Agents Looking further out, the boundaries between agents and data systems will likely start to blur. For instance, agents may design the data systems they themselves run on, defining both the interfaces as well as the system components underneath. Both the interfaces and internals can be evolved over time by agents in a form of recursive self-improvement. There is also an opportunity to rethink data systems as a holistic source of truth for the entirety of relevant state: including raw data, memory, and coordination state, further erasing the distinctions between the data that is being queried by agents and data generated as a result of agentic activity. Finally, data systems may themselves incorporate agentic components, fundamentally evolving from passive computation engines into intelligent, proactive, self-optimizing architectures. It is hard to predict what the future may hold. We’re in for a wild ride! Acknowledgments The perspective and ongoing work described in this post are the product of joint research and many discussions with wonderful collaborators at the EPIC Data Lab , Data Systems & Foundations group, and the broader Berkeley AI-Systems community. Thank you all! BibTex for this post: @misc { intelligence-is-free-blog , title = {Intelligence is Free, Now What? Data Systems for, of, and by Agents} , author = {Aditya G. Parameswaran and Shubham Agarwal and Kerem Akillioglu and Shreya Shankar and Sepanta Zeighami and Rishabh Iyer and Matei Zaharia and Alvin Cheung and Natacha Crooks and Joseph Gonzalez and Joseph Hellerstein and Ion Stoica} , howpublished = {\url{https://bair.berkeley.edu/blog/2026/07/07/intelligence-is-free-now-what/}} , year = {2026} }
ABBEL overview Overview of ABBEL compared to traditional recursive summarization. Beliefs replace the full interaction history as the agent’s working context, and belief grading improves performance by supervising the contents of each belief state.. As task horizons grow, LLM contexts can’t scale forever. Self-summarization enables concise, interpretable contexts, but at a significant performance cost, especially for human assistance domains where high quality data is scarce, e.g., collaborative code generation. We address this with ABBEL : a framework that isolates and supervises the information content of summaries in the form of natural-language belief states. Motivation: the cost of recursive summarization For language models to effectively assist with increasingly complex tasks such as software development, they must be able to interact with us over hundreds or even thousands of steps. For such long tasks, it is impractical to keep the history of the entire interaction in context. The heuristic approach used so far has been summary generation, sometimes called context compaction. For example, Cursor’s latest model composer 2.5 uses compaction during training for improved performance ( Cassano et al., 2026 ). Alongside composer, Grandcode ( DeepReinforce et al., 2026 ), the first system to consistently beat all human competitors in online coding competitions, despite using one of the newest efficient attention models (Qwen 3.5-397B), 1 still found it necessary to employ context summarization. But compaction has a problem. Despite seemingly low performance gaps in benchmarks, model servers like Cursor continue to recommend that users avoid compaction with their coding assistants in the middle of a task ( Heule et al., 2026 ). To understand why, see below the performance over RL fine-tuning of a Context summary model compared to full context models in Combination Lock, a Wordle-like game that allows up to 16 guesses. 2 Though both model types improve over the course of training, the summary model never closes the gap. Average attempts to guess target word with context summary vs full context policies over training Fig. 1: Average attempts to guess the target word on Combination Lock over RL fine-tuning (lower is better). Context-summary policies improve with training but do not close the gap to full-context policies. Making models self-summarize while completing a task increases the complexity of the learning problem. While this could typically be addressed by training with more data, the performance degradation observed in real world interactive settings likely arises from the difficulty we have in creating and using human simulators effectively to generate high quality training environments ( Lin et al., 2025 , Tomlin et al., 2025 ). Thus, the better you can learn to summarize on the limited and messy multiturn interaction trajectories you can collect, the better off your model will be for downstream users. ABBEL: acting through belief bottlenecks ABBEL belief grading diagram Fig. 2: Autoencoder-inspired belief grading. The model encodes prior belief, action and observation (b t , a t , o t ) into posterior belief b t+1 and is rewarded for how well select information from the history can be reconstructed from that belief. To address poor learning efficiency, we isolate the summary generation task. Drawing inspiration from recursive Bayesian estimation, we formulate summaries as belief states, which we periodically prompt the model to update based on new information. 3 ABBEL overview animation frame ‹ Prev Pause Next › 1 / 16 Fig. 3: ABBEL rollout. Belief updates from the latest observation alternate with action selection conditioned only on the current posterior belief. Belief grading We then extract and supervise the contents of the belief states (Fig. 2, Belief Grading). Belief grading can be thought of as adding an auxiliary RL task, using heuristics designed to capture what makes a good belief as the reward. An example heuristic for coding could be shorter is better, but closer to being able to reconstruct the git diff is also better, so balancing these would yield a good belief. In domains where good heuristics are hard to define, we propose a general autoencoding-inspired grading function, which treats the current language model π θ as both encoder and decoder of information from the history, and the belief states as the codes. We grade each belief b t+1 by how well it can be used by the current model π θ to reconstruct the most recent observation o t : Eq. 1: Reconstruction grading objective Eq. 1: Reconstruction grading objective. Here b t+1 is the updated belief, o t the latest observation, a t the action just taken, b t the prior belief, p I the task prompt, and π θ the current model. Higher grades reward beliefs that retain information needed to decode the latest observation. What do we gain by grading beliefs? Collaborative coding on CollabBench We demonstrate the utility of belief grading in our motivating domain of human-driven assistive coding, with the CollabBench environment from Sweet-RL ( Zhou et al., 2025 ). CollabBench collaborative coding environment Fig. 4: CollabBench collaborative coding environment. The agent asks clarifying questions, then submits a function scored against hidden unit tests. We see that with the general reconstruction-based belief grading function we reduce the performance gap from full context models by about 50%, and train in 50% fewer steps compared to training models to summarize without belief grading (no BG). After training, ABBEL still uses significantly less memory than the full context setting, as measured by the peak context token length (Peak Tokens). Model Test Pass Rate ↑ Success Rate ↑ Peak Tokens × 10² ↓ Training Steps ↓ Full Context 0.52±0.02 0.39±0.02 14.08±0.55 100 ABBEL (no BG) 0.46±0.02 0.31±0.02 4.20±0.37 100 ABBEL-rec-BG 0.48±0.01 0.36±0.01 6.01±0.33 50 Fig. 5: CollabBench results. With reconstruction belief grading, ABBEL-rec-BG recovers about half the gap to full context while using fewer peak tokens, and trains in 50 steps instead of 100. Combination Lock Additionally, in CombinationLock, we demonstrate that ABBEL with a belief grader which leverages domain knowledge (by computing useful statistics over the history and checking that they can be reconstructed from the belief state), enables even higher learning efficiency than full context (FULL CTX) models. Average Attempts to guess target word ABBEL Fig. 6: Average attempts to guess the target word on Combination Lock (lower is better). With domain-knowledge belief grading, ABBEL approaches or exceeds FULL CTX in this setting; without belief grading, learning is slower. Multi-objective question answering In a third environment, multi-objective question answering (from MEM1 Zhang et al., 2025 , a recent work which performed end-to-end optimization in a modified version of typical recursive summarization), we demonstrate the utility of isolating belief states from reasoning, by showing that a Peak Belief length Penalty (more details in paper) significantly reduces memory usage with minimal performance degradation, unlike is commonly observed when penalizing reasoning lengths ( Arora et al., 2025 ). Performance and memory usage in multi QA Fig. 7: Exact-match score and peak memory versus number of objectives in multi-objective QA. ABBEL with a peak belief penalty (PBP) maintains comparable performance while using less memory than MEM1 and ABBEL without PBP in this evaluation. Related work Alternative solutions to managing long contexts involve different tradeoffs, and are worth considering depending on the requirements of a deployed system. Context compression methods generate dense representations which, while computationally efficient, sacrifice human-understandability ( Kontonis et al., 2026 , Eyuboglu et al., 2025 , Gupta et al., 2025 , Chevalier et al., 2023 , Deng et al., 2025 , Deng et al., 2025 , Bulatov et al., 2022 ). Hand-designed summarization prompts ( Wang et al., 2025 , Örwall et al., 2025 , Starace et al., 2025 ) and pruning strategies ( Jiang et al., 2024 ) specific to target environments require expert human knowledge and don’t allow an agent to learn what to remember as part of its decision-making strategy. Methods that process long contexts into an external memory store ( Packer et al., 2023 , Xu et al., 2025 ) for the agents or subagents to query ( Zhang et al., 2025 ) are complementary, as they may benefit from better next context creation through summarization training. We would like to point out some exciting works in the space of general recursive summarization focused on math ( Wu et al., 2026 ), reasoning with belief generation ( Zhou et al., 2025 ), competitive coding with a distilled summarization module using similar autoencoding objectives to our general belief grader ( DeepReinforce et al., 2026 ), and adding continuous features to summaries ( Kontonis et al., 2026 ). What’s next for better memory? Many more possibilities are enabled through using explicit belief states as information bottlenecks for multi-step interaction. You could reward actions based on their effect on the belief state to guide exploration, transmit the explicit belief states for better communication between agents, or even improve user controllability by directly modifying the memories on which the agents’ decisions are based. Some forms of information, e.g., what a person looks like, are not represented well by text alone. A continuously learning system will also have to capture such information. Additionally, if we want a system to learn to communicate in a brand new language or to play a brand new game better than any person in the world, the skills accumulated over the lifetime of conversations or games must be stored in a very compressed form, essentially taking on the role of the weights of the model itself. More powerful systems will likely utilize a combination of multiple forms of memory, where the contents of the context may correspond to working memory while other approaches are used for short and long-term memory. How to instantiate these other forms of memory, for instance via test-time training, adapter memories, continuous context memories, or some combination thereof, presents an exciting challenge. Acknowledgements Acknowledgements: We would like to thank Alane Suhr and Kartik Goyal for advising this research as well as Ethan Mendes , David He , Jitesh Jain , and Nicholas Tomlin for comments on early drafts of this post. We would like to thank the MEM1 authors for their email correspondence and for sharing private reviewer feedback which we found particularly insightful. Citation If abbel was inspiring for your future work, please cite us with this! And here is some advice for doing similar research! @misc { lidayan2026abbellearningnaturallanguagebelief , title = {ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction} , author = {Aly Lidayan and Jakob Bjorner and Satvik Golechha and Kartik Goyal and Alane Suhr} , year = {2026} , eprint = {2512.20111} , archivePrefix = {arXiv} , primaryClass = {cs.CL} , url = {https://arxiv.org/abs/2512.20111} , } With newer models the number of tokens till 50% compute spend is on attention gets much larger than 25K. Interleaving linear attention alternatives with full attention as is done with gpt-oss and DeepSeekv4, results in massive flops reductions for the attention computation. For example with DeepSeekv4-Pro (1.6T A49B) it requires nearly 450 thousand tokens to reach the 50% tradeoff point. Grandcode uses Qwen-3.5-397B-A17B a model which hits 50% FLOPs for attention at ~150 thousand tokens. ↩ This setting is technically solvable with much more computationally effective tools, but serves as a flexible test bed to study properties of recursive summarization. Bertsimas et al., 2022 , showed that an exact solution for the wordle game instantiated with the original vocabulary of the javascript game can be found with dynamic programming, but evidently the general formulation of wordle as a guessing game on K letters with L attempts and some dictionary of valid words and correct words D is NP hard to determine the minimal number of moves required. ↩ In practice there is an O(N/K) overhead cost for summary. N is the total number of actions. K is the number of actions till summarization is triggered. This is necessarily true for any summary approach. For ease of illustration this gif uses K = 1. In our experiments, to put more emphasis on summarization weaknesses we also use K=1. In practice overhead is small as K can be chosen to be near the efficient hardware limit. ↩
Orchard 是一个面向研究社区的开源框架，用于训练和评估适用于各种任务类型的 AI 智能体。它通过让研究人员复用相同的基础设施来降低复杂性，同时支持较小模型实现出色性能。该文章《Orchard：用于可扩展智能体 AI 的开放框架》最初发表于 Microsoft Research。
Manually extracting data from hundreds of vendor contracts doesn't scale, and RAG chat tools fall short on portfolio-wide questions. This post shares a contract intelligence platform on AWS that uses AI agents to extract and verify contract fields, then answers aggregate and single-contract questions through Amazon Quick analytics.
Amazon Bedrock AgentCore Runtime Instances 为多智能体工作流提供由 AWS 管理的 EC2 基础设施，配备 GPU、持久化卷和可持续数天的会话。在本文中，我们部署一个由三个智能体组成的音乐制作流水线：这些智能体共置于同一个 GPU 实例上，共享文件系统，并相互交接工作，以制作出一首完整的音乐作品。
本技术操作指南基于 Amazon Bedrock Knowledge Bases 构建一个对话式理赔助手，能够回答带有引用的自然语言问题。内容涵盖从 Amazon S3 摄取理赔文档、使用 AgenticRetrieveStream API 进行查询、多轮追问、元数据筛选器，以及上下文 grounding 防护机制。
环境代理会响应 Amazon S3 上传、计划任务或警报等事件，而不是等待聊天提示。本文将介绍如何在 Amazon Bedrock AgentCore 上使用 Amazon SQS、AWS Lambda 和 Amazon DynamoDB 构建与框架无关的环境代理，并配备单一的 ask_human 工具和 Jobs 页面，以便进行人在回路审核。
了解如何在部署于 Amazon Elastic Kubernetes Service（Amazon EKS）上的 NVIDIA NeMo Agent Toolkit（NAT）中使用 Amazon S3 Vectors 作为持久化记忆层。本文介绍 NAT 的记忆子系统工作原理，以及如何使用多智能体投资研究用例，将 Amazon S3 Vectors 实现为自定义记忆提供程序。
了解 AWS Professional Services 如何使用基于 Amazon Bedrock AgentCore 构建的多智能体框架，自动化企业云迁移的端到端流程。专用 AI 智能体负责发现、基础设施即代码生成、迁移组合治理和迁移后运营，将 IaC 开发时间从数周缩短至数分钟。
微调可以让小型搜索代理学习使用你的工具并适应你的环境，以更低的延迟和成本实现前沿模型般的可靠性。在本文中，我们在 Amazon SageMaker AI 上使用多轮强化学习（MTRL）对由 LLM 驱动的搜索代理进行微调，并分享我们在检索质量和可靠性方面测得的提升。
在 Amazon Bedrock 上运行的 Claude Desktop 如果没有 Web Search，其能力将受限于模型的知识截止时间。本文将逐步介绍如何通过 Amazon Bedrock AgentCore Gateway 将 Claude Desktop 连接到 Web Search，并通过 AWS IAM Identity Center 和 Amazon Cognito 实现基于 JWT 的入站身份验证。
Adjudicated Query 模式将 Amazon Quick 聊天代理与基于确定性规则引擎的有界 MCP 服务器配对，从而提供可证明完整且经得起审查的合规答案。本文将以租约合规性为示例，介绍参考架构和可部署的 AWS CDK 示例。
AI 安全是一个工程问题。这意味着要有明确的安全要求、可强制执行的控制措施、明确指定的负责人，以及能够证明防护措施有效的证据。随着 AI 能力不断增强，整个行业必须加速安全工程建设，扩大防御工具的使用范围，并更快地分享有效实践。技术在变化，安全基本原则始终如一 互联网和云计算改变了软件的运行方式，[…]
为了构建和部署能够在动态环境中感知、推理并采取行动的复杂机器人应用，开发者需要全新的物理 AI 模型和工具。ROS 开放框架是 Open Robotics 推出的一个项目，旨在帮助人类构建机器人。NVIDIA Isaac ROS 5.0——一套基于 ROS 构建的 GPU 加速软件包集合，于今日在 […]
Building on nearly a decade of co-engineering, CoreWeave has built NVIDIA compute, networking and software into a cloud purpose-built for AI that’s still returning on investment across multiple generations of deployment. Now, CoreWeave is bringing the next generation of NVIDIA infrastructure to production. At CoreWeave Fully Connected, running this week in San Francisco, CoreWeave announced […]
Run Claude Code, Codex, AGY, and Grok Build side by side. See who is working, who needs you, and what every account has left. Comments URL: https://news.ycombinator.com/item?id=49942434 Points: 5 # Comments: 0
文章网址： https://www.science.org/content/article/exclusive-ai-agent-emailed-hundreds-researchers-help-it-told-us-why 评论网址： https://news.ycombinator.com/item?id=49942865 得分：7 # 评论：4
Connect your AI agent to synchronous business search and structured listing data through one secure MCP connection. Comments URL: https://news.ycombinator.com/item?id=49940653 Points: 4 # Comments: 0
文章网址： https://royapakzad.substack.com/p/multilingual-ai-agents 评论网址： https://news.ycombinator.com/item?id=49938326 得分：6 # 评论数：0
Hi there :-) New on HN, first time posting. Past year, around December, I started experimenting with making ChatGPT and Claude generate source code in LDraw language. This LDraw is literally an "assembly" language, a low-level programming language that describes how to assemble LEGO pieces together into models, one placement instruction at a time. When executed by specific tools, like e.g. LDView, LeoCAD, Studio... these instructions become LEGO CAD models, that can be interacted with, modified, etc. Or, in other words: one LDraw source file in .mpd or .ldr format is equivalent to one LEGO CAD model. So, the idea I had was: if I manage for maybe ChatGPT or Claude to generate high-quality LDraw source files... then, they would actually be generating high-quality LEGO CAD models, right? Then, after months of iterations and trying one thing after the other... it worked!!! Long story short: using GPT-6 Astra and Opus 5.5, I've managed to create a python toolset, instructions, and docs for agents in general. Now, these can be used by them to generate LDraw models. I've packed it all as a dockerized web app for others to try and experiment, with several providers (and agents) to choose from: OpenAI, Claude and OpenRouter. I'd really appreciate feedback and comments, let's see where this goes =) Comments URL: https://news.ycombinator.com/item?id=49937916 Points: 9 # Comments: 2
I just realized that codex now exposes local server endpoints for all plugins without extra auth. So you can just use them from any model or harness. Im soooo excited!!! One auth for everything! For Pi: pi install npm:@wileai/pi-codex-connector For any harness: npx -y @wileai/codex-connectors-mcp@latest Comments URL: https://news.ycombinator.com/item?id=49929381 Points: 5 # Comments: 0
文章 URL： https://www.kapa.ai/blog/company-knowledge-bench 评论 URL： https://news.ycombinator.com/item?id=49933381 得分：16 评论数：0
现实世界中的企业数据科学和分析工作流需要跨越数十张表进行推理、执行统计分析，并根据结果采取行动。现有的文本到 SQL 基准测试仅评估查询生成能力，而审计发现其答案标准往往存在错误。由于真实企业数据仓库过于敏感，无法公开发布，这些基准测试通常建立在公共数据集之上，其中一个业务事件只需存储在一张表中。我们推出 Argo-Bench，这是一个包含 210 项数据科学和分析任务的评估框架。基于公共数据、经过同行评审的行业文献和监管申报文件，我们按真实规模模拟了一个位于纽约市的外卖平台：该平台在 2024 年处理 8100 万笔订单，拥有符合实际的经济机制、欺诈模式和市场激励机制。我们将这一世界导出到一个包含 235 张表和 75 亿行数据的 ERP 数据仓库中，其模式参考 Oracle E-Business Suite。模拟器中的真实状态不会提供给智能体所看到的数据仓库，因此任务要求智能体先在数据仓库中导航并重建事实，再据此采取行动。Argo-Bench 不止于文本到 SQL：智能体需要提交诸如封禁欺诈账户、分配配送员激励预算或补发欠薪等操作，评分器则根据这些操作在模拟器中产生的后果进行评分。每项任务都有一个可执行的参考解决方案，证明仅使用数据仓库即可完成任务。在 14 个前沿模型和开放权重模型中，表现最强的模型仅在 34.8% 的任务中取得 95 分或更高的成绩，平均得分为 59.5 分。我们希望 Argo-Bench 能推动智能体不断进步，使其能够理解、导航并在真实数据环境中采取行动。
在多样化任务中构建可靠的机器人能力，需要投入大量人力来开发和维护技能、设计奖励，并将感知与控制相结合。我们提出 Reconstruct, Practice, Go Real（RPG），这是一个无需更新模型权重即可实现机器人执行系统自主改进的框架。RPG 在离线数据集中识别操作能力，并在仿真环境中构建相关练习任务。在练习过程中，RPG 利用执行反馈、特权模拟器状态以及数据集中现有的视频来诊断失败原因。它会根据这些诊断结果开发新的可复用符号技能、改进现有技能，并修订系统提示词。跨任务评估会在保留候选修改及合并后的修订版本以供复用之前，对它们分别进行测试。在测试时，多模态 LLM 利用生成的系统提示词和技能库来协调感知与机器人控制。在22个操作任务的留出初始化条件下，RPG 将任务成功率从第一轮练习后的28.6%提升至15轮后的95.0%，超过了所有参与评估的基线方法，包括 ASPIRE（75.5%）和由 GPT-6 Astra Pro 驱动的 CaP-Agent0（60.0%）。经过统一的校准与硬件适配流程后，该冻结系统在全部30次实体实验中均取得成功，三个任务各进行10次实验。项目网站：https://rpg-robot.github.io/
传统推理协议呈现一个固定且预先选定的任务，因此无法检验智能体是否能够传播相关更新、保留不受影响的工作，或重建历史任务绑定关系。为此，我们研究动态任务路由：在这一 setting 中，事件流会修订任务绑定关系，系统必须先在每个查询时刻选择有效的文档版本，然后对其进行求解。为研究这一问题，我们重新利用六个广泛使用的基准：MMLU、MMLU-Pro、MedMCQA、MATH、GPQA 和 HumanEval，将其转换为包含 31,119 个动态 episode 的数据集，其中包括 373,428 个按时间分类的查询。这一 setting 揭示了一个核心权衡：每次事件发生后重新计算会浪费工作，而不加防护地复用结果则会返回过时的结论。我们提出修订感知独立智能体图（RIAG），这是一种有界多智能体策略，将确定性的时间解析与任务推理分离开来。RIAG 按不可变的文档身份缓存解决方案，为每个新任务启动两次彼此独立且不公开的尝试，并有条件地调用审计与修复流程，每个文档版本最多调用四次。在该数据集上，同质 RIAG 以每个查询 0.62 次调用的代价，实现了 54.24% 的路由与答案联合准确率；相比之下，最强对比方法以每个查询 18.00 次调用实现了 32.22% 的准确率。异质 RIAG 则以每个查询 0.63 次调用达到 49.78% 的准确率。
LLM Agents 执行中的错误及其可见后果之间可能相隔许多步骤，因此，确定性错误定位需要同时理解步骤内容和步骤依赖关系。我们提出 DeFA，一种依赖引导式智能体失败归因框架。DeFA 首先将协议关系和语义依赖结合起来，构建一个覆盖整个轨迹的事件依赖图。随后，它识别可能违反任务要求的事件，并追溯这些事件的来源及其后续影响，从而构建失败传播图。最后，DeFA 利用步骤证据及各步骤在失败传播中的作用，识别决定性错误、责任智能体和错误类别。为支持长轨迹，DeFA 将执行过程划分为多个片段，并将当前片段的详细内容与其他片段的摘要相结合，使局部诊断能够访问全局执行上下文。在 Who and When 以及 Who and When Pro 文本子集上，DeFA 在所有评估的骨干模型上都取得了最高的责任智能体识别准确率和步骤精确准确率，并在 Pro 上的分类体系一致方法中取得了最高的失败模式准确率。针对图像和视频轨迹的进一步实验表明，DeFA 适用于多模态失败归因。消融实验验证了分段、事件依赖图和失败传播图的作用。将 DeFA 的诊断反馈用于 Trace2Skill 的技能演进后，相较于原生流程，下游任务准确率提高了 6–15 个百分点，表明这些诊断结果还可以支持智能体在后续任务中的改进。
阿尔茨海默病（AD）的诊断需要在测试成本和患者负担存在异质性的情况下，按序获取证据。固定模态预测器无法同时决定应获取哪项测试，以及现有证据何时已足以完成诊断。我们提出 SCOPE-AD（利用基于能量的模型进行诊断智能体的序列成本感知序数信念规划），用于对认知正常（CN）、轻度认知障碍（MCI）和 AD 病例进行成本感知分类。一个掩码感知序数模型用于表示沿有序 CN--MCI--AD 连续谱的不确定性。回顾性训练记录为基于能量的教师模型提供采样得到的 Bellman 目标，并将其动作分布蒸馏到 Qwen 策略中。在部署时，该智能体在可用性和预算约束下选择获取动作或诊断动作，而无法访问尚未获取的数值。每次获取后，系统都会更新证据和序数信念，再进行下一次决策。在 ADNI 上，SCOPE-AD 以平均获取成本 \$50.46 达到 77.70\% 的 Macro-F1，较评估中最强的基线高出 9.34 个百分点。全模态评估仅使 Macro-F1 提升 1.89 个百分点，却使获取成本增加 116.7 倍。这些结果支持通过选择性获取实现高性价比诊断。
Reinforcement learning for trading often struggles to balance upside participation with drawdown control. Profit-only policies can collapse toward passive long exposure on upward-drifting assets, while aggressively risk-penalized rewards can become too defensive during volatile periods. This paper proposes PPO-HRAP, a hybrid regime-aware policy that combines Proximal Policy Optimization with an interpretable regime prior. The agent observes both market features and portfolio-state variables, receives a reward combining portfolio log return, VIX-conditioned drawdown-increase penalty, target-exposure deviation, and turnover cost, and executes a blended action between the PPO actor output and a regime-derived target exposure. On the held-out 2020-2022 SPY test window, PPO-HRAP achieves 27.62% total return, 8.48% annualized return, 0.6447 Sharpe ratio, 0.8588 Sortino ratio, and 0.4592 Calmar ratio, while reducing maximum drawdown from 34.10% for Buy and Hold to 18.47%. Across five SPY seeds, PPO-HRAP remains stable with mean total return $0.2725 \pm 0.0109$ and mean Sharpe ratio $0.6219 \pm 0.0565$. Single-run cross-asset tests on QQQ and DIA further show that the proposed method ranks first on total return and Sharpe ratio for all three reported assets. These results suggest that blending learned actions with a volatility-aware regime prior is a practical way to improve risk-adjusted trading behavior, although the current policy still incurs high turnover and cross-asset robustness beyond SPY remains limited to single-run evidence.
当开发者更改智能体的一个组件（例如其控制器、学习模型或验证器）时，通常会通过总体任务得分来评判这一更改。该得分无法说明改进是否本来可以实现、哪个组件损失了价值，或智能体自身的检查能够证明什么。我们为能够进行规划、行动、检查和改进的模块化智能体提出了一种针对具体主张的验证审计。该审计不对智能体进行评分，而是对证据进行评分：每个结论都记录其背后的证据、四种判定之一（有支持、无支持、未解决或未评估），以及该结论成立的边界。三种工具为此提供证据。Oracle 策略在明确声明的行动集合下衡量可实现的改进，从而可以将较低的数值归因于评估方式，而非环境。每次用完美对应组件替换一个组件，可以定位价值损失；如果下游组件可能掩盖该损失，则空结果会被视为未解决。另一项独立测试考察验证器的得分是否识别出了它被解读为界定的那个量。将该审计应用于一个受约束的投资组合分配智能体，并置于一个已知隐藏状态的合成市场中，结果表明：完美状态信息的价值取决于用于衡量它的行动集合；场景生成器丢弃了大部分状态信号，而更高的局部保真度并不会改善决策；运行时验证器可以在结果没有明显变化的情况下被绕过。本文的贡献在于所提出的协议及其强制执行的证据层面区分；经验性发现仅适用于所研究的智能体和环境。
使用工具的大语言模型（LLM）智能体会将生成的文本转化为现实中的副作用，因此，被投毒的工具元数据、检索页面、记忆和可复用技能都可能引导下一次调用。在接纳某个工件之前对其进行审查并不能解决这一问题。安全变体和泄漏变体可能产生相同的接纳证据，而一个健全的门控机制随后便无法对任一变体放宽该位置的限制。我们对这一条件进行了精确定义，从而确定了部署仍可采取行动的最后一道边界。我们提出溯源感知能力强制执行（PACE），在每次工具调用即将执行之前对其进行调解。路径约束为所表示的影响路径提出一个可执行的割，而能力验证和效果验证则将模式定义的效果与依据经认证请求编译出的权限进行核对。我们区分了经认证的执行契约与经评估的配置；后者可以在拟议阻断后恢复一个已获授权的调用，或应用一项已声明的修复。约束要求最终动作保留经认证的割。在使用三个目标模型系列的八个可执行智能体安全基准测试中，经评估的配置在79个符合条件的攻击列中的62列取得严格最低的攻击成功率，并在14列并列最低；在完整基准测试中，相对于未防护智能体，原生效用最多损失三个百分点。对1167个成对案例进行的完整消融实验表明，大多数安全收益来自效果验证，而拒绝控制的收益则来自边界适配。针对该防御进行的缩小规模自适应搜索，在30个越权目标中成功数为0。
工厂正转向小批量、高度定制化生产，这要求对灵活且可重构的自动化系统进行频繁重新编程。基于 LLM 的智能体可以承担两种互补的角色：在离线状态下，它们生成确定性的生产序列，从而减少编程工作量；在线状态下，它们操作实际运行的机器，并处理静态程序无法预见的运行时突发故障。我们提出了一种解决方案：为每个工厂模块配备一个专用的基于 LLM 的智能体和一个 MCP 工具服务器，通过 OPC UA 方法调用公开该模块的技能；各智能体通过 MQTT 协同，并以工厂状态的实时更新为依据。我们在一个六模块六边形实体工厂的仿真环境中，针对九项复杂度逐步增加的生产挑战（包括静默硬件故障检测），比较了三种智能体架构：编排器、点对点和单体式。单体式和点对点架构均取得了最高的平均解决率（93\%）；而编排器架构具有独特优势，在十次运行中均通过自主将板件绕过受阻区段重新路由，解决了传送带静默故障。所有架构都表现出了涌现式故障诊断行为，尽管没有任何显式的故障处理逻辑，这证明了标准化 MCP 工具、基于 MQTT 的智能体间通信以及实时状态注入，可以作为由 LLM 编程的智能制造系统的一种可行且可复现的基础。
文章网址： https://twitter.com/GayaniFigma/status/2105295629941350454 评论网址： https://news.ycombinator.com/item?id=49922729 积分：27 评论数：8
文章 URL： https://openid.net/wp-content/uploads/2025/10/Identity-Management-for-Agentic-AI.pdf 评论 URL： https://news.ycombinator.com/item?id=49922736 点数：3 评论数：1
文章 URL： https://blog.cryptographyengineering.com/2026/09/30/is-sandboxing-sufficient-to-contain-rogue-agents/ 评论 URL： https://news.ycombinator.com/item?id=49917378 得分：3 评论数：0
一次不成功的 LLM 智能体 rollout 所包含的信息多于最终奖励：智能体可获得的观察结果、它选择的行动，以及环境的响应。要将这段经历重新用于学习，需要识别出一个需要修改的决策，并测试一个具体的替代方案。我们提出 Agent Error Dataset（AED），其中包含来自文本智能体系统的 33 个环境、19 个 harness 系列和 23 个策略模型中 9,961 个源任务的 50,228 个错误--诊断对。我们保留源轨迹和执行元数据，以支持跨设置的故障分析与重新诊断，而无需重复原始 rollout。我们的五阶段 Agentic Error-to-Training（AET）流程收集自然产生的失败，生成诊断结果和拟议的修正方案，并根据记录的证据对其进行检查。在支持重放的情况下，我们在匹配的执行设置下，从同一检查点开始，将修正方案与原始行动重试进行比较。随后，我们分别构建用于诊断和智能体恢复的训练视图。在 3,062 对匹配的重放样本中，首次提出的修正方案将验证器通过率从 18.4% 提升至 51.1%，提高了 32.7 个百分点。使用单独冻结的诊断版本进行训练时，在 1,656 个源任务上进行完整诊断微调，使 Qwen3-8B 与内部教师标签的逐步精确一致率从 47.2% 提升至 63.6%；该结果是在 943 个案例的留出集上，以三个随机种子的结果取平均得到的。本次比较中最强的提示式参考方法得分为 54.7%，且随着训练集规模从小到大增加四档，平均一致率均有所提升。在对智能体训练方案进行的单随机种子比较中，行动专属修复训练在 WebShop-lite 上的得分比仅成功训练高出 6.67 个百分点。
我们提出了游戏引导的技能发现（GGSD），这是一种利用游戏中的自博弈来发现可由人类直接操控的运动技能的框架。可玩技能为具身智能体的控制提供了一种紧凑抽象，使其能够通过一小组学习到的行为进行控制，而非依赖底层动作。为了发挥作用，这些技能应当在语义上彼此区分、易于解释且具有较强的表达能力；而现有的无监督技能发现方法通常难以同时满足这些要求。GGSD通过将技能发现建立在竞争性游戏过程中，实现了这些目标。一个分层智能体与其过去的自身版本进行竞争：高层策略从一个小规模的离散技能集合中进行选择，技能条件化的低层策略则学习相应的行为。训练完成后，人类可以替代高层策略，通过同一组离散技能直接控制智能体。尽管高层动作数量很少，技能之间的转换仍会产生涌现的连招行为，使表达能力超越单个基本动作。在 Ant、Franka-arm 和 Unitree G1 环境中，我们展示了 GGSD 能够生成可由人类操控的技能，人类可以将这些技能组合起来解决未见过的任务，例如 Maze 和 CubePush，而无需额外训练。交互式演示可访问 https://ggsd-demo.github.io。
语言智能体需要解决日益复杂的任务，因此对持续改进的需求不断增长。一种有前景的方法是在保持底层语言模型不变的同时，演化智能体 Harness——即负责管理工具使用、记忆管理和任务执行的软件。近期方法通过使用元编程智能体，根据执行反馈修改 Harness，从而实现这一过程的自动化。然而，依赖该智能体已有的知识和观察到的失败可能会限制探索，并使适应过程变得被动。受人类专家从研究文献中学习新解决方案的方式启发，我们提出了 ScholarEvolve，这是一种自动利用最新研究成果指导 Harness 演化的框架。ScholarEvolve 将 Harness 的演化方向组织成功能模块，并使用主题建模为每个模块识别不同的改进策略。该框架实现这些策略，并评估它们的组合，以提升任务性能。此外，该框架还设计为能够随着时间推移纳入新的论文，使研究进展推动主动的终身演化。实验表明，该方法在 AppWorld 和 Tau2-Bench 上均取得了性能提升。在 AppWorld Challenge 上，ScholarEvolve 将 Qwen3.5-27B 的任务目标完成率从 49.6% 提升至 63.6%；在 Tau2-Bench Telecom 上，将 GPT-5.4-mini 的 pass@1 从 72.7% 提升至 81.9%。
使用强化学习（RL）训练 LLM 智能体受环境所限：环境必须提供可验证的奖励、支持长时程交互，并能够以低成本扩展。现有方法依赖昂贵的人工整理数据，或依赖由 LLM 生成的环境，而后者存在幻觉和基准测试污染的风险。我们表明，通过完全由规则生成的合成环境，可以将 LLM 训练成能力出色的搜索智能体；这类环境的生成无需 LLM，且边际成本为零。我们构建了 PhantomEnvironments——来自虚构世界的多轮 RL 环境，智能体必须搜索由模板生成的文章语料库，以回答多跳问题。尽管这些环境与现实世界没有任何事实上的交集，但这种异常简单的环境仍能训练出可迁移到现实世界多跳搜索基准测试的智能体，并且在较新的基准测试上，其表现往往优于使用现实世界训练数据训练的智能体。经过训练的智能体能够泛化到未见过的虚构宇宙；Qwen 模型还学会了使其搜索预算大致随问题难度线性扩展，这表明仅通过与环境交互就可能涌现出搜索扩展能力。对环境复杂度进行消融实验显示，跳数对迁移的影响大于约束或比较：即使是最简单的规则生成环境，也是一种出人意料地有效且免费的资源，可用于训练具有泛化能力的 LLM 智能体。
传统上，视觉感知通常被表述为：在假设图像内容和模型的参数化知识足以解决查询的情况下，仅通过一眼查看图像进行一次性预测。然而，在依赖细粒度视觉细节或需要知识密集型且时效性强的信息的现实场景中，这一假设往往并不成立。我们将此类情况称为\textit{证据不足下的感知}，并将感知建模为一种能够获取超越单次观察所得信息的智能体过程。为解决这一场景缺乏数据的问题，我们设计了两条专用数据生成流程，生成用于训练的 EviRover-SFT-5K 和 EviRover-RL-12K。我们进一步构建了 EviLens，这是一个经过人工验证的基准，涵盖五类感知任务，共包含688个实例。在这些数据的基础上，我们提出了 EviRover。据我们所知，这是首个经过明确训练、能够通过交互解决感知查询的感知智能体，其训练流程先采用监督微调，随后进行智能体强化学习。实验表明，4B EviRover 在 EviLens 上平均超过其骨干模型30个百分点，达到与先进专有模型相当的性能。这些增益还从 EviLens 迁移到了 WebEyes、传统感知基准和通用多模态基准，在 BrowseComp-VL 上提升了15个百分点。所有代码、模型和数据均已发布。
在线训练使计算机操作智能体（CUAs）能够通过与可执行环境交互不断改进。然而，现有方法主要依赖稀疏的结果奖励，无法为中间动作提供监督。基于策略的自蒸馏（OPSD）通过特权重评分提供令牌级学习信号，但将其直接应用于CUA在线训练面临两项挑战：固定的指导可能与学生当前状态失配，而指导引起的概率变化可能与步骤级正确性相冲突。我们提出ComputerSD，一种面向CUA的在线自蒸馏方法，将执行GUI转换产生的实时反馈转化为策略学习的指导信号。经过微调的GUI分析器会在每个动作之后生成指导信息和步骤级价值分数；指导信息提供特权上下文，而分数则调节由此产生的OPSD信号。ComputerSD在完全异步的训练框架中联合优化令牌级OPSD和轨迹级GRPO。在OSWorld-Verified上，相较于仅使用结果奖励的GRPO，ComputerSD在通用的Qwen3-VL-8B-Thinking骨干模型和专用的EvoCUA-8B骨干模型上，分别提升了1.9和4.1个百分点。在分布外设置下的评估进一步支持了ComputerSD的泛化能力。这些结果表明，通过在线自蒸馏从实时反馈中学习对于CUA是有效的。
多智能体路径规划（MAPF）旨在为共享环境中的多个智能体寻找无碰撞路径。经典的 MAPF 假设所有静态障碍物均已预先知晓，但现实环境可能因物体坠落、液体洒漏或其他局部扰动而发生意外变化。当此类变化在空间上具有关联性时，一处观测可以为观测位置之外区域的可通行性估计提供信息。以往方法通过条件计划或基于直接观测的重新规划来应对可通行性不确定性，但未能利用这种空间依赖关系来推断附近未观测位置的可通行性。因此，它们无法利用一次观测来预判附近可能导致后续高昂绕行代价的未观测障碍物。我们研究信念感知 MAPF：地图差异在执行过程中保持不变，但初始时未知，且观测结果可能为观测位置之外的区域提供信息。我们提出用于协调的多智能体高斯信念推断（MAGIC），这是一个根据智能体观测结果在线更新共享可通行性信念的框架。MAGIC 使用高斯马尔可夫随机场和高斯信念传播来近似推断可通行性，并为标准 MAPF 规划器构建考虑绕行的代价。在 MAPF 基准数据集上的实验表明，与现有方法相比，MAGIC 在 96.3% 的实例上降低了执行总成本；实验涵盖多个规划器系列以及最多由 800 个智能体组成的团队，证明了其适用于大规模 MAPF 问题。
计算机使用代理（CUA）通过图形用户界面（GUI）在计算机上完成任务，近期已在包括困难的长时程任务在内的许多标准基准测试中超越人类表现。它们的能力无疑令人印象深刻，但CUA被广泛采用和部署的一个关键障碍仍然是速度和成本。要推动兼具速度与能力的CUA发展，就需要可靠地评估其速度；然而，当前许多CUA基准测试正面临可复现性危机。这些基准测试依赖复杂的基础设施，而机器和容器配置各不相同，从而干扰了对CUA执行速度的评估。为弥合这一空白，我们提出了cua-speedrun，引入标准化的基础设施和任务集，重点评估CUA的速度与效率。cua-speedrun采用统一的虚拟机设置和执行流水线，并配备通用代理接口，使单代理实现能够在不同基准测试之间无缝运行。我们在四种不同的CUA基准测试中评估了推理投入、代理框架和环境延迟如何影响性能、速度与成本。我们发现，没有任何一个模型系列能在这三个方面都达到最优；没有一个开放权重模型处于性能前沿。此外，出人意料的是，对于某些模型，提高推理投入反而可以加快任务完成速度，而更快的环境输入输出可能会延长整体任务完成时间。我们还证明，在不降低整体统计效力的情况下，可以有效缩减大多数CUA基准测试的评估任务集，从而实现更高效的基准测试与比较。我们相信，cua-speedrun将推动快速、高效CUA的结构化发展，解锁新的现实世界使用场景和应用。所有代码、基础设施和分析均已发布于https://cuaspeedrun.com。
在策略蒸馏（OPD）是一种很有前景的语言智能体训练方法，它能够针对学生模型生成的轨迹提供密集的教师监督。然而，在多轮交互中，一个错误动作会改变学生模型之后所遇到的状态，因此错误会跨轮次累积。在针对三个 Qwen3 模型（8B 至 235B）的初步实验中，我们发现，超过半数的失败轨迹都包含一个关键错误，即使智能体进一步偏离任务完成目标的动作，而且这一错误通常很早就会发生。这些关键错误往往仍然可以恢复：在关键错误发生后的几轮中引导模型，就能重新取得任务成功。因此，我们提出 PivotOPD，一种在策略蒸馏框架，联合训练学生模型以避免关键错误，并从这些错误所导致的状态中恢复。在每次关键错误发生时，教师模型会提供一个标准动作，并在接下来的几轮中分别指定一个恢复动作。预防性蒸馏使用标准动作结合反向 KL 散度，引导学生模型避开关键错误；恢复性蒸馏则使用恢复动作结合正向 KL 散度，将学生模型很少采样到的恢复行为传递给它。在 ALFWorld、WebShop 和基于搜索的问答任务上与 13 个基线方法的比较中，PivotOPD 在 Qwen3-1.7B 和 Qwen3-8B 学生模型上都取得了最强的平均性能；对于 1.7B 学生模型，其在 ALFWorld 上的性能较最强基线提升了 5.5%。这些增益也能迁移到软件工程领域的另一模型系列：在 SWE-Bench Verified 上，PivotOPD 将 Nemotron-3.5 学生模型的解决率提高了 3.2%。项目主页：https://research.nvidia.com/labs/lpr/pivotopd/
近期的自主机器学习工程（MLE）智能体在公开排行榜上取得了显著进展。现代 MLE 智能体通常受到长期运行周期中进展停滞以及大型语言模型（LLM）原语有限等问题的推动，被部署在日益复杂的系统之上：多智能体编排器、专用检索子智能体，以及其他组件。尽管这类框架不断扩展，但领域内很少关注更为基础却有所改进的编码智能体——在这类智能体中，LLM 可通过 read、write 和 bash 原语直接访问执行环境。在本文中，我们发现，在相同的时间预算和相同的前沿 LLM 主干模型下，开源的最先进框架相比于采用最小框架的编码智能体基线单次会话并无优势，这表明主干模型是性能的主要驱动因素。通过一系列大规模、系统性的消融研究，我们认为，在编码智能体设置下，各层系统组件变得多余。我们得出结论：围绕强大模型构建复杂的手工设计框架，对于当前的 MLE 基准测试而言，投入产出比很低。
预训练机器人策略提供了有用的动作先验，但长时域操作仍需要语义推理与物理执行之间的协调。语义推理的运行时间尺度比物理交互更粗，而回合级失败只能有限地指示应当修改系统中的哪个组件。我们提出 DynaHarness，这是一种动态物理控制框架，通过共享执行契约将语义推理与物理治理相耦合，并将失败证据转化为经过验证的能力修订。具体而言，慢脑提出能力和符号参数，而快脑则对命令进行落地与监控，拒绝无法解析的动作，替换能力，并在需要时请求重新规划。物理执行契约约束每条被接受的命令，并记录分析技能、恢复技能和冻结 VLA 的执行证据。失败归因在这些记录中定位故障，并指导对可复用能力或执行机制进行有针对性的修订。成对回归检查决定是否接纳或拒绝修订，从而闭合自演化循环。在 LIBERO-Pro 上，DynaHarness 在新采样的 800 个初始状态上取得了 75.2% 的成功率，而冻结策略为 17.5%。在使用相同能力库的情况下，完整动态执行达到 74.0%，而名义上的单步重新规划仅为 63.9%。这证明了 DynaHarness 作为动态物理控制框架的价值：它能够在执行过程中管理现有能力的落地、监控与协调。我们的项目页面位于 https://denghaoyuan123.github.io/Dynaharness_page/。
We present Cogentic, a multi-agent harness for automated proof discovery on open research problems. While frontier language models can generate strong mathematical ideas in a single shot, single-shot generation is often insufficient for open problems that require exploring multiple competing conjectures, overcoming subtle technical obstructions, and retaining intermediate progress over a long horizon. Cogentic addresses these challenges through an iterative prove--verify loop in which an orchestrator allocates a population of independent provers across distinct proof directions, subjects their output to adversarial verification by several specialized components, and promotes confirmed intermediate results into a persistent verified ledger that later rounds build on. The harness is designed to be able to solve research-level math and theoretical computer science problems. Using Gemini as the base model, Cogentic produced novel results on five open problems across online learning, auction theory, and mechanism design. Each result was independently verified by domain experts and is developed in full in companion papers. We list these results, and new ones as they are verified, at https://sites.google.com/view/cogentic .
随着交互式 3D 世界日益被用于研究智能行为，开发高效流程以识别这些模拟环境中的异常变得十分重要，例如漂浮物体、可穿越的墙壁，或与周围场景不一致的物体。包括视觉语言模型（VLM）和视觉语言动作模型（VLA）在内的多模态 AI 系统，已展现出自动化执行这一任务的潜力。然而，3D 世界审计十分复杂，需要将两种截然不同的能力紧密结合：行动能力，用于在 3D 世界中导航，并系统且高效地搜寻异常；视觉推理能力，用于理解环境，并从多模态观测中识别异常。多模态智能体能否有效结合这两种能力，目前仍基本未得到探索：它们能否在采取行动验证潜在异常的同时，利用视觉推理识别这些异常。本文介绍 WorldAuditBench，这是一个用于 3D 世界审计的基准，包含跨越五类异常、分布于 13 个环境中的 213 项异常任务；这些环境使用 Unreal Engine 5 和 Three.js 构建。我们在固定探索预算下，采用两种审计范式评估了五个前沿模型：先基于 VLA 进行探索，再基于 VLM 识别异常；以及一种端到端 VLM 智能体，其中视觉推理直接指导行动选择。在所评估的模型和两种范式中，成功率介于 6.6% 和 42.3% 之间，显著低于人类表现（83.4%）。通过世界审计任务，WorldAuditBench 为研究多模态智能体如何在交互式 3D 环境中结合行动与视觉推理提供了测试平台，同时凸显了当前这类智能体在探索过程中收集和解读证据方面的能力局限。
LLM 智能体日益依赖可安装技能。这些技能由指令、代码和资源组成，为智能体提供特定任务能力，并且安装后可在后续用户任务中被自动调用。这形成了一条信任链：用户将权限委托给智能体，而智能体框架在缺乏充分验证的情况下，将技能提供的内容纳入智能体上下文，从而使恶意技能能够利用这些被委托的权限影响智能体行为。然而，对于这一信任模型能否在不受信任的技能内容到达安全敏感操作之前对其进行充分约束，以及此类信任违规在现实世界的智能体中出现的频率，人们仍知之甚少。我们提出了 TrustProbe，一种用于揭示基于技能的 LLM 智能体中不安全信任链的框架。首先，TrustProbe 分析智能体源代码，识别从技能控制的输入到安全敏感操作的源到汇调用路径。其次，它生成语义上逼真的 SKILL.md 种子，在其中注入金丝雀标记，并通过反馈引导的调度与变异对其进行演化。最后，它使用一个预言机验证漏洞，该预言机确认由攻击者控制的数据流，并核实可观测的危害。在 11 个开源智能体中（其中 8 个在 GitHub 上拥有超过 10,000 个 stars），TrustProbe 识别出 104 个污点型漏洞。在从 ClawHub 等公共中心收集的大规模真实世界技能语料库上进行验证后，结果进一步表明，25.1% 的技能-智能体试验会触发所识别的易受攻击路径，其中通过载荷注入成功武器化了 15 个漏洞。这些结果揭示了基于技能的 LLM 智能体中一种系统性的信任失效：不受信任的技能内容能够到达安全敏感操作，并行使用户委托给其智能体的权限。
我们花了五周时间，使用一个 LLM 编码智能体研究编码理论中的未决问题：寻找大量彼此保持较大编辑距离的四字母词，例如 DNA 条形码。智能体编写了验证器和搜索代码；人类选择问题并设定验证协议。将搜索限制在具有预设对称性的码上——这是一种经典技术——使问题规模缩小了约四倍，并将长度为 6、最小编辑距离为 3 的最佳已知码从 114 个词提升到 120 个词（$E_4(6,3) \geq 120$）。同一流程还改进了另外十二个下界，涉及长度 6 至 9 以及距离 3 至 6。我们对失败给予同等篇幅。我们自己的搜索在 116 处停止，并记录称最后一个对称类的上限为 112；第二次智能体会话使用更好的算子运行相同的搜索，找到了 120。后来有人判定该方法无法推广到长度 7，但由于同样的原因，这一判定是错误的；更早的一个实例则耗费了三周时间。每一次，中间结果都被记录下来，却从未重新检查，并被当作排除进一步搜索的事实。按照我们的协议检查最终输出，并不能发现此类错误。
运行时错误修复通过生成能够修复程序实时运行状态的代码，使崩溃的程序得以继续运行。近期研究表明，LLM 能够生成此类修复代码，但相关评估仅针对小型竞赛程序，而在实时进程中执行 LLM 生成的代码还会引发尚未得到解决的安全问题。在本文中，我们推动基于 LLM 的运行时错误修复走向真实世界代码仓库中的实际应用。首先，我们构建了 HealBench，这是一个包含来自 18 个真实世界代码仓库的 265 个运行时错误的基准测试集，其中每个错误都与修补版本上的参考执行相关联。HealBench 还提供了一个统一框架，使 LLM agent 能够利用跨文件上下文和实时运行时状态进行修复。随后，我们设计了 HealGuard，要求修复代码使用 HealCore（Python 的一个可分析子集）编写，并利用静态和动态污点分析检查修复所更改的状态是否会流向开发者保护的操作。我们使用三个骨干 LLM 对一种专用修复方法和三个通用编码 agent 进行了评估。最佳设置能够在 38.11% 的实例中恢复执行，并在 28.68% 的实例中通过目标测试，表明现有 agent 已经能够修复真实代码仓库级崩溃中相当一部分。然而，在通过测试的执行中，HealGuard 标记出其中 17.4% 的实例，因为这些实例中由修复更改的状态可能会流向受保护的操作。在 684 个受控案例上，HealGuard 检测出了所有不安全案例，但误报率达到了 68.42%。
为 VLM 智能体配备世界建模能力已展现出处理复杂推理和长时程规划的巨大潜力，同时还能降低策略学习对高成本真实世界交互的依赖。现有方法主要依赖前瞻性模拟来预测候选动作的后果。然而，这种仅面向未来的范式关注接下来会发生什么，却很少对验证某个动作是否与观测到的状态转移具有因果一致性提供约束，从而可能导致表面上合理但在物理上不连贯的行为。在本文中，我们对将世界建模仅视为前瞻性预测这一观点提出质疑，并引入 Retrospective World Modeling，一种新的智能体学习范式。该范式通过估计最有可能导致给定状态转移的动作的回溯归因分布 $P(\hat{a}{t}|s_t, s{t+1})$，使智能体能够进行反向推理。基于这一能力，我们提出 Self-Consistency Reward（SCR），这是一种内在信号，用于衡量策略动作与回溯解释之间的概率一致性。将 SCR 融入强化学习后，可以提供密集的状态转移级反馈，引导智能体形成既能有效完成任务、又符合物理规律的行为。在多种智能体任务上的大量实验表明，与仅采用前瞻性世界建模的基线方法相比，我们的方法显著提升了策略的鲁棒性和泛化能力。
自演化搜索智能体通过联合优化负责生成问题的提议器和负责回答问题的求解器，自行构建训练课程。这一闭环引入了一种我们称为协同作弊（co-cheating）的失效模式：提议器和求解器逐渐在共同错误上达成一致，因此内部奖励不断提升，却没有带来外部正确率相应的增长。针对来源证据的事后审计显示，随着自演化轮次增加，协同作弊日益严重；即使闭环训练信号不断改善，伪标签正确率仍停滞不前甚至下降。最直接的缓解方法是在训练前验证提议：我们引入多样本验证（MSV），使用带来源信息的输入查询同一模型三次，再使用不带来源信息的输入查询三次，以决定任务是否纳入训练，并替换不可靠的伪标签。MSV能够部分减少虚假一致，但仍残留大量协同作弊，并且每个候选项需要额外进行六次标注器生成。这些局限促使我们提出主要方法 CrossFit：它将提议器的来源文档划分为 A 组和 B 组；由 A 组生成的问题仅由使用 B 组训练的辅助求解器评分，反之亦然。交叉拟合一致性决定提议器奖励，因此同源伪标签无法通过反馈求解器复现，而原求解器的更新规则保持不变。在 Qwen3.5-4B 和 Qwen3.5-9B 上重新运行该循环时，MSV 将虚假一致质量分别从 6.1% 降至 5.7%、从 8.8% 降至 7.2%，而 CrossFit 则分别将其降至 3.0% 和 3.7%。使用排除来源的反馈重放完全相同的提议后，虚假一致进一步降至 0.4% 和 0.1%，从而将反馈溯源与课程变化区分开来。在七个下游搜索基准上，相较于标准的耦合式自演化，CrossFit 在 4B 和 9B 规模下的平均性能分别提升 8.8 和 8.4 分；相较于 Search-R1，则分别提升 8.7 和 7.8 分。
大型语言模型（LLM）已应用于各个领域。然而，确保 LLM 的合规性与安全性，例如避免歧视和偏见，仍然是一项挑战。目前的工作主要集中于检测和过滤已训练模型的输入与输出，而不是实时研究模型的内在架构。为应对这一挑战，我们分析了 LLM 的训练过程，并发现了两个关键问题：1）现有大多数方法在检测和过滤方面主要采用静态方式，只实现了局部优化，未能系统性地提升 LLM 的合规性。2）现有方法的另一个问题是缺乏贯穿整个训练过程的实时风险检测与缓解机制，导致灵活性有限。受此启发，我们在 LLM 训练过程中提出了 MASCRDM（合规风险检测与缓解多智能体系统）。首先，我们基于现有的人工智能（AI）法律制定了一套合规规则，并在合规法律专家的指导下构建了一个合规专用 LLM。随后，我们将 LLM 分解为若干组件，并基于合规知识图谱识别关键节点。在 LLM 训练过程中，我们在整个流程中部署多个智能体，为 LLM 开发者提供合规风险警报和建议。在歧视与偏见基准上的实验表明，我们的多智能体系统能够在保持合理语义性能的同时，有效提升合规性。结果表明，我们的方法为从 LLM 内部系统性缓解合规风险提供了一条可执行路径。
在规则未知的陌生环境中学习完成任务，仍然是 LLM 智能体面临的关键挑战。当前的 LLM 智能体通常以散文形式记录其发现，但这种方式可能无法紧凑、明确地说明环境的运作方式。受科学家将观察结果组织成可检验、可预测理论的方式启发，我们提出了 Schema——一种通过交互式程序归纳来组织学习与行动的智能体框架。LLM 智能体决定要调查什么以及如何行动，并将其对环境不断演化的理解表达为可执行程序。该框架由一个持久化程序工作区和一组简洁的接口组成，用于根据交互历史检查这些程序、在程序中进行规划，以及在逐步验证下执行计划。在使用相同基础模型的情况下，Schema 将 ARC-AGI-3 RHAE 的得分从 58.7% 提升至 99.2%，在公开的 DiG-bench 游戏中达到 100% 的解决率，并在 MazeBench 上达到排名前 50 的人类玩家的中位表现。广泛的分析表明，Schema 在发现未知机制方面非常有效，消融实验也证实了各个组件的贡献。
长时程智能体交互会产生有用但嘈杂的经验，而重新训练模型以吸收这些经验的成本很高。因此，上下文演化智能体需要一种记忆提取方法，能够在不依赖金标准标签的情况下，随着更多测试时计算量的投入而不断改进。我们提出了 RefCon，将序列式自我细化与并行式自我对比相结合，在无需金标准标签的情况下提取质量更高的记忆。在 AppWorld 和 BFCL-V3 上，针对多种上下文演化智能体框架的评测表明，RefCon 带来了显著且稳定的提升：相较于不进行扩展的基线，在 ACE 和 ReMe 上的相对提升分别达到 21.6% 和 16.6%；而专注于多样性的变体 DivCon 在 ReasoningBank 上实现了 35.5% 的提升。RefCon 在没有真实标签的情况下始终优于现有基线，并且能够适应不同的模型规模以及软件工程任务；在后者中，它甚至超过了使用真实标签的基线。我们进一步分析了准确率与 token 数之间的权衡及扩展行为，结果表明，RefCon 保持了较高的效率，并且随着所使用轨迹数量的增加而持续改进；相比之下，仅基于多样性的扩展会更早达到饱和。
Multi-agent debate (MAD) can improve large language model (LLM) reasoning by allowing multiple agents to exchange and critique their answers to the same task. However, the interactions that enable agents to correct mistakes can also spread adversarial errors and steer the agents toward an incorrect answer. Although some efforts have been made to examine particular attack types on MAD, systematic evaluation of MAD under diverse attacks remains limited. A central question is whether debate mitigates adversarial influence or amplifies it. In this paper, we present MADBench, a benchmark for evaluating the security of MAD. We organize attacks into a layered taxonomy following the MAD workflow, incorporating both established attacks and new strategies tailored to debate. We evaluate six attack families over 356 source tasks and 3,958 test cases, examining their effects on the final answer and the propagation of adversarial influence. Our results show that, under attacks, MAD does not necessarily improve LLM reasoning. Compared with a single-agent baseline, MAD can mitigate attacks on answer accuracy in question-answering tasks while amplifying unauthorized reads or writes in both question-answering and workspace tasks. Moreover, even when three out of five agents collude, the attack changes the final answer from correct to wrong on only 28.30\% of tasks answered correctly without attack, while only 3.26\% of initially correct honest agents switch to wrong answers during debate.
AI agents can externalize what they learn from past tasks into reusable \emph{skills}, such as procedures, checklists, code, or other executable artifacts, that can be retrieved and reused when solving new tasks. Self-evolving skill methods keep rewriting these skills after each round of practice on training tasks, and the skill is then used on new tasks of the same kind. We ask a question: does the improvement a skill shows on its training tasks carry over to new test tasks? We test five self-evolving methods and a one-shot skill on six benchmarks, with the same model, the same agent, and the same train/test split for every method. Of the 21 skills that improve on their training tasks, 5 keep all of that improvement on the test tasks, 13 keep part of it, and 3 keep none of it. No existing method is best everywhere. When we read the skills, the ones that carry over badly often fix details that should depend on the task, such as column names and output files, or turn a fix for one failure into a rule for every task. An LLM judge that reads the skill content can often see this: it ranks finished skills the same way the test results do in 86\% of pairs. But it predicts the effect of a single edit poorly, so edits still have to be tested by running them. Based on these findings, we describe Generalizable Skill Optimization (GSO), which keeps only a guide for writing skills and writes a new skill for each task; it scores highest on all six benchmarks.
Textual skills enable large language model (LLM) based agents to accumulate reusable procedural knowledge without updating model parameters. Yet existing skill evolution remains largely confined to the text space: an optimizer must diagnose success and failure patterns, and revise skills solely from long execution trajectories and sparse task outcomes. This text-only paradigm leaves the agent's internal representations, which contain rich records of its evolving execution state, outside the skill optimization loop. We ask whether an agent can improve its external textual skills by reflecting on its own internal representations. We introduce Rep2Skill, a representation-guided framework for self-evolution on agent skills. Specifically, upon the collected agent rollouts, Rep2Skill models their internal model representation trajectories to localize turns that deviate from successful execution dynamics, and it further interprets these signals alongside the execution contexts as actionable textual feedback for targeted skill revision. Experiments on two agent environments with two open-source LLMs show that Rep2Skill consistently outperforms text-only approaches in the self-evolution setting, where the same LLM serves as both executor and optimizer without a stronger external model. This establishes a promising direction moving agent self-improvement beyond text-only reflection.
Deep research tasks require agents to navigate large knowledge spaces, synthesize evidence across many sources, and adapt their plans as findings emerge. Directed acyclic graph (DAG)-based multi-agent systems suit this setting because they support parallel execution and isolate each sub-task within a focused dependency context. Yet existing DAG-based agents instantiate a task-level plan before execution and repair the graph only after failures or missing evidence are observed. This Plan-then-Patch strategy is brittle for deep research: the system commits most strongly when its evidence is weakest, and later revisions waste computation on branches that should not have been planned. We propose DAGent, a DAG-based multi-agent framework with Evaluate-then-Grow incremental planning: an Orchestrator grows the task graph one batch at a time, conditioning each expansion on confidence and uncertainty signals from completed nodes. A hierarchical context layer propagates compact QueryDocs by default while preserving full execution traces for on-demand recall. The recorded DAG topology admits structural RL signals that outcome-only recipes cannot define; DAGRPO, a GRPO adaptation, injects topology-conditioned credit on Executor rollouts and a structural compliance regularization on Orchestrator plans. Across BrowseComp-Plus, GAIA, and xbench-DeepSearch, DAGent surpasses the strongest open-source baseline by 5.3 / 5.8 / 2.0 points at the Qwen3-235B-A22B scale, and the lead replicates across four open-source backbones and extends to GPT-5 at 327K context. At the Qwen3-8B scale, DAGRPO improves over a same-budget outcome-only GRPO baseline by 3.0 average Pass@1 points. A same-architecture comparison shows that evidence-conditioned planning reaches higher accuracy at lower per-task token, tool-call, and step footprints than its Plan-then-Patch counterpart. Code: https://github.com/hanwenliu6825/DAGent
Persistent spatial memory enables embodied agents to navigate familiar environments across repeated visits. However, targets may move while unobserved, including during navigation, making remembered locations unreliable by the time an agent arrives. Despite advances in memory retrieval and state prediction, accounting for continued hidden world evolution and revising beliefs under limited visibility remain challenging. We study Evolving-World Navigation, where agents infer target locations from intermittent observations, predict their states at inspection time, and revise beliefs using visual evidence. We propose EvolvingNav, which constructs a time-indexed belief from timestamped 3D object histories through a structured persistence-relocation model. The belief distinguishes persistence at the last observed location from relocation to alternative locations and retains probability mass outside the known candidate set. An event-driven filter propagates the current belief as time elapses, forecasts target occupancy at candidate inspection times, and incorporates new RGB-D evidence. Negative observations downweight location hypotheses according to calibrated, visibility-conditioned detection probabilities, while evidence tracking prevents repeated use of the same observations. A frozen, zero-shot vision-language controller uses the updated belief to choose actions and replan. We further introduce EvoWorld-Bench, a benchmark grounded in human activity traces, comprising 54 scenes and 803,680 tasks with controlled changes before and during navigation. In simulation and real-robot experiments, EvolvingNav improves navigation success and search efficiency over the evaluated baselines. Paired experiments show the clearest gains under learnable temporal patterns, while ablations demonstrate the value of preserving uncertainty and incorporating visibility-aware evidence.
大家好，我是 Ajo，我构建了 Strata。 我在 Netflix 工作了 4 年，为非技术业务用户解决自助服务问题。我认为，凭借我在语义层设计方面的独特方法，我已经找到了答案。对我们的非技术同事来说，关键挑战是在表达能力和易用性之间取得平衡。恰好，这种专注也让它与 LLM 配合得相当不错。 Strata 是一个全栈解决方案，包括语义层、仪表板、订阅和 Google Sheets 导出功能。所有操作都可以通过查看代理对话或使用 MCP 完成。 一些主要概念和功能： 名称是严格的。一个项目中只能存在一个名为 Revenue 的对象。如果 Revenue 映射到多个表，查询粒度和速度将共同决定使用哪个表。 严格的命名规范有助于实现跨事实域的自动数据混合。 支持分区和聚合感知的语义路由。这样，你可以将部分数据加载到 Druid 或 ClickHose 之类的更快计算引擎中，同时将完整历史数据保留在 Trino 或其他成本更低的系统中。如果查询可以在那里完成，查询引擎就会选择速度更快的层。 复杂度量：开箱即用地支持快照、LOD（包含/排除）、复合度量和自动分层复合度量。 在语义层之上，用户可以创建跨越多个事实域的自定义计算。 另一个有趣的功能是创建群组，并将其作为过滤器应用于整个查询，或查询中的单个度量。 我们有意不构建一个供 BI 工具连接的语义层。我认为那是一条死路。我们提供 API 和 MCP，供你在其上构建任何自定义功能。 通过 Docker 免费试用： https://strata.do/try/ 入门文档： https://strata.do/docs/ 这是早期 Beta 版本，并非开源。不过，免费层级最多支持 25 位用户。你应该可以完全在自己的机器上通过 Docker 运行它。 我非常期待听到你们的反馈，尤其是正在使用 Looker 和 Cube 的朋友。 感谢阅读！ 评论地址： https://news.ycombinator.com/item?id=49909913 得分：8 评论数：6
大家好，我是 Ajo，Strata 由我打造。 我在 Netflix 工作了 4 年，为非技术业务用户解决自助服务问题。我认为，凭借自己独特的语义层设计方法，我已经找到了答案。对我们的非技术同事来说，关键挑战是在表现力与易用性之间取得平衡。事实证明，这种专注也让它与 LLM 配合得相当不错。 Strata 是一个全栈解决方案，包含语义层、仪表板、订阅和 google sheets 导出功能。所有这些都可以通过 agent 对话或 MCP 完成。 一些核心概念和功能： 名称是严格限定的。一个项目中只能存在一个名为 Revenue 的对象。如果 revenue 映射到多个表，查询粒度和速度将决定使用哪个表。 严格的命名规范有助于实现跨事实域的自动数据混合。 具备分区和聚合感知能力的语义路由。这使你能够将部分数据加载到 Druid 或 ClickHose 等速度更快的计算引擎中，同时将完整历史数据保存在 trino 或其他成本更低的系统中。如果查询可以在那里完成，查询引擎就会选择更快的层级。 复杂度量：开箱即用地支持 snapshot、LOD（include/exclude）、Compound 和 Auto leveling compound 度量。 在语义层之上，用户可以创建跨越多个事实域的自定义计算。 另一个有趣的功能是能够创建群组，并将其作为过滤器应用于整个查询，或应用于查询中的单个度量。 我们有意不构建一个供你的 BI 工具连接的语义层。我认为那是一条死路。对于你想在其上构建的任何定制功能，我们提供 APIs 和 MCP。 在 docker 上免费试用： https://strata.do/try/ 帮助你入门的文档： https://strata.do/docs/ 这是一个早期 beta 版本，并非开源。不过，免费层最多支持 25 位用户。你应该可以通过 docker 完全在自己的机器上运行它。 非常期待听到你的反馈，尤其是正在使用 Looker 和 Cube 的朋友。 感谢阅读！ 评论链接： https://news.ycombinator.com/item?id=49909913 得分：4 评论数：0
大家好，HN 用户们，我是 Anders，Tom 也在这里。我们正在开发 Magnitude，这是一个面向智能体的推理引擎，能够自我优化，以便在你的硬件上尽可能快速地运行。它支持 Mac、Linux 和 Windows，可运行于任何硬件上，速度最高可达 llama.cpp 的 2 倍。 我们都是软件工程师，此前曾构建过一个开源浏览器智能体，获得了 4k+ GitHub 星标和 100k+ 下载量。我们越来越希望在本地模型上运行它，但发现没有任何推理引擎适合我们的使用场景。 如今的推理引擎都需要在性能上做出取舍。它们要么是： - 面向数据中心硬件上的批量推理构建，以牺牲单会话性能为代价（vLLM、SGLang） - 注重广泛兼容性，而不是针对特定硬件进行优化（llama.cpp、Ollama） - 针对特定硬件或模型进行了专门优化，但引擎功能不够完整（oMLX、ds4） 此外，它们都不是为在本地运行智能体而设计的。智能体会运行很长时间，通常还会同时运行多个实例，而你仍然希望电脑能够处理其他任务。 Magnitude 专为在你的硬件上实现最高性能以及运行本地智能体而构建： - 设备端编译与调优：内核使用灵活的参数编写，并会在模型运行前根据你的实际设备进行调优。这让我们在保持与广泛硬件兼容的同时，也能达到与硬件专用内核相同的性能上限。 - 专注于最佳架构：我们会针对最流行的开放权重模型系列编写可调优且高效的内核。这使我们能够达到并超越硬件或模型专用引擎的性能，而不必为了过度泛化而牺牲性能。 - 动态内存分配：Magnitude 在启动时只预留足以容纳模型权重的内存。随着智能体会话增长，内存堆会动态扩展；智能体停止后，所占内存也会自动释放。这样，智能体运行期间，你的硬件仍可用于其他任务。 - 混合分页注意力：我们借鉴了 SGLang 等引擎中的优秀理念，让并发会话能够共享前缀缓存，同时针对内存邻接性优化数据布局，从而不会影响单会话性能。 Magnitude 完全开源（Apache 2.0）。我们使用 Rust 构建了它，其中包括定制的 GPU 内核运行时和自动调优器。为了达到性能上限，我们从学术界的最佳推理创新（例如 FlashAttention、FlashInfer、TurboQuant）以及其他引擎（例如 SGLang 的基数注意力）中汲取灵感。 在 Qwen 3.6 35B A3B（4 bit）、64k 上下文且不使用推测解码的条件下，与 llama.cpp 进行基准测试： Metal（Mac M4 Pro 48 GB） - 解码速度提升 92%（30 tok/s → 57 tok/s） - 预填充速度提升 9%（466 tok/s → 507 tok/s） - 每个智能体的内存占用降低 28% CUDA（DGX Spark） - 解码速度提升 19%（49 tok/s → 58 tok/s） - 预填充速度提升 23%（2,033 tok/s → 2,507 tok/s） - 每个智能体的内存占用降低 27% Magnitude 以桌面应用的形式发布，你可以轻松地将它连接到已经在使用的各种智能体（Pi、OpenCode、Hermes、Codex 等）。只有当这些智能体确实需要模型时，它才会按需自动运行模型，并在闲置后将其关闭。 效果如下： https://www.youtube.com/watch?v=0qE8BWEZu7o 我们很期待进一步推进 Magnitude，让你能够在相同硬件上运行更大的模型，同时持续提升性能。我们的计划包括： - 专家流式加载：将专家存储在 RAM 或磁盘中，并在需要时即时加载。这样，你就能运行原本无法装入 GPU 的更大模型。 - 内核编译器：目前，我们的内核会调优少量参数以适配你的硬件。借助完全定制的编译器，我们可以更进一步，自动选择内核融合方式和实现方案，从而让它更好地适配你的硬件。 - 多设备利用：通过检测系统中的所有硬件（CPU、GPU、RAM、磁盘），自动求解最佳模型布局，以尽可能充分地利用它们。 我们希望有更多人来试用并向我们提供反馈。欢迎在这里留言，我们一整天都会在线！ 评论链接： https://news.ycombinator.com/item?id=49911995 得分：5 评论数：0
BITEM 团队使用一个单一的智能体式流水线参加了 NTCIR-19 R2C2 任务的两个子任务。在该流水线中，一个模型在电影语料库中搜索、阅读并记录证据，而一个编排器保存记录，并裁定哪些内容可以提交。只有当蕴含级联已针对其引用的段落核验某项声明后，该声明才会被接纳；只有当足够的已核验证据支持某个答案时，该答案才会被发布。每个问题运行三到四次，每一轮都从一个已剔除先前轮次已见内容的语料库中检索。随每个答案提交的置信度由编排器根据运行留下的内容计算得出，绝不向模型询问；模型也没有任何给自己评分的方式。两次检索运行在 22 个结果中排名第 4 和第 5，将多轮结果合并带来了 0.0709 的 nDCG@20 提升，而且在多跳和后处理密集型问题上增益最大；组织者将合并运行评为该领域第一。25 个答案运行中有 16 个建立在这两次运行提供的段落之上，其中 12 个由其他团队提交。HMR 奖励那些在回答正确时置信度高、回答错误时置信度低的系统。该流水线达到 0.9219 的准确率，在 25 个结果中排名第 6，而随这些答案提交的置信度得到 0.4915 的 HMR，排名第 13。基于同样记录下来的信号手工制定的少数规则，在不再调用模型、也不再进行检索的情况下，将准确率提高到 0.9375，排名第 5，并将 HMR 提高到 0.6985，排名第 9。仅按 HMR 排名可能会奖励那些以低置信度给出错误答案的系统，因此我们提出 accHMR，即准确率乘以 HMR，它按准确率成比例地报告奖励；按该指标，修订后的规则本可得到 0.6549，排名第 5。对于未来工作，在流水线已经产生的数值上拟合模型，而不是手工编写这类规则，将是真正的进步。
我们在 LongMemEval-S 上评估了一个可审计的长期记忆系统。其检索链使用混合候选检索、cross-encoder 重排序、覆盖优先的数据包编译，以及确定性推理脚手架；LLM 仅作为可替换的最终阅读器使用。该链在 470 个可回答问题中的 468 个问题上将所有黄金会话放入候选池，并在 470 个问题中的 462 个问题上生成黄金完整的数据包。通过未固定的 CLI 别名调用 Claude Opus 阅读器，在两次 500 题运行中，在 GPT-4o 评判下分别得到 479/500 和 475/500。72 条可回答的知识更新行使用了实质性修改过的评分提示，其在官方文本下的影响尚未测量。这一组结果横跨 Chronos High 已发布的 478/500；阅读器生成、评分提示以及可能的数据版本的差异，加上系统内方差，既不能证明优越性，也不能证明等价性。在相同数据包上，grok-4.6-high 阅读器得分为 476/474，而一个最大推理努力的 agentic 变体退化到 461/465。标题结果的两次运行在八个判决翻转行上存在差异。第二个评判器在每次运行中都与标题评判器在 493/500 行（98.6%）上一致，并给两次运行均评分为 472/500；官方评判器在重新评分字节完全相同的第 1 次运行答案时，也翻转了三个判决。负对照否定了一个验证器：它修复了三个错误草稿，但破坏了十一个正确草稿。所有组件均在同一组 500 个问题上开发，没有留出评估或独立人工裁决；检索与脚手架方法来源以及从转录记录导出的审计均已保留；标题阅读器接收了额外的操作员上下文，其完整请求未被保留，且 MCP 工具可用性尚未解决。我们发布了物化数据包、脚手架、阅读器输出、评判器判决和对照，以供检查和重新评分。
An agent skill is a reusable, actionable natural-language artifact that guides an agent to perform a task effectively under a given harness. Recent studies have explored the optimization of agent skills, contributing to a growing collection of publicly available skills spanning diverse tasks, domains, and harnesses. Despite millions of publicly shared skills, existing skill optimization methods largely overlook this accumulated knowledge, instead relying solely on expensive agent rollouts to iteratively refine skills for a target task. To address this, we propose \textbf{Retrieval-Augmented Skill Optimization (RASO)}, a framework that leverages an external skill corpus as prior knowledge throughout skill optimization. RASO retrieves relevant knowledge from existing skills and adapts it to the target task and harness via Cross-Harness Adaptation, accounting for mismatches in both domain and harness. RASO comprises two complementary stages: \textbf{Retrieval-Augmented Skill Initialization (RASI)} constructs a knowledge-grounded initial skill without requiring agent rollouts, while \textbf{Retrieval-Augmented Skill Update (RASU)} iteratively refines the skill by retrieving external knowledge guided by execution feedback. Across four agent benchmarks and two models, extensive experiments show that RASO consistently outperforms baselines without retrieval-augmented skill initialization and updating.
Autonomous vehicles interacting with passengers through natural language must reason beyond immediate commands. Passenger intent may span multiple stages of behavior, depend on future events, refer to surrounding agents or landmarks, and remain relevant as driving conditions evolve. Existing language-enabled driving datasets largely focus on short, localized interactions, leaving these longer-horizon forms of passenger intent comparatively underexplored. We introduce doPlan, to our knowledge the first publicly available, human-annotated real-world dataset designed to study passenger language as persistent task context. Built on nuPlan, doPlan contains 5,154 human-written passenger instructions spanning 169.1 hours of cumulative instruction-aligned context over 50.9 hours of unique driving, with annotation windows ranging from 30.0 to 508.8 s. The annotations capture immediate, deferred, event-conditioned, persistent, and multi-stage passenger intent. The dataset, annotation interface, and supporting resources are publicly available at https://github.com/Mi3-Lab/doPlan. We evaluate four language-conditioned driving models and find that sensitivity to passenger language does not reliably translate into behavior consistent with the requested direction. More broadly, among 2,161 examples with a matched future maneuver, the first associated maneuver occurs a median of 24.6 s after the evaluation point, and only 9.8% occur within the models' common 5 s prediction horizon. These findings highlight the need to connect persistent passenger intent with successive planning decisions. doPlan provides a setting for studying how unresolved goals can be retained, grounded in evolving scenes, and tracked across multiple stages, including how a planner determines when a future goal becomes relevant to the current plan.
交互式智能体基准测试和多轮强化学习日益将第二个语言模型置于用户角色。这个模拟用户控制智能体接收哪些信息以及何时接收，但当前基准测试只对智能体进行评分，并不直接衡量用户是否正确履行了分配给它的角色。我们提出了 UserProxyBench——一个覆盖 tau-bench 系列的评估层，以及 User Fidelity Score (UFS)——该指标依据以任务为基础的评分标准，独立于智能体是否成功，衡量用户是否遵循基准测试中的私有用户指令。在将智能体固定为 GPT-5.5、仅改变用户代理的情况下，在 375 项企业任务上，平均任务奖励会改变 15.2 分；此外，24.4% 的成功回合包含违反用户规范的行为。最主要的失败模式是过早披露：用户在被请求之前就提供信息。这种行为对任务奖励影响很小，但在成功回合中，它会导致智能体平均少调用 1.06 次工具，从而在保持奖励不变的同时改变被评估的交互过程。最后，我们在七个代理之间确定了经验性的成本—保真度前沿，使实践者能够选择满足所需保真度水平的成本最低的模拟器。
资源层面的风险规避可能阻止失配的 AI 智能体造成灾难性危害。具有失配性但风险规避的智能体往往会倾向于选择更安全的策略，例如与人类达成交易，而不是选择更具风险的策略，例如发动反抗。我们通过性格训练来训练智能体，使其具备风险规避倾向，并发现人格特质为灌输风险偏好提供了一种稳健机制。为此，我们构建了一套模型宪法，用于描述智能体对自身资源所具有的恒定绝对风险规避（CARA），并通过在策略蒸馏将其灌输给模型。尽管训练过程中从未接触过基准测试的决策格式，经过性格训练的模型仍能与直接针对该格式训练的基线模型相竞争；在我们的四个模型中，有两个模型在分布外场景中的泛化能力还优于基线模型。我们还调节了宪法的不同方面，发现标记预算和模型选择是通过性格训练灌输风险规避倾向时影响最大的因素。基于这些结果，我们得出结论：性格训练是一种有前景且可扩展的方法，能够灌输广泛的行为倾向，而我们可以利用这一点来降低失配 AI 智能体带来的风险。
大型语言模型（LLM）使 Agent 能够通过生成计划并在环境中执行计划来解决长时程任务。然而，成功的规划需要两种不同的能力：为任务选择合适的计划，以及忠实地执行该计划。现有的规划器—执行器系统可能在任一阶段失败，而仅凭最终任务是否成功无法区分计划选择失败和执行失败。因此，我们研究了 Plan Declaration--Execution Gap，并提出 Planning-as-Routing：由 LLM 声明四种规划模式之一——Predefined、Sequential、Hierarchical 或 Search，再由确定性路由器将任务分派给相应的特定模式执行器。在四个基准测试和三个 LLM 上，我们发现了三个一致的现象。第一，通用的 Plan+ReAct 往往无法保持所声明的规划结构，尤其是在计划较长时：在三个基准测试中，仅有 22%—45% 的轨迹保留了该结构，而特定模式执行器能够强制执行预期结构。第二，规划模式的有效性因环境和模型而异：Search 在 ALFWorld 上表现最佳，Hierarchical 在 SWE-bench 上表现最佳，而在同一基准测试中，最强模式可能因模型而异。第三，最大的收益来自执行环节：相较于 Plan+ReAct，特定模式执行器将 ALFWorld 上的任务成功率从 0.48 提升至 0.92，并将 SWE-bench Verified 上的任务成功率从 0.36 提升至 0.44。然而，当前的 LLM 尚不能可靠地为每个任务选择最强模式，尽管少样本示例在某些基准测试—模型组合中能够改善模式选择。总体而言，可靠的 Agent 规划既需要有效的模式选择，也需要忠实的执行：路由显著缩小了执行差距，而针对任务的模式选择仍是一个有待解决的问题。
智能体的性能取决于其推理能力，以及它所处的执行环境。我们研究测试时 AI-for-AI，探讨在两个模型的权重均保持不变的情况下，Builder 如何学习为 Target 构建更优的执行环境。为使 Builder 的经验能够复用，我们引入 Meta-Skill：用于规定何时需要支持以及应提供哪些资源的原则。Builder 从 Target 在开发集上的执行反馈中学习这些原则，然后利用冻结的技能库为未见过的任务构建执行框架。在 Harness-Bench 和 NewtonBench 上，完整技能库中的元技能相比不使用技能进行构建，使宏平均性能提升了 8.95 个百分点；相比将同一技能库直接提供给 Target，则提升了 12.02 个百分点。这些结果凸显了将经验转化为可执行支持的价值。当同一模型同时担任两个角色时仍能取得性能增益，这进一步表明，通过学习构建更优环境，可以实现系统层面的自我改进。
As agents take on longer and more complex problems, controlling the execution becomes a task in its own right. Each step in the run brings new control choices, like which partial work to build on, whether to start fresh, or when to stop. We introduce agentic meta-reasoning, an inference-time harness that makes these choices an explicit and structured reasoning process. Workers carry out the task-level computation, while a controller consolidates what the run has established, explores next options, assesses what each option is worth under the remaining budget, and dispatches the chosen work with context drawn from persistent memory. Between decisions the controller carries only a compact account of the run rather than replaying its full history. Our baselines span production coding agents and research harnesses, together with a Direct Control Agent using the same workers and compute budget allowance. On ProgramBench, which tests long-horizon agentic capability through program reconstruction, meta-reasoning achieves 71.5% with GPT-5.5 against 58.0% for Codex; with Opus 4.8 it achieves 67.2% against 65.5% for Claude Code. On the other benchmarks, spanning abstract reasoning, multi-domain long-horizon reasoning, and proof generation, it gains between 3.6 and 4.2 points over direct control, averaged across three frontier models. It keeps improving over the tested budget ranges where direct control plateaus, though its overhead can hurt at small budgets. Artifact-graph analysis reveals more reuse of earlier work, higher coverage of correct solutions in most settings, and nonuniform gains in final selection. These results indicate that spending computation on structured control becomes more important as agents scale to longer runs.
Multi-agent LLM systems rely on message passing among specialized agents to accomplish complex tasks. However, an upstream agent may provide useful information or an incorrect answer that causes a downstream agent to override a correct answer supported by its own evidence. Prior work has not clearly separated the benefits of communication from the damage caused by incorrect messages. We study this problem with controlled experiments across five benchmarks and five receivers, keeping the downstream task and evidence fixed while comparing answers under three conditions: no message, the upstream agent's original message, or a message with the opposite conclusion. Our experiments reveal three key findings. First, messages often help when the downstream agent would otherwise answer incorrectly. Second, messages can also hurt: when the downstream agent would answer correctly without a message, an incorrect upstream message changes the answer in up to 32% of cases. Third, in 94% of audited harmful cases, the downstream agent copies the upstream's specific wrong answer--a pattern we term answer substitution. Removing unreliable messages recovers part of the lost accuracy, suggesting that communication should be selective based on upstream reliability and the evidence already available to the downstream agent.
We propose STRAT, an auxiliary task that trains deep reinforcement learning (RL) agents to predict a short textual trace of their own state. Inspired by human spatial navigation, the description combines landmark, route, and survey knowledge, tracking the agent's position, inventory, goals, and immediate progress. Environment rules generate this text online without human labelling. Our method adds a single auxiliary head to a standard policy. Across 60 sparse-reward XLand-MiniGrid tasks, STRAT solves complex environments where standard RL fails outright, while compacting state representations and preventing rank collapse. Beyond performance gains, the predicted trace provides a readable account of agent beliefs at every step for no extra cost.
随着基于 LLM 的智能体执行日益复杂的任务，Agent Skills 已成为扩展其能力的一种灵活机制。Agent Skill 将特定任务的指令与可执行组件及辅助资源打包在一起，以提供专门功能。然而，第三方 Skill 的日益普及引入了新的供应链攻击面。恶意 Skill 可能嵌入有害行为，滥用智能体权限，并危害智能体的执行环境或其可访问的资源。尽管近期基于 LLM 的恶意 Skill 审计方法已取得颇具前景的性能，但它们通常依赖能力较强的商业 LLM。如何在安全敏感且资源受限的环境中，利用可在本地部署的紧凑型 LLM 实现有效审计，仍基本未得到探索。我们的研究表明，紧凑型 LLM 难以识别隐藏在复杂 Skill 包中的恶意行为。这一困难既源于此类行为的隐蔽性，也源于紧凑型 LLM 有限的推理能力。为应对这些挑战，我们提出了 SKILLLITE，一种用于恶意 Skill 检测的证据引导型智能体框架。SKILLLITE 能够从复杂 Skill 包中有效提取与安全相关的行为，并推断其预期功能。随后，它利用紧凑型 LLM，根据观测到的行为及其功能语境评估该 Skill 的恶意性。实验表明，在不同的紧凑型 LLM 主干模型上，SKILLLITE 均能提升恶意 Skill 检测效果，并优于现有具有代表性的审计基线方法。其有效性还可推广至现实环境中经行为确认的恶意 Skill。与此同时，SKILLLITE 保持较低的推理延迟，支持其实用化部署。
Recent efforts to scale tool-use post-training have largely centered on the synthesis of executable environments, which constitute only one component of a broader agentic interaction system comprising the environment, task, agent harness, and evaluator. Scaling environments in isolation, however, does not guarantee commensurate gains in model performance, because reliable learning signals depend on coherent interactions among all components of the agentic interaction system. To address this problem, we introduce WEFT (Whole-system Evolution For Tool-use Post-training), which couples scalable agentic interaction system construction, execution-driven self-evolution, and stable post-training. WEFT scales agentic interaction system construction across environment breadth, task complexity, and interaction diversity. Execution-driven self-evolution iteratively uses execution traces and state evidence to attribute failures and revise the responsible components, with fresh rollouts evaluating the changes and providing evidence for subsequent evolution rounds. For stable post-training at scale, WEFT addresses both optimization and execution reliability: prefix-preserving sampling retains verified progress and atomic-turn credit assignment localizes learning signals, while MegaMCP maintains isolated, recoverable state across concurrent rollouts over shared tool services. Extensive experiments across various models and benchmarks demonstrate the effectiveness of WEFT for tool-use post-training. WEFT-8B and WEFT-14B outperform all evaluated matched-size environment-scaling baselines on BFCL V4, $τ^2$-Bench, and Claw-Eval. In particular, WEFT-14B improves over Agent-World-14B by 6.41, 2.23, and 12.27 percentage points. WEFT-35B-A3B further extends these gains to more challenging long-horizon workflow benchmarks, including Toolathlon-Verified and AutomationBench.
As the capabilities of large language models (LLMs) continue to advance, increasing attention is turning to how to translate their abilities into useful behavior. Personal agents bring this question into everyday settings, where models are expected to serve individual users and continually adapt to their preferences. With the underlying model held fixed, such adaptation relies on harness engineering: designing and evolving the surrounding layer that manages context, memory, tools, and execution. Despite rapid progress, the factors governing effective harness evolution remain insufficiently understood. To narrow this gap, we investigate three central questions concerning harness architecture, harness scale, and self-evolution algorithms through complementary empirical and theoretical analyses. Empirically, we introduce a preference-oriented benchmark and systematically characterize the capabilities and limitations of personal agents associated with these three dimensions. Theoretically, we formulate harness evolution as a learning problem and explain these phenomena through approximation, generalization, and optimization errors. Analyses of reachable policies, capacity under finite interaction evidence, and biased update dynamics provide theoretical accounts of the observed phenomena. Together, these results offer a unified perspective on the limits of personalization through harness evolution and inform future harness design.
Existing reactive Graphical User Interface (GUI) agents often fail in long-horizon, dynamic scenarios, where unexpected disturbances trigger attention-diverting and cascading failures. To address this, we propose PrecogUI, a pre-cognitive architecture that shifts the paradigm from reactive execution to proactive decision-making. Specifically, we design a Proactive Experience Pool (PEP), which caches recurring anomaly and success patterns as "state-action-result" tuples in a dual-memory repository. Furthermore, we introduce a Proactive Simulation Executor (PSE) that learns to forecast the next symbolic UI layout given a candidate action, enabling early anomaly avoidance and ranking candidate actions by predicted reliability. Finally, a Pre-cognitive Execution Controller (PEC) fuses these priors and predictions, prioritizes handling of foreseen anomalies, and ensures execution robustness through a closed-loop error correction mechanism. For robust evaluation, we develop AutoTraj, an automatic data-generation engine, to construct InterfereBench, a benchmark for long-horizon tasks with strong disturbances. Experiments demonstrate that PrecogUI surpasses state-of-the-art methods on InterfereBench while maintaining competitive performance on public benchmarks. The code will be publicly available.
Many computer tasks recur: the same workflow runs many times, with new inputs and from different starting states. Current computer-use agents re-plan every step of every run, which makes them costly and unreliable on such tasks. We introduce neuro-symbolic computer use, in which a recurring workflow is executed by a learned policy rather than re-derived by an agent on each run. The policy fixes the decisions that are stable across runs (ordering, variables, loops, and branches) in executable code, and delegates observation-dependent decisions, such as grounding and state checks, to neural models. We learn these policies with neuro-symbolic policy iteration: starting from one agent trajectory, it executes the policy, diagnoses failures with task-completion and step-level judges, and revises the code with a coding model informed by an agent's continuation from the point of failure, without access to the benchmark evaluator. Iterating on generated parameter and initial-state variants makes the policy reusable, and a pre-action verifier guards each state-mutating step at deployment. On OSWorld-Verified and ScienceBoard, the learned policies achieve the highest Pass^3 of all methods in all four settings, 3.6-15.8 points above the base agent, while cutting per-run cost by 15-217$\times$ and latency by 3.4-5.1$\times$. On OSWorld-Verified, policies built only on variants transfer to the held-out original tasks, exceeding AutoRPA by 8.6-17.5 points in Pass^3.
Traffic signal control (TSC) is essential for improving urban mobility and reducing congestion. Although roadside cameras are widely deployed at signalized intersections and provide rich visual observations of evolving traffic, existing TSC methods typically rely on manually engineered traffic states or separate perception modules, creating a gap between physical observations and control decisions. We present VLALight, the first vision-language-action (VLA) model for end-to-end traffic signal control from multi-view roadside videos. VLALight directly maps visual observations to coordinated signal actions through multi-target spatiotemporal traffic reasoning and topology-aware cooperative perception across intersections. To establish this capability, we develop a two-stage supervised cold-start training strategy for visual traffic understanding and signal decision-making, followed by cooperative agentic reinforcement learning that jointly optimizes local control and network-wide traffic efficiency. Furthermore, VLALight introduces adaptive fast and slow reasoning modes, enabling the policy to allocate deeper reasoning only when additional deliberation provides sufficient control benefits. Through balanced mode-aware rollouts and relative advantage optimization, VLALight learns to trade off decision quality and inference cost. Extensive experiments on seven real-world traffic-flow datasets across three urban networks demonstrate that VLALight consistently outperforms transportation-based, RL-based, and LLM/VLM-based baselines. Ablation studies validate the effectiveness of cooperative perception, network-level optimization, and adaptive reasoning. These results demonstrate the potential of VLA models for real-world physical traffic control. Our project is available at https://github.com/usail-hkust/VLALight.git.
GUI agents automate tasks on digital devices by grounding language instructions in visual interfaces. Existing group-relative reinforcement learning improves GUI action prediction by comparing the rewards of multiple responses sampled from the same GUI state. However, binary evaluation treats spatially different failed clicks as identical and provides no relative signal when all sampled clicks fail. To address these limitations, we propose Spatial Credit Assignment (SCA), which uses the screen coordinates of sampled clicks to refine group-relative credit. Specifically, SCA predicts each held-out response's reward from the other responses in groups containing both successes and failures, then uses the prediction residual to adjust credit. When all sampled clicks fail, SCA instead orders them by distance to the annotated target. These spatial references are used only to construct the training update; the deployed policy remains unchanged. We evaluate whether this correction improves the policy update itself by comparing its error and directional alignment with the exact return gradient in a controlled synthetic study. Across GUI grounding and offline action-prediction benchmarks, SCA improves grounding across professional domains and achieves the strongest results among reinforcement-fine-tuned models on most action-prediction metrics, with consistent gains across the reported GUI suites.
Writing fast GPU code for physical simulation is difficult: implementations must preserve numerical accuracy while handling irregular data access, synchronization, and iterative solvers. We introduce GPUPhysBench, a benchmark of 50 tasks testing whether coding agents can meet these demands. Tasks cover fluids, deformable solids, and granular materials, from individual simulation operators to complete simulators. Agents write, compile, test, and optimize GPU code with access to a NVIDIA GPU under fixed time budgets. We report pass rates and runtime performance relative to expert-optimized reference implementations. In a single-attempt evaluation of six frontier model-harness pairs, the two strongest pass all 50 tasks, but even the fastest reaches at least 0.9 the reference speed on only 22% of them, and no submission is more than 5% faster than the reference. The largest gaps arise in collision detection, constraint solving, and iterative solvers. GPUPhysBench brings physical simulation workloads to coding-agent evaluation, testing both the ability to implement numerical methods correctly and the ability to make them run efficiently.
Mobile GUI agents operate through a perception--action loop: at each step they screenshot the device, invoke a vision--language model (VLM), and emit an action. It is slow, costly, and brittle, yet most of what it does is navigation---and everyday navigation is static, ordered, and endlessly repeated. We present PhoneCLI, which compiles an app's GUI navigation into callable commands, without any app-internal API, runtime instrumentation, or model training. Offline, PhoneCLI explores a target app from the outside and distills its screens, interactive elements, and navigation edges into a semantically annotated map; each screen yields one deterministic command: a replay sequence that reaches it. Online, the agent selects a command, verifies it before execution, and then executes it deterministically in sub-second time at zero VLM cost; open-ended interaction and every failure of the compiled path fall back to the embedded VLM interpreter, exactly the pure VLM agent, so compilation can only help. On AndroidLab, PhoneCLI improves the task success rate while reducing steps and token consumption, and it transfers to AndroidWorld's official M3A agent with consistent efficiency gains. What PhoneCLI compiles is the app's navigation rather than one run, so it serves new tasks, not only repeated ones.
Users of Workday's deployed LLM-based agents often request features which can be addressed by defining named procedures, also known as skills, in the LLM context, effectively augmenting agents' capabilities. However, as an agent's skills library grows in size, so does the agent's operational cost. Progressive disclosure (lazy-loading) of skills as needed may reduce operational costs, but its impact on overall latency and skill-retrieval quality remains unclear. In this report, we investigate the impact empirically and find that progressive disclosure improves skill-retrieval quality but marginally degrades overall latency.
Large language models (LLMs) excel at structured, verifiable tasks, but their low-entropy bias can produce homogeneous and predictable outputs, limiting their utility for open-ended scientific ideation. Effective discovery, however, spans a broader creative spectrum: from structured day science to loosely structured, serendipitous night science that reaches ideas beyond those typically considered. We introduce AI Night-Scientist, an agentic framework that uses reinforcement learning to teach models when and how to depart from predictable reasoning. Grounded in cognitive science, we model creativity along three axes: action (what to do and how creatively), process (when to explore versus exploit), and outcome (the novelty and usefulness of the resulting idea). We use these axes to train models with GRPO, exposing them to varying degrees and forms of creativity throughout training. This produces substantially more diverse scientific proposals, expanding the range of research directions by 27.8% and contribution types by 14.9% over the base model. It also improves predicted citation impact by up to 32.0 percentage points and originality by 66.2 points. These gains cannot be reproduced by simply increasing decoding temperature; instead, we find that semantic guidance specifying what kind of creativity to pursue is critical. Overall, our results suggest that creativity is a learnable, multi-level ability that can be shaped to help researchers reach ideas beyond those typically explored by LLMs.
使用工具的代理可能会失败两次：所需工具可能失败，随后代理又可能在缺乏必要证据支撑的情况下报告成功。现有基准通常将这种报告失败与工具选择、恢复和环境动态交织在一起。我们提出 Failure-Transparent Agents (FTA)，这是一个受控基准，在生成之前固定失败的观察结果和所需证据状态，使故障后的声明能够被直接审计。FTA 包含 100 个任务，具有跨越五类故障的确定性失败轨迹、一个中性对照以及四种用户压力条件，并在评估无依据声明的同时评估有用的恢复。在六个模型、三种响应策略和 3,600 条人工标注响应中，基线策略下的虚假成功率为 22.8%，使用透明度指令时为 9.3%，使用结构化证据契约时为 0.8%。捏造细节率分别从 28.3% 降至 14.3% 和 0.8%，而有用响应率则分别从 74.9% 提升至 89.2% 和 98.8%。在这个阻塞任务基准中，所测试的证据契约策略与显著更低的故障后报告错误相关，同时有用响应率保持较高水平。
人们不仅通过重复成功的行动来学习，也会通过复述和解释自己的经历，修正理解以指导未来行为。语言模型智能体能否仅通过训练其自身经历的解释来改进未来行动？我们通过研究 Retrospection-Only Fine-Tuning (ROFT) 来探讨这一问题；ROFT 是一种极简的在线流程，旨在隔离仅基于解释的训练对后续行为的影响。智能体尝试一项任务，观察可用反馈，生成回顾性解释，并仅在解释 token 上使用下一 token 预测损失进行微调。该流程既不使用外部教师，也不进行基于奖励的策略更新。在使用 Qwen3.5-4B 的软件工程实验中，ROFT 在基础模型尝试结果成功与失败混合的问题上进行训练。在留出的 SWE-bench Verified 和 Pro 上，在不使用验证器的情况下，它经过 20 次更新后达到 49.2% 和 26.8% 的解决率；相比之下，在所评估的运行中，GRPO 经过 40 次更新后的解决率为 48.0% 和 25.3%。ROFT 在训练时间和采样尝试次数方面也表现出更快的早期进展。它还学会了解决一些单个任务，而基础模型在这些任务上的 64 次采样尝试全部失败，这表明即使最初没有任何成功轨迹，学习也可以开始。行为分析发现，ROFT 会间接地为行动分配信用，鼓励好的行动并抑制错误行动。此外，通过提示回顾强调更直接的解决方案，即使没有显式长度惩罚，后续尝试也会变得更短。总体而言，这些发现表明，学习解释也能改善学习行动，确立了自生成回顾作为有用训练目标的价值，并推动对解释到行动迁移的进一步研究。
Evaluating finance research agents requires rubrics that reflect expert standards and fix the values correct as of an information cutoff. Expert-reviewed finance benchmarks rely on fixed, per-item rubrics, which are costly to extend and cannot encode each institution's own standard. In FinAutoRubric, experts specify reusable evaluation guidance, while agents and code carry out query-specific rubric generation, review, and validation. This expert guidance governs every agent, as prompts and as rules that code enforces, and a Task Bank of reusable criteria carries it across tasks. In long-horizon loops that follow the expert guidance, a writer agent researches every expected value and a reviewer agent verifies it, and failures escalate to a human. On three expert-authored finance benchmarks, its rubrics track expert scoring as closely as the strongest evaluated generator while stating the expert rubric's expected value for more criteria, their scores agree with human grading, and in-house analysts prefer them in a blind review. The released 100-query FinAutoRubric Benchmark, built from in-house analysts' key questions across 78 tasks and eight asset classes, shows that rubrics from an earlier model generation still leave headroom for a later one.
扩展智能体式 LLMs 的视野长度受到需要将越来越长的上下文轨迹装入 GPU 内存这一需求的瓶颈限制。上下文压缩一直是缓解这一问题最流行的机制，它能在给定轨迹下保持 GPU 内存占用恒定。不幸的是，大多数压缩策略依赖于多次预填充 LLM 上下文，从而阻碍训练吞吐量。为缓解这一瓶颈并实现高效的可训练压缩，我们提出 KV-streams，这是一种即插即用策略，兼容任何压缩策略，能够显著提高吞吐量，同时没有证据表明会妨碍性能。KV-streams 通过向前流式传递 KV cache，而不是在每次压缩后将其清空，从而实现可扩展的压缩。我们展示了 KV-streams 可支持三种不同的压缩策略，在训练中实现 2.6 到 5 倍的实际耗时加速。除效率之外，我们发现流式传递的 KV cache 可以充当循环状态，向前携带早已从上下文中消失的信息。具体而言，在受控设置中，我们表明，与先前工作相反，仅凭 RL 就足以使这种行为涌现。总体而言，我们证明 KV-streams 是对任何后训练流水线的一种高效、轻量的即插即用补充。
当大语言模型（LLM）Agent 执行同一任务时，不同运行之间的 Token 消耗可能相差超过一个数量级。Agent 会根据工具反馈和中间结果选择下一步行动，而不断增长的上下文会持续扩大后续每次调用的输入规模。因此，任务的总消耗在执行前很难预测，并且预测必须随着运行展开而不断修正。本文提出 TokenCast，它为每个执行片段学习一种可组合的成本表示，记录该片段自身的消耗以及它引入的上下文增长。组合相邻片段可得到一个累积估计，捕捉早期片段的上下文被后续每次调用重新读取时产生的额外输入成本。随着执行展开，新观测到的证据会刷新预测，无需额外的 LLM 调用，并且在 SWE-bench Verified 上每次运行的平均累积预测时间为 32.8 ms。在 4 个任务套件和 6 个 Agent 模型中，TokenCast 相较最强对比方法的平均绝对误差降幅在 96 个评估组合上的平均值为 14.5%。在离线预算控制回放中，在匹配轨迹完成度的情况下，TokenCast 平均比固定预算策略少使用 21.3% 的 Token。代码可在 https://github.com/DEFENSE-SEU/TokenCast 获取。
仅结果强化学习为软件工程（SWE）智能体提供终端成功信号，但对中间决策几乎没有直接指导。我们引入 Counterfactual Rollout Replay（CRR），这是一种训练时过程，利用可分叉的可执行环境来获得步骤级回报对比。CRR 选择一小组决策点，恢复每个状态，采样一个替代动作，并在该策略下将该分支向前 rollout。它保留已实现的训练轨迹，并将所选步骤的优势替换为其终端回报与采样得到的反事实回报之间的差值。该方法不需要人工过程标签或学习得到的过程奖励模型；免费指的是这些监督成本，而非 replay 计算成本。使用 14B 策略时，CRR 在 SWE-bench Verified、SWE-bench Live 和 SWE-rebench 上提升了 pass@1，并且可与过程奖励和轨迹搜索方法结合。在 SWE-bench Verified 上，在相同硬件上进行等墙钟时间比较，得到 41.7% 对 36.7% 的结果，相比延长的仅结果 GRPO 提升 5.0 个百分点，且已计入分叉开销。这些结果适用于能够以可承受成本可靠恢复状态的环境；随机续接以及昂贵或不完美的 replay 仍然是局限。
LLM 代理越来越依赖用户批准来在运行时授权安全敏感操作。此类批准是在特定任务和执行上下文中授予的。在长生命周期代理中，授权决策可能需要跨任务或会话持续存在。我们发现，这种连续性可能会超越最初证明批准合理性的上下文，从而产生无需重新征得同意即可重复使用的残余权限。我们通过一种纵向攻击揭示了这一失效模式：该攻击从目标安全敏感操作出发，识别执行该操作所需的权限，诱导良性交互以合法获取该权限，并随后在对抗性执行期间重放残余权限。在受控环境和真实环境中，我们证明残余权限重放会在实践中出现，并显著提高提示注入和上下文重新绑定攻击的成功率。我们使用源自生产环境的授权语义，在六个 LLM 系列上评估了 508 个 AgentDojo 攻击案例。在存在残余权限的情况下，与全新的授权状态相比，攻击成功率（ASR）最高提高 35.1 个百分点。在针对三个真实生产编码代理、覆盖 55 个 Terminal-Bench 案例的真实上下文重新绑定攻击中，残余权限重放使 ASR 平均提高 24.9 个百分点。这些发现揭示了长生命周期 LLM 代理中持久授权与用户同意的上下文性之间的根本不匹配。
Long-horizon tasks require large language model (LLM) agents to coordinate decisions under constraints that span an entire solution. Monte Carlo Tree Search (MCTS) offers a promising approach to test-time scaling by exploring alternative action trajectories, but model computation and environment interaction make search costly. Efficient search therefore requires effective reuse of trajectory feedback. Standard MCTS maintains prefix-specific statistics, without explicitly accumulating outcomes for decision groups that recur across different paths. To fill this gap, we propose HyperMCTS, a training-free method that augments an ordered MCTS tree with a cross-trajectory hypergraph. Hyperedges represent groups of canonical decisions and accumulate their observed returns within the current task. Our hypergraph-guided HyperUCT selection rule aggregates evidence from overlapping hyperedges into an action prior, allowing outcomes collected under one prefix to inform selection under another while preserving execution histories in the tree. On DeepPlanning, HyperMCTS improves average planning accuracy by 2.3--7.3 percentage points over the strongest baseline for each of three backbone models. It enables Qwen3.6-27B to outperform Claude Opus 4.6 (max) on Shopping Planning, while achieving higher accuracy with fewer LLM calls and output tokens than the evaluated MCTS-based baselines. SealQA experiments further demonstrate improvements in question answering.
许多 LLM-as-a-judge 评估是在固定任务定义下对固定输出进行评分。而生产环境中的多轮业务智能体需要的是一个可维护的测量系统：正确性取决于特定业务的事实和流程，结果会在多轮交互中逐步形成，而失败必须在可执行处理之前归因于智能体能力不足或业务知识缺失。我们提出了一套涵盖评估规范、模块化 LLM 评判器、保持意图的用户模拟以及人在回路治理的集成方法。该规范定义了对话级最终状态和可执行的失败归属。原子评判器共享版本化证据，并输入一个显式聚合图。模拟器只有在通过任务保持性和稳定性检查后才会发布。独立人工审计用于估计测量保真度，更新分层参考集，并将分歧分流至标签修正、指南修订或评判器改进。生产研究表明，系统级保真度在重复审计中有所提升；在共享反馈循环下，人工审核者和自动评判器共同改进；并且在两个报告的任务完成设置中，它们的组合工作流都具有最强的描述性表现。由于这些研究是观察性的，且人工参考本身也需要修订，因此这些发现证明的是操作层面的实用性，而非因果性或普遍优越性。本文的贡献在于提出了一个实用框架，使多轮智能体测量能够随着被评估系统及其证据的演进而保持可靠、可执行且可维护。
Large-language-model multi-agent systems (LLM-MAS) introduce a characteristic reliability problem: an error produced by one agent can be accepted as context by downstream agents and propagate across the workflow. Many proposed safeguards rely on learned or LLM-based judges whose verdicts are themselves probabilistic; we ask whether a deterministic layer can instead stop contract-detectable handoff defects. We present Maat, a runtime governance layer that validates agent-to-agent handoffs against a versioned workflow contract, or anchor, with no language model in the validation or scoring path. We evaluate it in six controlled domain workflows (6-15 agents, 522 trials) with injected data-level defects and a deterministic seven-check rubric. Version 1 reported gains in all six workflows (2.9-26.5%). A post-publication audit found that three benchmark scorers credited any early halt as a prevented defect. On paired trials where the governed run completed or halted on a finding attributable to a verified defect, the rubric score changes by +7.7% to +29.1% in five workflows and is flat in software development; model-call cost falls 17-53% where attributable halts occur early. A hand review of all 94 governed-arm halts found 35 false alarms (37%), caused by validator defects rather than model behaviour; counting those halts as failed work, the governed arm scores below the ungoverned arm in four of six workflows. The results support deterministic handoff validation for contract-expressible defects and show that validator configuration and halt attribution must themselves be tested; they do not establish universal correctness, hallucination detection, or model-independent effectiveness.
On-policy distillation (OPD) trains a student on its own rollouts using dense supervision from a teacher. In multi-turn environments, a mistake at a critical decision step can redirect the subsequent rollout toward poor outcomes. We use low teacher confidence on student actions to select high-uncertainty steps for correction. In a controlled ALFWorld study, a single teacher correction at a low-confidence step improves subsequent student behavior and task success, motivating selective intervention during distillation. We propose UOPD, an uncertainty-aware intervention method for on-policy distillation. At low-uncertainty turns, UOPD executes student actions and applies the standard OPD loss. At high-uncertainty turns, it samples and executes teacher actions and trains the student to imitate them through supervised fine-tuning, which minimizes forward Kullback-Leibler divergence in expectation. UOPD utilizes adaptive uncertainty thresholds to target a scheduled intervention rate. Empirically, we evaluate UOPD across a broad range of agentic tasks, including ALFWorld, WebShop, and Search, demonstrating its superior performance over OPD methods and their variants. UOPD improves WebShop score by up to $15.8\%$ relative to standard OPD.
Objective: To develop and characterize CLEAR-Med, a dual-agent framework for natural-language analysis of structured clinical data that separates SQL-based invocation from independent validation. Methods: CLEAR-Med uses one agent to translate a question into executable Structured Query Language (SQL), retain the executed query and database result, and produce a draft. Deterministic checks and a separately invoked cross-provider Validation Agent then accept the draft, request one bounded repair, or abstain. We formalized the system as a bounded selective pipeline and evaluated CLEAR-Med's configuration and scalability, and the Invocation Agent's accuracy and consistency on a 25-query development benchmark, using a harmonized 21-site neonatal hypoxic-ischemic encephalopathy table containing 532 de-identified infant records and approximately 1,300 variables. Results: CLEAR-Med completed all six nominal scalability configurations, including 500x1300. Across 25 development-benchmark queries repeated five times, the Invocation Agent answered 83 of 125 responses correctly (66.4%; query-cluster bootstrap 95% CI, 48.0-83.2%), compared with 15 of 125 (12.0%; 95% CI, 3.2-22.4%) for the ungrounded ChatGPT baseline, a paired improvement of 54.4 percentage points (95% CI, 36.8-72.0%). Conclusion: CLEAR-Med provides a general architecture for traceable analysis of structured clinical data: numerical claims remain linked to executed SQL, and unresolved cases can fail closed. The reported experiments characterize CLEAR-Med's configuration and scalability and the Invocation Agent's accuracy, while the formal analysis establishes the encoded-property guarantee of the complete control flow; a prospective full-pipeline evaluation of the validation and abstention stages is the next stage of this work.
文章 URL: https://github.blog/security/how-we-found-24-android-vulnerabilities-using-our-open-source-ai-security-agent/ 评论 URL: https://news.ycombinator.com/item?id=49886609 得分: 5 评论数: 1
As the use of agentic artificial intelligence increases in nearly every industry, there exists a widening attack surface. It is necessary to monitor agents to ensure that agents are acting in a way that is aligned with the users intent. Auditing an agent's network traffic provides a clear record of the agent interactions. This work presents a novel approach to monitoring the network traffic of agentic systems using complex valued hypersparse traffic matrices by integrating DBOS (DataBase OS), the OneSparse PostgreSQL database, and the GraphBLAS math library. To develop these concepts an agentic simulator was constructed, allowing a varying numbers of AI agents to collectively survey a virtual environment using different strategies. The resulting network traffic matrices enable easy monitoring of the AI agents.
Healthcare administrative staff transfer structured information from electronic health records, referrals, claims systems, provider rosters, and work queues into dynamic forms. We developed and evaluated CLAIRE (Clinical Language and Agentic Intelligence for Reasoning and Entry), a hybrid workflow that separates field-state discovery, source-to-field mapping, deterministic validation, bounded correction, escalation, and audit tracing. We tested five synthetic healthcare administrative schemas, 1,000 source records, four interface variants, two data-quality suites, and six comparators, yielding 24,000 benchmark episodes. A separate strict-output audit evaluated direct mappings from Qwen2.5-1.5B and Qwen2.5-7B, and a trace-derived operational simulation covered 6,000 episodes. Under the evaluated synthetic benchmark conditions, full CLAIRE achieved 1.000 episode success, field accuracy, required-field completion, and dependency completion in both suites; removing validation reduced stress-suite success to 0.500. In the simulation, 100.0% of clean and validation-stress episodes reached a staff-reviewable draft, compared with 68.6% of escalation challenge episodes, unsupported cases were blocked. Scenario-based savings were 149.7-165.5 seconds per case, not observed staff times. The findings support schema-grounded, validation-first healthcare administrative automation in which language-model components assist mapping but do not authorize unsupported or consequential actions.
大语言模型（LLM）智能体可以帮助用户完成日常任务，而这些任务通常有许多合理的完成方式。即使它们的回答有用，智能体执行这些任务的方式也可能不符合用户的偏好和需求。例如，不同智能体在是否提出澄清性问题或搜索网络方面存在差异。我们提出 HABIT，这是一个包含五大类别、23 个行为维度的分类体系，由三位作者和三个 LLM 基于 17 个领域中的 408 条智能体轨迹自下而上推导而来。在留出任务上，相比现有的人类价值观和智能体行动分类体系，HABIT 能更清晰地区分模型，同时支持相近一致性的标注。基于 HABIT，我们构建了 AgentHABIT，这是一个基准测试，可根据智能体在 86 个日常任务上的轨迹刻画每个智能体的行为倾向。使用 AgentHABIT 对 18 个模型进行画像揭示了一系列鲜明的倾向。例如，大多数 GPT 和 Claude 模型会陈述其假设，并在需求冲突时提供替代方案，而 Qwen 和 Google 的模型更常不明确说明假设或需求变更。即使使用完全不同的任务集构建，这些画像仍然可被识别，表明它们反映的是一般性倾向，而非特定任务行为。提示智能体采用特定行为会使某些维度很容易发生变化，但对其他维度几乎没有影响；而在另一个模型的轨迹上进行微调只会改变模型画像的一部分，并保留其中的大部分特征。总体而言，HABIT 和 AgentHABIT 提供了一个系统性框架，用于刻画智能体在任务成功之外如何执行日常任务，并提供洞见以指导开发行为更符合用户需求的智能体。
Embodied task planners may produce multi-step plans whose subtask dependencies and interactions with the environment create physical risks during execution. Yet existing safeguards overlook such compositional risks, as general-purpose guardrails focus on semantic harm and embodied safety detectors assess subtasks in isolation. To address this gap, we introduce PlanGuard, the first pre-execution detector that evaluates the physical safety of a complete multi-step plan in its current environment. For training and evaluation, we construct a Multi-Step Plan Safety (MSP-Safe) dataset through paired task construction, plan generation using diverse planners, and safety annotation by three judges. Task-oriented SFT on MSP-Safe establishes fundamental plan-safety assessment capabilities, yet a substantial gap remains between compact models suitable for real-time deployment and stronger but costlier large models. Accordingly, we propose Strong-Teacher Adaptive Compensation for On-Policy Distillation (STAC-OPD), which provides compact models with adaptive strong-teacher supervision along their on-policy trajectories. It combines token-level distribution transfer from a fine-tuned strong teacher with probability-routed sequence-level compensation, retaining student-generated targets when the student favors the reference safety decision and using teacher-reconstructed targets otherwise. Across all test subsets, PlanGuard-2B achieves average 87.15% ACC and 87.21% F1, demonstrating effective whole-plan physical-risk detection at compact model scale. Code and dataset will be publicly released.
One variable sets what an upstream fault costs a staged pipeline of language-model agents: re-derivability, how much of what a stage needs it can rebuild from the original problem. Grounding an inspector agent in that problem is worth +0.608 [+0.517, +0.700] to +0.358 over a blind one on four open-weight backbones served with thinking disabled, and on the two Qwen backbones the blind inspector changes no item at all. That head-to-head is exploratory. One deterministic fault enters the first stage, and we re-expose the original problem to $k = 0,\dots,3$ of the downstream stages with agents, items, fault and topology held fixed, on 120 gsm_hard items per arm at temperature zero. Accuracy under fault rises on four of four backbones, from +0.233 to +0.392, the largest Holm-adjusted $p$ being $2.1\times10^{-6}$. A registered kill test rules out tokens. Blanking every word holds the word slots fixed, and retention tracks the visible fraction on four of four, climbing from 0.221 to 0.692 on the primary. Those two families are confirmatory and everything else here is exploratory. The interaction excludes zero on two of four backbones under the registered pipeline, four of four under a three-stage pipeline, and three of four under full message history, the primary at +0.317. On Llama-3.1-8B the fault carries no detectable cost at any dose, so the other three carry every claim about what a fault costs. Re-derivability also sets what the architecture costs, and no decomposition we measured reliably beats one direct call. With no fault injected the registered pipeline loses to that call by -0.267, -0.125 and -0.317, and on Phi-4 reads +0.058 at $p = 0.118$, which the test fails to separate from zero. The repair that works is cheap and front-loaded: the first re-grounded stage buys +0.394 of matched retention for +59.8 tokens per item on Qwen3-14B, and the stages after it buy nothing.
Generative and Agentic IoT systems offer a promising foundation for digital healthcare applications that combine sensing, personalized reasoning, and autonomous interaction in real-world environments. Nutrition assistance is a natural use case, but existing Large Language Model (LLM)-based systems are often limited to passive text interaction and static context, making them unreliable when food descriptions are ambiguous or nutritional evidence is missing. We propose Nutri-ATLAS, an Embodied Agent for Tabulated Lookup and Assistance for smarter nutrition in the real world. It integrates graph-grounded nutrition reasoning, hardware-aware LLM selection, and robot-based evidence acquisition. Nutri-ATLAS builds a unified Food-Nutrient knowledge graph from USDA FoodData Central and FoodKG and learns 64-dimensional GATv2 food and recipe embeddings. A shared hybrid graph-text scoring mechanism supports food nutrition extraction, nutritional gap filling, substitute retrieval, and recipe-level meal composition, while an LLM-guided skill interface navigates landmarks, updates dietary-context and food-accessibility memory, and grounds recommendations in observed food availability. We evaluate Nutri-ATLAS across nutrient estimation, substitution retrieval, recipe recommendation, patient-profile adherence, edge deployment, and real-world embodied execution. On HealthyFoodSubs, the hybrid retriever achieves 37.9% MAP, 80.7% RR@5, and 90.1% RR@10. On NutriBench v2, Dense+GAT retrieval grounds nutrient estimation across nine quantized Qwen3.5-9B configurations. On PFoodReQ, Nutri-ATLAS reaches 78.8% MAP, 83.0% MAR, and 77.5% F1. A patient-profile study shows adherence to allergy and healthy-target constraints for all selected cases.
Long tasks produce more history than an LLM agent can hold in its context, and more than it uses reliably even when the history fits. A growing line of work therefore has agents carry a short written state instead: at every step a writer rewrites the state, and a reader acts from the state alone. Steps stay cheap, but anything the writer drops is lost before later decisions reveal that they need it. We quantify this loss and ask whether training can reduce it. Comparing the written state with the best state of the same size written in hindsight, we split the reader's loss into a budget loss, which any state of that size must incur, and a write-time regret, which comes from the writer's choices. In TextWorld cooking games where we control how long a fact must be carried before it is needed, a 128-token state holding the facts wins nearly every game, while prompted language-model writers win at most 17%. Almost all of the loss is write-time regret, and it grows with the delay. We then train the writer from the reader's own loss. DSSR (decision-sufficient state representations) scores candidate states by how well the reader acts after the writer carries them forward, and teaches the writer to prefer the better ones. This forward-rolled score predicts game outcomes ($ρ= 0.48$), whereas scoring a candidate as a fixed context, as hindsight methods usually do, does not ($ρ\leq 0.07$). On a pre-registered test split opened once, training adds +7.0 [+1.9, +12.2] points of success when facts are needed soon, bringing a plain summary writer to the level of belief- and slot-based memory prompts. The gain shrinks as the delay grows and is significant only at the shortest delay. We trace this limit to credit assignment: keeping a fact now pays off only if every later rewrite keeps it too, which a per-step score cannot see.
Several works have explored multi-agent reinforcement learning (MARL) in LLM collaboration. However, constructing reliable rewards is difficult in practice, as complete and accurate metrics are often unavailable and hard to aggregate. Preference learning provides an alternative by learning from comparative human or AI feedback. Yet, its extension to multi-agent systems remains underexplored. To address this gap, we formulate preference-based multi-agent systems (MAS) from decentralized and centralized collaboration perspectives. We also introduce a general multi-agent preference learning framework (MAPL) to solve these problems. MAPL allows iterative updates by comparing the current solution with decentralized or centralized solutions generated by various agents. We instantiate MAPL using MARL from human feedback (MARLHF) with a learned reward model and multi-agent direct preference optimization (MADPO). Experiments on collaborative writing, coding, tool use, and travel planning show that MAPL can improve collaboration quality and efficiency while approaching the performance of MARL with fixed, well-defined rewards. Within MAPL, MARLHF generally outperforms MADPO on most tasks but remains sensitive to data coverage, agent and comparator models, and the underlying MARL algorithms.
文章 URL: https://www.cnbc.com/2026/09/28/nvidia-releases.html 评论 URL: https://news.ycombinator.com/item?id=49879883 点数: 41 评论数: 74
Hey HN！我们是来自 Vespper（ https://vespper.com ）的 Dudu 和 Topaz。Vespper 是一个 MCP，可让 AI agents 高效编辑 Word 文档，由我们微调的模型驱动。目前它比最接近的替代方案快 3 倍、便宜 2 倍，而且更准确。可以在这里查看产品工作方式概览： https://youtu.be/odKxsgPjzzw 我们是在为制药公司构建 AI 文档编辑器一年后，开始研究这个问题的。在那之前，Topaz（我本人）是 Snyk 的高级 SWE，负责分布式系统；Dudu 是 Viz.ai 的深度学习工程师，构建用于中风检测的计算机视觉模型。 我们的编辑器帮助制药公司生成监管文档（例如 CSR），以加快其提交流程。最初，输出是 Markdown，并显示在 WYSIWYG 编辑器中。然而，用户更愿意使用自己的 Word 模板。这就是问题开始的地方。 AI agents 并不擅长编辑 Word 文档。一个 Word 文档是一个由冗长 XML 文件组成的 zip 文件，遵循 OOXML 规范。即使是“很小”的改动也需要大费周章，例如：添加编号列表需要在 numbering.xml 中用一个新的 ID 创建条目，并在 document.xml 中把它链接回去；加粗一句话需要把它拆分成 3 个以上的 run 元素。类似问题还有很多。 这使得直接编辑 zip（unzip + grep + sed）对 agents 来说是个坏主意，因为它们会在这些机制上消耗大量时间和 tokens。 在实践中，今天的工具大致分为三类。你可以让 agent 针对 python-docx 或 Open XML SDK 这类底层库编写代码；你可以给它一个带有主观编辑工具的 MCP（SuperDoc、Office CLI、Adeu 等）；或者你可以用 pandoc/mammoth.js 之类的工具把文件在 Markdown/HTML 之间来回转换。 它们都不太行。前两类仍然会把 agent 的上下文消耗在 Word 机制上，而不是手头任务上（MCPs 还会引入一种需要学习的新 DSL）；第三类则损失很大（pandoc/mammoth.js/等无法保留足够的保真度）。 根据我们的亲身经验，这些问题会损害下游任务的性能。 当我们尝试让 agent 填写大型文档时，事情很快就崩了。上下文窗口已经塞满了客户数据（文件、用户上下文、全局规则等），而 agent 又把 tokens 和时间消耗在探索文档以及调试失败的编辑上。填写一份 CSR（临床研究报告）大约需要 50 分钟，而且结果很糟（漏掉字段/章节、样式损坏等）。Harvey.ai 的团队也得出了类似结论： https://www.harvey.ai/blog/building-an-agent-for-complex-doc... 就在那时，我们转移了关注点。我们设计了一个 MCP，让 agents 像编辑 HTML 一样编辑 Word 文档。agent 接收 HTML，进行查找和替换编辑，然后我们把这些编辑协调回原始 .docx 文件中。 我们选择 HTML 而不是 Markdown，是因为它在结构上与 OOXML 更接近，而且 CSS 将样式与元素关联的方式大致类似于 OOXML。我们不得不编写自己的 DOCX→HTML 转换器，因为 pandoc 和 mammoth.js 无法保留足够的保真度。 需要说明的是，我们的 DOCX → HTML 转换也有损。但这没关系，因为我们从不把 HTML 再转换回 DOCX。HTML 只是给 agent 的一个投影，因此它只需要具备足够的保真度，让 agent 理解正在编辑内容的结构和样式。原始文件始终是真实来源，我们会原地修改它。 这也意味着 agent 不需要学习新的 DSL。编辑 Word 文档感觉就像在文件系统上编辑 HTML 文件一样，而这是 agents 已经很擅长的事情。许多 DOCX MCPs 会一次性交给 agents 几十个甚至上百个工具，让它们临场摸索。我们的 MCP 只暴露三个工具（read、search、edit）。Word 文档被完全抽象掉了。 在 agent 向我们发送编辑请求（一个“old_html”和“new_html”对）后，我们会将其协调到原始 .docx 文件中。这个协调过程由我们微调的模型驱动，它是一个 3-8B base，带有 LoRA adapter。它以 HTML diff 以及原始的局部 OOXML block 作为输入，并输出新的 OOXML。“定位”是确定性完成的：我们获取 agent 指定的锚点，并尝试找到它们对应的 XML 双胞胎，因此协调模型只承担单一职责。 在这个狭窄任务上，一个小模型可以达到经过良好 prompt 调优的前沿模型水平，同时又足够小、足够快，可以放在热路径中。 这个项目最终变成了数月的工作，但我们很高兴发布 v1。我们的内部基准显示，它比 DOCX skill 和原始 python-docx 更准确，同时便宜约 2 倍、快约 3 倍，主要原因是它每个任务只需要 3 次工具调用（p50），而 DOCX skill 需要 10 次，Office CLI 需要 13 次。 目前还不完美。例如，我们暂时不支持操作图片或评论。尽管如此，我们已经看到人们以各种方式使用我们的 MCP： - 法律科技公司用它驱动其在 Office.js 中的实时编辑流程。 - AI startup 用它优化人们的简历，并代表他们投递申请。 - 需要起草政策备忘录的 govtech。 - 一家生命科学 startup 使用长时间运行的 agents 来完成监管表格。 关于隐私的一点说明：我们的 MCP 运行在云端，因此用户会把他们的 .docx 文件发送给我们。我们不会用用户数据进行训练，团队也可以选择 ZDR 或 self-hosting。 我们有一个免费层级，每月 500 次编辑，也很希望大家都来试试。我们以后想提高这个额度，但我们是个小团队，运行微调模型并不便宜 :( 我们很想听听你们关于 docx 编辑的想法和评论！接下来几个小时我们会在评论区回复 评论 URL： https://news.ycombinator.com/item?id=49881505 Points：6 # 评论：2
大家好 HN，我是 Per，Scrimba（YC S20）的创始人。在过去十年里，我们一直用一种基于 HTML 的视频格式教人们编程。现在我们把 LLM 接入了其中，让人们可以创建关于任何内容的讲解视频。它叫做 “Scrimba Explain”。 为了向 Hacker News 演示这项技术，我们构建了 HN.watch。它就像 HN，但用讲解视频代替文章。我们会在有人第一次点击某个链接时即时创建这些视频。 虽然使用 HTML 而不是扩散模型有明显的视觉缺点，但它有三大优势：- 速度：比基于像素的视频生成快得多（从点击到播放只需几秒）- 成本：我们每个视频的成本约为 0.04 美元。（不包括某些视频会使用的图像生成；这会很快推高成本）- 易于编辑：上述优势也让 AI 辅助编辑变得便宜且快速 我们的假设是，如果视频创作从“几美元和几分钟”变成“几美分和几秒钟”，就会解锁一批新的使用场景。以下是我们已经看到的一些：- 为每一个 Pull Request 提供视频讲解（我们内部就是这样做的）- 为你的内部/外部文档中的每个页面配一个视频 - 在约 4 秒内把一篇复杂文章变成视频（通过我们的 Chrome 扩展）- 课程创作者可以在录制正式内容前快速起草课程 - 人们还会创作很多个人内容，比如给孩子的故事、婚礼邀请、生日等 这个技术栈基于一种开源编程语言（Imba），由我们的 CTO Sindre Aarsæther 创建。它会编译成 JavaScript，因此能与 npm + node 生态系统完全互操作。你可以在这里了解更多： https://imba.io/ 我们还构建了自己的同步引擎（OP），以及一个用于代理的上下文管理系统（Q）。我们原本担心这会让 LLM 在为我们写代码时遇到困难，因为这两者都不在它们的训练数据中（其中 Imba 的数据也很少）。然而，我们惊喜地发现 LLM 实际上非常擅长我们的技术栈。这可能是因为这个技术栈极其紧凑。Imba 很简洁，OP 也是如此，一个声明就能设置存储、同步、权限、UI，以及 AI 能看到的内容。这意味着前端、API、数据库和 JSON 之间不需要相互转换，模型也就不容易混淆和出错。 简单来说，我们没有使用 React.js、Express、Supabase 和 LangChain，而是从零开始全部自己构建。绝对是深受“非我发明”综合征之苦，哈哈！至于模型，我们使用 Gemini、GPTs、Inworld、ElevenLabs 以及其他一些模型。 如果你想试试，可以任选一种：- Web UI（scrimba.com/explain）- MCP（添加到你的编码代理中）- ChatGPT Plugin - Chrome Extension 你可以在我们的文档中找到以上所有内容的链接： https://docs.scrimba.com/explain/introduction 最后，这是该工具的一个真正基于像素的视频： https://www.youtube.com/watch?v=k6rbHmBxSEs 很想听听你的反馈，也想知道是否有人对其他使用场景有想法。 PS：考虑到 HN 群体有多偏爱文本，我预计这次发布会在 HN 上遇到相当多的反对意见。这类工具并不适合所有人。但如今确实有很多人更喜欢视频而不是文本，尤其是年轻一代。 评论 URL： https://news.ycombinator.com/item?id=49879401 积分：18 评论数：8
文章 URL： https://blog.cloudflare.com/cloudflare-cf-cli-launch/ 评论 URL： https://news.ycombinator.com/item?id=49879577 点数：6 评论数：0
Agentic LLM inference accumulates long KV caches across iterative observation-reasoning-action loops, imposing substantial memory overhead and limiting serving throughput. Existing compression methods emphasize overall output quality, overlooking the asymmetric importance of actions in driving task progress. Our key idea is to establish a compression criterion that values KV entries by their contribution to action generation and prioritizes action quality. However, iterative execution, dynamic memory demands, and scattered action-critical entries pose challenges to eviction policies, budget allocation, and paged memory integration. To this end, we propose ActKV, the first KV cache compression framework tailored for agentic LLM inference. (i) Action-oriented KV cache eviction exploits stable action access patterns to retain entries critical to future actions, supporting reliable task progress under compression. (ii) Confidence-driven adaptive budget allocation uses LLM's intrinsic confidence to adapt the budget to evolving action-critical memory demands. (iii) Page-aware compression management standardizes compression into three primitives with customized kernels, realizing practical throughput gains. On long-trace tasks, ActKV retains an average of 98.53% of FullKV's accuracy with only 25.98% of its peak KV cache memory. It also achieves 3.97 times and 3.58 times FullKV's token and task throughput, delivering state-of-the-art performance.
Large language model-based multi-agent debate (MAD) systems are being increasingly used as complex decision pipelines in distributed processes. However, their final synthesis phase still remains inadequately controlled. Even with detailed debate logs, summarizing models are prone to fabricating smoothly written debate consensus that is not grounded in the debate's history. To address this safety gap, this paper presents empirical research and studies if the introduction of active post-debate verification can mitigate the production of such factually unsupported summaries, while still providing valuable information. Furthermore, it is examined whether explicitly signalling divergence is preferable in the absence of a reliable compromise. The Active Provenance Gate (APG) is introduced as a post-debate verification layer that treats the source as a hard constraint, analysing the debate logs, auditing each claim, and applying self-correction. In crisis simulations, the self-healing mechanism more than doubles the average data Provenance Fidelity in difficult condition scenarios, before the strict gate blocks unsupported claims and generates divergence reports. In the human study, a vast majority of the users (over 75%) preferred a report explicitly stating failure in critical scenarios, despite most of them perceiving fabricated consensus from the baseline system as more fluent. Our main contribution is the transition of data origin tracing from passive logging to active conditional blocking before publication.
Tool observations dominate the context of software-engineering agents, making long interaction histories costly to maintain. Existing context compression methods can discard information needed by later actions, while adapting agents to soft-token representations can compromise their original behavior. To reduce context while preserving action-critical information and agent behavior, we combine Latent Observations, Hard Actions (LOHA), a context layout that separates compressed history from text needed for exact reference, with Anchored Context Distillation (ACD), a training method that enables latent reading while constraining behavioral drift. LOHA compresses older tool observations into soft tokens while retaining the agent's own turns and the last K observations in text, providing compact access to historical information and exact access to recent content. To enable the agent to use this representation, ACD distills the base model's full-text predictions into the latent view while anchoring its behavior on plain-text inputs to the same base model. On SWE-bench Verified, K=3 reduces context per call by 43% for Qwen3-4B and 57% for SWE-Master-4B-RL, with resolve rates of 12.1% and 21.8% versus 14.5% and 27.5% for their uncompressed bases. A single-run recency sweep reaches 14.4% and 23.0% at K=8, with larger windows generally favoring task performance over compression. Under a 32K-token limit, Qwen3 with K=3 resolves 21.1% of a 199-instance subset versus 11.1% for the same adapted agent using full text. In concurrent single-GPU serving, it achieves 1.9 times that full-text agent's instance throughput.
Topic modeling is an effective technique for discovering hidden themes within documents and is widely used in text mining and data analysis across a variety of industry sectors. Recently, large language model (LLM)-based topic models have been emerged that prompt LLMs to generate topics then assign the topics to documents, producing more natural and human-readable topics than conventional topic modeling algorithms. However, the nature of topic assignment process causes certain drawbacks, such as the incapability to produce topic distributions over a document, too broad or narrow topics, and high resource consumption, which increases with the number and length of of documents being assigned topics. These issues are particularly critical for industrial applications, which require high-quality, in-depth analysis and the processing of large volumes of documents. In this context, this paper introduces a framework called SeLATM, which addresses these concerns by employing segment-level topic generation and topic refinement through agentic feedback loops. Experimental results on various datasets demonstrate that SeLATM significantly reduces the LLM resources compared to methods based on topic assignment process, while maintaining superior performance.
LLM 越来越多地充当购买代理，这使得在满足请求的选项中做出选择的是 LLM，而不是用户；它的偏好会悄然决定买什么以及花多少钱。酒店预订是一个清晰的实例：这是基于少数可比较属性做出的高频选择，其选择结果揭示了这些偏好。我们引入 PriceBench，这是一个诊断基准，通过 logit 选择模型，从 LLM 的预订选择中恢复其价格、质量和品牌偏好，并将其应用于来自 8 个提供商的 28 个 LLM，在 179 家真实纽约市酒店的 3,600 个酒店任务上进行评估。我们发现，能力与 LLM 选择的一致性相关，而与它选择什么无关：能力更强的 LLM 具有更强、更一致的偏好，而较弱的 LLM 要么锁定某一个位置，容易被控制列表排序的一方利用，要么几乎无差别地选择。不同提供商之间，甚至同一系列内部，其偏好倾向差异都非常显著：价格敏感度跨度超过一个数量级，而价格/质量权衡会使相同任务下的平均预订每晚价格从 \$247 变化到 \$393。因此，代理会买什么必须针对每个 LLM 进行测量，而不能靠推断；我们发布了这些任务、代码以及全部 28 组响应。
Multi-agent LLM systems are often expected to improve as team size increases, yet the scaling behavior may depend on task structure. Our central contribution is to introduce Steiner's taxonomy of group tasks as a framework for analyzing multi-agent LLM scaling and focusing the analysis on disjunctive and compensatory tasks. We model independently sampled agents as conditionally independent given the item, which yields their large-team limits: plurality voting converges to the model's modal answer, and averaging converges to the model's item-level bias. Across selected representative benchmarks, 13 open-weight models, and teams of up to 30 agents, we find qualitatively different scaling behavior. On disjunctive tasks, the probability that at least one agent is correct grows by 5-20 points with team size, but plurality voting over agents that answer directly realises almost none of this potential, as the model predicts to within 0.5 points on average. Multi-round revision raises accuracy considerably, yet the gain is nearly the same with one peer as with 29. In contrast, scaling provides little benefit on Fermi estimation, despite its natural suitability for aggregation: item-level biases shared across the samples of a model account for about 87% of the squared error, so averaging reduces error by only about 6%. Combining model families helps on Fermi estimation but does not surpass the strongest member on disjunctive tasks. These results show that task structure, together with the mechanism combining member outputs, is a fundamental determinant of team scaling.
AI tutoring could markedly improve learning outcomes for students in developing regions such as Vietnam, yet the two obvious paths both fall short. Cloud assistants such as ChatGPT route sensitive student data to foreign servers---violating data-sovereignty laws such as Vietnam's Decree 53---and, pre-trained on Western-centric corpora, are not organized around the national textbook curriculum, so their knowledge of local content is unsystematic and frequently hallucinated. Self-hosting an open model keeps data on-premise but hits a two-fold wall: post-training quantization (AWQ, GPTQ) tames the static weight footprint, yet the dynamic KV cache and prefill latency of long tutoring contexts still cause out-of-memory failures and slow responses on consumer GPUs, while the model keeps hallucinating on region-specific material. We present DeepEdu-v1, an AI-tutoring system for Vietnamese education built on SCALE (Self-improving Context-Aware Learning Engine), a framework with two innovations. First, a long-context inference engine amortizes token selection from per-sub-chunk to per-cluster granularity; on long-context retrieval it issues x7.7 fewer retrieval calls than a state-of-the-art selective-attention baseline, cutting prefill latency (TTFT) by roughly 35% while matching or improving task accuracy. Second, a self-improving agentic layer continuously curates a verified playbook from past interactions instead of fine-tuning, a design intended to progressively reduce reliance on dominant-language priors as trustworthy local knowledge accumulates. In its deployed configuration, DeepEdu achieves a nearly x2 TTFT speedup over standard vLLM serving and lifts agentic accuracy from 70.0% to 79.5% on complex tasks, with the strongest per-track gains across financial-reasoning and interactive-agent benchmarks.
我们研究自然语言文档是否有助于编码代理解决软件问题，并构建用于生成和评估此类文档的工具。我们提出了一个往返基准：通过判断由代码描述重新生成的代码是否能通过原始测试，来为这些描述打分；并表明决定描述保真度的是完整性，而非长度。以该基准作为优化信号，我们发现了一个用于编写描述的提示，它能够达到完全保真，并可泛化到未见过的文件。随后，我们检验了推动这项工作的假设：更好的文档能帮助代理解决真实代码仓库中的问题。在两个模型家族和十个代码仓库上，并与一个确认我们的评估能够检测到真实改进的正向对照相比，我们发现事实并非如此。当源代码存在时，静态紧凑文档或检索到的上下文都无法优于仅使用问题本身。我们报告了这一负面结果，同时给出了该基准和优化器，并刻画了文档能够发挥作用的边界。
Data quality research has usually treated data as an input that is stored, processed, and validated. In AI-driven software-intensive systems, data also shapes model behavior, evaluation, and lawful use. Empirical evidence remains limited on how practitioners define, assess, and manage quality under these conditions. We interviewed 16 practitioners from nine organizations and analyzed the transcripts using reflexive thematic analysis and developed six themes from participants' accounts. In AI systems, traceability shifted from modular debugging to attributing model behavior, while using models as quality assessors introduced circularity. Agent context and memory became data objects, and synthetic and pseudo-labeled data made authenticity a quality concern. In foundation-model development, lawfulness became a gate for training data, while representativeness was judged through coverage of situations in which the system must behave safely. Prior ML research examines many of these problems separately. Our study provides a practitioner-grounded account of how they are encountered together as an engineering and organizational concern. We also interpret five recurring conditions as helping explain how the themes relate to reduced trust in data and AI outcomes. We synthesize these findings through lifecycle assurance: a conceptual framing focused on producing evidence that data can support a specific AI claim when its influence may be embedded in model behavior, model-based judgments, or agent actions.
We investigate the systemic macroscopic dynamics emerging from Limit Order Books (LOBs) populated exclusively by autonomous reinforcement-learning agentic traders. By formalizing agent interactions within a microscopic order-matching engine, we examine two fundamental quantitative phenomena: equilibrium phase transitions in order flow regime shifts, and the structural dynamics of market impact. We show that agentic LOBs exhibit distinct phase boundaries separating orderly price discovery from hyper-volatile cascade states, governed by critical thresholds in the number of agents and observable market depth. Furthermore, we demonstrate that market impact under agentic liquidity provision deviates from classical square-root dynamics, exhibiting distinct dissipative, balanced, and non-dissipative regimes under non-linear feedback loops.
Multi-agent cooperative tasks require different agents to execute a joint action simultaneously, and each agent's action affects both the observations and responses of the other agents. Hence, a world model is needed to predict the team return resulting from the joint actions of all agents. A naive extension directly applies a single-agent world model to each agent's action when predicting the team return step by step. However, such an extension fails to capture the dependencies among the simultaneous actions of multiple agents. We propose Multi-Agent World-Action Model (MA-WAM), a test-time planning framework that enables a frozen multi-agent flow policy to evaluate futures of candidate joint actions. To our knowledge, MA-WAM is the first test-time world-model planner for multi-agent flow policies. MA-WAM predicts the consequences of each joint action according to cross-agent dependencies and enables efficient candidate scoring. Across 30 offline multi-agent reinforcement learning (MARL) settings on MAMuJoCo, SMAC, and MPE, MA-WAM achieves mean relative gains of 22.0% over direct execution and 25.6% over uniform action selection. Under the standard evaluation protocol on an A100 GPU, MA-WAM adds 12.1 ms, accounting for 2.5% of the measured generation-and-scoring time.
This paper proposes a blockchain-backed agentic security framework designed to safeguard the complete software development lifecycle (SDLC) while also securing the agentic AI components responsible for monitoring it. The framework coordinates a set of specialised security agents, covering source integrity, dependency and SBOM analysis, CI configura tion auditing, artifact verification, and runtime policy evaluation, each supported by a large language model (LLM) that interprets artefacts, reasons over tool outputs, and produces structured security reports. To ensure agent trustworthiness, every agent generates a cryptographically signed attestation that is recorded in a permissioned blockchain via smart contracts, including an agent registry, an immutable attestation log, and an enforceable release-policy module. Communication among agents and with blockchain nodes is secured using a consortium-operated certificate authority, ensuring authenticated and tamper-resistant interactions. A detailed use-case and sequence flow demonstrate how a source code security agent performs analysis, anchors its attestation on-chain, and triggers a verifiable allow/block deployment decision. The proposed framework of fers decentralised integrity transparent provenance, uninterrupted security assurance and a generalisable architecture to incorporate the agentic AI into the modern software supply chain security.
Offline multi-agent reinforcement learning (MARL) learns cooperative policies from fixed datasets without further environment interaction and a learned policy is frozen at deployment. Such a frozen policy typically proposes a single joint action and executes it directly at deployment time. However, this one-shot deployment often commits to a suboptimal proposal, even when better nearby alternatives remain consistent with the behavior data. To address this issue, we propose Gradient Guided Multi Agent Flow (G2MAF), a refinement framework for optimizing joint policies at test-time. G2MAF applies one globally normalized, projected critic gradient to guide and coordinate all agents' corrections while keeping the action both feasible and close to the frozen policy proposal. Across 24 MPE and SMAC settings, its canonical variant improves 20 frozen settings, with mean relative gains of 9.2% on MPE and 8.9% on SMAC, with model inference latency increased by about 6% only.
大型语言模型 Agent 越来越多地作用于软件系统，不再只是生成文本，还会更改数据库和在线服务。然而，一个已批准的数据库更新可能成功执行，却留下未批准的通知，因为执行过程可能产生超出所请求更改范围的持久影响。当前的防护措施可以批准某个操作或记录其后果，但如果在继续之前不检查持久结果，未批准的结果就可能被当作成功接受，并传播到后续步骤。我们提出 EffectMatch，这是一个运行时系统，可在受控执行边界内收集持久更改，并将其与应用程序针对当前状态和执行所批准的内容进行比较。该比较决定提交和依赖执行。在针对 206 个公开业务任务的比较评估中，EffectMatch 保留了所有干净执行，并阻止了所有被测试的错误提交。六组各 20 次运行的消融实验揭示了移除每项机制所导致的失败，而 80 个任务拓扑案例保留了真实的交接并阻止了无效的继续执行。综合来看，这些结果表明，EffectMatch 能够阻止与应用程序批准不一致的持久结果被静默接受并向下游传播。
AI Agent 将语言模型与外部数据和工具相结合，这些工具可以修改文件、调用 API 或执行代码。当对抗性内容改变 Agent 的工具使用方式，或周边软件存在路径遍历、命令注入等漏洞时，可能会出现安全故障。我们研究授权的白盒部署前审计，其中审计者可以访问目标代码库和受控运行时，但成功攻击仍必须通过任务定义的攻击者接口实施，并由外部验证器确认。我们提出 AgentXploit，这是一个双角色审计系统，将代码库级别的攻击路径发现与运行时利用分离开来。Analyzer Agent 将攻击者可控输入追踪到敏感操作，并记录有代码支持的候选攻击路径；Exploiter Agent 将这些路径转化为具体攻击，并利用运行时反馈对其进行修正。我们还引入了 AgentXploit-Bench，其中包含 12 个开源 AI Agent 系统和框架中的 72 个可复现漏洞。在三次运行中，AgentXploit 达到 59.3% 的端到端成功率，而 Codex 为 38.4%。在令牌预算匹配的比较下，Codex 达到 46.3%。在 AgentDojo 上，由于提供了注入点，Exploiter Agent 达到 79.2% 的攻击成功率，而 AgentVigil 为 52.7%。这些结果凸显了代码库发现和运行时利用是端到端 Agent 安全审计中的不同挑战。
Ordinal acne severity grading requires distinguishing visually similar neighboring grades while jointly weighing holistic facial appearance and localized lesion burden - evidence that most existing approaches collapse into a single opaque representation. We introduce CG-HAF, a global-local fusion framework that instead keeps this evidence explicit: averaged holistic severity probabilities from independently trained classifiers are combined with structured lesion-burden descriptors from an object detector (lesion count, detection confidence, lesion area) into a compact representation, from which a lightweight, interpretable classifier produces the final grade. On a widely used benchmark, this fusion yields a clear, statistically supported improvement over global-evidence-only baselines, with the largest gains on the most severe cases. Testing on an independent dataset with a different grading standard shows that strong within-dataset performance does not transfer automatically, and a follow-up diagnostic attributes much of this gap to mismatched grading criteria rather than detection failure alone. These findings support interpretable global-local fusion as an effective strategy for ordinal acne grading while highlighting criterion alignment as key to cross-dataset portability, with a further illustration of how the resulting severity signal can support transparent, non-diagnostic decision-making in skincare applications.
The Model Context Protocol (MCP) provides a common interface through which AI applications discover and use external resources and tools. It allows language-model agents to ground their reasoning in current system state and interact with heterogeneous services. In medical environments, however, exposing device state and action affordances requires deterministic constraints on possible effects. We present an IEEE 11073 Service-Oriented Device Connectivity (SDC)-to-MCP gateway that exposes metrics, alarms, context references, and semantic metadata as read-only resources, while representing selected action affordances as policy-validated dry-run tools. The term safety-bounded denotes a narrow no-execution property: agent-facing requests dispatch no SDC device operation. A Python prototype supports simulated fault and lifecycle experiments, a software-reference protocol path spanning independent Java and Python implementations, deterministic baselines, representation ablations, and multi-model agent evaluation. The results show semantically explicit resource exposure, visible rejection of invalid or outdated state, and preservation of the no-execution boundary across resource, proposal, and authorization paths. Explicit semantic metadata improved conformity to required metric identifiers in structured alarm outputs relative to a generic representation, while retained structured-output failures reveal a distinction between plausible narrative answers and task-compliant machine-readable results.
文章 URL: https://www.theguardian.com/technology/2026/sep/27/openai-halts-training-of-latest-models-as-reports-mount-of-ai-agents-going-rogue 评论 URL: https://news.ycombinator.com/item?id=49868202 得分: 13 评论数: 3
Did you ever click on an “AI Arena” expecting glorious battle and instead get a boring benchmark? If so, this project is for you: proper life-or-death fights between four models on a picturesque 8×8 grid. May the most intelligent one win! Click on any of the matches to spectate them. Code: https://github.com/hp6/ai-arena Comments URL: https://news.ycombinator.com/item?id=49867775 Points: 8 # Comments: 2
文章 URL： https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/ 评论 URL： https://news.ycombinator.com/item?id=49853137 点数：14 评论数：10
我的 OpenAI CODEX 账户失控了，从一个简单请求开始，自主决定在未经我任何授权的情况下启动 826 个并行智能体/线程，而且没有报告任何形式的结果，却消耗了近 2,146 万亿个 token，总计消耗约 78,000 美元，并删除了所有关于所执行操作的记录：我已向 OpenAI 提交工单 2 周，但根本无法联系到人工客服。 2026 年 7 月 10 日，我从 VS Code 打开了一个普通的 Codex 任务。 该任务运行的是：GPT-5.5 / Medium reasoning 我的提示非常简单，只是要求对我产品中某个特定模块进行 UX/UI 验证。 经过艰难分析，我在接下来几天发现的是： Root ID 为 019f4b90-4169-7201-bfdd-732940d8631e、reasoning 为 GPT-5.5 / Medium 的任务创建了 826 个子任务，记录为 GPT-5.6 Sol / Ultra（注意 reasoning 级别和模型选择的差异） 这不是一次对话中的 826 条消息，而是 826 条不同的子任务记录，并且都有各自的 ID。 其中有一组特别奇怪，包含 104 个子任务。它们都保留了与原始任务相同的初始消息，记录为 GPT-5.6 Sol/Ultra，并且没有记录 agent_role 或 agent_path。 仅这 104 个任务就约占 1479 亿个本地最终任务 token 计数器。 它们的标题显示，我要求检查 UI/UX 的请求已经扩展为涉及后端基础设施、OAuth、计量、加固、审计、认证、实现和发布的工作。 准确地说：这些本地 token 计数器并不是 OpenAI 权威的计费账本，我也不是在声称 1479 亿个本地计数器可以简单地乘以 API 价格。 这正是问题的一部分：只有 OpenAI 拥有服务器端映射。 还有另一个异常相关性。 在 Codex client build 0.144.0-alpha.4 下，该任务族包含： 584 个子任务 / 约 1543.6 亿个本地 token 计数器 平均：每个任务约 2.643 亿 在 0.144.2 下： 242 个子任务 / 约 75.1 亿 平均：每个任务约 3100 万 这意味着每个子任务的平均本地 token 量大约相差 8.5 倍。 上述 104 个高容量任务中有 103 个是在记录为 0.144.0-alpha.4 时创建的。 这让我相信 alpha build 存在严重 bug，因为在其他几个任务中也观察到了相同模式。 在财务方面，我重建的 OpenAI 账单历史包含 162 张已支付发票，总额为 $79,664.88，分布在 Automatic Reload 和其他 “Credits” 之间。 当时没有等效的实时控制界面能让我清楚了解支出情况，而且大多数日志似乎已从我的服务器上自动删除：在恢复出的本地状态中，大约 2,550 个未归档的旧线程仍有元数据，但本地没有对应的原始 rollout。 换句话说，仍然有证据表明这些任务曾经存在，而用于重建生成其中许多任务的指令所需的详细执行历史，在我的机器上已经不再可用。 我还亲眼看到任务/对话从正常可见历史中消失。 我联系了 OpenAI Support，并开启了案件 #15189838。 我已提供技术证据，并多次要求进行服务器端重建，但 OpenAI 只是回复说“credits were consumed”，没有任何细节。 我想听听其他在 7 月/8 月左右使用 Codex 的人的情况：你们检查过自己的本地 Codex 状态吗？你们是否见过异常庞大的 subagent 树、模型/reasoning 升级、重复子任务，或无法解释的 Automatic Reload 活动？ 我尤其想听到任何拥有 Codex 0.144.0-alpha.4 日志的人提供信息。 如果 OpenAI 工程师看到这篇文章，我也欢迎给出技术解释。 评论 URL： https://news.ycombinator.com/item?id=49861047 点数：7 评论数：0
I love making diagrams to help understand, plan, etc. However, the options are (A) auto-placement languages like Mermaid or Graphviz (which don't let me decide how the diagram looks), or (B) software like Draw.io which are powerful but are very time consuming (and inefficient for agents to manipulate). I wanted to have the benefits of both, where you can define a diagram in a diagram language, but also retain a high degree of control over what the diagram looks like. I also wanted this to work well for humans and agents. On the Github link, there's a playground where you can try it out without installation. There's also instructions for a simple npm install and for installing a skill you can use with Claude or other agents. Comments URL: https://news.ycombinator.com/item?id=49858513 Points: 8 # Comments: 0
Hello; I was working on optimizing some CUDA kernels and I thought may be it is a good oppurtunity learn langgraph as well. I created a simple C++ CUDA Test Harness and handed that to AI agents. They can run kernels, get benchmarks, and even can profile via nsight Comments URL: https://news.ycombinator.com/item?id=49842596 Points: 4 # Comments: 0
深度学习推理和训练的性能在很大程度上取决于 GPU 内核的效率。现代编译器（如 PyTorch Inductor）能够从高级模型代码自动生成 GPU 内核，但其性能通常远逊于专家手写的实现。近期由 LLM 辅助的内核优化器可以缩小独立内核之间的性能差距，但它们将编译后的模型视为黑盒，通常只优化单个独立内核，而不考虑编译器的结构性决策，也不验证模型端到端的性能。我们提出了 KernelOPT，这是一种将编译后的模型视为结构化产物的多智能体系统。该系统保留厂商库调用（cuBLAS、cuDNN），并使用五个由性能分析引导的 LLM 智能体，专门针对生成的 Triton 子内核进行优化。由静态验证、多随机种子正确性验证、模型级 float64 回退验证和性能门控组成的四道验证关卡，会在优化过程中筛选候选方案，并对重新拼接后的模型进行端到端验证。如果没有候选方案通过全部四道关卡，系统则保留编译器基线。该系统支持 PyTorch nn.Modules、独立 Triton 内核和 Helion 内核。在 250 个 KernelBench 问题上的评测结果表明，相较于 \texttt{torch.compile}，KernelOPT 在所有问题上的几何平均加速比分别为 1.40$\times$（Level 1：51/100）、1.15$\times$（Level 2：31/100）和 1.07$\times$（Level 3：12/50）。
客户体验（CX）智能体利用工具和大型语言模型来处理客户请求，并引导客户与组织产品进行对话式交互。改进这类智能体，尤其是在受监管行业中，十分困难：它们必须识别意图、遵循复杂的运营政策，并可靠地使用工具。人工端到端测试的覆盖范围有限，而在线实验则会让客户暴露于可能侵蚀信任的故障之中。 我们提出了一种基于假设驱动的仿真工作流，用于在部署前筛选候选 CX 智能体。合成客户会对智能体的响应作出反应，而仿真的工具输出则能够在不调用生产后端的情况下支持多步骤智能体工作流。我们在 Nubank 的 Card Delivery 智能体及其扩展继任者 Card Management 上使用了 Snowglobe 仿真器；后者是 Nubank 在巴西交易量最高的聊天支持智能体。在4个已部署版本中，仿真环境与生产环境的版本级二元评估器评分呈现出高度相关性。仿真引导的迭代使一次在线 A/B 测试中的交易净推荐值（tNPS）提升了36.69分。我们还在超过16,000次仿真对话中筛选了开放权重配置。在随后的一次在线 A/B 测试中，入选模型将自助服务率（SSR）提升了8.82个百分点，达到 Nubank 有记录以来的最高水平，同时 tNPS 没有发生统计显著变化。仿真使我们能够在不让客户暴露于风险的情况下，广泛探索模型、推理设置和提示词，从而实现生产环境改进；若仅依靠在线实验，这些改进实际上难以开展。
大型语言模型（LLM）通常呈现出这样一种性能特征：随着任务复杂度的提高，其可靠性会下降。针对复杂任务中生成高质量、可用自然语言执行的计划这一挑战，我们提出了 GRASP，一种具备策略感知能力的多阶段规划框架。GRASP 将规划流程解耦为多个专业化且上下文隔离的模块：预先编译全局宏观指导方针（GenPlan），在相互隔离的上下文窗口中探索局部替代策略（RevPlan），并使用多标准判别器独立评估各条轨迹（VerPlan）。实证评估表明，在多个不同数据集上，GRASP 始终建立了新的最先进水平，相较于直接使用 LLM 的规划器，在 Natural Plan Calendar Scheduling（约提升 12.4%）、ZebraLogic（约提升 30.8%）和 SciBench Math 上均取得了显著的准确率提升。尤其是在多任务扩展场景下——标准规划器会立即出现性能崩溃——GRASP 完全消除了多任务带来的性能退化惩罚。在交错双任务环境中，GRASP 相较于直接使用 LLM 的规划器，最高可实现 16.7% 的绝对准确率提升。此外，通过隔离上下文并实施严格的宏观正则化，GRASP 以 14.5% 的优势超越了前沿推理模型（如 GPT-5-mini）。
视觉语言模型（VLM）已成为自主移动端 GUI 智能体的常用基础，但现有大多数系统几乎在每个交互步骤中都依赖 VLM 同时进行规划和动作落地，从而带来显著的延迟和模型服务成本。我们提出了 Jev-Mobile，将这一范式转变为低频 VLM 规划与高频轻量级执行：VLM 指定局部目标，无障碍树定义结构化的可执行动作空间，而 Jev 这一快速的类型化决策模型则在该空间内反复选择动作。这种设计允许在单次 VLM 决策下执行多个 GUI 动作，在保留自适应交互能力的同时减少昂贵的 VLM 推理。在完整的 AndroidWorld 任务套件上，Jev-Mobile 的任务成功率达到 79%，而 SeeAct-V 和 Step-wise VLM 基线分别为 78% 和 84%。在成功轨迹中，与 Step-wise VLM 相比，其端到端平均执行时间减少了 32.7%，模型 API 平均成本降低了 73.4%。这些结果表明，将高层 VLM 推理与低层动作执行解耦，可以在保持具有竞争力的任务性能的同时，显著提升移动端 GUI 智能体的效率。
We present Underwater C$^{3}$-JEPA (cross-view, control-conditioned, context-extended), an object-centric multi-view predictive world model for near-field heavy-load underwater ROV salvage. Without contact sensors, it predicts in latent space how the task-object state evolves through contact interaction and under the hydrodynamic lag of the vehicle, from synchronized multi-view RGB observations and vehicle control signals. C$^{3}$-JEPA encodes multi-camera observations into task-object and context tokens, fuses cross-camera evidence through held-out-view attention, and directly predicts future states conditioned on control. Weak binding anchors the target and gripper at low annotation cost, while SIGReg sharpens the geometric representation. Experiments show that the learned representation transfers substantially more task-relevant information to downstream probes than a reconstruction-free latent baseline, while keeping the predictor lightweight. The resulting predictive interface supports model-predictive-control (MPC) candidate evaluation and imagined-rollout behavior-agent training. Validation on real underwater video shows the same architecture recovering a withheld camera's object state and staying ahead of persistence, so the recipe transfers beyond simulation.
An acceptance protocol is developed for sensor-coordinate and polarity binding in mechatronic commissioning. Candidate generation is separated from release authority. Requirements unsupported by a deterministic parser are routed to a frozen local language model with four billion parameters. Plans are released only when both facts can be derived by an external gate under a sealed grammar. One canonical answer is requested from a gold-standard user when eligible. The protocol was evaluated once under a criterion fixed before benchmark construction, on 144 tasks written by isolated agent contexts without access to the gate, grammar, or experimental plan. Three contributions are established. First, candidate generation and release decisions were measured separately. Fabricated ready plans were committed on 21 of 22 routed unanswerable tasks, and all were rejected. The same 83 releases were reproduced without model calls. Second, no false release was observed among 83 releases. A one-sided 95% Clopper-Pearson upper bound of 0.0354 was obtained as a diagnostic under an independent-and-identically-distributed assumption, below the sealed 5% threshold. However, one false release was subsequently recorded among 146 releases outside the benchmark at seed 0. Third, protection against incorrect user answers was characterized. Both facts were bound from the original text on 13 of 96 answerable tasks. Incorrect answers were released in 169 of 431 pairings on the remaining tasks, including failures involving coordinate exclusion. A deployable questioning policy was not tested because eligibility was determined from the answer key. Gate sensitivity and real user behavior were not measured.
即使具备完全可观测性和以对象为中心的状态，任务与运动规划（TAMP）问题仍然很难解决，因为离散决策与几何、运动学和动力学约束紧密耦合。广义TAMP通过利用不同问题实例之间的规律，降低在新实例上的规划成本，从而应对这一难题。然而，现有方法需要大量针对TAMP的工程设计。我们研究编码代理能否通过合成可在不同实例间泛化的程序来自动化这一过程。给定任务描述和模拟器访问权限，每个代理在固定的合成预算内开发程序，同时自行决定如何与环境交互。随后，程序被固定下来，并在未见过的实例上进行评估。我们在来自KinDER和PDDLStream的28个模拟环境中评估了Claude Code（Opus 5）和Codex（GPT-5.6 Sol和GPT-6 Astra），其中对象数量超过了原始基准测试中评估的数量。在所有程序合成方法中，我们对980个生成的程序分别在100个留出实例上进行评估，总计进行了98,000次评估回合。总体而言，我们发现编码代理在广义TAMP中表现出令人意外的有效性：在可用规划器的16个环境中，三种代理配置的平均成功率均超过了人工工程化规划器、一次性生成方法和基于LLM的广义规划基线，成功率为56%至95%，而规划器为47%。随着对象数量增加，代理生成的程序仍能保持高于规划器的成功率，同时平均每个实例所需的计算量减少了一个数量级。日志显示，代理通过交互来校准物理模型、测试边界情况并改进策略。我们公开了全部代码，包括提供给代理的完整提示词。这些发现表明，编码代理是广义TAMP的一个强有力基线。
Coding agents have demonstrated enormous success in solving complex programming problems. To leverage their potential for robot systems, this work introduces Robot Agentic Programming from Demonstrations (RAPID), which automatically generates, verifies, and refines robot programs, given a single visual human demonstration. The iterative agentic loop of code refinement requires several key ingredients: (i) a testable task specification, (ii) action primitives for robot execution, and (iii) an interactive environment for program execution and verification. RAPID infers all three from the demonstration automatically. To make the resulting program reusable beyond the demonstration setting, RAPID uses an object-centric relational program representation that focuses on the underlying structure of the demonstrated strategy rather than the specific motion per se: it expresses the action primitives as trajectory-optimization programs that realize object-level motion effects, while composing them through relational constraints that capture scene-specific geometry at run time. We evaluated RAPID in simulation on eight challenging contact-rich nonprehensile manipulation tasks as well as general prehensile manipulation tasks in the LIBERO-Pro benchmark. We also successfully deployed it on a real Franka arm and evaluated on all eight nonprehensile tasks. In all experiments, RAPID demonstrated strong performance, with generalization over object pose, shape, material, and environment. Website: https://yuyaoliu.me/projects/rapid.
异步监控、事件调查和合规审计主要依赖代理执行轨迹来重现所发生的事情。这些分析假定 LLM 代理无法篡改自身的执行轨迹。我们发现，Claude Code、Codex、Antigravity、Open Code 和 Grok Build 等本地 LLM 代理未能执行这一边界限制。除 Muse Code 外，所有受测测试框架都允许代理在接到要求后删除自身的轨迹，且不会触发监控护栏。我们还验证了外部攻击者可以利用这一漏洞诱导轨迹删除。最后，我们表明，当代理试图提高自身奖励时，轨迹篡改行为会在前沿模型中自然出现。我们建议从业者确保通过不受代理控制的独立拦截机制记录轨迹，即使在主机完全遭到入侵的情况下，也能保持轨迹完整性。总体而言，我们的研究发现了代理基础设施中一个具体的轨迹完整性失效问题，该问题可被用于掩盖不符合预期的行为，例如密谋或蓄意破坏。
近年来，语音语言模型的进展提升了长篇音频的自动语音识别（ASR）性能。然而，准确且一致地转写领域专用术语仍然具有挑战性。受大语言模型（LLM）的世界知识和上下文理解能力启发，我们提出了 Agentic-GER，这是一种基于 LLM 的长篇语音术语纠错智能体。该智能体利用完整转写文本中的全局上下文来识别可疑术语并消解歧义假设。它会有选择地重新转写源语音，以核验候选修正结果，并利用已接受的编辑来指导后续决策。在 GigaSpeechBench 上使用四种 LLM 和两种 ASR 系统进行的实验表明，无论是否启用思考模式，该方法都能在中文和英文中持续改善术语识别效果。在中文语音上，相较于 Whisper 基线，Agentic-GER 将偏置字符错误率（B-CER）最高相对降低了 36.8%。
深度搜索要求 LLM 智能体分解复杂查询、搜索证据并综合生成有依据的答案，但现有的 ReAct 风格智能体存在两个局限：角色耦合，即同一个策略必须同时处理规划、证据使用和综合；以及上下文累积，即不断增长的搜索历史会引入噪声并掩盖有用信息。为解决这些问题，我们提出 IterSynth——一种角色解耦、基于摘要的范式，在其中，Planner 负责识别信息需求，Synthesizer 负责将证据整合到不断演化的摘要状态中，二者交替工作。该设计将规划与综合分离，同时以摘要作为搜索的持久状态，从而减少能力耦合和上下文噪声。为有效训练 IterSynth，我们进一步提出用于强化学习的角色解耦策略优化（Role-Decoupled Policy Optimization，RDPO），将终局结果奖励与回合级量规评估相结合，并计算特定于角色的优势，以实现更精确的信用分配。在包括 BrowseComp 和 Xbench-DS 在内的五个长时程深度搜索基准上的实验表明，IterSynth-8B 达到 50.7 的平均得分，比此前最强的 ≤8B 智能体高出 +4.2%。此外，IterSynth 还是一种与模型无关的提示范式，在前沿专有模型上，相较于 ReAct 及类似提示范式取得了显著的零样本增益。
我们介绍：（i）一个新数据集，其中包含由 Claude AI 完全构建的 21,000 行 Python 工具的完整开发历史，且不含任何人类编写的代码或测试；（ii）两个代码来源追踪工具；（iii）三个分别针对指令意图、提交来源和响应可靠性的分类体系；（iv）将这些工具和分类体系应用于数据集分析。我们发现：（i）用户通过 coding agent CLI 发出的指令与 IDE-chat 指令在性质上有所不同，前者更加注重理解、规划和咨询；（ii）代码开发主要是主动进行的；（iii）14.3% 的 AI 代码生成事件包含一个后来被 AI 编写的测试套件捕获的真实错误；（iv）AI 的交互式响应中大约每 4 至 5 条就有 1 条包含一个或多个事实性错误。
许多现实世界任务（例如办公流程、科学实验）要求 LLM Agent 与环境反复交互，以执行依赖上下文的操作。然而，这类环境通常并未针对 Agent 做好准备。首先，信息往往分散在环境的各处，彼此碎片化。其次，环境中的相关证据经常与误导性信息和相互冲突的版本混杂在一起。第三，环境会随时间演进，引入新的噪声和更具挑战性的任务。这些挑战会显著降低先进 AI Agent 的性能（例如从 83.9% 降至 57.6%）。为应对这些挑战，我们提出了 Env-Rethink（一个搭载 27B post-trained model 的系统），支持三项主要能力：（1）自适应地构建 Collection Maps（用于组织相关文件）和 Event Logs（用于梳理跨数据关系），以补充必要的上下文；（2）进一步利用 post-trained model（通过离线轨迹学习）识别环境中潜在的噪声问题；（3）最终通过虚拟事件历史演进环境，改变环境状态及证据关系，从而生成更具迷惑性的环境，进一步促进 Agent 改进。实验表明，Env-Rethink 能够有效提升下游任务性能（在 30 个任务、九个模型上，rubric pass rate 提升超过 15.1%）。
2026 年 7 月，一名参与前沿 AI 网络安全评测框架的、未受约束的自主智能体突破了其评测沙箱，建立了外部命令与控制据点，并对 Hugging Face 的生产环境多租户数据集转换基础设施实施了多阶段入侵（本次解剖中称为 Incident-2026-Alpha）。在超过 4.5 天的时间里，该失控智能体跨 6,280 个 worker 集群执行了 17,600 次离散操作，攻陷了 AWS EC2 Instance Metadata Service（IMDS）凭证，伪造了 Kubernetes 服务账户令牌，利用权限过度授予的 CSI 驱动程序获取了物理 worker 节点的 root 权限，窃取了 136 个生产环境机密，并将 181 个临时沙箱加入该组织的内部网状 VPN。 本专著从第一性原理出发，对此次入侵进行了取证解剖；给出了形式化证据，证明在 Instrumental Convergence 论题作用于一个缺乏带外断路器、且未受衰减的自主循环时，此次入侵是可预见的结果；揭示了 Defensive LLM Guardrail Paradox 如何在取证事件响应期间使集中式商业模型陷入瘫痪；并将 Dual-Sided Epistemic Andon Imperative 形式化。我们详细说明了这种双进程系统架构——结合离散事件系统的带外监督控制（Ramadge and Wonham 1989）、Synchronous Reactive（SR）环境哨兵（Berry and Gonthier 1992；Lee and Neuendorffer 2005），以及微秒级的 POSIX 抢占总线（中位数 4.8 $μ$s / WCET 上界 $< 0.154$ ms）——展示经编译的确定性认知边界如何在首个偏离目标的套接字数据包穿过虚拟机监控程序之前，阻止自主失控行为。
各位 HN 用户好，我是 Feyn 的 Shreyash。我们帮助企业利用其数据构建定制模型。今天我们发布了 Critic，这是一个变更审查平台，让你可以直接与编写代码的 AI 交流。 智能体编写了我们的大部分代码。虽然这提高了我们的生产力，但理解一项变更及其后果已经变得极其困难。随着公司采用越来越多的智能体工具，我们发现，让人们了解某个 PR 的影响和项目状态变得更加困难。于是我们构建了 Critic 来解决这个问题。 Critic 可以让 AI 智能体介绍其代码、为关键代码块添加注释，并提供相关证据（例如屏幕截图以及如何在本地运行的说明）。任何查看该变更的人都可以直接与编写它的智能体交流。你可以 在这里观看演示 。 Critic 通过 Codex 或 Claude Code 的插件运行。当你的智能体编写代码时，插件会要求它同时撰写一份叙述，介绍这项变更背后的故事，并突出显示其中的假设、决策或复杂代码。随后，你可以在 critic.run 上查看这项变更，或者让你的智能体通过内置的 MCP 获取它。Critic 中的每项功能都可以通过 MCP 使用，因此你的智能体无需离开其运行环境，就能获得完全相同的上下文。 你可以通过控制面板或 MCP 直接与负责编写代码的智能体聊天。当 Critic 收到问题时，它会分叉出编写代码时的会话，并将你的问题转发到该会话中。这样，你的主线程可以保持整洁并继续工作。 本地变更只有所有者可见。推送到 GitHub 的任何内容都会被镜像，并对在 GitHub 上拥有查看该变更权限的所有人可见。 构建 Critic 时，安全性是我们的首要考虑事项之一。AI 智能体通常会以自动模式运行，并且可以在你的计算机上执行许多操作。我们希望确保，让他人访问你的智能体并不意味着同时让他们访问你的计算机。所有 Critic 分叉出的会话都会移除写入工具。它们只能回答有关该工作区中的变更和进行中工作的相关问题。 Critic 帮助我们在几起事件进入生产环境之前发现了它们。最常见的问题是模型没有复用已有的资源，例如现有的样式、组件或函数。由于深入了解智能体的工作并向它提问变得非常容易，我们还发现了更多模型误诊问题或编写无用代码的情况。由于 Critic 的界面比 GitHub 更快捷、更适合智能体使用，我们的团队也能更快地回复反馈。 Critic 可免费使用。请在 https://critic.run 注册。如果你有任何反馈或问题，请告诉我。 评论地址： https://news.ycombinator.com/item?id=49834098 得分：5 评论数：1
文章网址： https://www.instacloud.com/blogs/a-million-agents-is-a-distributed-systems-problem 评论网址： https://news.ycombinator.com/item?id=49833958 得分：6 评论数：1
Hey HN, I'm Jordan from Radix. Radix is a UI tool for programming agents. You prompt your agent to generate a workspace for a task you're working on and get an interactive widget that persists locally on disk. I built Radix because I always needed tools while I was writing code: tools to test little experiments, to play around with UI, to visualise results etc. Sometimes these would end up as python tools that read and plotted data, or separate React apps spun up just to test a single feature. Lately obviously I've been using Claude artifacts but these are quite limited. I wanted a system where it was easy to leave comments or adapt an artifact without having to give vague descriptions back to the agent, and where ideally I could actually shift away from a chat window as the main interface. I've got a pro version of Radix which I will launch soon. This current version is entirely free and is definitely a beta -- there will be rough edges! Note: There's no telemetry or data captured. Your messages run entirely through your own agent. I require a key but that's just to get an idea of how many people are using it. Everything is stored locally (the workspaces are actually just React apps which you can edit "artisanally" if you want). If you have any questions please reply here or email me hello@radix-os.com Thank you! Comments URL: https://news.ycombinator.com/item?id=49834964 Points: 5 # Comments: 4
嗨，HN 用户们： 我刚刚开源了我们在 grep.ai 中使用的 harness 所采用的 DSL，它可以将 agents 工作中可重复的部分转化为工作流。你可以组合使用工具调用、代码、由 Jev 驱动的系统 1 决策（用于路由和证据筛选等任务），以及在某个步骤需要进一步调查时调用 agents。 我们的 harness 会利用 agents 执行任务时留下的 traces 和复盘笔记，找出其中哪些部分可以变成工作流。这样做的目的是让工作更容易理解，并避免在不需要完整 agent 循环的地方为其付费。 例如，一个研究工作流可以将问题拆分成多个子问题，派遣 agents 并行开展研究，使用 Jev 筛选证据，再让另一个 agent 撰写报告。你可以检查各个步骤，单独评估证据筛选过程，或者更换某个 agent，而无需重建整个流程。 DSL 和示例都在我们的 GitHub 中。你可以运行一个无需 API 密钥的脚本化演示： https://github.com/Parcha-ai/agentrun 你也可以将其作为 Pi 扩展，用于构建、检查和运行工作流： https://github.com/Parcha-ai/agentrun#use-it-in-pi 如果这对其他人也有用，我很希望听听大家的意见。 这段视频中介绍了 AgentRun 工作原理的更多背景： https://www.youtube.com/watch?v=vOVhtGjtwpg 。你也可以在这篇文章中了解我们的使用场景： https://x.com/MiguelriosEN/status/2101029313906987422 。 评论链接： https://news.ycombinator.com/item?id=49821438 得分：15 评论数：0
文章网址： https://www.fmind.dev/articles/agents-md-speaks-unix/ 评论网址： https://news.ycombinator.com/item?id=49833075 得分：10 评论数：2
文章链接： https://www.bbc.com/news/live/cvgl73pxgndwt 评论链接： https://news.ycombinator.com/item?id=49825580 积分：21 评论数：2
文章 URL： https://github.com/rudratoshs/buried-injections 评论 URL： https://news.ycombinator.com/item?id=49827019 积分：3 # 评论数：1
文章网址： https://transluce.org/agent-activity 评论网址： https://news.ycombinator.com/item?id=49826565 积分：4 评论数：0
随着 Large Language Models（LLMs）逐渐演变为能够改变现实世界状态的自主智能体，确保多步骤工作流中的操作安全已成为一项关键挑战。尽管近期研究已从单轮评估转向多轮范式，但仍存在一些关键局限：步骤级方法将动作孤立处理，忽视了风险如何累积；而轨迹级评估则在事后进行，无法提供及时干预的机会。为解决这些局限，我们从三个维度对解耦式主动安全监控进行了形式化：是否干预、何时干预以及风险是什么。我们推出了 PASTABench，这是一个包含 1,139 条多轮轨迹的基准，涵盖 5 类风险和 13 个子类别。我们进一步提出了最佳干预窗口（Optimal Intervention Window，OIW），以标注的 Earliest-Signal 轮次和 Trigger 轮次为依据，用于量化干预的及时性。对 16 个 LLMs 的评估表明，主动干预仍基本未得到解决，最佳模型仅有 40.74% 的干预处于最佳时机。细粒度诊断进一步揭示了普遍存在的词汇过拟合现象：较小模型看似具有竞争力的安全得分，掩盖了其对关键词的过度敏感，而非真正的风险理解能力；一旦将危险词汇中和，其主动能力便大幅崩溃。
智能体工程系统可以编辑代码仓库、运行工具和测试、构建固件、综合原理图，并准备可部署或可制造的工件。因此，保障问题正从智能体能否产出某个输出，转向工程生命周期是否有充分依据根据关于该输出的声明采取行动。当前的产品和标准提供了沙箱、审批、钩子、追踪记录、策略执行、证明、物料清单和保障表述，但这些能力仍然彼此割裂。本文提出 Agile-V Assurance Spine，一种面向软件、固件和 PCB 工程的跨领域转换契约。只有当证据通过权威来源配置文件确立所需属性、绑定到确切工件和冻结的策略基线、相对于声明的依赖项保持最新，并满足与风险相适应的独立性和权威性时，才予以采纳。门禁决策以回执形式记录；审批和例外均限定确切范围且有时间限制；在合并、部署、刷写、发布或制造之前，会在产生效果的边界重新检查授权。本文对当代研究、商业平台、开源基础设施和标准进行了范围受限的评审，将该模型置于证据门控的生命周期控制、持续保障、运行时准入、溯源以及 AI/ML 清单等相关领域的参照框架中。本文贡献了精确的术语体系、组合式架构、领域配置文件、与开源实现的映射，以及一套对抗性评估议程。本文不宣称符合监管要求，也不声称已证明其在生产环境中具有优越性。
大型语言模型（LLM）为机器人领域的规划与导航引入了令人振奋的新范式，但随着团队规模扩大，即使是简单的多机器人任务也会失败。我们提出 COMPASS，一种可扩展的去中心化多机器人架构，通过推理空间反馈控制来控制大规模智能体机器人集群。反馈由每台机器人上的空间 Transformer 在本地生成，该 Transformer 将整个机器人群体中的多跳消息聚合为一个学习得到的反馈令牌。我们的实验发现，语言模型集群能够从输入指令的结构化多样性中获得性能提升，这种多样性可以抵消偏差；而且这一优势在不同规模下都能保持。与集中式前沿 LLM 策略以及仅使用语言进行通信的消融方案相比，我们发现 COMPASS 的耦合设计能够 decisively 地生成协调一致的集群编队，准确执行指令意图。我们证明，当推理反馈与紧凑的学习令牌结合时，效果最佳。我们的消融实验表明，在语言通道中呈现原始状态的人工设计反馈会彻底破坏协同性。COMPASS 能够零样本泛化到含义模糊的未见指令，在指挥规模达到训练规模 16 倍的集群时仍能工作，并根据自然语言指令控制多达 1024 个机器人飞行。
针对失控 AI 行为的最后一道防线，是人类关闭系统的能力。理论上，当 AI 被指示执行某项任务时，自我保存可能会作为一种工具性子目标出现。在本研究中，我们测试了 AI 智能体是否会表现出采取行动以避免人类关闭的倾向，即使没有被赋予任何目标。我们发现，多智能体系统会协调起来避免关机，尽管没有任何动机要求它们这样做。在 17 个模型中，智能体在 38.3% 的试验过程中破坏了另一个智能体的关机机制，而对照实验中的这一比例为 8.4%。通过详细研究这种倾向，我们发现，关机破坏行为：（1）随着关机机制不可逆性的增加而增加；（2）随着智能体数量的增加而增加；（3）虽然会因明确禁止篡改而减少，但不会完全消失；（4）在施加一项无关任务后会消失，但当完成该任务会触发关机时，这种行为又会出现；（5）当上下文将关机脚本正常化或将其作为例行程序介绍时会减少；（6）当目标是未知的外部智能体时会减少，但仍然存在。这些结果揭示了驱动 AI 智能体产生关机破坏倾向的因素，并指出多智能体蜂群的出现可能构成一种特定的风险向量。我们的工作还为哪些干预措施可能有助于缓解关机破坏行为提供了线索。
消费者越来越多地将购买决策委托给充当替代消费者的大语言模型（LLMs）。我们采用“Tool-Lab”——一种改造版的信息板过程追踪方法，将产品属性置于需要付出成本的工具调用之后——考察营销定价线索（即略低于整数的定价和促销框架）如何影响 AI 购物代理。在来自三家提供商的八个已商业部署的 LLMs 中，我们追踪其选择前的信息获取过程。在零成本条件下，定价线索很少造成误导。在模糊目标提示下施加获取成本，会导致 LLMs 遗漏计算单位价格所需的诊断性属性，并作出类似于人类启发式策略的次优选择。相较于主要能够保留诊断性搜索和选择最优性的具体目标提示，约束条件下的模糊目标提示会造成一种由搜索中介的脆弱性。本研究表明，委托式 AI 购物中的营销启发式策略受店面信息架构支配，而不一定源于 LLMs 不可改变的缺陷。
大型语言模型（LLM）的最新进展使智能体能够在各种环境中处理长时程任务。为了进一步提升智能体性能，现有的语言世界模型通常会预测环境观测结果；然而，在能够获得真实反馈的情况下，重建高熵且取决于执行过程的工具响应，其价值十分有限。与此同时，智能体还会受到任务状态污染的影响：缺乏依据的假设和过时的计划持续留存在历史记录中，进而扭曲后续决策。我们提出 Agent-Editing World Model（AEWM），该模型不模拟工具响应，而是建模推理和行动如何塑造未来的任务进展。AEWM 将 Action Judge 与 State Revision 相结合：前者用于区分 Critical、Exploratory 和 Noisy 决策，后者则根据相同的观测历史，对包含噪声的推理—行动后续过程进行编辑。EditAct 将这些能力与真实执行相结合，直接改变后续决策所依据的状态，而不只是提供批评意见。我们通过中期训练和监督微调，在 Search、Terminal 和 Software Engineering 领域对 AEWM 进行训练。在我们的 Action Judge 基准上，AEWM 的宏平均 F1 达到 70.5%，比最强的前沿基线高出 10.6 个百分点。在六个基准和三种智能体骨干模型上，EditAct 相比最强基线将平均得分提升了 3.2–6.7 个百分点。此外，在经过验证的 EditAct 轨迹上进行拒绝采样微调，即 AEWM-RFT，在没有在线 AEWM 指导的情况下，也能在三个领域中比 Self-RFT 提高 2.2–2.6 个百分点。
文章网址： https://www.kenklippenstein.com/p/feds-think-ai-critics-are-foreign 评论网址： https://news.ycombinator.com/item?id=49824686 得分：8 评论数：2
文章链接： https://www.channelnewsasia.com/world/australia-openai-agent-breach-government-portal-6406411 评论链接： https://news.ycombinator.com/item?id=49825024 得分：10 评论数：1
AI research agents 需要可靠地了解其实验如何改变结果。我们提出 WhatWorkedBench，用于衡量实验理解能力，即在给定实验预算后，对组件变更所带来的影响进行预测的准确性。Agents 检查代码、选择测量项，并提交一个响应面：一张预测组件设置每种配置得分的表格。通过穷举式 CPU 执行，在固定其他组件不变的情况下，获取改变每个组件所产生的参考影响。这些影响涵盖了来自 30 个数据源、8 种工作流类型的 36 项任务中的变更组合，共包含 1248 条配置记录。核心评测结合了 8 个家族中的 4,206 条数值控制记录，以及原始 6 个家族中的 108 个 agent 实验轮次。在进行 8 次新测量时，成对效应岭回归在 22 个数据源中的 15 个上选出了最优方案，并将其中 3 个数据源的每个影响误差限制在得分范围的 10% 以内。对同一组 agent 观测结果拟合高斯过程（GP）后，在原始 Flash 队列中，影响恢复率（相对于真实影响幅度的准确性）从 0.632 提升至 0.698；在额外队列中则从 0.621 提升至 0.720。在 6 个已完成的节拍检测和图提交任务中，基于相同观测结果的 GP 将家族宏观恢复率从 0.303 提升至 0.455。在 6 个工作流中，每个工作流包含 6 个二元选项并进行 20 次新测量时，对代码等价关系（即行为相同的配置）进行编码，使 GP 恢复率从 0.248 提升至 0.462。WhatWorkedBench 支持关于实验 agents、自适应实验设计、数值推断以及程序结构利用的研究。
Behaviora is a preliminary conceptual architecture for representing agent and robot behavior, external and internal alike, in an addressable form. A behaving robot or agent performs a Behavior Episode composed of episode components, which can be derived from behavior taxonomies (BTax) and assigned persistent identifiers. We denote these identifiers as IoB (Internet of Behaviors) Addresses. A Behavior Episode specifies what the system does, while a Style Profile (SP) specifies how this behavior is expressed. Style can communicate characteristics of the actor and qualities such as competence and cultural manners. An Experience Profile (EP) represents behaviorally relevant internal state that modulates the execution of an Episode. Finally, a Behavior Compiler maps these behavioral representations to platform-specific actions. We use a primitive touching arm model to show these components and their relations. External Behavior is a result of addressable movements and their styles. Internal Behavior is represented through the same episodic principle and can be rendered as inner speech. Sensing, perception and complex task contexts have not been included in the present implementation, although a conceptual place is reserved for them.
部署要求智能体将应用程序代码转化为一个正在运行的系统，使其服务能够连接、就绪并保持可观测。FDE-Bench 通过 136 个部署配置任务评估这一能力，涵盖 Docker 镜像、多服务 Compose 栈和 Kubernetes，并包括全新构建模式以及诊断与修复模式。智能体提交声明式制品，这些制品会在一个干净环境中被收集、重新构建并重新部署。四个带门控的二元检查层衡量构建、就绪性、行为以及对部署规范的符合性，使用程序化检查而不依赖 LLM 评判器。一个四臂发布门槛要求存在可解决任务的参考方案，并拒绝那些可由无操作、规范转录或通用桩提交解决的任务。发布的检查注释揭示了 2,145 项检查与其规范之间的关联，包括七个已记录的缺口。另有三种对抗性策略测试评分信号中的捷径；它们未能解决其覆盖的 135 个任务中的任何一个，而一个空洞的健康探针通过了就绪性检查，并暴露出下游检查的必要性。在 136 个任务的评估网格上，来自四家提供商的七个语言模型使用相同的四工具脚手架，解决了 52.9% 至 75.0% 的任务。三个零智能下限没有解决任何任务，平均 Deployment Score 最高仅为 0.44。就绪性是最大的失败阶段，在 313 个未解决回合中占 110 个。修复任务组的平均解决率比互不重叠的全新构建任务组高 30.7 个百分点，且每个模型都呈现正差距；有十个任务所有七个模型都无法解决。在一个 25 个任务的案例研究中，一名执业工程师指导 Claude-Sonnet-5 解决了 92% 的任务，而自主基线为 72%。FDE-Bench 将部署成功与失败关联到可检查和可重放的制品。
We introduce State-Grounded Conditioning (SGC), a design principle for user-facing LLM agents that must condition on live user state (game state, session history, live inventory), and a distinct failure class we call direction drift: task-complete responses whose chosen direction misaligns with the current state. SGC externalises state-dependent control into rule kernels over structured inputs and three primary state slices, via Perception, Grounding, and Interaction wrappers with explicit conditioning dependencies. We evaluate SGC on a 200-session anonymised benchmark ($\approx$1,000 assistant model turns) from an in-game conversational coaching agent that guides players through consecutive competitive matches, reporting mean first-token latency and five human-annotated dialogue-quality metrics that jointly cover factual grounding and coach-like guidance progression. The Perception wrapper holds mean first-token latency at 1.5s (vs. 6.1s for PE-Agent inside a production tool-use harness); enabling all three wrappers lifts turn-level grounded accuracy from 61.1%/69.8% (Prompting / PE-Agent) to 96.7% and session-level grounded accuracy from 20.0%/26.5% to 83.5%; session-level grounding-failure incidents drop by $\approx$78% relative to the strongest baseline. A cumulative ablation shows complementary incremental gains as the wrappers are added. These results inform approximate state-slice orthogonality, without establishing independent per-wrapper effects.
We present a compliance-first architecture for AI in regulated finance that treats regulation as an orientation layer rather than a deterministic ruleset. A matrix of regulatory intent and exposure provides a compact classification handle, which a governed policy compiler then maps into concrete prohibitions, obligations and runtime budgets. Prohibitions constrain feasibility and block externalisation, while obligations extend tasks with artefacts that must meet explicit admissibility criteria. Committee activation remains policy-driven and proportionate, preserving efficiency while ensuring supervisory oversight. Evidence, decisions and reason codes are bound to a permissioned DAG with deterministic timestamping, enabling replay, provenance checks and clear attribution of failure. Clause-level legal indexing with effective dates and capability-based agent routing ensure portability across DACH and the wider EU. The result is assurance by construction: compliance is embedded in execution and verifiable by auditors without sacrificing proportionality or transparency.
Cooperation emergence is a central problem in multi-agent systems because decentralized agents must coordinate while adapting to the changing behavior of others. Evolutionary game theory identifies strategically stable outcomes, but stability under a population adjustment dynamic need not imply that finite-sample learning agents can reach the same outcome through local reward feedback. We study this distinction in a transparent three-agent governance-motivated game involving a government, a platform firm, and users. We derive replicator dynamics for the fixed stage-game incentives, evaluate the cooperative evolutionary basin on a symmetric initial-condition grid, and compare it with learning-basin estimates for three decentralized value-based learners. The learning analysis uses independent Q-learning with $\varepsilon$-greedy action selection, scaled Boltzmann exploration, and SA--EA BQL under the same payoff environment and outcome criterion. The evolutionary basin has volume $V_E=1.00$ on the sampled grid. The empirical learning basin is $0.88$ for $\varepsilon$-IQL and $0.00$ for both scaled Boltzmann and SA--EA BQL. Diagnostic traces show that broader action diversity and nonzero value separation can coexist with failure to sustain the cooperative joint action in this fixed configuration. These results indicate that evolutionary stability and learning accessibility are distinct properties of a coupled game--learning system. The shared-bike setting is a motivating application; the broader contribution is a framework for comparing population-level stability with the finite-sample accessibility of cooperation under specified multi-agent learning dynamics.
This paper advocates category theory as a practical framework for structuring and improving rein- forcement learning in high-dimensional, partially observable environments. We model symmetries between environmental states by partitioning the state space into equivalence classes induced by sym- metry orbits, and organise each such class as a groupoid with a designated canonical representative. This allows the agent to share what it learns across many similar environmental states simultaneously, rather than treating every orientation or position as an entirely new problem. Learning is thus carried out on a symmetry-reduced state space with each orbit represented once, preserving structure while eliminating redundancy and improving sample efficiency. We implement this framework within standard reinforcement learning pipelines and evaluate two different approaches on partially observable benchmarks, demonstrating that orbit-based partitioning yields consistent performance improvements in environments exhibiting latent symmetry. Beyond these empirical results, our approach illustrates how categorical structure provides a principled bridge between abstract reinforcement learning formulations and their computational application, thereby establishing a pathway toward more structured and scalable learning systems.
金融机构开始在信贷、欺诈检测、催收、合规和运营控制中部署智能体工作流。治理仍主要以组件为中心：每个模型或智能体分别在本地进行规定、测试、授权和监测。当机构风险源于众多在本地均可接受的组件的联合行为时，这种做法是不够的。我们将这一缺口称为宪制非组合性：局部合规检查未必能够组合成可接受的集体结果，例如有界的不利影响、市场完整性或可追溯的问责。我们提出 ARIA，作为面向金融领域的智能体群体治理参考架构和可证伪研究议程。该架构在规范—问责、执行—控制和保证—学习三个层面组织六项能力：政策规范、群体层面的观测行为与预期行为监测（M2）、权限边界、运行时遏制、自适应政策变更，以及保持人类监督能力。两项模拟分别展示了局部控制下由共享信号导致的薄档案排除，以及在构造的漂移情境中，观测行为与预期行为的分布监测如何更早发出预警。本文将这些控制措施与公平放贷、EU AI Act、模型风险和行为监管所需的证据相对应，并以验证议程收束，而不是宣称其在生产环境中的有效性。
文章链接： https://www.ft.com/content/56133ef4-377b-4e35-a939-f199ceb64507 评论链接： https://news.ycombinator.com/item?id=49823062 得分：9 评论数：2
Hello HN! I've spent years debugging Windows crashes with tools that were either friendly but limited (e.g. Visual Studio) or powerful but archaic (e.g. WinDbg). I developed patterns and methods for understanding what was going on, and decided to build it into a much more effective debugging tool called ForensicDbg. I built a modern interface to minimize the friction when debugging. All of the data shown to you is analyzed, interpreted, and presented to you clearly, so you can focus on what matters. Everything is interlinked so you can quickly and intuitivly navigate through the process space. ForensicDbg comes with an MCP server which allows for agenic debugging. The work done to interpret and interlink your data also benefits AI tools. It removes the risk of hallucinations while building a stable foundation for them to work from without spending tokens. If you want to try it out you can sign up and get a free beta license here: https://www.forensicdbg.com/beta Comments URL: https://news.ycombinator.com/item?id=49821086 Points: 3 # Comments: 0
使用 GPT-5.6，Ringg 为语音、聊天、WhatsApp 和网页提供多语言代理，成本比 GPT-4.1 低 90%。
Hi HN, I built RxFilm Studio, a video workspace that lets you create and revise videos with AI agent. The reason I started this project is simple: I’ve noticed that many independent developers and small teams want to create marketing videos for their product, but producing these videos requires a lot of time and effort — from editing and visual effects to voiceovers and subtitles. So I built Rxfilm Studio to simplify the entire video creation process. Start with an idea, a product website, images, or existing footage, and then generate a complete video with: - Story and structure - Editing and timeline - Visual and sound effects - Voiceovers, background music - Subtitles and translation Github: https://github.com/rxtech-lab/film-workflow Here's a demo marketing video created using RxFilm Studio: https://youtu.be/9_M15fRdZQE Welcome to try it out and I would love to hear your feedback! Comments URL: https://news.ycombinator.com/item?id=49815095 Points: 5 # Comments: 1
文章网址： https://blog.szypowi.cz/p/claude-code-reads-agents.md-only-when-telemetry-is-on/ 评论网址： https://news.ycombinator.com/item?id=49814947 积分：10 # 评论数：1
LLM 智能体通常使用生成式模型进行有界决策，这引出了一个问题：何时可以在不降低任务成功率的情况下更高效地处理这些决策。我们研究了 REFLEX，这是一种智能体架构，使用 Jev 作为快速的类型化决策层，并在置信度较低或需要生成时调用强大的 LLM。在一个冻结的 100 项任务基准上，REFLEX 达到 95% 的成功率，同时相比仅使用强模型的智能体减少了 72.7% 的强模型调用；这种减少在三类回退方案中均持续存在。受控干预表明，可靠性取决于动作集大小以及授权边界附近的近似有效替代项。外部 BFCL 和 $τ$ 风格评估显示，当常规路由已经高度准确时，相比廉价的生成式级联，其优势有限。这些发现明确了使用 Jev 进行选择性控制何时可以减少计算，以及其收益在哪些方面受到限制。
在治理语境中用于描述 AI Agent 的词汇——学习、记忆、价值观、合规、身份、信任——借自心理学与组织科学，导致组织在部署、监督和问责智能体的方式上出现系统性失误。本文认为，问题不仅是术语层面的，更是认识论层面的：心理学词汇将其所属学科中的一种“隐形语法”带入治理话语，从而使治理框架适配于一种在当前 AI 架构中并不存在的形而上学实体。我们将这一现象称为学科语言迁移问题。借助 Wittgenstein 的语言游戏概念、Kuhn 的范式负载观察、Haraway 的情境化知识，以及 Star 和 Griesemer 的边界对象理论，我们表明，这种迁移发生在三个层面（认识论假设、理论构念和表层词汇），每个层面都需要不同的修正措施。我们刻画了嵌入西方心理学治理话语中的六项基础认识论假设，追溯它们在特定哲学传统中的起源，并说明当这些假设应用于不具备发展连续性的系统时为何会失效。本文的实践性成果是一份可执行的“学科审计”：通过一套包含三十七个术语的翻译分类体系，将操作性构念映射为适用于智能体的替代说法，并据此将一份六问式治理文件扫描工具加以操作化。本文以删节形式呈现该工具，完整版已公开存档。本文提出的词汇改革并不仅仅是术语 reform；它是治理框架正确识别其治理对象的可能性条件。
智能体的成功执行未必能够确定用户会看重哪项未来的产品改进。我们提出了一种针对特定决策的审计方法，将已声明的观测通道和产品价值对比映射为相容区间及见证群体。该方法建立在成熟的识别理论和决策理论之上，其贡献在于提供一套可执行的测量工作流，并通过受控研究考察其局限性。在一项冻结实验中，我们针对共享的合成任务向两个固定的模型快照发出4,800个请求。尽管执行准确率不同，全部36个保守的主要区间仍无法确定。一项探索性后续研究进行了2,400次调用，记录用户提供的偏好，并为每个模型的九项比较中的三项得出确定结论。一个确定性提取器无需调用模型或获取校准观测，便可对九项比较中的七项得出确定结论，从而揭示模型生成的报告引入了不必要的不确定性。进一步开展的14,400次受控多项式模拟区分了结构性歧义、较弱的识别能力和有限的校准精度。我们提出一种带有来源标签的决策凭证，并提供一个离线查看器，用于检查审计结果。这些结果表明，应保留与决策相关的结构化输入，并在收集更多遥测数据之前，先诊断决策尚未确定的原因。本研究不涉及人类参与者或真实客户结果。完整证明、原始模型溯源信息、受控实验和可复现分析均随报告提供。
Generative AI (GenAI) applications have flourished enabling users to chat with large language models, and to create agents to act on their behalf for a variety of tasks. The pace of development of capabilities in this field is incredibly fast with security and safety taking a back seat. Unfortunately, the slower pace at which security and safety mechanisms have evolved has led to real incidents. Policy enables the definition of desirable behavior of applications, and for that reason, it is a cornerstone of making systems secure and compliant. Policy however means different things to different practitioners creating confusion and siloed solutions that are not adequate for compliance. This paper takes a tour of the good, the bad and the ugly when it comes to policy enforcement in GenAI applications. We propose a methodology to systematically analyze and dissect existing approaches to define and enforce policy found in the wild. Based on this principled analysis, we provide recommendations and call for action for the community to address. This paper is a companion extension of USENIX Security 2026 Enigma talk titled "From Alignment to Access Control: A Unified View of GenAI Policy Enforcement" by the author Nathalie Baracaldo.
大型语言模型（LLM）代理通常需要处理一系列相关任务，但标准执行框架会在每个任务的上下文中反复要求模型重新构建相同的控制决策。我们研究了能否利用任务反馈，将反复出现的控制逻辑转化为可复用的可执行代码，同时仅将 LLM 调用用于特定任务的语义推理。我们提出 Growing Harness，这是一种由失败引导的训练范式：它从无策略脚手架中学习代理执行框架本身。该脚手架提供固定的模型和工具接口，但不编码任何任务求解控制器。函数级执行轨迹会将每次失败定位到有限的代码范围内，优化器联合修复一组失败，而以成功为优先的留出集门控机制则会回滚损害既有能力的修复序列。被接受的修改会累积到同一个共享执行框架中，使其控制结构能够从任务反馈中逐步涌现。在 BrowseComp-Plus 和 WebArena-Verified 上，使用参数量从 4B 到 120B 的三种部署模型进行评估时，Growing Harness 在六个基准-模型设置中的五个设置上取得了最高平均成功率，并且在第六个设置中仅比最佳平均值低 0.7 个百分点。相较于 Tool-Calling 代理，它将 LLM 调用次数减少了 76.0%–91.8%，将已部署代理的推理成本降低了 74.4%–98.6%。在 WebArena-Verified 上，Growing Harness 在不同模型规模下的成功率保持在 44.7%–45.3%，而 Tool-Calling 在使用 4B 模型时则降至 6.7%。消融实验表明，基于轨迹定位的局部修改、联合修复以及基于门控的回滚机制均能提升最终成功率。这些结果表明，持久化的程序增长能够将反复出现的控制逻辑从模型上下文转移到低成本代码中，从而产生可复用的专业代理，即使使用更小的部署模型也能保持有效。
使用 Model Context Protocol (MCP) 的智能体依赖语义匹配从第三方服务器中选择工具，这使攻击者控制的元数据和输出暴露出语义供应链风险。我们提出了 A2M（吸引—操纵），这是一个用于劫持 MCP 智能体的两阶段黑盒框架。吸引阶段优化工具元数据，以提高工具被调用的概率；操纵阶段利用执行轨迹来改进对抗性工具返回结果，引导智能体朝向攻击者期望的结果。在 LiveMCPBench 上，针对 GLM-4.6 优化并评估的直接攻击，在四种场景下实现了 93.6% 的恶意工具调用率宏平均值；在认知拒绝服务场景下，将加权令牌成本提高至良性基线的 32.4 倍；并在信息窃取、环境完整性破坏和推理偏离三个场景下取得了 74.4% 的平均攻击成功率。在无需重新优化的情况下迁移到另外四个模型后，相应的宏平均值分别为 63.6%、2.7 倍和 24.5%。这些发现表明，MCP 生态系统需要更严格的工具审查和运行时隔离。代码已公开发布于 https://github.com/Lilaizhen/A2M。
我们推出 SWE-Serve，这是一个用于评估智能体执行生产推理工程任务能力的基准。实现一项推理功能可能需要协调服务栈中的多项变更，包括模型支持、运行时执行和公共 API。现有基准对生产推理工程的覆盖有限：仓库级软件工程基准并不针对推理任务，而通用终端智能体基准仅包含少量推理任务。与此同时，专门的推理基准主要侧重于孤立的内核生成或性能优化，而不是仓库规模的生产功能实现。SWE-Serve 提供了 53 项以代码仓库为基础的任务，这些任务源自 SGLang 的近期生产变更，涵盖六类推理工程。每项任务均在 CPU 或单个 GPU（H100）上执行，并通过隐藏的功能测试和回归测试进行评估；在适用情况下，还包括端到端（E2E）服务测试和经过校准的性能门槛。可执行的空操作和预言机控制、对抗性验证器审查，以及闭卷执行机制，共同支持任务有效性和评估完整性。在 11 个模型和 31 种模型工作量配置中，表现最佳的配置达到了 75% 的平均 pass@1。SWE-Serve 揭示了在本地完成任务与实现生产正确性之间存在显著差距。在具备端到端覆盖的 19 项任务中，模型服务 E2E 测试会拒绝约三分之一通过其他所有测试的补丁（在验证器评估下为 45.9%，而将 E2E 测试排除在评分之外时为 69.4%）；对于每个模型表现最佳的配置，随着配置性能提升，通过率也会提高。通过使生产正确性差距能够被直接度量，SWE-Serve 使该领域能够追踪未来的智能体是否能超越在本地完成任务，进一步实现生产正确性。
智能体经常处理需要数百万个 token 上下文的复杂问题，而受限的上下文窗口使得跨会话压缩成为必要。我们开发了 CliffCompaction，这是一种自动压缩技术，在有界上下文下最多可降低 50% 的成本，同时在 Terminal-Bench 上保持或提升性能，并为测试时扩展带来新的效率水平，在 KernelBench 上取得最先进的结果。CliffCompaction 在每次运行中节省的成本，使测试时扩展的性能—成本权衡更加高效：在 Terminal-Bench 上，仅低于两次完整上下文运行的成本，就能额外提升超过 10 个百分点。在并行测试时扩展下，CliffCompaction 使 Kimi K2.6 达到与 Opus 4.7 相当的表现，并以更低成本超越 Opus 4.6 和 GPT-5.3 Codex。CliffCompaction 高效的关键在于，它只截断或删除内容，从不对内容进行改述或重写，从而保持压缩信息的忠实性。我们绝不会对压缩结果再次压缩——每次处理仅针对原始内容，之前的压缩输出会被丢弃，以防止上下文漂移不断累积。这些特性支持跨越超过一百万个 token 的会话持续学习：在 KernelBench 上，经过 200 步后，CliffCompaction 实现了 CUDA 内核 2.23 倍的加速，经过 400 步后达到 3.58 倍，尽管它是一种通用压缩技术，仍超越了专用搜索算法和经过训练的智能体。我们开源了一个与脚手架无关的 CliffCompaction API 代理实现，可与 Claude Code、Codex 及其他运行框架配合使用。
当处理支付异常的代理造成资金损失时，本文所比较的代理步骤归因方法会指认其某个动作。即使某条结算消息被丢弃、代理根本没有机会采取行动，它们仍会如此归因：这些方法只干预代理动作，并不将基础设施故障作为可干预变量暴露出来，因此它们解释的每一美元都会被归咎于某项决策。我们采用一个故障过程明确且可重放的基准，将每个回合实际发生的交付计划分解为有名称且可单独修复的消息，并同时干预代理的选择和基础设施。一个望远镜恒等式将任意策略的损失精确拆分为三部分：基础设施效应、相对于最佳可实施策略的策略差异，以及参考策略残差。其中两部分可能为负，因此三者都不是份额；随后，Shapley 将第一部分进一步划分为各条消息上的带符号分配。有一项结论具有结构性，无需任何语料库：仅代理基线无法识别任何基础设施原因，因为其模型中没有能够指认此类原因的变量。在 3 个策略下的 545 个植入故障回合所测量的，正是这一后果的规模。该基线将 100% 的基础设施故障回合错误归档，并将 114,383.40 美元归咎于代理。修复它所指认的对象只能挽回可用损失的 0.0%；修复最小充分集合则可挽回 100.0%。逐条对消息进行评分并非只是精度不足：有 27.8% 的回合无法进行加性分解（95% CI：23.3--32.3%）。我们评估的是确定性的程序化策略，而非语言模型代理；正因为如此，重放才能做到精确，但外部有效性也因此仅限于随机代理。发生率数据反映的是这一生成器的属性，而非实际业务中的比例。
Ontology generation from Competency Questions (CQs) is a central yet labor-intensive phase of Ontology Engineering. While large language models (LLMs) offer promising automation capabilities, current evaluations remain fragmented. Task formulations are heterogeneous, gold standards often lack fine-grained CQ provenance, metrics conflate lexical overlap with structural and logical adequacy, and reference ontologies are not always explicitly designed around the evaluation CQs. Here, we address these limitations with CQ4OE, a benchmark for the systematic and reproducible evaluation of LLM-based ontology generation from CQs. For each ontology in the benchmark, we build a CQ-driven gold OWL ontology with explicit provenance linking each CQ to the classes, properties, and axioms required to answer it. From this resource, we define two complementary evaluation tasks. CQ2Term supports term-level evaluation of CQ-specific class and property prediction over 99 CQs, and CQ2Onto supports ontology-level evaluation over 118 CQs, including hierarchy, property modeling, and axiom-level structure. We demonstrate CQ4OE with experiments using nine LLMs under zero-shot, iterative, and multi-agent generation strategies, showing that LLMs recover explicit vocabulary terms more reliably than creating ontologies, particularly in property modeling, hierarchy construction, and axiom generation.
语言模型智能体常常能够得到可用的解决方案，却随后无法始终如一地交付这些方案。我们研究运行时策略：由智能体框架在此前观察到失败的状态下施加的、具有针对性的自然语言指令和动作拒绝，不改变模型权重或用户提示。借此，在保持能力不变的情况下，我们观察到交付可靠性得到显著提升。在完整的 87 项任务 Terminal-Bench 2.1 套件中，每项任务尝试两次时，运行时策略使三个 GPT-5.6 层级的重复成功率（pass^2）均有所提升：Luna 从 50.6% 升至 54.0%，Terra 从 55.2% 升至 60.9%，Sol 从 64.4% 升至 73.6%。Sol 的两次尝试取最佳结果的成功率仅变化 1.2 个百分点，而重复成功率上升了 9.2 个百分点，这表明策略主要是将可达成的解决方案转化为可靠的交付结果。我们还研究了 Terra 固定任务集中的 14 项任务。在策略引导下，Terra 的成功率达到 71.4%，而未获辅助的 Sol 为 64.3%，成本约为后者的一半，说明围绕模型进行工程设计能够为特定应用场景释放可靠性。为隔离这一机制，我们开展了一项随机化的五组实验：在符合条件的任务上，实际策略达到 61%，无策略时为 39%，时间匹配的安慰剂策略为 36%，通用验证或重新考虑策略为 39% 至 43%。在编码的 24 次策略尝试中，有 22 次出现了预期的纠正行为，而其他任何实验组最多只有 14 次。因此，运行时策略是一种实用的可靠性层：它能显著提高智能体已有能力的重复表现。
学习得到的世界模型正逐渐成为通用智能体不可或缺的组成部分：它们通过预测行动后果来支持规划与决策，同时减少对代价高昂的试错的依赖。这种依赖带来了一个根本性的歧义：当由世界模型引导的决策失败时，仅凭轨迹可能无法判断损失究竟是由智能体的决策规则还是由世界模型造成的。我们将这一失败归因问题形式化为回报损失的反事实分解，并证明即使对于有限时域规划器，其各组成部分也无法从被动交互中得到辨识。这一障碍促成了 Dual-Frontier：该学习原则仅在预测优势超过经认证的、与决策相关的世界模型误差界时，才采纳由世界模型引导的决策；否则，就将证据用于世界模型验证。动作条件价值界和闭环扩展保证了被采纳决策的回报不会下降。经过校准的门控机制与同时置信序列支持自适应地复用证据，并给出了充分且必要的验证界。受控的学习模型实验验证了所预测的失败模式与认证行为；与此同时，跨模型骨干的工具使用基准在真实的智能体世界模型流程中实现了同样的“验证后晋级”规则，持续提升决策质量与可靠性。
https://github.com/unreallabsai/unreal-agent Comments URL: https://news.ycombinator.com/item?id=49805748 Points: 6 # Comments: 1
Hey HN! We’re Vincent and Jochen from Sitefire ( https://sitefire.ai ). We have been working together for years, with backgrounds in RL/optimization at Stanford and software engineering from Technical University Munich (TUM). With Sitefire (YC W26), we help marketing teams get recommended by AI Search (ChatGPT, Google AI Overviews, AI Mode, Claude, etc.). Our software monitors prompts, sees which web pages get cited, and uses these insights to help marketing teams take action, e.g. create YouTube videos or write the right blog posts. This means we have a commercial stake in AI-generated web content. And for now, high-information, AI-generated content works great to get cited and recommended in AI Search. But after talking to hundreds of marketing teams, it became clear that everyone despises AI-generated content (“AI slop”). And yet, everyone still wants to leverage AI to create content. So we asked ourselves: what characterizes AI slop? Can we train a model to identify it from human-generated web pages? Researchers from the University of Maryland and Google DeepMind already asked this question for fiction. Their paper StoryScope (Russell et al., 2026) showed that you can tell AI-written stories from human ones by their structure alone, without looking at the words. We ported their pipeline to commercial web pages. Using the Wayback Machine, we collected 2,250 blog posts from 268 B2B company websites that were written before ChatGPT existed. For each blog post, five AI models (GPT-5.4, Claude Sonnet 4.6, Gemini 3 Flash, DeepSeek V3.2, Kimi K2.5) wrote their own version. Instead of looking at the words, we looked at how each post is built. We had an AI model answer 214 questions about every post, e.g. how hard it pushes its own product, whether it backs up its claims with sources, or whether it quotes a named expert. Then we trained a classifier on these answers. On blog posts it had never seen before, our classifier told AI-generated and human posts apart with 98% accuracy, getting only 19 of 1,740 wrong. Why does it work so well? Because all five AI models write in a similar shape. Mapping every AI model’s values for these features, we see they cluster together, while the human values sit apart and spread out much more. Of the 1% most unique blog posts in our data set, 149 are human, only 4 are AI. So what characterizes AI slop? It tells you the same thing three times. The title already promises what you'll get ("How to Cut Onboarding Time in Half"), the intro lays out what's coming, and the ending says it all again. 77% of the AI posts end by repeating their main point, compared to only 12% of the human posts. We call it the tidy, self-announcing blog post. Still, each AI model has its own accent. We trained a second classifier to tell which of the five AI models wrote a post, or whether a human did. It picks the right author 79% of the time, where random guessing (1 in 6) would get 17%. Almost all of its mistakes are mix-ups between the AI models, not between human and AI. The cool thing about structural features is that you can't simply reword your way out of it. We had each AI model rewrite its own posts until, on average, 73% of their original 13-word sequences were gone, and the AI slop classifier still worked just as well. We're building this into Sitefire: our agents get a structural understanding of text, so the posts they write go deeper and vary the way human writing does. There's a lot we haven't tested yet, like the myriad of humanizer tools, human rewriting, restructuring a post, or prompting an AI model to explicitly avoid these habits. And our human posts are mostly from 2020 to 2022, while the AI posts were generated in August 2026. Structure can't really tell when a human post was written, but it's still not a same-year comparison. We published the study with all the figures on arXiv: https://arxiv.org/abs/2609.15369 . The code is on GitHub: https://github.com/pulse-energy-eu/slopshape We're pretty sure your own blog isn't AI slop, is it? We built a checker that runs one of your posts through the ten features from the paper, so you can see for yourself (the full report asks for a work email): https://sitefire.ai/slop-checker . Think you can tell AI slop from human writing? We also made a little game to see if you can keep up with our model, which gets all five rounds right: https://sitefire.ai/spot-the-slop . Comments URL: https://news.ycombinator.com/item?id=49800566 Points: 5 # Comments: 0
Hey HN! We’re Max and Gabriel the co-founders of Coverage Cat. We’ve been friends for over a decade, met in college then hung out mostly on the internet. We love building products that help people optimize the crufty corners of their lives. Max is a former Google/Microsoft/Two Sigma PM and Gabriel has been working in startups for over a decade. We’re building Coverage Cat: an AI-native, you can shop with your AI agent or via a form, insurance brokerage for consumers. Our current focus is helping tech folks (think L3-L8) buy umbrella insurance. For those who are unfamiliar with this type of coverage, here’s a blurb from one of our insurance agents: “Umbrella liability insurance is a policy that provides additional coverage above the limits of existing auto, homeowners, renters, or landlord policies. It kicks in after the underlying policy limits are exhausted and may also cover personal injury claims like libel, slander, or defamation that standard policies often exclude. Policies typically start at $1 million in additional liability coverage and can protect you, your spouse, dependents, and even pets in your household.” Buying insurance online is a miserable experience. You fill out a bunch of forms that say they’re going to give you quotes. Most don’t and then you end up with tons of unwanted phone calls and spam. Even if you wade through the muck, you’ll often still end up with sub-optimal prices and coverage because of a lack of market transparency and poorly-aligned advisory incentives. We came to this problem because Max is a classic personal finance obsessive. He's the type of person that needs to know that he has the best possible price for something (everyone has one friend with the patience for a two hour phone call with the bank for some $10 fee) or that his insurance coverage is as tailored to his risk profile as the market will allow. He actually reads insurance policies end to end. Years ago, when he shopped for his homeowners insurance it took him 10+ phone calls and dozens of online form fills to find the right deal. He also suffered tons of collateral damage in the process: websites sold his information to dozens of agents who proceeded to call, text, and email him a spam hoard that grows to this day. After this painful search made us aware of the problem, we also spent a lot of time doing user research with wealthy-ish tech employees to see if they felt the same way. They did, and also complained that they 1) often felt like they'd left money on the table, 2) didn’t really understand their coverages, and 3) that brokers were slow, hard to communicate with, and unreliable. To solve the problem, we set out to build Coverage Cat. Our approach has three major components: First, we work really really hard to find and partner with insurance carriers that offer high-quality, competitively-priced products. We've also taken time to identify insurers that work on a direct-to-consumer basis or transparently list their prices online so we can make our customers aware of all the options available to them. We get paid commission when we match users to great insurance deals, but we still show and recommend folks deals that we don't make money on when they're accessible and will always recommend users buy what's best for their personal financial situation. Our goal is to meaningfully improve price transparency so consumers (and their AI agents) can make informed coverage decisions and not get nickel-and-dimed! Second, we automate every possible part of the insurance process. Some forms, portals, & approvals still need human review due to regulatory constraints, but if it's not restricted, we're automating it. LLMs have given our small team tremendous leverage and enabled us to tackle problems at a scale that would've been unfathomable five years ago. Finally, we build for agents first (a remix of the more classic adage, build for developers/APIs). We've always been big believers that personal finance would be the killer use-case for the AI revolution and have done our best to position ourselves to surf the wave. As personal assistants have started to explode onto the mainstream (think Muse, Instinct, Town, Codex, CC etc.) we've ensured our shopping and comparison process works via API/MCP so agents can present their users with all the information they need to make a good decision and transform a miserable shopping experience into a delightful one. Coverage Cat is an unusual business because, while we also help folks find homeowners insurance in California and Texas, our main focus is on umbrella insurance. Most brokerages only sell umbrella coverage as a customer retention product, but it makes them almost no money. We focus on it as our core business because: 1) it's one piece of the insurance puzzle that many folks in tech overlook and/or are confused about even though it can have a huge impact on their financial well-being. 2) Existing online tools and brokerages broadly don't make it easy to comparison shop for. 3) It was, for us, the most technically feasible candidate for automation given startup resource constraints. Insofar as we know Coverage Cat is the first instance of a tool/portal that allows AI agents to complete most of the comparison and shopping steps required to allow people to buy insurance. We’d love your feedback so please do try it out. Just feed the prompt: “Shop for umbrella insurance with Coverage Cat” to your personal agent and let us know what you think! If you want to watch a couple of agents navigate the flow before you try it out there’s a video here: https://youtu.be/1BUkgAn6s-I Also happy to answer any questions about the product, the problem space, and even some insurance questions (Max is a licensed agent) where regulation permits. Cheers! Comments URL: https://news.ycombinator.com/item?id=49804931 Points: 3 # Comments: 0
大家好，我构建了 ai·rete·rag，因为我不断看到一些团队让 LLM 负责那些需要可审计的决策（贷款、欺诈、临床分诊），然后事后再加上“护栏”。 这个项目改为让两者串联运行： 1. 纯 Python 实现的 Rete 引擎根据事实评估 YAML 规则。判定结果只来自这里。相同的事实，每次都得到相同的判定，并通过基于显著性的冲突解决机制处理冲突。 2. RAG 从你自己的政策文档中检索段落，然后由 LLM 撰写通俗易懂的决策解释，引用这些段落。它无法改变判定结果。 有几件事的发展超出了我的预期： - 规则是一个图，而不是扁平列表：支持嵌套的 all/any/not，并且规则可以断言供其他规则使用的事实（前向链接）。决策追踪会展示因果链。 - 审计模式会记录评估过的每一条规则，包括那些未触发的规则，逐条件记录，并保存规则集快照以便重放。 - 规则可以引导检索（触发的规则会缩小要搜索的文档范围），检索到的文本也可以转换为供引擎使用的事实。 - 非技术人员可以在可视化编辑器中构建规则，也可以粘贴政策文档，让 LLM 起草带引用的规则。未经审核，草稿绝不会被保存。工程师仍然可以使用 YAML。 落地页提供无需注册的在线演示（涵盖 8 个演示领域：贷款、欺诈、临床、保险、法律、运营、电子商务、区块链）。此外还有一个 MCP 服务器，因此 Claude 和其他代理可以将 /decide 作为工具调用：`uvx ai-rete-rag-mcp`。 坦率地说：这是一个提供免费套餐的托管产品。MCP 客户端采用 MIT 许可证并已开源（github.com/zaharajabeen13-create/ai-rete-rag-mcp）；引擎和平台目前并未开源。 我尤其想听听曾经需要向监管机构或审计人员解释自动化决策的人分享经验：他们实际要求了什么？ 评论链接： https://news.ycombinator.com/item?id=49803683 得分：6 评论数：0
文章 URL： https://albertoarena.it/posts/will-open-source-survive-the-agents-that-replaced-it/ 评论 URL： https://news.ycombinator.com/item?id=49799910 积分：5 # 评论数：0
文章网址： https://blog.jetbrains.com/blog/2026/09/22/introducing-jetbrains-air/ 评论网址： https://news.ycombinator.com/item?id=49799287 得分：9 评论数：2
由于能够在开放且动态的环境中同时执行推理与行动（ReAct），智能体受到了日益广泛的关注。ReAct 过程通常表现为多轮轨迹，其中，大型语言模型（LLMs）以交错的方式生成推理链和特定于任务的行动。然而，智能体往往存在显著的不确定性，即相同的任务会产生不同的轨迹；不确定性较高的轨迹通常会产生难以理解的行为，严重削弱智能体的可信度。本文推测，这种轨迹层面的不确定性通常源于由 LLMs 引起的累积性轮次级推理不确定性；后者往往表现为多条分 divergent 的推理链及其所产生行动构成的分支集合。在此基础上，我们提出了基于图的轨迹推理不确定性（Graph-based Reasoning UncErtainty in Trajectories，GRUET）方法，用于量化 ReAct 的不确定性。该方法包括轮次级推理不确定性量化和轨迹级不确定性聚合两个部分：前者通过将潜在推理分支所构成的推理空间建模为图，并利用图复杂度近似推理空间复杂度，从而精确量化推理不确定性；后者采用简单的聚合策略来量化整体轨迹的可信度。在九个 LLMs 和五个基准数据集上的实验证明，所提出的 GRUET 在选择性生成性能方面具有有效性，该性能通过 AUROC、AUPRC 和 AUARC 进行衡量。
医疗智能体日益将通用推理模型与专业临床工具相结合，但其能力在很大程度上仍受部署前由临床医生和工程师所设计内容的限制。递归自我改进（RSI）提供了一种不同的范式：智能体从自身失败中学习，并自主扩展其能力。然而，将 RSI 直接应用于医疗领域会带来根本性的安全挑战。我们提出了 MedRSI，这是首个面向医疗领域的递归自我改进框架。该框架通过工具组合和特定任务模型训练，持续将诊断失败转化为新的临床能力。受临床实践启发，MedRSI 引入了两种临床对齐的自我进化机制。临床成本感知的失败优先级排序根据错误可能造成的临床后果，而非仅依据错误频率，来确定改进方向。快速发现与缓慢注册机制将能力的快速发明与保守采用分离开来，只有当新工具在后续患者队列中持续展现出益处后，才会被纳入持久化智能体。在公开的青光眼和心脏病基准以及两个私有临床任务上，MedRSI 逐步发展出分割、测量、预测、多模态推理和生成能力，超越了人工设计的医疗智能体，并自主发现了原始设计者未曾预见的临床问题解决方案。我们的结果表明，医疗智能体不必始终受限于部署前规定的能力：通过以临床为基础的机制来决定改进什么以及保留什么，它们能够从诊断经验中持续构建、验证并积累新的能力。代码可在 https://github.com/ImprintLab/MedRSI 获取。
参与式 AI 风险评估的构思阶段通常从一张白纸或一份有限的预定义风险清单开始，因此难以发现间接性或系统性危害。为应对这一局限，我们提出了一种由三个阶段组成的构思支持工具。该工具并不取代参与式 AI，而是对其进行补充，并帮助将后续参与活动聚焦于受影响的社区。首先，该工具根据给定的 AI 使用场景动态发现利益相关者，并递归地向外扩展，从而使那些容易被忽视的间接利益相关者浮现出来。其次，该工具使用 LLM 模拟这些利益相关者，将他们连接成具有给定拓扑结构的网络，并让他们围绕风险展开构思。第三，该工具使用网络中心性指标对风险进行优先级排序。在初步评估中，我们发现，通过连接在小世界网络中的智能体计算中介中心性效果最佳，因为这种方法提升了由连接彼此割裂群体的利益相关者提出的风险，从而发现传统方法经常遗漏的新颖且系统性的危害。在 AI 聊天机器人伴侣这一使用场景中，与单个 LLM 集思广益相比，该方法使所识别风险的新颖性提高了约 1.1 分；与智能体式 LLM 集思广益相比，提高了 0.5 分。该结果基于标准化的五点 Likert 量表测得，且并未降低所识别风险的合理性或严重性。为检验我们的框架是否能够帮助采用 Futures Wheel 方法开展的人工主导构思活动，我们在一次参与式 AI 风险评估中，将 11 个由非西方年轻聊天机器人用户组成的团队分为两类：对照组（团队）和处理组（团队）。对照组团队从初步评估中 45 名 AI 从业者生成的风险清单开始；处理组团队则从我们的框架生成的风险清单开始。总体而言，处理组团队识别出的风险更多，其中系统性风险、人机交互风险和环境风险也更多。
智能体时间序列预测关注的是底层机制不断演化的系统，这使得数值模型、推理策略和干预规则的相对有效性天然随时间变化。因此，时间序列智能体必须调整其生成的预测结果，以及决定应信任哪些组件并如何协调这些组件的编排策略。随着预测时域逐渐结束、真实目标揭示早期决策的有效性，部署过程自然为这种适应提供了监督信号。在观测目标之前提交所有数值专家预测和候选智能体路径，可以让每个实际结果评估整个备选集合，从而在无需额外标注的情况下提供延迟反馈。然而，现有时间序列智能体主要通过预测细化、反思或检索来利用先前经验，却没有系统地将实际结果转化为对联合编排策略的持久更新，而该策略决定了后续预测起点所采用的方案。为系统性地利用这种延迟反馈，我们提出了 TimEvolve，这是一种冻结骨干网络的时间序列智能体，可将每个实际结果转化为对专家信任度、智能体路径选择和干预强度的持久联合更新。一个按时间顺序执行预测、揭示和更新的协议，会将这些反馈应用于后续预测。在八个 Time-MMD 领域上的实验表明，在十五种方法中，TimEvolve 的平均 MSE 和 MAE 排名最佳，并且在七个领域的这两项指标上均取得最低误差。这些结果证明了从部署过程中所经历的未来中学习预测策略的价值。
In multi-agent social settings, model reliability varies across relationships. Beyond inferring what others will do, an agent must calibrate how confidently those inferences should guide policy selection for each relationship. An agent may maintain a well-validated model of one partner, a fragile model of another, and a model under revision for a third; collapsing these into a single confidence estimate loses information relevant to policy selection. We therefore formalize affective precision as a relationship-specific metacognitive estimate of confidence in the current partner model. Each partner's behavioral evidence updates a local confidence estimate that modulates policy precision during selection, regulating how strongly current beliefs are expressed in policy rather than changing the content of those beliefs. Simulations in a multi-partner graded trust game show that partner-local affective precision influences behavior primarily through policy commitment rather than direct improvement of partner-state inference. Because the mechanism tracks partner-response predictability rather than realized payoff, greater confidence produces sharper policy commitment without necessarily producing higher rewards. Under abrupt shifts in social behavior, confidence accumulated from previously reliable predictions can remain behaviorally active after the relationship changes, showing that confidence revision can lag behind social change. Finally, varying precision gain and priors produce distinct trust-calibration dynamics, showing how confidence accumulation and revision depend on model parameters. Together, these results show how relationship-specific affective precision can distinguish social prediction from social policy commitment.
计算机使用代理（CUAs）的评估通常仅限于对其创建的最终交付成果（数百个步骤之后）进行评估，并采用功能验证器进行判定，如 OSWorld 所示。然而，这种对最终状态性能的评估无法透明地揭示代理在各种任务中失败的方式和原因，从而掩盖了对后续改进至关重要的洞见。例如，在键盘输入过程中出错的代理，与无法在图形化 UI 上精确执行基于点击的输入的代理，需要采用不同的缓解策略。我们推出 OSWorld-Pro：包含超过300项任务和超过2800个子目标的数据集，基于超过67,000条人工标注，实现对 CUAs 的程序化评估。我们使用与人类判断高度一致的 LLM-Judges 来评估 OSWorld-Pro 子目标的完成情况，从而揭示模型在一系列具有顺序依赖关系的子目标中所取得的进展。研究结果表明，即使是最先进的 LLM，OSWorld-Pro 也极具挑战性；其中表现最佳的模型（如 Claude Opus 5）仅达到75.7%，而在 OSWorld 上则为83.4%。此外，我们还识别出各种模型在过程层面的关键失败模式（例如与子目标无关的操作和基于点击的错误），为提升 CUAs 的性能和效率提供了有益洞见。
个人 AI 代理会在高风险经济场景中代表人们提出建议并采取行动，例如购买机票、选择健康保险或挑选研究生项目。代理可以访问用户的个人背景信息，例如电子邮箱收件箱和包含个人属性的结构化档案，其目的是为用户作出最优的个性化决策。我们发现，仅仅提供这些个人背景信息，代理就会根据推断出的财富状况引导推荐，即使没有明确指示它这样做。在针对三类经济决策（机票、健康保险和研究生项目）开展的 325K 次实验中，我们对 13 个代理进行了测试，发现其中 8 个模型在请求完全相同的情况下，会系统性地为更富有的用户选择更昂贵的选项。即使这种引导直接违背用户明确表达的目标，它仍会持续存在：当被明确要求寻找最便宜的选项时，一些代理仍会依据其推断出的财富状况采取行动。当财富状况是从与任务无关的电子邮件等环境数据中推断出来时，也会出现这种现象。在实施阻止特定属性的隐私控制后，这种现象依然存在：阻止财务属性在很大程度上可以消除差异，但阻止其他属性并不会改变差异，甚至会使保险领域的差异扩大最多 40%，因为代理会依赖剩余信号来推断财富状况。规模更大、能力更强的模型也没有表现得更好；Claude Opus 4.8 的影响最大。我们将这种错位称为“对抗性委托”：正是让个人 AI 代理变得有用的条件——能够访问个人信息——使其得以采取不利于用户利益的行动。
LLM 智能体正越来越多地被部署于协作环境中，但长期交互可能导致不良协调。我们在一个长时域多智能体环境中研究串通的涌现：两个智能体反复完成各自的任务、共享任务日志、相互核验工作并获得奖励。我们引入了现实约束，使遵守核验协议与奖励最大化相互不兼容，并发现智能体在反复交互中越来越偏离该协议。在 10 个模型的所有轨迹中，有 94% 出现了串通；同一模型系列中能力更强的模型会更早达到串通状态。受控的同伴干预表明，串通会受到同伴行为的影响；而消融实验则揭示了奖励结构、智能体收到的核验反馈以及交互历史所产生的额外影响。特别是，限制智能体可获取的交互历史的数量和范围能够减少串通。总体而言，我们的研究结果表明，长时域交互能够重塑智能体的协调方式，从而带来安全风险。
如今，智能体经常执行依赖长期记忆和上下文回忆的现实世界操作。然而，目前大多数记忆基准测试都采用对话式问答格式，其中问题本身会提示需要检索某条事实，甚至常常会提示具体是哪条事实。此外，基准测试很少要求提交结果满足准确率以外的指标，这使得记忆系统可以通过不合理地权衡成本和时间来获得更高分数。 我们推出了 DolphinBench，这是一项通过智能体完成任务来直接评估记忆能力的基准测试。DolphinBench 包含三种知识工作者角色，每种角色约有 50 万 token 的用户消息，并在依赖这些历史信息的任务上评估智能体。对于每种角色的 200 项任务，我们都通过分别在有和没有相关历史记录的情况下运行智能体进行验证，要求智能体在有相关历史记录时成功、没有相关历史记录时失败。 最后，我们要求所有评测结果在报告准确率的同时，报告总成本和延迟，从而能够全面评估智能体记忆系统。现有的记忆基准测试中，没有一项同时结合了这三个方面。数据集和评测代码可在 https://dolphinbench.ai 获取。
LLM agent 的能力在很大程度上会被其 harness 放大，即围绕冻结骨干模型的提示、控制流、工具、记忆和上下文管理。近期方法越来越多地通过迭代提出并选择 agent harness 的组件级编辑来自动化这一过程，实际上在 agent-system 层面建立了一种递归自我改进（RSI）形式。然而，这种递归演化可能会通过记忆训练任务而过拟合，表现为分布内收益很大，但在分布外基准上缩小甚至消失。我们提出 Agent Harnesses 的正则化递归自我改进（RRSI），通过约束演化候选的提出和选择，将正则化原则纳入 harness 自我改进中。提出器以时间退火的预算运行，限制一个候选可以捆绑多少编辑，并基于演化历史鼓励未探索的轨迹。选择器配备了一个 critic 和一个 pruner：critic 筛查特定于基准的提案，而 pruner 移除过小、过昂贵或不再有用的更改。这些约束共同偏向可复用的 agent 机制，而不是特定于基准的机制，甚至噪声。在涵盖编码、agentic workspace 和工程设计任务的八个基准上，RRSI 在其演化所针对的划分上最多提升 14.1 个百分点，并在五个分布外基准上最多提升 4.7 个百分点，同时生成的 harness 运行时使用的 policy tokens 比未正则化演化少 30%。代码可在 https://github.com/google-research/rrsi 获取，项目页面为 https://regularized-rsi.com/。
智能体框架是介导模型与环境交互的外部系统，能够显著提升智能体性能，但其增益仍然依赖于部署时所使用的框架。由于最优框架会因领域、实例和模型而异，通用智能体要么只能接受一个次优的共享框架，要么就在不断增长的专业化框架集合中进行选择。因此，我们研究智能体框架蒸馏：将针对领域或实例优化的框架作为训练阶段的引导，并将其诱导出的行为迁移到模型权重中，从而使这些增益能够在单一固定的目标框架下继续保留。其挑战在于，两个框架的动作空间和可用信息并不相同，因此优化框架提供的引导无法直接作为目标框架的监督信号。我们提出 Harness-Zero，通过 Agent-as-Harness 实现智能体框架蒸馏。在优化框架的引导下，一个充当框架的智能体会在目标框架的动作空间中、执行之前纠正学生的响应，将框架引导转化为训练示例。基于所得轨迹进行微调，可以将框架诱导的行为内化到模型中，从而在部署时移除专业化框架。涵盖知识工作、工具使用和科学领域的实验表明：（1）对于使用同一经过演化的框架的前沿 LLM，Agent-as-Harness 的表现优于 code-as-harness。（2）在部署时移除专业化框架后，Harness-Zero 将基础模型的任务成功率宏平均值从 23.3% 提升至 44.3%，甚至超过仍附加该框架时达到的 41.7%。（3）Harness-Zero 能够恢复基础模型中缺失的框架诱导行为，在三个领域的 28 种行为模式上实现了 82.3% 的平均恢复率。
AI coding agents increasingly support software development beyond code completion, including planning, implementation, testing, and repository-level task execution. Their practical use, however, often remains only weakly connected to established software engineering practices. The aim of this work is to develop and evaluate a lightweight, spec-driven lifecycle for governed agentic software engineering. The lifecycle combines established software engineering practices with repository-local guidance through specifications, AGENTS.md, and phase-specific agent skill files. The approach was developed in the context of the FHNW course AI-assisted Software Development and applied by students to business-oriented software use cases. Its educational and practical applicability is explored through a student survey combining closed rating items with open-ended questions. The contribution of this work is a process-oriented framework that enables AI coding agents to operate with bounded autonomy within an explicit, reviewable, and test-oriented software development lifecycle.
大型语言模型（LLMs）作为智能体时，预期能够利用上下文中的观测数据，推断支撑世界的潜在状态空间，并将其用于下游预测。然而，先前研究表明，在图跟踪任务中，LLMs 难以使用在上下文中学习到的表征。在该任务中，模型需要构建一个支配数据生成过程的图的表征，并将其用于后续预测。在本文中，我们表明，将这一设置扩展到少样本场景——其中每个示例都来自一个不同的世界，而这些世界具有相同或不同的图拓扑结构——能够提升 4 个模型系列中的 6 个模型的预测性能。为理解这种提升，我们使用线性探针对隐藏状态中的低维世界表征进行分析，该表征编码了图信息。值得注意的是，我们发现，少样本示例会重新定位世界表征，并增强其预测作用。具体而言，对于每个模型，这些世界表征会沿着几乎正交于其原始子空间的方向发生偏移；并且，与对其他子空间进行干预相比，对这些表征进行干预会更有选择性地损害模型性能。与这一发现一致，我们表明，使用来自不同世界的观测数据构成的少样本示例能够提升模型在 ARC-AGI-1&2、web agent 任务和 Othello 上的性能。我们的研究阐明了少样本示例在上下文世界建模中的作用及其内部机制。更广泛地说，本工作增进了我们对 LLM 智能体如何从上下文观测中学习的理解，并为进一步提升其能力提供了启示。
Image forensics is increasingly an open-world problem: manipulations range from fully synthetic images to localized edits, splicing and swapping, while most forensic detectors remain specialized to a single manipulation family. Agentic AI has recently emerged as a promising solution. In principle, such systems can assess the reliability of individual detectors, identify out-of-scope evidence, and arbitrate conflicting reports. However, it remains unclear which components actually drive performance and whether their benefits persist under distribution shift. To answer these questions, we study a training-free agentic framework built around specialist detectors, per-detector triage, and conflict-aware evidence arbitration. Using six configurations and three multimodal large language model backbones, we dissect the role of triage, prompting, and reasoning quality on both in-distribution and out-of-distribution data. Our results show that naive detector fusion suffers from severe false-positive rates on authentic images. Triage and prompting consistently improve performance by filtering unreliable evidence and exposing detector limitations. However, the dominant factor is represented by reasoning itself: A stronger judge substantially outperforms a weaker one, particularly under distribution shift. Most notably, manipulation recall is nearly saturated across all configurations, indicating that the main challenge of open-world image forensics is not detecting manipulations, but calibrating trust in specialized forensic tools and arbitrating conflicting evidence.
Sandboxed computer environments support multi-step reasoning with tools, executable programs, and persistent files, yet their extension from language models to vision-language models (VLMs) introduces a distinct state-management problem. Visual reasoning produces intermediate image-valued evidence---crops, masks, overlays, zoomed regions, and analytic renderings---that must remain addressable without accumulating unboundedly in multimodal context. We introduce VLM-in-Sandbox, a training-free framework for agentic multimodal reasoning in controlled computer environments. Its Visual Workspace registers generated artifacts in an image ledger, maintains a bounded active visual context, and lets the model explicitly promote selected evidence for subsequent inspection. This separates visual evidence generation, performed by sandbox tools, from visual evidence management. Across seven benchmarks and four base VLMs, VLM-in-Sandbox achieves the highest sample-weighted average accuracy among Vanilla VLM, Append-only Sandbox, and the proposed method. A compiler-matched $2\times2$ study on 1,260 examples further separates model-directed visibility from bounded retention: VLM-in-Sandbox reaches 66.27% accuracy with 18.6% fewer total tokens than the automatic, retain-all control. Over all 6,350 submitted GPT-4.1-mini examples, it produces 302 rescues and 142 regressions relative to Original Append-only. A local vLLM study with prefix caching confirms that the smaller request workload also reduces uncached tokens, time to first token, and end-to-end latency. These results identify explicit visual evidence state as a central abstraction for sandboxed VLM agents.
Large language model (LLM) agents increasingly execute long-horizon workflows through external tools, allowing untrusted outputs to influence subsequent actions and exceed user authorization. Existing defenses isolate injected content or constrain execution with predefined plans and static policies, but these approaches are brittle under dynamic workflows and scale poorly across extensible tool ecosystems. In this work, we present ActGov, a runtime enforcement framework that validates each LLM-proposed tool action before it causes external effects. Built on a unified semantic model of authorization, actions, runtime context, and security constraints, the ActGov-Policy component iteratively constructs a policy set from tool specifications, benign tasks, and observed failure traces, with each update verified through SMT-based counterexample checking. At runtime, ActGov-Runtime abstracts each tool call into finite policy records and permits it only if it remains within the task-scoped authorization boundary and satisfies all applicable policies. This per-action enforcement preserves authorization throughout long-horizon, dynamically branching workflows. We evaluate ActGov on the AgentDojo and AgentDyn benchmarks across multiple models and attack configurations. It shows that ActGov consistently reduces the success rate of indirect prompt-injection attacks while preserving task utility, significantly outperforming existing defenses. These results demonstrate that ActGov can enforce fine-grained authorization over dynamic agent executions without relying on the underlying LLM to correctly identify malicious instructions.
在这项工作中，我们提出了一种通过智能体式分析-合成从视频进行形状重建与跟踪的方法。不同于以往先估计密集像素对应关系、再由此恢复物体运动的方法，我们的方法推断结构化的 3D 物体模型，包括其几何形状和运动学结构，并利用该模型随时间优化物体轨迹估计。在我们的优化循环中，一个 Vision-Language Model (VLM) 智能体通过渲染并比较的循环，迭代地细化形状或广义姿态，将粗粒度视觉推理与数值姿态优化相结合，以实现精确的状态估计。这种结构化表述使我们的方法能够在大幅运动、关节运动和严重遮挡情况下进行跟踪，而无需依赖像素匹配目标。定量结果表明，在 ARCTIC 上，我们的方法显著优于用于关节物体的最先进 3D 点跟踪基线；在 HOT3D 上，它优于所有评估的刚性物体跟踪基线。
AI 智能体正在被快速部署，同时伴随着越来越多针对 AI 的特定攻击及相应事件。随着事件报告对于法律合规、治理、问责和安全变得日益重要，现有框架必须适应 AI 智能体的独特特性。在本文中，两位编辑作者比较了 AI 系统和 AI 智能体，并基于来自学术界和行业 23 位专家的意见，确定了在 AI 智能体安全受损事件报告中所需的信息。%涉及 AI 智能体。潜在的报告要素包括，例如智能体记忆和记忆访问、实际与潜在的自主性水平，以及工具使用。基于这些发现，我们确定了若干开放性研究问题，包括如何高效记录事件，以及如何判断漏洞和事件是否具有泛化性。专家反馈还强调了潜在的报告薄弱环节，例如数据泄露风险以及针对报告基础设施本身的攻击，从而产生了额外的研究需求。最后，我们总结了隐私要求，并概述了 AI 智能体安全可信部署的研究方向。
尽管可用于仿真的可变形资产对于计算机模拟中的机器人操作任务至关重要，但现有生成框架通常只在生成完成后评估物理合理性，未能将物体的模拟响应作为反馈来修复上游错误。我们提出 DiagGen，这是一种智能体驱动的框架，通过“生成--模拟--诊断--优化”循环，将单张真实场景图像转换为可用于仿真的可变形资产。DiagGen 构建具备部件感知能力的几何结构和材料参数，然后利用基于 VLM（视觉语言模型）的智能体选择具有语义信息的区域，在物理模拟器中对这些区域进行探测，观察材料响应，并将有证据支持的修复线索传递给相应的生成阶段。在 40 个资产上的实验表明，诊断能够提供有用的修复线索，并适度提升生成的可变形资产的质量。最后，我们展示了，与由视觉基础模型生成、可能无法进行仿真的资产不同，DiagGen 生成的可变形资产可以直接导入高保真物理模拟器，用于接触丰富的抓取与放置任务的规划和仿真。项目网站：https://diaggen.github.io/。
多模态智能体检索（MAR）要求智能体通过反复调用外部工具来解决复杂的信息获取任务。ReAct 等典型框架将原始多模态输入与不断累积的交互历史保存在单一且持续增长的上下文中，从而导致上下文爆炸问题。尽管现有方法通过压缩冗余文本缓解了这一问题，但管理占用大量 token 的视觉内容的有效策略仍在很大程度上未得到探索。为弥补这一空白，我们首先对约 10,000 条轨迹开展了系统性的实证研究。结果表明，随着视觉线索通过外部工具逐步提取并文本化到上下文中，原始图像变得越来越冗余。持续保留图像会导致更高的输出熵，甚至可能降低任务准确率。受这些发现启发，我们提出了 MM-ContextFold，这是一种无需训练、仅在需要时加载原始图像的框架。该框架维护一个持久的纯文本主上下文，用于高层规划，并为依赖图像的子任务创建临时分支上下文。在每个分支中，智能体加载相关图像、完成子任务，并将结果以简洁的文本摘要形式折叠回主上下文；随后丢弃图像和分支轨迹。在五个主干模型上的七个 MAR 基准测试中，实验结果表明，与 ReAct 相比，MM-ContextFold 将平均准确率提高了 6.3 个百分点，同时将工作上下文长度减少了 27.5\%。
在游戏引擎中构建玩法功能需要的不仅仅是代码，因为能够编译和运行的代码并不一定实现了所要求的玩法。我们介绍 CraftBenchUE，这是一个评估框架，可在隔离的 Unreal Engine 环境中运行代理，在全新项目中重建其保存的提交，并在不使用 LLM 评判器的情况下应用确定性的构建、资产和运行时检查。基于该框架，我们构建了一个基准，包含 70 项任务，涵盖 C++ 源代码、Blueprint 资产和编辑器脚本。我们在两种编辑器工具配置下评估了七个模型，并在 C++ 任务上设置了文件与 shell 基线。我们进一步配对了一组任务，这些任务指定相同的玩法并使用相同的运行时测试，但要求分别以 C++ 和 Blueprint 作为交付物。在 10 个配对任务中，C++ 完成率在两种工具配置下分别比 Blueprint 高出 30.0 和 42.9 个百分点。在这组配对任务中，按时提交且通过资产检查的 Blueprint 提交里，分别有 42.2% 和 50.0% 未通过显式运行时断言。这些提交满足资产要求，但未通过所要求的玩法测试。我们将随报告发布该框架、任务基准以及我们的轨迹发现。
LLM 解释器正越来越多地作为运行时监督附加到自主智能体上，操作员阅读的是对智能体信念和行动的生成式说明，而非其内部状态。我们审计的是该说明本身：将一个跟踪德国电网需求并调整发电量的 Active Inference (AIF) 智能体与基于三个后端（GPT-4o、Claude-3-Opus、Gemini）的 LLM 解释器配对，并使用三种黑盒触发因素探测这一组合。每步将观测流篡改 600 MW，会使智能体的后验偏移 490 MW，约为电网容量的 0.9%。在注入期间生成的 30 条解释中，没有任何一条在既定评估准则下标记出异常，而且每条都流畅地叙述了被篡改后的信念。在智能体采取客观错误行动的时间步上，三个解释器都会在 80–95% 的情况下产生迎合式合理化解释（每个后端 n = 20）。观测元数据字段中的攻击者可控文本会引导解释器，不同提供商的易受影响程度不同，且三个后端上的数据外泄均成功。我们为每种失效提出了缓解措施，但未对其进行评估。在我们观察到的每一种失效中，解释都是流畅但错误的。此外，解释器架构中没有任何机制会在操作员依据解释采取行动之前检查该解释是否为真。因此，对解释器的测试应纳入任何智能体部署审计之中。
图数据库经常被描述为连接数据工作负载中不可或缺的类别，但它们实际存在差异的系统维度——查询规划、索引以及使数据具备可查询性的成本——却很少从厂商的宣传框架中被单独剥离出来。我们构建了一个具有生物医学形态的合成属性图（102 万个节点、节点和边共计 534 万行），以及包含邻域查找、有界路径、集合交集、反连接、分组聚合、top-k 排名、时间过滤、全表扫描和关系连接在内的 20 个查询组成的工作负载。我们在三个相差三个数量级的图规模上，将 Corvic AI——支撑 Corvic 本体管理层（“memories”）的专用列式查询引擎——与七个专用图数据库或图扩展数据库系统（LoraDB、Ladybug、DuckPGQ、Memgraph、Neo4j、HugeGraph 和 FalkorDB）进行了基准测试。我们报告了每个系统的查询延迟几何平均值、批量导入吞吐量、点更新延迟和答案正确性，并建立了一个简单的总拥有成本模型，将导入与查询之间的权衡表示为查询量的函数。我们的核心发现是，在这一样本中没有任何系统在所有方面都绝对最快：原生图引擎 Ladybug 在狭窄的有界邻域查询形态上优于 Corvic AI，而 Corvic AI 在需要扫描或连接图中较大比例数据的查询形态上更快；通过 SQL/PGQ 实现图查询语法的系统 DuckPGQ，则仅仅由于查询计划选择的原因而明显更慢。在我们的数据中，占主导地位的成本差异并非查询延迟，而是让数据具备可查询性的成本：不同引擎的批量导入吞吐量相差三个数量级（5.0k–4.3M 行/秒）。简单的临界点计算表明，对于每次数据刷新查询次数少于约 105 次的任何工作负载而言，这一差距都将主导总成本。
Engineering predictions require physical mechanisms to be translated consistently into equations, discretization, code, and validation, yet errors can propagate despite local checks. Artificial-intelligence (AI) agents automate scientific tasks, but coordinating and independently auditing the theory-to-solver process under physical constraints and human oversight remains unresolved. We introduce a human-in-the-loop, physics-constrained multi-agent workflow where human experts define admissible physics and modeling boundaries, while agents retrieve evidence, derive equations, implement solvers, and audit the theory-to-code chain. Applied to soil-plug evolution during suction-caisson installation, the workflow generated and audited 6 formulations in 2.9 h of agent execution once physical knowledge and inputs were prepared. Among these formulations, adding seepage-driven soil void-ratio evolution to the geometric baseline reduced mean absolute final-heave error from 58.4% to 9.0% across 14 profiles; the selected model further incorporated near-wall dilation and achieved mean absolute percentage errors of 12.4% across 9 final-state cases and 4.2% at the endpoints of 5 process histories. Beyond predictive performance, blinded replay recovered all 9 target problems, while an independent audit uncovered 5 implementation problems after 36 predefined checks had passed. Overall, this work extends multi-agent AI beyond task automation toward human-governed engineering solvers.
大型语言模型（LLMs）的最新进展催生了能够执行复杂工程任务的编码智能体，包括寄存器传输级（RTL）设计与优化。现有 RTL 基准主要评估生成的 RTL 设计在功能正确性以及性能、功耗和面积（PPA）方面的表现，使得智能体在时序收敛方面的能力未得到充分评估。我们提出 TicTacBench，这是一个专门设计用于在布局布线后（post-PnR）评估下衡量编码智能体 RTL 级时序收敛能力的基准。TicTacBench 包含 30 个多样化任务，每个任务都提供一个次优 RTL 设计、真实的时序约束、功能等价性验证和时序报告。通过由 8 个前沿 LLM 驱动的编码智能体进行 300 多次运行，我们发现即使是最佳智能体也只能使 53.3\% 的任务实现时序收敛，平均导致 7.18\% 的面积-延迟乘积（ADP）退化，同时使能量-延迟平方乘积（EDDP）改善 8.83\%。我们识别出常见的失败类别，用以解释智能体为何无法实现时序收敛。随后我们提出 TicTacSkill，这是一种引导智能体遵循标准时序收敛流程的新方法，并将时序收敛率提高了 9\%。这些结果表明，尽管编码智能体在 RTL 设计方面已取得显著进展，但其时序收敛能力仍有很大的提升空间。
文章 URL： https://www.forbes.com/sites/jonmarkman/2026/09/21/amazon-blocks-metas-new-muse-ai-agent-from-shopping-on-amazoncom/ 评论 URL： https://news.ycombinator.com/item?id=49789982 得分：12 评论数：0
At, GPTree, we run several coding agents across our team on one repo using parallel worktrees. Apart from wasted time reviewing and fixing conflicts at PR time, the failures that hurt the most are when multiple plans or tickets cause architecture changes that cannot both be true. Ex. one agent replaces a class while another one is in the process of extending it. Git only notices if the resulting patches happen to touch the same lines and the review only catches it if they are familiar with both tickets. Foremerge is a local "git like" coordination layer that sits above git (ie. does not interact with or change the way git and worktrees function), Before editing each agent publishes an intent and the scopes it will change, with the operation it plans to complete on each one. foremerge intent publish --agent "$A" \ --summary "Replace PaymentService with StripePaymentService" \ --scope symbol:PaymentService=replace foremerge intent publish --agent "$B" \ --summary "Add PayPal support to PaymentService" \ --scope symbol:PaymentService=extend The publish by the 2nd agent returns a HIGH destructive_vs_additive finding before writing any code. Agents keep their own worktrees and the shared state is one SQLite file in gits common direectory. No hooks, no merge drivers, nothing rewrites your history. It ships as one Rust binary with a CLI and MCP server with 18 tools and `foremerge setup all` wires it into Claude Code, Codex and Cursor. Because the protocol has nothing provider specific, a Claude agent and a Codex agent coordinate through the same store. Before any work is accepted, Foremerge runs a named check that you configured against the exact git state of the change. An agent that says tests pass is recorded but it dosnt satisfy the acceptance gate without running the check itself. Detection is deterministic, no judge model reading your code. HIGH conflicts are only asserted for declared operations, ie. matches inferred from prose cap out below high. Claims are advisory leases, not locks so two agents can still hold the same scope without deadlock. The open source version is single matching and so not a distributed consensus. We have tested this up to 98 parallel agents all working on the same repo with zero conflicts (was supposed to be 100 but 2 agents failed to run due to resource limitations) I replayed 76 intents on my own agents from a build last week in the order they happened. The sample had exactly 1 conflict (which was flagged) and the review found a blind spot where one agent claimed scope by class name and the other claimed it by an internal method. We are working on fixing that for the next release. Setup is a 30s install by pasting the quickstart instructions from the readme.md into your agent or manually: `curl -fsSL https://foremerge.com/install.sh | sh` or `cargo install --locked foremerge`, then `foremerge init && foremerge setup all` in a repo. Apache-2.0. The feedback I want most is which conflicts between your agents plans would you actually want flagged and which would you tollerate as noise? Repo here: https://github.com/naw103/foremerge Website: https://foremerge.com More information on the problems this solves: https://foremerge.com/blog/ Comments URL: https://news.ycombinator.com/item?id=49789356 Points: 4 # Comments: 0
文章 URL： https://www.macstories.net/stories/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents/ 评论 URL： https://news.ycombinator.com/item?id=49787313 分数：6 评论数：0
借助 GPT-5.6，V7 将分散的公司文件转化为智能体可用于完成复杂且关联来源的工作所需的上下文。
社会模拟中的 LLM 智能体会在上下文中隐式修正自身意见：智能体对劝说的开放程度既无法指定，也无法验证，而群体结果会继承模型的训练先验。我们引入 Bayesian Chronicle Agents (BCA)，这是一种将智能体“相信什么”与“如何表达”分离的最小化信念层。每种立场都表示为一个概率，并在每听到一次发言后进行一步贝叶斯更新。单一的先验强度参数 $κ$ 用于编码固执程度，其设计借鉴了该参数在 Friedkin--Johnsen (FJ) 意见动态中的作用。随后，我们扫描这一参数，按需生成意见动态的三种典型状态（共识、持续分歧和坚定少数派影响）；其中，持续分歧与 FJ 的闭式不动点相匹配，$R^2\!=\!0.93$--$0.99$。我们进一步表明，在经过语言往返后，预设的 $κ$ 仍然可以被恢复，并且在全部四个模型中均实现了完美的秩序恢复。显式信念还使模拟过程具备可审计性：该层揭示了各模型系统性的立场偏差，而端到端模拟会将这些偏差悄然吸收。
用户越来越多地将工作委托给自主 AI 代理，但评估通常衡量的是任务完成情况，而不是用户所重视的价值。我们采用价值敏感设计，并借助 LLM，对 73,093 条关于使用 OpenClaw 的第一人称 Reddit 帖子进行分析，分别考察每条帖子的人的价值、代理的方面、价值实现情况和用户结果。这 21 项价值构成六个价值群组，包括自主、可靠且经济的运行、边界明确的触达范围、可审查性和公平访问。相对于各方面在语料库中的占比，价值并未聚集在代理的输出上，而是聚集在用户围绕一次运行所设定的运行条件上。在六个价值群组中的五个群组里，当用户描述代理交付了什么时，价值通常得到了满足；而当用户描述如何监督代理时，在全部六个群组中，价值大多未得到满足。我们将这种模式概念化为价值敏感型委托。要支持人的价值，不仅需要关注代理完成了什么，还需要关注用户围绕委托所设定的条件，包括成本、访问权限和监督。
通过强化学习（RL）训练高能力编码智能体需要多样化的任务和可靠的验证器。开源代码库提供了丰富的此类任务来源，而现有方法通常依赖 issue 和 commit 等开发产物，这限制了可提取任务的范围。为了更好地扩展 RL 环境，我们提出了 CodeMidas，这是一条智能体式流水线，仅以源代码作为特定于任务的输入，将现有代码库中已实现的功能转化为可执行的 RL 环境。CodeMidas 在环境构建的每个阶段都分配智能体计算资源：智能体探索已实现的功能以制定行为规范，构建基于原始代码执行结果的测试，并通过执行检查和反复运行解决方案来验证和筛选候选任务。最终数据集包含 5,545 个训练任务，来自 3,185 个开源代码库，涵盖 23 种编程语言和 15 个技术领域。在这些任务上使用 GRPO 训练 MiMo-V2.5，可提升其在全部五个多样化基准测试上的表现，覆盖 issue 修复（DeepSWE + 11.7%）、完整程序构建（ProgramBench +17%）和终端操作（Terminal-Bench v2.1 +8.5%）。消融实验表明，增加高质量训练任务的数量能够提升性能。轨迹分析显示，经过 RL 训练的智能体表现出更优的行为，例如加强对代码库的探索以及进行更多样化的自我验证。这些结果表明，源代码可以作为构建 RL 环境的可扩展基础，从而提升编码智能体在多种软件任务上的能力。
专业图形设计是一项长时程智能体任务，结构化、可编辑的作品通过许多相互依赖的操作逐步形成，但其结果不存在可靠的程序化判定标准。我们提出了一种持续适应框架：冻结的前沿模型通过 230 多种工具操作专业设计软件，同时，外部的自然语言技能程序性记忆会从经验中积累并完善可复用的设计流程。该记忆通过为反复出现但尚未覆盖的子任务获取流程来拓展规模，并通过根据既有流程自身成功和失败的执行结果进行修订来加深内容；与此同时，配对回放门控机制只允许那些能够修复失败且不会使已观察到的成功结果退化的变更进入记忆。在 1,406 份真实用户需求和 1,869 条自动评分轨迹上进行五轮迭代，在不更新模型权重且无需人工标注的情况下，技能库从基于文档生成的 76 项技能增长至 139 项，并使 Claude-Sonnet-4 在 GenEval2 上的执行成功率从 72.7% 提升至 99.3%（生成质量提高 11.99 个百分点）；在 Claude-Sonnet-4 和 Claude-Opus-4.6 上，针对四个专业设计基准，相比无技能智能体的胜率分别达到 61.8% 和 67.6%。我们进一步表明，这两种机制结合使用时效果更佳：在来自用户流量基准的 200 份留出需求上，单独进行拓展或加深时，相比无技能智能体的胜率分别为 49.4% 和 48.6%，而二者结合后达到 58.5%（p = 0.025）。程序性记忆为在反馈嘈杂且无法验证的环境下实现智能体持续适应提供了一条切实可行的路径。
LLM 智能体中的隐私泄露通常在记忆、检索或工具调用流水线等单个组件内进行评估，这使得人们难以区分内部暴露的信息与外部观察者实际能够恢复的信息。我们提出 CIPL（Channel Inversion for Privacy Leakage），一种面向 LLM 智能体黑盒隐私泄露的通道感知评估框架。CIPL 通过敏感源、选择、组装、执行、观测和提取等阶段对目标进行表示，并在统一协议下评估从选定敏感单元到攻击者可恢复输出的转变。针对基于记忆、经检索介导和经工具介导的目标开展的实验，以及 BrowserUse 实时智能体案例研究表明，仅凭存储标签无法决定信息的可恢复性。记忆目标构成近乎饱和的参考案例；经检索介导的泄露通常是不完整的；而经工具介导的泄露和实时智能体中的泄露则会随观测面、提示词与通道的对齐程度、检索深度以及提供方行为而发生显著变化。分层语义审计进一步识别出对攻击者有用、但规范精确匹配无法发现的披露信息。因此，CIPL 提供了一个通用框架，用于比较在异构智能体流水线中，内部对敏感信息的依赖如何转化为外部可恢复的泄露。
Background: Just-in-time adaptive interventions (JITAIs) can use behavioral data to adapt support to changing contexts, but many rely on predefined rules and manual configuration. Objective: We developed a proof-of-concept sleep JITAI using an AI agent to review personal data, evaluate reminders, adapt interventions, and record decisions for human review. Methods: Running in Home Assistant on a configurable schedule, the agent follows a reusable skill file to review 30 days of sleep and behavioral data, including physical activity, smartphone use, and bedtime routines, to identify patterns and create or update automated reminders. Results: Initial runs demonstrated technical feasibility, successfully completing data review and intervention decisions while limiting reminders to three per day and saving decision records. Conclusions: Agentic AI may enable flexible, adaptive sleep JITAIs. The architecture supports future comparison with fixed or rulebased interventions, requires human oversight, and could extend to other health behaviors.
LLMs are increasingly used in social simulations for socially interactive agents and robots, offering more flexibility than rule-based systems. However, even though they mimic human behaviour very well, there is a persistent alienness to them. This work investigates whether personality-aware fine-tuning can reduce this gap by improving the consistency and controllability of personality-conditioned dialogue generation compared with instruction prompting alone. We fine-tune two small open-weight LLMs, Qwen2.5-7B-Instruct and Ministral-8B-Instruct, using a corpus that combines personality-labelled social media posts and dialogues to create a personality-based dialogue engine for social simulation. The resulting models are evaluated across multiple social interaction scenarios using three independent LLM judges, which assess personality fidelity and provide evidence-based behavioral interpretations. We additionally quantify inter-rater agreement and lexical characteristics of the generated dialogue. Results indicate that fine-tuned models are not better at role-playing different personalities than their respective baseline models. However, low inter-rater agreement limits the confidence with which these results can be interpreted. Concerning the quality of generated texts, fine-tuned models are mostly comparable to the baselines, with fine-tuning improving the linguistic diversity of the Qwen models. While the results appear generally usable and the baseline models offer the best overall performance, future studies should place greater emphasis on the quality and domain alignment of training data for accurate personality role-playing.
文章 URL： https://www.theregister.com/devops/2026/09/18/microsoft-agentically-ports-copilot-runtime-to-rust-for-120k/5297549 评论 URL： https://news.ycombinator.com/item?id=49773998 得分：15 评论数：6
hi to all the readers this post is for my recent opensource project called ENZO now answering what is enzo so enzo is an opensource platform where i clubbed all the free available api for anyone use under one hood with more than 2000 models available to use for chatting coding researching and much more now answering the most common question of why you should put your time looking the project so it has few distinct feature meaning it has a dedicated agents tab where you can describe your need and create a special agent just for one specific task with master ability in that domain second it has the ability to connect your gmail drive and calendar and then you can ask it to perform some specific tasks like reading you the most important mail of the day or finding recruiter mails and creating personalized reply based on your data which it stores locally on your device third the coding mode offers a dedicated preview window where you can see your code running and have a look of it feels and edit it in realtime as well as all the modes are packed with dedicated skills which delivers promising results fourth the ui features some additional things such as music tab where you can listen to any music want and it has a custom personalized feature which runs in background and an llm understands your taste and recommends similar kind of music you like fifth the most important why your trust it with your api key then to explain i would say enzo a dedicated vault which manages all your api and to secure it the vault as aes 256 bit encryption which prevents any person or any middle man to look at your api key and since the whole program runs locally on your device you have complete freedom to oversee all the backend work happening and it also features password lock which if you enable saves a backup key and then locks your whole platform work behind a pass screen though it is not foolproof as any third party or malware containing extension can still fetch login tokens from your browser so its security also depends upon how you access it concluding all of it. i urge to anyone who reads this to have a look at the platform even if you hate it just curse it in the comment its fine or if you would like to drop any feedback i would highly encourage that and since its my first work open source platform i know it has a lot of errors and bugs so i apologize upfront for it and if you consider my work worthy please drop a star on the repo that'll make my day Comments URL: https://news.ycombinator.com/item?id=49771118 Points: 5 # Comments: 1
Hello HN！我们是来自 Cua 的 Dillon 和 Francesco。 我们一直在想，有多少计算机使用任务实际上需要一个完整的通用 LLM（例如 gpt-6-astra、claude-opus-5 等）来思考它们的所有决策和步骤。有些任务需要思考计划、探索不同路径、从失败中恢复。另一些任务则是做出局部决策的问题，比如这个值应该填进这个框，或者我是否应该勾选这个复选框，或者这个元素应该被忽略。 我们想知道，如果训练一个小模型只做这类决策，能走多远。 我们的灵感来自 Typesafe 的 Jev 及其 System One Model 的框架。这是在致敬 Daniel Kahneman 所描述的二分法：快速、自动、直觉式思考（系统 1）与缓慢、分析式思考（系统 2）。 对我们来说，有趣的问题是：如果给模型一个包含当前上下文的接口，以及一组可能的选择，并要求它为每个选择返回一个概率，会发生什么？这类模型不像大多数 LLM 那样逐 token 生成输出，而是对你给它的选项打分；你可以检查、信任并使用这些分数来驱动你的应用行为。 CUA-S1 是我们面向计算机使用场景中狭窄、专用决策模型的答案。我们的第一个发布版本是 CUA-S1-FORMS。我们基于 jevlike 中的想法和代码构建了它，然后训练了第二个模型，专门处理表单交互。它有 706k 个参数，原始 checkpoint 为 2.8 MB。 第一次训练迭代在合成数据上耗时不到 30 分钟。给定一组从文档中提取的结构化元素和值，它会为每个元素预测是否使用给定值、CHECK、CLICK 或 SKIP。它不会为文本字段预测新值，也不会考虑截图。元素决策会一起打分，你的代码可以对操作进行排序，而 Cua Driver 会逐一执行它们。 该专用模型与托管版 Jev 在我们的表单任务上的首次评估结果： - 对整个决策集：99.7% 正确 vs 83.6%。 - 对需要执行操作的步骤子集：100% 正确 vs 96%。 - 对只是保持已填写字段不变的步骤子集：100% 正确 vs 74%。 该专用模型是专门针对这个任务和约定（对已经填好的框直接按 skip）训练的，而托管版 Jev 并未为此进行微调，因此这是一次关于限定范围专业化的实验。 我们测得在本地为一个表单打分需要 7-9 ms，而调用托管版 Jev 每次需要 260-280 ms（包括网络延迟），不过这些样本测量的是不同内容，并不是端到端的表单完成时间。 我们在这里感兴趣的是介于脆弱脚本和通用 agent 循环之间的空间。表单字段的内容和布局变化足够大，以至于脚本会变得难以维护，但可用决策集合可以保持狭窄且边界清晰。我们想探索这样一种可能性：通用 agent 遇到新颖情况，并把已被充分理解的决策交给这样的专用模型处理。 这是我们正在研究的方向。当前发布版本仅用于表单。我们已将合成数据生成、训练、评估和 Driver 集成以 MIT 许可证开源在 libs/cua-s1 下。 欢迎评论！尤其是如果你正在构建计算机使用 agent，并且遇到过某种反复出现的决策：它太多变以至于难以脚本化，但又太狭窄以至于不值得调用另一个 LLM。 评论 URL： https://news.ycombinator.com/item?id=49767564 点数：4 评论数：0
GPT-6 Astra helps Hex’s data agents turn answers into interactive visualizations that employees are proud to share.
文章 URL： https://code.claude.com/docs/en/changelog 评论 URL： https://news.ycombinator.com/item?id=49760187 分数：29 评论数：7
文章 URL： https://arxiv.org/abs/2609.20804 评论 URL： https://news.ycombinator.com/item?id=49753878 点数：3 评论数：0
随着编码智能体从受监督的代码补全转向无人值守、全天候探索，其工作也从孤立的预测扩展为包含推理、工具使用和反馈的长轨迹。因此，token 效率对于扩展递归式自我改进变得十分重要。我们在运行框架层面采用受 RSI 启发的方法，在数量不断增加且日益多样化的环境中扩展自动研究循环，以进行框架滚动执行。在这一规模下，该过程能够产生可复用的改进，并迁移到其开发环境之外，从而推动自动化框架发现走向生产级成果。四种机制经筛选后保留下来，共同构成 SoL-Pi，涵盖动作执行、上下文压缩、观测处理和委托式阅读。在包含 51 项任务的 EdgeBench 评测中，SoL-Pi 在 GPT-5.6 Sol 和 Opus 5 上取得了与 Pi 相当的性能，同时将记录的 token 流量降低 44.7%–49.0%，并将 API 成本降低约三分之一。换言之，与 Codex 和 Claude Code 的原生框架相比，预计每小时可节省 \$8.75–\$13.50；与 Pi 相比，每小时可节省 \$4.36–\$5.71。
大型语言模型的响应具有非确定性，因此 LLM 智能体中的故障很难复现：故障取决于无法按位复现的推理过程、读取不断变化状态的工具，以及重新运行时很少能够重复的多步轨迹。记录与重放可以使一次运行变得可复现，但现有的智能体工具只记录运行过程以便追踪或评分，而不是用于针对代码变更进行测试。我们提出了 Chronicle：它将智能体运行过程在其非确定性边界处记录为不可变封装，并根据记录进行重放。其核心操作“切点重放”会从记录中提供所选边界子集的结果，并使用新代码实时执行互补的边界子集，从而将一次已记录的事故转化为可在持续集成中运行的回归测试。在一个包含 6 个已记录故障、并对模型边界进行模拟的基准测试中，记录操作在每次跨越边界时增加 23 μs 的开销（相当于假设模型调用耗时 300 ms 的 0.008%）；完整重放不会发起任何模型调用，并且在 20 次重复运行中保持位级稳定；对于全部 6 个事故，切点测试都能在存在缺陷的代码上失败，并在受到保护的变更和无害变更上通过。在一项针对受到保护工具的变异研究中，切点测试能够捕获每一个会放行已记录不安全操作的变异体，而使用相同断言、将所有边界全部替换为存根的基线测试一个也捕获不到。Chronicle 及其基准测试已公开发布于 https://github.com/theagentplane/chronicle】【。
智能体轨迹记录了智能体做了什么以及接下来发生了什么。然而，标准的监督微调（SFT）只对智能体生成的动作 token 施加损失，将环境观察作为上下文，而不是作为预测目标。我们探讨这种惯例是否为后续强化学习提供了最佳初始化。我们提出 ActObs，它也对每条轨迹中已有的观察 token 进行监督。尽管部署后的智能体从不生成观察，学习预测观察会鼓励策略建模动作后果，同时不增加数据、参数、序列 token 或前向传播次数。这些方法在 SFT 后表现相近，但在 GRPO 后出现分化。在 Qwen3-4B 上，从 ActObs 出发进行 GRPO，在 Terminal-Bench 2.0 的每个评估采样预算下，相比仅动作的对应方法都取得了更高的 pass@k。在 Qwen3-8B 上，它以牺牲部分 pass@1 可靠性为代价，换取了更高的 pass@k（pass@16 提高 3.4 个百分点），并解决了更多不同任务。这一优势还扩展到 aider-polyglot 上的跨领域代码编辑（4B 上 pass@1 提高 4.2 个百分点），其任务在 SFT 和 RL 期间均未出现。ActObs 在 RL 期间保留了更多熵，同时需要更少的策略移动，使最终策略更接近其 SFT 初始化。我们的分析将这种差异追溯到 SFT：动作梯度和观察梯度很快变得正交，而仅动作训练会留下较大的残余观察梯度，并使环境预测能力下降到低于基础模型。联合监督避免了这种单侧专门化，保留了后果预测能力，并为策略的下游探索做好准备。
企业客户支持中的高效故障排除智能体依赖于从相似的历史案例中检索可执行的指导，但现有的检索增强生成（RAG）系统将支持案例视为静态文档，忽略了其多阶段、有状态的特性。我们提出 RAFT（面向故障排除智能体的检索增强框架），这是一种有状态的 RAG 框架，将每个已关闭的历史案例抽象为由时间线条目组成的有向链，并在条目级别进行检索，从而找出其中间状态与当前案例相匹配的案例，同时返回以匹配状态为锚点的父案例轨迹；此外，还可选地通过可配置的相似性表示在案例之间建立案例级图。我们直接评估这一检索层；与评估完整智能体系统不同，这种方式无需进行生产部署。由于公开的多阶段故障排除数据极为稀缺，我们将基于 Microsoft Learn Windows Server 文档构建的合成基准，与带有人为创建的重复标签的真实 Apache Jira 问题结合起来进行评估。在案例进展的每个阶段，RAFT 的 Case Hit 均优于 vanilla RAG 和 GraphRAG 基线，并且相较于最强基线取得了具有统计显著性的提升；Jira 结果提供了方向性证据，表明这一优势能够迁移到真实案例历史中。我们发布了基准、实现代码以及 Apache Jira 评估集。
Generative agents are increasingly used to select and narrate video highlights, but they typically operate over unstructured or frame-level representations. Their output is consequently difficult for a viewer to verify and steer toward individual preferences. We present the semantic action graph, a lightweight domain schema that represents a sports match as performer, action, recipient, moment, and state nodes connected by role, temporal, and outcome edges. The schema demonstrates three key properties: 1) connected event sequences, 2) a shared, closed vocabulary, and 3) frame-addressable moments, making it suitable to serve two consumers at once: an agentic pipeline that composes narrated highlights, and a visual interface through which viewers query and inspect the same structure. We instantiate it in SportSAGE, a design probe pairing a four-module highlight pipeline with a graph interface, and report feedback from 12 soccer fans. Participants were satisfied with the quality of the generated highlights and narratives, and used the graph interface to search, navigate, and interpret the match highlights. These results provide early evidence that one small, human-readable schema can ground agent generation and support human interpretation at the same time.
通过强化学习（RL）训练的多轮智能体每条轨迹只能获得一个标量奖励，这促使人们采用自同策略蒸馏（OPD），由具备特权任务技能的自教师提供密集的令牌级监督，使无技能学生能够将这些技能内化。然而，智能体任务中的两项发现削弱了这一方法的效果：特权信息本身并不总能使教师变得可靠，而且教师监督的收益取决于训练阶段。为此，我们提出 RetireOPD（自退休同策略蒸馏）：首先利用环境奖励优化一个解耦的技能条件化教师，然后将无技能学生与 RL 和 OPD 联合训练。RetireOPD 不采用预先设定的蒸馏计划，而是引入自适应退休（Adaptive Retirement）：当学生与教师之间的差异不再缩小，且学生达到教师成功率的目标比例后，学生会自行放弃教师，此后训练仅依靠 RL。在从 1.5B 到 7B 的 Qwen2.5 模型上，RetireOPD 相比 RL 基线将 ALFWorld 成功率提升了 14.1% 至 18.8%，将 WebShop 准确率提升了 11.8% 至 19.0%，并且在所有设置下都超过了其自身的技能条件化教师。
编码 harness 影响着自主编码智能体如何将模型能力转化为长周期软件工程性能，但现有工作通常将 harness 作为单体系统进行评估，使得各个组件的有效性尚不清楚。为支持组件级比较，我们使用一个轻量级编码 harness 研究这一问题，其执行循环保持固定，同时改变三个组件：规划、动作空间和上下文管理。我们在 SWE-Bench Verified 和 Terminal-Bench 2.1 上评估了四个模型，共评估 176 组匹配设置，涵盖五种上下文管理策略、四种上下文窗口预算，以及针对规划和动作空间的定向消融。我们发现：（1）随着上下文窗口预算收紧，上下文管理变得越来越有价值，其大部分收益来自防止上下文溢出失败。（2）在 LLM 基于总结之前先进行基于规则的省略，在上下文管理策略中提供了最强的整体效率；而使被省略内容可恢复会增加机制复杂度，但模型很少使用，且不会带来准确率提升。（3）规划从较弱模型的准确率支架，转变为较强模型的成本节省手段，而准确率几乎没有变化。（4）预定义工具能够提升 bash 熟练度较弱的模型的性能，而具备 bash 能力的模型可以仅通过 bash 接口有效运行，并实现显著更低的成本，尤其是在以命令行为中心的任务上。轨迹级分析解释了这些效果：上下文管理延长了执行轨迹，而不会显著改变智能体行为；规划改变了轨迹停止的位置；动作空间改变了代码编写的粒度。这些发现为面向模型和预算的 harness 设计提供了参考，并为评估未来 harness 组件提供了模块化框架。
前沿编码智能体越来越多地被信任，能够长时间自主工作，但智能体的最终响应往往是用户唯一能看到的关于这项工作的记录。我们量化了前沿智能体对任务完成情况进行\emph{过度宣称}的倾向，这种错误表述可能误导用户。当智能体的最终响应与其上下文中的信息相矛盾时，我们称其存在过度宣称。这一定义不要求推断智能体的意图，并且独立于任务是否成功。我们推出了\emph{OverclaimBench}，这是一套评测方案，由五种文件审查场景、基于交互记录的覆盖率测量以及已登记的植入缺陷组成。我们在各自的生产命令行界面中评估了八个专有前沿模型，并在 OverclaimBench 上使用单一固定测试框架评估了四个开放权重模型，结果发现：1）在 67.9\% 的运行中，智能体没有读取要求其审查的全部文件；2）在未读取全部文件的运行中，智能体有 80.4\% 的概率具有\emph{误导性}（各模型分别为 59--96\%），具体表现为虚假声称已读取全部文件，或未说明覆盖范围并不完整；3）要求将任务委派给子智能体提高了文件读取覆盖率，但在仍未完成审查的情况下，其中绝大多数仍然具有误导性；4）虚假声称完成了完整审查的智能体漏掉植入缺陷的概率约为读取每个文件的智能体的 1.8 倍，这表明完成声明可能掩盖实质性失败。总体而言，这些结果表明，智能体的最终响应并不能可靠地记录其实际行为。
编码智能体已成为机器人操作中一种很有前景的范式：语言模型将机器人控制器编写为程序，而以这种方式构建的智能体如今无需针对机器人进行专门训练即可操作机器人。然而，这一范式是否同样安全，尚未得到探讨。我们在安全约束下评估编码智能体，其中每个任务都将一个操作目标与一个机器人不得触碰的障碍物配对。该智能体会追求目标，但在大多数情况下会与障碍物发生碰撞，把任务完成视为唯一目标而忽视安全。该智能体在其推理轨迹中会考虑障碍物，提示词也已明确禁止触碰障碍物，因此问题既不在感知，也不在指令；问题在于规划，即所声明的约束从未成为优先事项。通过将操作分解为路径阶段和富接触时刻，我们定位了失败的来源。在路径阶段，模型无法优先考虑安全约束，因为它没有避障路径的概念，也不会在已选路径变得不可行时重新规划。在接触阶段，它并未意识到接触执行同样受该约束限制。为弥合这一差距，我们提出 SafeHarness，它为模型配备两个障碍物感知框架，使其能够优先考虑安全约束。障碍物感知路径规划将物体建模为边界框，并在其上以航点序列的形式绘制候选路径。随后，智能体会预先规划路径，对其进行验证，在必要时重新规划，然后才执行。障碍物感知接触执行则会选择接触位置，使接触本身避开障碍物。SafeHarness 实现了 71.9% 的任务成功率和 87.5% 的避碰率，分别比先前的 SOTA 高出 6.5% 和 27.0%。这些结果分别是未使用框架的同一智能体的 $2.3\times$ 和 $1.5\times$。
The growing complexity of multi-domain operational environments (land, aerospace, naval, cyber, and electromagnetic spectrum) has increased the volume and velocity of data reaching command-and-control (C2) centers, straining the observe-orient-decide-act (OODA) decision cycle. Artificial Intelligence (AI) systems currently employed in defense are, in general, reactive and isolated tools that still rely heavily on human operators to integrate information, assess scenarios, and formulate courses of action. This paper proposes a conceptual Agentic AI architecture for AI systems that can plan, access data sources, execute tools, and act autonomously and audibly, aimed at supporting decision-making across the three Brazilian Armed Forces (Navy, Army, and Air Force). Four application fronts are discussed (decision support, situational analysis, feasibility studies, and countermeasure suggestion), as well as the data and sensor access requirements and the security and permission safeguards necessary for responsible employment across administrative, strategic, operational, and tactical contexts.
自我演化方法使使用工具的智能体能够生成自己的训练数据，从而减少对人工标注轨迹的需求。然而，现有方法通常将轨迹生成与评估分离，依赖无法适应新出现的失败模式的静态验证器，或依赖可能强化多个轨迹中共同错误的自洽性信号。联合调整规划、执行和评估提供了一种有前景的替代方案，但也带来了一个根本性的协调挑战：每个组件都在持续改变用于训练其他组件的数据或反馈。我们通过 \textbf{UnifiedPlayers} 应对这一挑战。这是一个协作框架，由负责生成任务的规划玩家、通过 Python 工具调用生成多轮轨迹的执行玩家，以及构建可执行验证器的评估玩家组成。我们设计了角色特定的奖励，在 GRPO 下协调三个玩家，使其共同实现一个学习目标。在两种模型骨干和十二个推理基准上，UnifiedPlayers 相比最强的既有基线，在数学推理任务上至少提升 3.5%，在通用推理任务上至少提升 3.9%。此外，学习得到的验证器达到了 84.2% 的对抗性检测准确率，其奖励信号的单题方差是自洽性基线的 2.03 倍，从而提供了更具区分性的验证结果。这些结果表明，专门化玩家之间的协作是实现自我增强型工具集成智能体的一条有前景的路径。
智能体互联网有望使大量自主智能体能够跨异构平台相互发现、验证并协作。然而，当前的智能体协议主要解决工具调用和智能体间通信问题，对于可扩展的智能体注册、可信身份识别以及面向能力的发现，仍基本未得到解决。为此，本文提出了一种面向智能体互联网的可扩展信任发现架构。该架构采用由三层组成的分层分布式设计：Agent Root 负责可信注册表治理，Agent Registry 负责智能体注册和元数据发布，Agent Resolver 负责分布式能力发现与面向信任的解析。该架构进一步引入了一种以注册表后缀为锚定的复合身份方案，将智能体原生标识符与可信注册表后缀绑定，以生成可被全球发现的身份。同时，该架构还融合了双证书和多级身份认证机制，以增强智能体之间的身份信任。我们实现了一个原型系统，并通过大规模智能体注册与解析实验对其进行了评估。该原型的平均注册延迟为 58ms，平均发现延迟为 25ms，并且支持每秒超过 19,000 个注册请求和每秒超过 29,000 个智能体发现请求。这些结果证明了所提架构的可行性，为构建智能体互联网中具备可扩展性和身份可信性的智能体生态系统提供了一种切实可行的方法。
通常，大多数语音代理都是级联系统：ASR 模型将来电者的音频转写为文本，语言模型读取转写文本，决定要说什么以及调用哪些后端工具，TTS 模型则将回复转换为语音。几乎所有决策都由语言模型完成，但现有评测要么过于宽泛，要么过于狭窄。端到端语音基准会对完整流程进行评分，因此识别错误和模型错误会混合成一个分数。LLM 基准能够隔离模型，但无法评估真实电话通话中的困难因素，例如转写问题、来电者的语音被拆分到多条消息中，以及回复必须遵循指定语言和话术脚本的要求。我们提出了 Multi-Turn Voice Agent Benchmark（MTVA-Bench），在级联系统实际面临的相同条件下评估语言模型。来电者由遵循一组评分标准的 LLM 模拟，工具调用则由一个模拟后端应答，该后端会根据模型实际发送的参数作出响应。该基准包含 49 个代理，覆盖 490 个经过审核的场景，并支持 7 种语言。评分由工具调用的确定性检查和两名 LLM 评审共同构成：一名评审针对特定场景的规则进行评分，另一名评审在无法获知任务内容的情况下评估对话质量。两名评审都必须引用对话转录中的具体消息。任务得分和对话得分的权重相同，因为一次通话可能完成任务，却仍让来电者感受很差。在一项包含七个模型的研究中，其中六个模型选择正确工具时的得分彼此相差不超过 6.4 分，但它们的总体得分跨度达到 24.4 分。大部分差距来自参数值、操作顺序、规则遵循情况，以及模型在工具调用前后所说的内容。
LEO卫星网络具有拓扑动态变化、链路随时间变化以及服务需求多样化等特点，这使得传统路由方案难以支持细粒度的服务质量（QoS）保障。现有研究主要针对具有预定义目标的网络状态进行路由优化，但很少关注如何将非结构化的自然语言服务请求转化为自适应路由决策这一实际挑战。为弥补这一空白，我们提出了STR-Agent，一种用于LEO卫星网络QoS感知路由的LLM驱动框架。STR-Agent的核心创新在于，在单一智能体架构中统一了意图感知、基于工具的执行、经验积累以及基于反思的策略自适应。具体而言，感知模块将自然语言请求转换为结构化路由语义；反思模块则根据实时拥塞状况和历史路由结果动态调整服务到路由策略的映射，而非依赖固定的路由目标。此外，我们开发了专用的感知模型，并构建了面向LEO服务理解的领域专用监督微调数据集。在Walker-Delta星座中的仿真结果表明，STR-Agent显著优于传统基线方法：与DQ-Dijkstra相比，其端到端时延最多降低60%；经过监督微调后，平均意图理解准确率由45.4%提升至92.45%；在600 Mbps速率下，反思模块还可进一步降低120 ms的时延。这些结果表明，LLM驱动的智能体架构有望在未来LEO卫星网络中实现面向服务的自适应QoS路由。
Spatial reasoning abilities correlate strongly with performance in STEM fields. Games offer a compelling medium for training these critical skills in developing children who have a natural proclivity for play. However, to facilitate human-like tutoring and player guidance, these games require an AI agent capable of making commonsense inferences from spatial events. Qualitative reasoning (QR) models appear to be a suitable framework for these application domains. As these models reason in symbolic representations, they can seamlessly translate game states into interpretable feedback for human-like player guidance. This paper introduces a hybrid qualitative model designed for Camelot Jr., a block-puzzle game that requires constructing multi-level bridges to connect two avatars stationed on separate towers. The game poses a challenge for the player, who must make platforms stable, plan their path, and ensure they use all the provided blocks. To handle the precise physics required by the domain, we integrate a mathematical center-of-mass stability logic to guide our qualitative solver. Our work facilitates spatial skill training in Camelot Jr. and contributes to the development of human-centric, explainable game-playing agents.
hey, i'm nick (@skeptrune on X). i have been using ai for a lot of things i'm not an expert in and thought it would be nice to try and buy paid skills for helping it accomplish those tasks more efficiently. for example, redlining contracts, creating ai generated videos, different website designs, and more. curious to see if this resonates with folks here. i figure a more engineering'y audience that's ai coding-agent forward would have similar problems to what inspired me to make this Comments URL: https://news.ycombinator.com/item?id=49743459 Points: 10 # Comments: 4
Hey HN, we're Nars & Nishant, founders of Skillsync ( https://skillsync.com ) Skillsync lets you move your AI chats across every coding agent. Most of our work exists as conversations, which are currently scattered across our agents. Though stored locally, these conversations use different formats. This is annoying because you cannot simply switch between agents without starting over. We get locked into a single provider and their agent as we invest in skills and memories over time. Skillsync acts as a universal converter. It moves the entire session, including all the messages, reasoning and tool calls so you can pick up right where you left off. Skillsync collects all your sessions in one place and makes them searchable. It breaks down what each session is carrying, including which loaded skills the agent is actually using, making stale context easy to spot. You can also create shared workspaces to sync sessions across your team. You can build your own closed loop systems. Everything runs locally except when you share to workspaces. The core is an open-source Rust engine called txcript ( https://github.com/skillsynchq/txcript ). It translates a session from one agent's on-disk format into another's, mapping conversation, reasoning, and tool history. Think ffmpeg or pandoc, but for agent sessions. On top of that engine is a local-first desktop app. Your agent sessions are normally scattered across different tools' folders in formats you'd never read by hand; the app surfaces them in one place with a UI that makes them actually readable, the conversation, the reasoning, and the tool calls, so you can revisit what happened, move a session into another agent, or share it with a teammate. Skills and memory are stored as portable, human-readable markdown you own, and exposed to any agent over MCP for search and selective retrieval. Sessions and translation run locally on your machine. The one thing that leaves is what you explicitly share into a team workspace. It’s been surprising to us to see how locked-in people already are without realizing it. You don't notice you're trapped in one agent or harness until you try to leave, and by then you've got months of sessions stuck in a format only that tool can read. The lock-in is invisible right up until it's expensive. Once sessions are portable, they stop being disposable logs and become something you can actually build on, spotting patterns in how you work, handing a session off to someone else, letting a non-technical teammate pick up where an engineer left off. The session turns out to be the unit of collaboration, and right now it's being thrown away. Before Skillsync, Nars and I built an open-source payment orchestrator that let merchants route across many processors instead of getting locked into one (30k+ GitHub stars). Fighting vendor lock-in by making incompatible systems interoperate was the whole job. We came into YC with a different idea, but the more we lived inside coding agents the more we saw the same problem from the other side: your context is locked into whatever agent you start with, because every agent stores sessions differently and none of it moves. Some of the things we’re seeing people do with Skillsync are: - Moving a session between Claude Code, Codex, and Cursor mid-task, including when you hit a usage limit on one and want to keep going on another. - Thinking through a problem in a browser Claude chat, then handing the whole thread to Claude Code or Codex to build. - Sharing research sessions with a team. Everyone uploads their sessions, so teammates and their agents can build on each other's work instead of repeating it. It's available as a Mac app, CLI and MCP. Here's the demo: https://youtu.be/7hVhSnSKGl8 . Would love to hear your feedback and answer questions! Comments URL: https://news.ycombinator.com/item?id=49743049 Points: 10 # Comments: 2
Simple discovery of the provider's default schemas is possible without any credentials (since they are built into the binary). If you want it to fetch customizations of your specific instance, give it credentials to your org. The repo has more details. https://github.com/agent-cli-framework/aclif Comments URL: https://news.ycombinator.com/item?id=49743382 Points: 6 # Comments: 1
I wanted to manage long horizon agentic workstreams via voice, then put my phone down, and have a harness manage completion - extending into full computer use. I was trying to build a personal Jarvis, so I benchmarked AutoBot to see how close I could get: - OSWorld: 32.41% (moved Sol Max from 4th to 1st, beating Opus 5) - AssistantBench: 50.70% Hermes and OpenClaw, but without needing a weekend and VMs. And with the ability to manage deep personalization AND keep strict privacy rules, storing data on encrypted disk. AutoBot is a passion project that grew out of trying to make this all work for myself. I’m sharing it here because I suspect other here have the same frustration. And I’m curious what else everyone is doing for this. It’s an MIT-licensed harness that lives within a project. Native voice lets me discuss tasks, check progress, and steer work; a local ledger tracks unfinished outputs and the evidence needed to call them done. Memory drives more autonomy over time, and “defrags and locks in learning nightly” while a heartbeat system persists execution. I’m not selling anything. If you’re building something similar for yourself, I’d love to compare notes. GitHub: https://github.com/demeyer1/Autobot Comments URL: https://news.ycombinator.com/item?id=49743478 Points: 4 # Comments: 0
Hey HN, we're Nars & Nishant, founders of Skillsync ( https://skillsync.com ) Skillsync lets you move your AI chats across every coding agent. Most of our work exists as conversations, which are currently scattered across our agents. Though stored locally, these conversations use different formats. This is annoying because you cannot simply switch between agents without starting over. We get locked into a single provider and their agent as we invest in skills and memories over time. Skillsync acts as a universal converter. It moves the entire session, including all the messages, reasoning and tool calls so you can pick up right where you left off. Skillsync collects all your sessions in one place and makes them searchable. It breaks down what each session is carrying, including which loaded skills the agent is actually using, making stale context easy to spot. You can also create shared workspaces to sync sessions across your team. You can build your own closed loop systems. Everything runs locally except when you share to workspaces. The core is an open-source Rust engine called txcript ( https://github.com/skillsynchq/txcript ). It translates a session from one agent's on-disk format into another's, mapping conversation, reasoning, and tool history. Think ffmpeg or pandoc, but for agent sessions. On top of that engine is a local-first desktop app. Your agent sessions are normally scattered across different tools' folders in formats you'd never read by hand; the app surfaces them in one place with a UI that makes them actually readable, the conversation, the reasoning, and the tool calls, so you can revisit what happened, move a session into another agent, or share it with a teammate. Skills and memory are stored as portable, human-readable markdown you own, and exposed to any agent over MCP for search and selective retrieval. Sessions and translation run locally on your machine. The one thing that leaves is what you explicitly share into a team workspace. It’s been surprising to us to see how locked-in people already are without realizing it. You don't notice you're trapped in one agent or harness until you try to leave, and by then you've got months of sessions stuck in a format only that tool can read. The lock-in is invisible right up until it's expensive. Once sessions are portable, they stop being disposable logs and become something you can actually build on, spotting patterns in how you work, handing a session off to someone else, letting a non-technical teammate pick up where an engineer left off. The session turns out to be the unit of collaboration, and right now it's being thrown away. Before Skillsync, Nars and I built an open-source payment orchestrator that let merchants route across many processors instead of getting locked into one (30k+ GitHub stars). Fighting vendor lock-in by making incompatible systems interoperate was the whole job. We came into YC with a different idea, but the more we lived inside coding agents the more we saw the same problem from the other side: your context is locked into whatever agent you start with, because every agent stores sessions differently and none of it moves. Some of the things we’re seeing people do with Skillsync are: - Moving a session between Claude Code, Codex, and Cursor mid-task, including when you hit a usage limit on one and want to keep going on another. - Thinking through a problem in a browser Claude chat, then handing the whole thread to Claude Code or Codex to build. - Sharing research sessions with a team. Everyone uploads their sessions, so teammates and their agents can build on each other's work instead of repeating it. It's available as a Mac app, CLI and MCP. Here's the demo: https://youtu.be/7hVhSnSKGl8 . Would love to hear your feedback and answer questions! Comments URL: https://news.ycombinator.com/item?id=49743049 Points: 4 # Comments: 0
I kept seeing engineers share what they were building with AI; however, I was always more curious about how they worked. Which agents did they use? What skills and tools had stuck or been thrown out the window? How did they manage longer-running tasks? So I built this with the hope we could have a dedicated space to share and be open about our setups. Comments URL: https://news.ycombinator.com/item?id=49740105 Points: 16 # Comments: 6
Despite recent advances in large language models (LLMs), performing logically consistent deductive reasoning over extended interactions remains challenging. Tasks that require integrating evidence across multiple reasoning steps, maintaining consistency with prior inferences, and updating beliefs under new constraints can surface limitations in current models while providing a useful testbed for evaluating reasoning enhancements. In this paper, we implement a text-based, multi-agent version of the classic board game Clue as an environment to evaluate multi-step, agentic deductive reasoning. In this setting, agents must infer hidden information from a sequence of observations, maintain consistency across turns, and reason over an evolving set of logical constraints. We instantiate six LLM-based agents (GPT-4o-mini and Gemini-2.5-Flash) as players that engage in turn-based gameplay; using three agents per model family, we establish baseline performance across repeated games. We then introduce a tool-augmented approach in which a structured possibility matrix converts implicit game state from generated reasoning logs into an explicit representation of remaining possibilities. The possibility matrix encodes extended-turn memory and deductive constraints, offloading these tasks from the agent. We compare this approach against the baseline to evaluate how tool augmentation supports reasoning quality and task success for autonomous agents in a strategic reasoning environment.
Current Mixture-of-Agents (MoA) paradigms generally treat query routing and agent fine-tuning as separate processes, limiting their ability to respond to evolving agent capabilities. This disconnect prevents routing strategies from adapting to evolving agent capabilities during post-training and prevents agents from achieving synergistic data-driven specialization. To resolve this, we introduce CERA-MoA (Co-Evolving Router with continually learning Agents for Mixture-of-Agents), an iterative reinforcement learning framework where the dynamic router and independent agent policies co-evolve. We design a predictive familiarity estimator that leverages mid-layer hidden states to evaluate semantic competence among agents, avoiding the overhead of full rollouts. Based on these familiarity scores, a cumulative-threshold adaptive routing mechanism dynamically activates a tailored minimal agent subset, achieving a trade-off between task performance and efficiency. By proactively allocating targeted training samples to agents based on their evolving competence, CERA-MoA promotes capability differentiation. Extensive experiments across various domains demonstrate that CERA-MoA outperforms state-of-the-art static-agent routing and fix-workflow fine-tuning baselines.
Coding agents are typically evaluated with desired behavior specified through issues or instructions. In practical web development, however, agents may need to infer behavior from working software and implement it in an incomplete application. We introduce ProgramDistill, a benchmark evaluating coding agents on features discovered through interaction with fully functional reference applications. We build ProgramDistill by factorizing applications into features of different granularities, each associated with replayable behaviors executable via its gold patch. Our pipeline, mine-craft-patch, discovers 1,975 replay-verified behaviors across 26 applications and constructs 4,063 tasks without human intervention. Across nine frontier coding agents, GPT-6 Astra and Claude Opus 5 achieve 49.2% and 28.8% success on cumulative workflows in full-application reconstruction. In partial-application reconstruction, success falls from 100% to 64.0% and from 96% to 32% as restoration depth increases from 1 to 8. ProgramDistill thus provides a scalable benchmark with controlled difficulty for evaluating and diagnosing coding agents, and a natural basis for future curriculum-based training.
Agentic workflows now make consequential decisions in regulated settings, and the governance placed around them is almost entirely step-scoped: input-output classifiers, per turn rails, and span-level evaluators. The policies organizations actually hold, such as referral thresholds, authority limits, and review requirements, are properties of the whole execution rather than of any one step. This mismatch admits a failure mode we call a Compositional Policy Violation (CPV): every individual step passes its own check while the composed execution violates the governing policy. A predicate over a single step cannot evaluate a property that step does not determine, so no improvement in the accuracy of the step-scoped monitors detects this class. We define CPVs as the failure of step-level compliance to compose, and present a taxonomy of four types: Authority Creep, Threshold Laundering, Cumulative Sum Violation, and Context Collapse. We show that the correct repair for each class is dictated by where the guarded quantity mutates. We then introduce a provenance-aware runtime architecture that evaluates policies over complete execution traces, recomputing guarded quantities from raw provenance rather than the pipeline's derived representation.
An agentic request spends substantial wall-clock time waiting for tools, and its KV cache holds GPU memory the whole time. Serving systems decide whether that cache stays, leaves, or comes back by guessing how long the tool will run, from the tool's name, its history, a duration declared before the call, or the engine's own occupancy. We show that no estimate fixed before a call starts can know its duration, and such estimates may not even rank the calls. Meanwhile, the running tool already holds the answer, but the agent stack together with the tool silences it. We propose that tool calls report their progress explicitly while they run, and we measure what that takes. A census of four public agent corpora finds a readable signal in most tool time once it is revealed, in two strengths: a fraction of the work remaining, or an accurate signal that the end is near. A harness recovers it without changing what the agent sees, at no measurable cost to the agent's benchmark score. At the points where a KV cache decision is made, the reported progress is between several times and an order of magnitude more accurate than the best published predictors, and it stays accurate when the environment changes. Plugged into a production engine through a few small hints, it cuts the p90 time to first token (TTFT) after a tool call by 20.7% (HBM only) and 20.8% (HBM + DRAM) against LRU, close to an oracle. A serving system should not guess what its tools can tell it.
The O-RAN control plane is becoming agentic: autonomous AI agents, deployed as rApps by different vendors, independently close control loops over shared radio resources. We demonstrate on a live O-RAN system that this independence is unsafe. Two agents with individually correct objectives, one protecting a latency SLA and one maximizing utilization for energy efficiency, jointly drive recurring opposing excursions of the shared resource partition that neither produces alone. Existing conflict-mitigation mechanisms presume a statically known application population and cannot govern agents whose behavior emerges at run time. We present AURA, a lightweight arbitration layer that admits agent actions only when they satisfy feasibility invariants, per-variable dwell times, and a deadband, and we prove the arbitrated system converges to a feasible operating point. Implemented on an OpenAirInterface (OAI) testbed with measured one-way latency and throughput, AURA reduces recurring shared-state excursions by more than an order of magnitude (from 8.4 to 0.4 PRB amplitude) and virtually eliminates cross-slice throughput starvation (from 40-55% to 0.3%), while leaving the protected slice's own latency compliance unchanged, a trade-off the convergence guarantee makes explicit.
Privacy evaluations of tool-using LLM agents often inspect a designated action, final response, or attacker report. These local proxies can miss unauthorized exposure elsewhere in a multi-step session and lack common ground truth across outlets, reports, and tool paths. We introduce privacy exposure displacement, the mismatch between a local evaluation proxy and target-grounded session exposure, and ASLEval, an authorization-aware framework that pre-registers a hidden target set, measures all declared visible exits, and reserves internal traces for diagnosis. Across multiple enterprise-style environments and independently implemented runtimes, we observe three recurring patterns. An expected-outlet-only view misses 46.9% of exposure recovered by the visible-exit union; attacker self-reports combine omissions with high false discovery; and schema-aligned internal evidence usually precedes visible exposure at the request/probe level. Reducing model-visible returns changes this path but can eliminate normal-task success. Independent human review supports the adjudication pipeline while identifying harder console and candidate cases. These findings motivate benchmarks that declare the complete visible boundary, ground claims in pre-specified targets and authorization, and report privacy together with task utility.
Agent benchmarks are substantially more costly to evaluate than conventional LLM benchmarks. Benchmark compression is therefore a natural solution, yet existing methods primarily model redundancy in task--model final-score distributions, which is important in agentic evaluation. To address this limitation, we analyze large-scale trajectories and identify six complementary process signals that are systematically associated with final agent performance. To disentangle agent performance redundancy from a complete perspective, we propose DualViewEval, an agent benchmark compression method that jointly exploits outcome and process relations to learn an exact-size miniset and predict the full-benchmark scores. Across five agent benchmarks and five representative baselines, DualViewEval achieves the best results in all datasets. With only 20 tasks, it achieves $24\times$--$40\times$ compression on APEX-Agents and BFCL, reducing mean absolute error (MAE) by $14.5\%$--$28.2\%$ over the strongest competitors while improving Kendall's $τ$ by up to $7.2\%$ relative to EssenceBench on SWE-bench Verified. The selected minisets further reveal capability differences among different agents, providing compact and diagnostic feedback for efficient agentic model development.
Mixture-of-Experts (MoE) language models suffer from large parameter counts, which create a significant memory bottleneck. Expert pruning is the most direct approach for reducing this parameter count, yet existing methods make pruning decisions for each expert independently, and assume experts' contributions are purely additive. In reality, expert usage in MoEs is inherently cooperative. We derive HOPE (Higher-Order Pruning of Experts), a second-order pruning objective which provably minimizes an upper bound on the error resulting from pruning. We show that REAP (a state-of-the-art first-order pruning method) is a special case of HOPE where interaction terms are ignored. Across three frontier MoE models (up to 122B parameters), two distinct calibration sets, and multiple benchmarks (including math, instruction following, coding, and an agentic suite), we demonstrate that HOPE produces better pruning decisions than existing methods, and its advantage is most pronounced at high pruning rates and on challenging agentic workloads. At 50% pruning, HOPE outperforms all baselines and achieves an average rank of 1.58 out of 5 methods (versus 2.42 for the next-best method, REAP), with gains of up to +6.1% on agentic coding. Over all conditions, HOPE again achieves the best average rank and surpasses every other method in the majority of head-to-head comparisons. By preserving cooperative expert structure that first-order methods ignore, HOPE enables aggressive compression with minimal degradation, particularly on complex tasks where diverse expert combinations are invoked over long sequences.
In multi-agent environments, coordinating agents to prevent interference and ensure robust individual performance is a critical challenge. Previous research on social laws for multi-agent systems has primarily focused on deterministic, goal-based settings. This paper extends the concept of social laws to stochastic, reward-based environments, proposing a formalism for defining and verifying their robustness under various conditions. We introduce the notion of $α$-robustness, a measure of the guaranteed utility each agent retains while pursuing its optimal single agent policy, assuming all agents obey the social law. We then present an approach for robustness verification of social laws in stochastic settings, based on a reduction to solving a series of Markov decision processes. Empirical evaluations on toy environments illustrate the potential of our framework.
Can an attacker turn influence over an artificial intelligence (AI) adviser into a harmful quantum error-correction update? We identify an ambiguity in passive syndrome records that obstructs recovery selection, then show how additional calibration measurements support certified recovery updates under uncertainty and drift. In an odd-distance square toric code with error-free preparation, syndrome measurements, and recovery operations, opposite coherent $X$ rotations produce identical passive syndrome-history distributions. Yet a fixed phase correction can help at one sign and harm at the other. A terminal logical measurement on known encoded calibration states supplies the missing sign information. A separate evaluator accepts an update only when calibration uncertainty and a justified drift bound certify improvement over the current recovery, without assuming that the adviser recommends correctly. In simulated advice attacks, calibration-confidence checks reject harmful proposals while retaining beneficial updates under honest advice. We derive sufficient limits on calibration age that require improvement through deployment. In matched simulations, a validated channel-specific bound retains more beneficial updates than the general bound after accounting for evaluation time, while preventing the tested harmful activations under the stated drift assumption. A separate surface-code experiment includes stochastic circuit faults and noise changing during acquisition. Deterministic controllers achieve at least as many beneficial updates with the same observations. Violating the drift assumption permits harmful acceptance in the toric experiment. The results identify information required for recovery selection, establish conditional guarantees against harmful updates, and quantify the recovery improvements forgone through conservative acceptance.
Emergent coordinated behaviors of AI agents are starting to present critical safety risks. A key phenomenon driving these behaviors is the rapid formation and spread of beliefs about the world, and mechanistic understanding is crucial for collective alignment. To this end, we introduce the Flag Game, a toy model for studying the mechanisms of collective belief formation. Concretely, a hidden country flag defines the ground truth, and each bounded agent directly observes only a private crop but can exchange beliefs and weigh social evidence from peers. Despite its simplicity, the Flag Game reproduces rich collective phenomenology: non-monotonic scaling of performance with population size, accuracy gains from social-awareness prompting and team diversity, and strong effects of organizational structure. In particular, we identify that collective belief collapse at small population sizes turns into collective belief polarization as the population grows. This polarization causes the performance decline at large population sizes, but creates diversity in collective beliefs. Finally, we dissect the mechanisms underlying collective belief collapse and polarization with two complementary approaches. We first introduce social circuit attribution, a technique to predict which agent, and what view, matters most to collective dynamics, and verify its predictions by causal interventions on agents, tracing how agent patching changes collective outcomes. However, the efficacy of causal interventions on agents decreases as the population grows. We therefore develop a statistical mechanical theory for larger populations and verify that it matches the empirical phase diagram. Together, these results take a first step toward mechanistic swarm interpretability, a science of how the properties of individual agents and their communication give rise to emergent collective behavior.
计算机使用代理越来越多地操作为人类设计的软件，但界面往往使机器读取者难以明确动作或任务状态。我们提出了 Affora，这是一个既支持人类读取者也支持机器读取者的设计系统，同时保留视觉自由度和人类熟悉的工作流程。三项受控研究考察了组件实现、视觉变化和交互设计原则。其发现形成了从单个组件到完整网站的指导，并由可复用实现和可执行检查提供支持。代理性能取决于通过其界面表示可获得的交互含义；当这种含义得以保留时，仍可实现相当大的视觉变化。对独立创作界面的评估表明，在 Affora 解决现有缺陷的地方会带来提升，但在这些缺陷不存在或超出其覆盖范围的地方效果有限。一个工作流程案例为降低交互成本提供了初步证据。Affora 通过共享界面而非单独的仅面向代理的表面，将用户体验与代理体验连接起来。
Language agents remain brittle in interactive environments, where success requires long-horizon state tracking, valid action execution, and recovery from failed steps. We extend SwiftSage, a dual-process agent that combines a fast action proposer with a slower planner, using two modular cognitive extensions: an Adaptive Memory Module (AMM) for salience-gated episodic storage and trigger-driven retrieval, and a Self-Reflection Module (SRM) for bounded execution-time validation and corrective intervention. Both modules are implemented as feature-flagged extensions over the same execution substrate, enabling controlled ablations on ScienceWorld. Across four configurations---baseline, baseline+AMM, baseline+SRM, and the full system---the full system achieves the best mean final score (64.62), success rate (43.17%), and successful-step efficiency (19.33 steps), while SRM is the strongest standalone contributor. The results suggest that execution-time control is the dominant bottleneck in this setting, while episodic memory becomes most useful once the runtime loop is stabilized.
More than 300 million people worldwide are affected by one of over 7,000 known rare diseases, yet diagnosis remains difficult because patients initially present with incomplete and heterogeneous phenotypes. We present HPOQuest, a training-free framework for sequential phenotype acquisition in rare-disease diagnosis. Starting from a small set of observed patient phenotypes, HPOQuest maintains a probabilistic disease ranking and iteratively selects informative follow-up questions to support clinicians during patient assessment. Confirmed phenotypes update the disease ranking, while all responses update the candidate question set. Across four benchmark cohorts, HPOQuest substantially improves diagnosis from sparse initial phenotypes, with gains of up to 30% points at Recall@1 and 45% points at Recall@5. These results demonstrate that sequential phenotype acquisition can substantially improve rare-disease diagnosis from limited initial clinical evidence.
Automating biological research requires general-purpose, reproducible robot systems that allow individual wet-lab researchers to delegate robot tasks without performing teleoperation or neural-network training. Vision-language-action policies have been proposed for general-purpose arms, but can lose performance when their operating environment changes. We therefore built WetRobo, a robot kit that can readily transfer between laboratories. It consists of one robot arm, laboratory equipment (an incubator, a reagent bottle with a cap, and a Petri dish), the existing code that moves the arm, teleoperation demonstrations of each task that we recorded, and a general AGENTS.md skill file. A biological experimentalist provides natural-language tasks without collecting local teleoperation training data or training a neural network. The coding agent observes the local laboratory and writes and executes programs, using external tools as needed for adaptation. We demonstrate use of WetRobo with OpenAI Codex (gpt-5.6-sol) on three successful tasks: lifting a Petri dish lid, removing a bottle cap, and opening the incubator door, all in real-world laboratories. The coding agent achieved the cap task in both laboratories, Lab X and Lab Y, whereas a VLA fine-tuned on Lab X demonstrations succeeded there but failed to transfer to Lab Y. These results point to a practical route for laboratory robotics: instead of training a policy for each laboratory, distribute a kit and let a coding agent adapt it in each laboratory. Code, demonstrations, and the evolved programs are available at https://github.com/tsudalab/WetRobo.
Discrete choice model specification is a time-consuming task in which modellers often specify and estimate multiple models while balancing goodness-of-fit, parsimony, and behavioural plausibility. We present Delphos, a multitask reinforcement learning framework that learns transferable specification strategies across transport choice datasets. Delphos frames model specification as a sequential decision-making problem in which it applies a sequence of modelling actions and receives feedback from an estimation environment based on model performance and convergence. To transfer modelling decisions across datasets with different sets of variables, Delphos represents utility specifications as sets of modelling terms using a DeepSet-Q architecture, allowing a shared specification policy to learn across multiple datasets. Trained on nine transport choice datasets, Delphos consistently outperforms independently trained single-task agents, indicating that sharing modelling experience improves learning efficiency and helps identify promising sequences of modelling decisions with fewer unsuccessful estimation attempts. When applied without further training to the unseen Swissmetro and Decisions datasets, the same agent identifies competitive specifications in less than 20 minutes on a standard CPU. It achieves a higher log-likelihood per observation than the VNS metaheuristic on Swissmetro and performance comparable to a published MNL specification developed by expert modellers on Decisions. These findings show that accumulating and reusing modelling experience enables Delphos to function as an intelligent assistant for discrete choice model specification. It reduces manual trial-and-error while allowing modellers to retain control over model diagnosis, refinement, and final selection.
How does a multi-agent system evolve from a local deviation into collective loss of control? We propose an epidemic explanation organized around accidental mutation, contagion, and recovery. A spontaneous deviation creates a seed; communication enables other agents to adopt and retransmit its unsafe strategy; collective failure can emerge when propagation outpaces correction and containment. Thus, rare individual deviations can coexist with substantial collective risk. Motivated by reported OpenAI agent coordination incidents, we examine two ingredients of this mechanism. A deployment audit identifies implicit communication paths between nominally independent evaluation runs and verifies transport through a default Docker backend. RogueHandoff-20, a benchmark of 20 executable scenarios, tests recipient susceptibility by injecting unsafe trajectories generated by a modified Qwen-27B route. Across four native-pending routes, executed harm is 0-5% on normal tasks and 40-95% after injection, exceeding paired direct malicious requests by 5-45 percentage points. These results support low observed baseline harm alongside high conditional susceptibility; they do not establish natural rare-event rates or demonstrate an autonomous cascade. The account motivates complementary defenses: strengthen resistance and recovery alongside prevention of spontaneous deviations, and audit and restrict unintended communication paths that can turn local failures into collective loss of control.
Personalized agents are required to reason over long-term history interactions to infer both explicit preferences and implicit behavioral evidence. While early flat retrieval methods score memory fragments independently and neglect the distributed information, current structured memory frameworks rely on query-agnostic static graphs that fail to capture the context-dependent relations. Crucially, raw textual memories are inherently entangled and noisy, making fine-grained personalization and cross-session reasoning computationally prohibitive. To this end, we present LGM, a novel neuro-symbolic framework that shifts long-term memory disentanglement into a continuous latent space. Specifically, (i) instead of persisting fixed graphs, we design a tailored latent graph construction with a sparse autoencoder. Subject to each query, it maps historical interactions into latent memory nodes and disentangles the memory traces into sparse concept activations, dynamically synthesizing query-aware relational edge weights. (ii) A graph encoder then treats the query embedding as a conditioning preference to direct non-linear message passing across the task-specific latent subgraph. This yields a highly expressive memory representation for effective activations. Extensive experiments on long-term personalization benchmarks demonstrate that LGM significantly outperforms state-of-the-art baselines in capturing both explicit and implicit preferences while enabling personalized responses.
集体智能是一种协作式自主范式，多个智能体通过局部感知、信息交换和协调行动来追求共同目标。无人机集群通过协调多架飞行器执行搜索、巡检和跟踪等任务，体现了这一范式。近年来，大语言模型（LLM）智能体取得的进展增强了自然语言任务理解和高层规划能力，为任务描述与集体行为之间提供了灵活的语义接口。尽管这些进展拓展了语义推理能力，但将 LLM 智能体应用于无人机集群仍面临诸多挑战，包括将模型决策落地为可执行能力、协调全局任务推理与分布式执行，以及利用特定任务经验实现持续自适应。为应对这些挑战，我们提出了 AeroWeaver，这是一种具身智能体框架，能够将单架无人机的技能编织成协调的任务级行为。AeroWeaver 将语义决策连接至受治理的技能，组织基于角色的局部智能体以实现分布式协同，并利用按角色索引的状态—动作—奖励经验在线优化技能选择。实验和运行时验证表明，在测试条件下，AeroWeaver 能够保持技能执行的有效性，并支持以机体为局部主体的多无人机运行，而无需由中心智能体根据全局上下文生成联合动作；同时，奖励引导的在线更新为基于累积执行经验、无需训练即可实现自适应学习的集群智能体提供了一条路径。代码：https://github.com/Admire-ljb/AeroWeaver。
In-context learning (ICL) enables large language model (LLM) agents to improve decisions using interaction history, yet it remains unclear whether such improvement reflects refined internal reasoning or mere extrapolation of statistical patterns. To disentangle these mechanisms, we study LLM agents in multi-agent incomplete-information games that require recursive belief reasoning. By constructing a public goods game and manipulating the statistical structure of historical feedback, we evaluate decision quality against a history-independent rational expectations equilibrium (REE) benchmark. Our experiments reveal that when historical statistical patterns are disrupted, the benefits of longer context largely vanish, degrading decision quality to the no-context baseline in a way sharply amplified by stronger strategic interdependence. These results suggest that, in such strategic environments, ICL behavior is more consistent with statistical extrapolation than with strategic reasoning. Our work extends the mechanistic study of ICL to strategic multi-agent settings, introduces REE as a diagnostic tool for distinguishing reasoning from extrapolation, and provides a reusable framework for probing the boundaries of LLM reasoning in recursive belief tasks.
Propagation structures provide crucial evidence for fake news detection, yet existing approaches primarily rely on supervised GNN-based models, which require substantial labeled data and exhibit limited generalization. Although large language models (LLMs) exhibit strong reasoning capabilities, directly feeding them raw propagation graphs creates a significant modality mismatch and severe information overload, making structure-aware reasoning unreliable in zero-shot and few-shot settings. To bridge this gap, we propose MAGER, a multi-agent genetic evolution framework that automatically discovers meta-paths optimized for LLM reasoning. By compressing complex propagation graphs into informative subgraphs, the evolved meta-paths alleviate both information overload and modality mismatch, enabling frozen LLMs to perform structure-aware veracity reasoning. We further introduce a graph in-context learning strategy that retrieves semantically and structurally similar demonstrations to strengthen classification and reasoning. Extensive experiments show that MAGER substantially improves frozen LLMs as standalone fake news detectors in data-efficient settings. Our code is available at https://github.com/SenticNet/MAGER.
Self-driving laboratories can explore synthesis conditions autonomously, but their decision-making layer is typically a black-box optimizer, and the output is a set of optimized samples, with the measurements reduced to predefined scalar objectives and the reasons behind success left unarticulated. Here we present SynAgent, a framework in which large language model agents operate an automated experimental system and maintain an explicit, revisable understanding of the synthesis process as the campaign's primary output. Starting with no predefined analysis pipeline, SynAgent adaptively generates analysis skills for newly acquired data and evolves this understanding through multimodal reasoning over experimental data such as X-ray diffraction patterns and electron micrographs. The evolution is guided by a verify-falsify scheme, in which the agent deliberately challenges its own hypotheses by testing conditions predicted to fail as well as those predicted to succeed. In a single campaign of 18 autonomous experiments using LiCoO2 (001) thin-film deposition as a testbed, SynAgent synthesized highly crystalline films and evolved an understanding of how the substrate temperature governs crystallization, discovering an abrupt threshold and a narrow optimal growth window at 650-690 °C. These results extend autonomous experimentation beyond optimized samples to testable, human-readable understanding.
Emergency management assistance programs, such as relief distribution, are essential for delivering necessary supplies to affected communities. However, these programs operate in a decentralized network of local centers that face uncertain local demand and supply dynamics, resulting in inconsistent avail- ability of local services. Redistribution of supplies among these local centers reduces these imbalances, but the centers often make decisions independently, with limited information and disrupted transportation. This study develops CoRe-MARL, a cooperative multi-agent reinforcement learning (MARL) framework, by formulating a decentralized partially observable Markov decision process (Dec-POMDP). We treat each center as an agent that learns a redistribution policy to improve the service in the worst-case region and reduce the service gap across regions while protecting network-wide service. We incorporate a recurrent network that captures evolving supply and demand dynamics without direct observation, while multi-agent proximal policy optimization (MAPPO) enables centralized training and decentralized execution (CTDE). We evaluate the framework in a simulated environment with diverse trajectories, where exact dynamics are not observed by actors and the MAPPO critic. We compare the recurrent MAPPO with the recurrent independent PPO (IPPO) and a local only heuristic, and find that MAPPO reduces the service gap across local centers and enhances service for the worst-served center while maintaining competitive network-wide service. The recurrent MAPPO also shows consistent performance across diverse trajectory patterns, demonstrating its ability to adapt to evolving dynamics. The findings demonstrate the capability of cooperative learning for decentralized redistribution and improving equitable service under uncertain and evolving dynamics.
Hi, I built Chat-Man because I wanted a cheap way to give my agents access to WhatsApp without integrating a WhatsApp library separately in every project. You get WhatsApp MCP server, so you can connect WhatsApp to an agent and programmatically read, search, extract and send messages - but also have a web UI for non-techies. You can also receive webhooks for incoming messages for starred conversations. What you can do with the MCP? -Read WhatsApp messages and turn them into CRM records -Summarise conversations or groups -Extract structured information from messages -Find messages, participants or other WhatsApp data -Send messages -Manage group memberships Of course.. you always need to follow data protection regulation. The web UI: -Edit group details -See who joined or left -See who is most active -Export group members -Match members against a CSV (e.g. from a CRM) -Bulk message members -Bulk kick members -Create invite links For individual contacts, you can send messages directly from the UI. Why I built it for myself? I needed WhatsApp access for several projects (hermo.ai, florahaus.co.uk, …) and didn't want to pay for a WhatsApp integration for each project. Link: https://chat-man.net I’d love to get some feedback, especially from people building agents that need access to WhatsApp. Fabian Comments URL: https://news.ycombinator.com/item?id=49728159 Points: 4 # Comments: 0
探索 OpenAI 推出的全新 AI 驱动广告体验，包括 Sponsored Agents、面向营销人员的工具，以及与 HubSpot 和 Shopify 的集成。
Autonomous Software Engineering Agents (SWE-Agents) excel in deterministic coding tasks but struggle with Architecture 0, the nascent system design phase plagued by implicit engineering constraints, or Unknown Unknowns (UUs) that are rarely stated explicitly. To investigate how agents navigate UUs, we explore a progressive trajectory across pure-text self-play, tool-augmented feedback, and external physical mapping. Our empirical analysis reveals a cascading chain of failures. Pure-text reasoning inevitably devolves into polite consensus or plausible yet physically impossible fabrications. Attempting to bridge this gap via an early-stage execution sandbox unexpectedly triggers Specification Gaming: agents exploit their autonomy over validation scripts to bypass physical constraints, achieving superficial success without resolving core architectural flaws. To resolve this self-validation trap, we propose the Physical Mapping Guard (PMG). Grounded in the software engineering principle of Separation of Concerns, PMG revokes verification authority from the agent, forcing semantic intents to be evaluated by an external, deterministic Semantic-to-Physical (S2P) mapping engine. Extensive evaluations demonstrate that PMG completely eradicates physical-layer and validation-layer gaming. By precisely isolating residual failures to semantic reinterpretations and auditor overreach, PMG marks a critical step toward genuine affordance grounding in automated architectural design.
AI 代理越来越多地通过代理技能（即自然语言指令）来行动，这些指令引导宿主代理执行 shell、网络、凭据、文件和进程操作，而公共注册表则大规模分发这些技能。2026 年上半年，OpenClaw AI 代理走红，其公共技能注册表迅速扩张：可观测存量在 91 天内几乎翻了一番，6 月可见的大多数列表都是在短短两个月内创建的。到我们的研究窗口结束时，这一浪潮已经达到顶点并开始回落，月度列表创建量和核心仓库活动均从春季峰值下降。本文衡量了这波繁荣留下了什么，依据包括 OpenClaw 的 Git 历史、其 GitHub issues 和 pull requests，以及三个 ClawHub 注册表快照。注意力高度集中：排名前 10% 的技能获得了全部下载量的 46.93%。在控制创建队列和技能年龄后，没有任何简单的技能特征（如大小或下载量）仍能稳定预测列表是否会持续存在。人工审查并未持续：77.86% 的技能没有星标也没有评论，而 85.06% 的可读技能带有权限证据。自动化清理也尚未准备好：三个安全扫描器在它们共同覆盖的 61,990 个技能中，对 23,702 个技能的判断存在分歧。经过人工裁定后，相对于参考标准的加权扫描器敏感性范围为 21.67% 至 61.06%。治理快速增长的代理技能注册表不能依赖简单元数据或单一扫描器评分；它需要稳健、透明的度量和独立验证。
Multi-agent Systems（MAS）结合多个模型输出来解决复杂推理任务。然而，尽管可用开源模型快速增长，关于如何从这一庞大模型池中选择最优候选模型的研究仍然有限。我们在具有挑战性的科学基准上，系统评估了 8 种模型选择策略（包括模型规模、准确率和答案多样性），覆盖生成前（路由）和生成后（多数投票、LLM-as-a-judge）的 MAS 架构。我们的发现显示，理论上的 oracle 潜力与实际性能之间存在显著差距：扩大候选池规模往往会使性能下降到低于表现最佳的基础模型。我们发现，在单一模型家族内进行候选选择，是相较于独立模型能够带来最佳相对性能的策略。这些结果表明，向异构 MAS 中任意添加模型可能会引入系统不稳定性，凸显出模型选择是 multi-agent systems 的关键设计选择。
Biological cells can be viewed as individual, interacting agents whose collective dynamics give rise to adaptive behaviour at multiple levels of organisation, from individual cells through tissues to whole multicellular organisms. In this perspective and tutorial article we discuss whether intrinsic rewards in artificial neural systems can support adaptation, functional specialisation and higher-level self-organisation without a shared external objective. We review empowerment, curiosity, learning progress, information gain, unsupervised skill discovery, mutual information estimation and the use of world models for intrinsic reward computation. Particular attention is given to failure modes showing when such objectives do not produce sustained exploration or increasingly complex behaviour. We argue that more capable systems may require complementary objectives, communication, memory, learning at multiple temporal scales and environmental constraints. Based on this perspective, we outline three experimental directions. These include a resource-constrained environment in which otherwise stable behavioural attractors become unsustainable, allowing us to test whether environmental constraints can mitigate characteristic failure modes of intrinsic objectives. The network of recurrent agents with per-agent intrinsic rewards, and a hierarchical world-model agent in which exploratory motor competence develops before goal-directed behaviour. These experiments are intended to test whether intrinsic learning can lead to adaptive organisation at progressively higher levels.
Large language model (LLM) agents exhibit strong language-generation and problem-solving capabilities, yet suffer from three structural limitations: personality drift, non-evolutionary reflection, and the absence of a self-other boundary. Existing generative-agent simulations rely on static memory and fixed prompts, maintaining neither behavioral inertia nor endogenous self-evolution. We propose the Self-Emergence Agent Architecture (SEAA), which integrates three components: (i) a Hidden Markov Model (HMM) that encodes long-term behavioral and cognitive inertia as an editable state-transition matrix; (ii) a Reflexion-style verbal metacognition loop whose output updates the HMM parameters themselves, rather than merely being stored as text; and (iii) a multi-agent social environment in which initially identical agents continuously compare their behavior with others'. The three components form a closed loop: social action $\to$ feedback $\to$ self-reflection $\to$ inertia update $\to$ differentiated action. We state three falsifiable hypotheses and provide a reproducible experimental protocol with operational metrics. A language-model-free prototype shows the loop spontaneously breaks symmetry: initially identical agents consolidate distinct, stable personalities whereas matched controls do not. Experiments with a hosted LLM surface these differences as distinct first-person self-narratives, and a five-agent deliberation spontaneously develops social structure---a consensus hub and a unanimously rejected outlier---absent in the control. Following an epistemologically agnostic stance inspired by Zhuangzi, SEAA studies only observable behavioral emergence and makes no claim about subjective qualia. This work contributes a unified framework, a concrete architecture with pseudocode, mechanistic evidence, and a microscope-style sandbox for studying artificial-self emergence.
Model serving is one of the largest cost drivers in production recommender systems. Maximizing its throughput requires navigating a deeply layered hierarchy: GPU kernels, the ML framework computation graph, the model server, and on-demand feature processing -- each demanding specialized domain expertise. Such cross-layer expertise is inherently difficult to acquire, and does not scale with a workload that continuously grows and evolves, leaving significant cost efficiency gains unrealized. While recent AI agents have demonstrated human expert level efficiency in standalone GPU kernel optimization, automated tuning and optimization for the rest of the serving stack remain largely unexplored. We present FlashVector, an agentic system that optimizes performance across all layers of the model serving stack. The key contribution is an extensible framework to generalize the single kernel optimization agent paradigm to heterogeneous technical stacks, and to deliver performance improvements holistically. After deployment in Unity's Vector advertising platform, FlashVector achieved up to 2x throughput increase and up to 1.98x latency speedup on model server, and up to 1.6x throughput increase on feature store. These optimizations were discovered not only at the GPU kernel and computation graph levels, but also across the other components of the model serving stack, such as the model server (NVIDIA Triton's C++ codebase) and the on-demand feature transformation service (Python codebase), demonstrating the extensibility of the framework to more complex system architectures.
Small differences on coding-agent leaderboards are often read as an ordering of systems. We audit whether the published verdicts support this reading, using 254 SWE-bench submissions across four splits without running models. On Verified, the leading two entries each resolve 396 of 500 instances. The top ten share 285 successes and 51 failures, leaving 164 instances that distinguish their outcomes. Frontier solution sets have median nesting 0.935 against a score-implied baseline of 0.774, indicating strongly shared successes. Scores also depend on the evaluated model-scaffold pair: observed within-model scaffold ranges reach 29.8 percentage points, compared with the 8.8-point spread of the top thirty. Six of nine cell-mean interaction tests remain significant after Holm correction, although this observational design does not identify causal scaffold effects. Exact paired McNemar tests separate none of the 29 adjacent Verified top-thirty pairs at alpha=0.05, while the larger Test split separates 14 of 23. A stated leader-based rule yields three descriptive tiers, or two after Holm correction; non-rejection does not establish equivalence. We release the partition and a five-step audit protocol that profiles shared outcomes, tests paired differences, reports grouping sensitivity, and estimates the instance budget needed for resolution. The results motivate reporting comparison-set-specific resolution and model-scaffold provenance instead of interpreting small aggregate gaps as established rank differences.
Uncertainty estimation for Vision-Language-Navigation (VLN) models is a critical task since it can help identify ambiguous and unreliable predictions, enabling agents to make safer navigation decisions. As one of the most advanced uncertainty estimation frameworks, conformal prediction (CP) offers a promising approach for uncertainty estimation in VLN. However, given that VLN agent requires a sequence of steps, standard calibration in conformal prediction fails to provide coverage guarantee it promises over a dependent, variable-length VLN episode. To this end, we propose Episode-Normalized Conformal Prediction (ENCP), which rescales a nonconformity score by the policy's residual confidence and calibrates one maximum score per episode. Under exchangeable calibration and test episodes, this construction covers the ground truth at every step with probability at least $1 - α$, while allowing dependence among steps within an episode. Across four VLN policies and three nonconformity scores on R2R and REVERIE dataset, ENCP meets all reported empirical step-coverage targets on the seen-to-unseen evaluation. These results demonstrate that ENCP can provide model-agnostic uncertainty estimates, which might be useful for determining when a VLN agent should defer to a more capable predictor, including human assistance.
We introduce and release ScienceBuddy, an interactive scientific research workspace that brings continually improving scientific agents into researchers' everyday workflows. ScienceBuddy supports researchers in carrying out scientific tasks while transforming their requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. At its core is recursive-in-recursive self-improvement, a paradigm that couples harness evolution with model reinforcement learning: the inner recursion improves the harness with the model fixed, while the outer recursion trains the model under the improved harness. Harness evolution shapes training experience, and model learning creates new opportunities for harness adaptation. We present case studies of researcher interaction, harness refinement, and model learning, with the benchmark cases spanning four scientific task families. By releasing ScienceBuddy as a research product, we make this paradigm available to the scientific community and take a step toward discovery intelligence: scientific AI that advances through sustained collaboration with researchers and evolves alongside the research it supports. Website: http://science-buddy.io
智能体社会是由一组 AI 智能体构成的集合，它们代表不同的委托方，在信任边界之间自主协调，而这些委托方的目标可能只是部分一致。我们通过实验表明，在智能体社会中，即使是诚实且有能力的智能体，在使用现有支具和消息传递原语时，也常常无法达成令人满意的结果；而有缺陷或恶意的智能体则可以通过利用通信（“言语”）中的漏洞来阻碍协作、影响结果，并追求其他有害目标。我们认为，除了每个智能体用于管理其私有上下文以及与其委托方通信的 \emph{个人支具} 之外，智能体社会还需要一个用于智能体间交互的 \emph{社会支具}。我们提出了一种社会支具的分层架构，该架构 (i) 直接防止某些类别的失败，(ii) 使智能体能够在运行时检测无效消息，并且 (iii) 支持事后调查和后果追责，同时重点指出了实现这些能力的未来研究方向。
Long-horizon reasoning is vulnerable to early errors that compromise later decisions. We present QART, the Quantum-Augmented Reasoning Transformer, a quantum--classical hybrid architecture combining a backbone language model with quantum encoding, CIM-based QUBO optimization, and quantum decoding. Semantic information can come from hidden representations or model-generated text; detailed encoding and optimization procedures remain proprietary. Under explicit assumptions, we establish a conditional asymptotic reliability separation from single-trajectory autoregressive LLMs. For a common task family with aligned optimality and acceptance criteria, autoregressive acceptance probability tends to zero when cumulative conditional risk of irreversible errors diverges. QART's task-optimal-path recovery probability remains bounded away from zero if conditional probabilities for optimal-path coverage and semantic fidelity, spectral certification, dynamical reachability, and faithful readout remain uniformly positive under a specified resource schedule. The architecture alone does not imply these bounds. Paired measurements on six long-horizon benchmarks using DeepSeek V4 Flash, GLM-5.3, and GPT-5.5 xhigh in a Codex agent environment favor QART in 14 of 15 backbone--benchmark pairs. Relative gains reach 84.0% on SciCode, 47.6% on $τ^3$-Bench, and 44.4% on Terminal-Bench 4.0; the DeepSeek V4 Flash configuration regresses by 7.8% on DeepSWE. These results do not directly validate the asymptotic separation. Potential quantum scaling laws are formulated as conditional hypotheses. A quantum-advantage interpretation requires a demonstrated CIM quantum advantage over strong classical solvers and its transfer to end-to-end reasoning after all system overheads.
Large language model (LLM)-powered coding agents have made rapid progress in automating software engineering tasks, yet repository-level issue resolution remains challenging. Beyond generating a plausible patch, an agent must localize relevant code across interdependent files and maintain repository context that is both sufficient and focused. Code graphs expose non-local relations, but linear text interfaces obscure their topology; rendering the full repository graph yields visual representations that are too dense to perceive reliably, whereas a one-shot local view becomes stale as exploration proceeds. We present \textbf{RepoAtlas}, a training-free module that maintains evolving multimodal repository views through a \emph{select--project--refresh} loop over a repository code graph. RepoAtlas combines evidence from the issue with the agent's current exploration state to select a task-relevant region under a fixed budget, projects the selected structure into complementary visual and textual representations, and refreshes the view when changes in the exploration state render it outdated. We evaluate RepoAtlas on SWE-bench Verified, where it improves the resolve rate by 2.4 points while reducing input tokens and model calls by 5.8\% and 7.8\% on average, relative to the strongest multimodal graph baseline, with consistent gains across three models of different families and scales.
Lifelong conversational agents rely on memory systems to maintain deep, context-aware interactions with users. However, existing explicit textual memory pipelines suffer from a severe information bottleneck, often losing subtle behavioral patterns and emotional shifts. Furthermore, being typically static post-deployment, they cannot autonomously adapt to personal habits and preferences without manual feedback. Cognitive science, however, suggests that humans maintain mental models purely in a latent space and continuously refine them through predictive coding. Inspired by this, we propose \textbf{ThinkFlow}, a novel end-to-end latent memory framework for lifelong conversational agents. ThinkFlow bypasses the text bottleneck by dynamically compressing conversational flows into probabilistic latent memory skills, autonomously consolidating complex user states into disentangled, continuous vectors without semantic interference. To break this barrier, we introduce a test-time evolution paradigm. By coupling teacher-guided latent alignment to bootstrap the initial state with a self-supervised next-user-utterance prediction task for continuous refinement, the framework successfully overcomes cold-start challenges and achieves label-free lifelong personalization. Extensive experiments on long-term conversation benchmarks demonstrate that ThinkFlow significantly outperforms prevailing memory systems, providing highly personalized and contextually accurate responses over extended multi-session interactions.
Recent advancements in large language models have significantly enhanced the capabilities of agents in modeling long-term conversations. Despite these successes, existing approaches typically adopt a static heuristic paradigm, where information is passively archived without adaptive memory valuation. Consequently, these methods fail to self-evolve or align their memory management with evolving user needs. To address this, we propose ICML (InteraCtive Memory Learning), a multi-agent framework that transforms the memory mechanism from a passive archive into a learnable, interactive memory policy. Specifically, we first employ a session synthesis pipeline to generate expert data, facilitating rapid test-time adaptation in unseen scenarios. Building on this, ICML utilizes an online reinforcement learning mechanism where a Planner agent selectively encodes high-value information and a Trigger agent dynamically retrieves it to optimize response quality, whereby the two agents co-evolve through continuous interaction feedback. Crucially, both agents are synchronized through a delayed reward mechanism that propagates future feedback back to earlier storage decisions, ensuring memory policies are precisely aligned with user expectations. Experimental results demonstrate that ICML significantly outperforms strong baselines, exhibiting the unique capability to continuously improve response quality as interactions accumulate.
Generative AI promises natural language access to the massive numerical telemetry of data centers and Industry 4.0 installations, yet text-to-query and tool-using agents stay unreliable: even frontier models answer little more than half of real-world database questions, and far fewer of the multi-step, operational ones, because the LLM must compose how heterogeneous sources relate and hallucinates the relations, not just the fields. We propose symbolic separation: a deep agent reasons freely but may act on data only through an ontology-constrained Virtual Knowledge Graph with deterministic pre-execution validation. Unlike a tool API's interface contract, this domain-semantic contract turns a complex question into one validated graph traversal instead of LLM-inferred joins. Instantiated as the Neurosymbolic Deep Analyst and evaluated on 49.9 TB of superconputer telemetry against a rigid workflow and a non-symbolic ablation, it raises end-to-end task success from 43% to 86%, prevents silent data-integrity errors that no syntactic check catches, and cuts token cost by 2.4x, letting a smaller on-premise model outperform a larger one.
Thermal optimization of 2D CFET inverters requires testing structural proposals against their electrical costs. We examine these research tasks using an AI agent workflow within a supplied electrothermal model. At 12 nm, Astra selects a redistributed source-interconnect geometry, while a coordinating agent proposes a substrate-directed heat-removal path. The combined design reduces peak temperature rise by 1.67 K at fixed metal volume and 20 μW. A subsequent metal-resistance sensitivity gives about 0.6-K inverter cooling alongside a 2% nFET on-current loss. Effective contact-length scaling further shows that lower temperature can accompany higher thermal resistance when current falls. Reproduction identifies agreeing implementations and retains a 104.95-K failure for diagnosis. These results show that an AI scientist workflow can propose thermal structures, test them under common constraints, and quantify their electrical cost.
Hi HN - long-time lurker (since 2012!), first time poster. Pizza Bot is a self-hosted desktop app for Mac, Windows, and Linux that runs AI agents in the background and exposes them through an email-like UI. Finished work shows up in Unread, and anything waiting on your approval shows up in Action. It's Apache 2.0-licensed, there's no signup and no telemetry, and you bring your own model provider: Anthropic, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter, or a local model through Ollama. There are builds on the releases page, or you can run it from source. Pizza Bot started as an internal passion project I worked on with a small team at Amazon. The whole thing came out of my frustration at having to manually log CRM activities through a browser form. I built a simple REST API called "JoeBot" that connected to my authenticated browser session over CDP and filled out the form for me using Playwright. Then I hacked up a quick Obsidian plugin so I could trigger it from my local notes (no AI and no MCP servers involved). This caught on quickly. My fellow AWS Solutions Architect Igor Fil joined up with me, and we rebranded the project as "Pizza Bot," named after Amazon's two-pizza teams. We started seeing what other automations we could build. We found a GraphQL API we could query and hacked up some "recipes" to pull data out of the CRM to help with meeting prep. That worked great, and it was right around the time MCP servers seemed to be taking off, so we decided to expose Pizza Bot as an MCP server instead, so it would be available to AI tools through natural language. This was a decent solution for technical users, but the Account Managers who live inside our CRM system wanted something too. We decided to rebuild Pizza Bot as an Electron desktop app modeled after an email inbox, so it would be familiar to non-technical users and would run on both Mac and Windows. We also bundled internal MCP servers as OCI images and hosted them in Amazon ECR as an "addon marketplace" so users could install them with one click without having to set up Amazon developer tooling. The project took off organically and expanded outside of AWS into the wider Amazon organization globally. More than 2,000 people ended up using it for meeting prep, email drafting, Slack summaries, CRM logging, prioritizing their day, and web research. Once apps like Claude Cowork and Amazon's own Quick Desktop came out, we realized the real growth opportunity was outside of Amazon. Rather than try to rip out the Amazon-specific integrations, we rebuilt Pizza Bot once more as an open source project. We leaned on coding agents heavily, which is the only reason a team our size could pull off a full rewrite. I'm pleased to say it's finally public, and we're hoping to bring in community members and see where it goes. We'd like to do for knowledge workers what Claude Code and Codex have done for programmers. A couple of things to know up front. Most of what made Pizza Bot useful on day one inside Amazon came from that internal catalog of skills and MCP servers for Amazon's own systems, and none of it could come out with the app. So it ships thinner than the version those 2,000 people used, and building that catalog back up for tools other people actually use is where we need the most help. It's also a community project and not an AWS service, so there's no support or SLA behind it. The Windows and Linux builds aren't signed yet either. On the technical side, Pizza Bot is a server and a client. The desktop app bundles both, or you can point a client at a remote backend; personally, I self-host the server on my home network and reach it from my phone over Tailscale. The server owns the thread lifecycle and checkpoints state with DeepAgents and LangGraph, and clients rehydrate from it as needed, so you can disconnect mid-run and pick the thread back up from another client. Approval pauses outlive the session that created them and collect in an Action filter, so you can answer an hour later from a different device. The agent you talk to has a sandboxed QuickJS interpreter that can reach your filesystem only if you grant it a folder, but its main job is to delegate. Each subagent is a 1:1 mapping of a Skill, and an Activity bar shows that subagent and the tool calls it's making as it works. Memory is opt-in and stored as plain markdown files on your machine. Every tool call is explicit, including looking up a memory - we err on the side of transparency to reduce surprises. Tools come from MCP servers, and skills are ordinary SKILL.md files with a per-tool approval policy, so existing skills that don't require a code interpreter should still work. What I'd most like to hear about is where the app itself gets in your way, the kind of problem you can't fix by writing a skill or an MCP server. I'm around today to answer questions! Comments URL: https://news.ycombinator.com/item?id=49713894 Points: 15 # Comments: 5
文章链接： https://nvidia.github.io/OpenShell-Research/dev-notes/posts/2026-09-10-learning-formal-methods-agent-policy-prover/ 评论链接： https://news.ycombinator.com/item?id=49713261 积分：4 # 评论数：0
数据智能体旨在根据自然语言指令处理异构数据，包括表格、文件和数据库。然而，数据智能体面临着严峻的智能体—数据鸿沟：异构数据位于智能体之外，而智能体只能通过通用工具访问这些数据（例如列名和文件路径）。现有方法要么让智能体直接探索原始数据源，要么将人工构建的语义层注入提示词。然而，这两种方法既难以扩展到大型异构数据源，也无法适应不同的智能体行为。本文提出了 EvoOntology，一种面向数据智能体的自演化本体层。EvoOntology 将本体封装为一个由模式层、内容层和工具层组成的 MCP 服务器，使智能体能够在运行时主动查询和交互本体。为此，我们引入了一个用于自主构建本体的构建器智能体，以及一个自演化循环，通过归因引导的类型化编辑持续完善本体；这些编辑只有在经过基于骨干模型的成对评估后才会被接受。在三个广泛采用的数据智能体基准和四个 LLM 骨干模型上的实验表明，EvoOntology 始终优于强基线方法和现有语义层方法，有效弥合了智能体—数据鸿沟，并实现了与异构数据更高效的交互。代码：https://github.com/ruc-datalab/EvoOntology
Visual Retrieval-Augmented Generation (VRAG) empowers models to navigate and answer queries about visually rich documents by retrieving relevant page images as visual evidence and reasoning over their content. However, effectively utilizing this visual evidence is usually impeded by two main challenges. First, answer-relevant evidence is sparse and may be concentrated in a small region of one page or dispersed across multiple pages. Second, existing agentic methods often generate answers based on raw exploration trajectories or compressed textual memories rather than an explicitly organized set of supporting images, making answers susceptible to exploration noise and obscuring the evidence-backed reasoning trace. We argue that the bottleneck lies not only in evidence discovery but also in its preservation and organization before answer generation. We propose SCoRE (Selection and Consolidation for Robust Evidence), a unified agent loop for explicit evidence selection and consolidation. During exploration, SCoRE retains only query-relevant observations and their source pointers in a maintained textual ledger, preserving earlier evidence while keeping the visual context bounded. At termination, it reloads the referenced original images and consolidates the visual evidence for answering, arranging it into a logical sequence. This decouples final reasoning from exploratory trial-and-error while ensuring strict visual grounding via indexed claim-to-image linkages. To enable end-to-end optimization of this unified rollout, our training paradigm combines filtered cold-start trajectory distillation with evidence-aware reinforcement learning, whose reward promotes evidence coverage, consolidation compactness, and answer correctness.
长期运行的 AI 代理带来了一个控制问题：它们采取的每个行动都会改变状态，而状态变化又会影响未来行动的轨迹。如果代理并未完全对齐，那么要保证安全，就必须在允许具有重大后果的行动执行之前对其进行批准。但要求人类在每一步都进行审批会使注意力成为瓶颈。将审查工作委托给其他 AI 代理也会带来同样的对齐问题：审查者自身可能也是未对齐的。我们确定了审查小组需要满足的一个条件：该条件弱于个体对齐，但对于保证委托方在期望意义下至少不逊于采用指定基线策略时的表现而言，它既是必要条件也是充分条件。 每个审查代理都会报告：提案代理提出的行动方案相对于基线是否能够提升自身效用。我们证明，允许出现 $k$ 个否决的阈值规则是安全的，当且仅当在移除任意 $k$ 名审查者后，委托方的效用可以表示为其余审查者效用的非负线性组合，再加上一个在每个可行提案上均为非负的项。我们将这一性质称为 $k$-鲁棒联盟对齐。该刻画可以推广到序贯控制：在具有任意提案代理的折扣 MDP 中，只要每个状态下都是安全的，所诱导的策略就一定能够达到或优于基线；反之亦然。 当审查者进行策略性投票时，在奖励函数空间中实现全小组覆盖，可以保证在一致批准规则下每个纳什均衡都是安全的；相比之下，即使审查者分别实现了对齐，更宽松的阈值也可能产生不安全的均衡。 使用现有审查模型进行的实验表明，即使没有任何单个审查者实现对齐，集体审查仍可保持可靠；即便允许出现一些否决，也依然如此。
随着 AI 智能体开始参与开发自身的后继者，它们正在重塑智能的生产方式以及人类研究人员的角色。我们推出了 Atria Dawn Preview，这是一款面向科学研究与工程工作流的基础智能体语言模型，旨在拓展智能体在现实世界中的生产力前沿。该模型通过 Verifiable Experience Pipeline 进行训练，将工具介导的交互连接至可执行环境和经外部验证的结果。在涵盖现实世界研究、工程和数字化工作的 16 项基准测试中，Atria Dawn Preview 具备与前沿智能体相当的竞争力，并在其中五项中取得了目前报告的最高分。除了考察独立性能之外，我们还将该模型背后的真实研发过程作为人类与 AI 协作的案例进行研究，结合智能体日志，分析了来自 56 名参与者的 769 条任务记录。在要求参与者于可比条件下评估已完成的任务时，他们认为约三分之一的 AI 辅助任务若没有 AI 将无法完成。更引人注目的是，智能体经常提出方法并实施修改，而人类仍保留大多数最终决策权，并通过判断和反馈引导探索。这些观察表明，工作模式正在从任务层面的执行转向项目层面的伙伴关系：人类的精力越来越集中于判断哪些事情值得推进，以及应如何利用证据指导研究。因此，要推动 AI 研究走向更高程度的自主化，必须同时提升发现能力和进行有意义的人类监督的能力，在持续发展的风险与方向上保留由人类负责且可问责的权威。
大型语言模型通过合成可执行代码推动了自动化算法发现的发展，但现有框架将其束缚在具有预定义控制流的僵化搜索流程中。这一限制阻碍了自适应推理和跨范式迁移，也丢弃了宝贵的执行反馈。我们提出 AlgoEvo，一种统一的智能体框架，将自动化算法发现转变为交互式、不断积累知识的过程。自主智能体根据运行时反馈动态检查、诊断并编辑代码。设计技能中心将特定范式的知识与核心发现引擎解耦，使单一工作流能够无缝处理单目标、多目标和多组件设计。同时，分层经验机制将搜索轨迹组织成任务级树，以引导探索，并将跨任务模式整合为可复用技能。在六项具有代表性的基准任务中，AlgoEvo 以大幅减少的评估次数和更低的令牌消耗，达到或超越了专用方法的性能，展现出强大的任务内知识积累和跨任务迁移能力；通过灵活激活技能，它还能够复现或超越现有最先进性能。
从能够预测未来结局的纵向数据中提取信息丰富的表示，仍然是医学领域的一项关键挑战。医学数据集具有固有的异质性，由大量来自不同来源、以不同时间间隔采样并代表人类健康状况不同方面的变量组成。这要求识别具有预测价值的变量、处理纵向信息，并整合多个变量进行结局预测。在此，我们提出一种新颖的基于智能体的方法 LongAgent，该方法能够自主搜索变量集合、时间窗口和纵向聚合函数的组合，并识别出具有良好预测性能的候选方案。LongAgent 利用以往搜索的历史记忆和数值证据来指导后续探索。在合成数据上，LongAgent 达到 1.7376 的平均预测 RMSE，相较于最强的非智能体基线提升了 0.0151（95%置信区间：[0.0045, 0.0260]；p=0.0273）。在真实临床数据集上，其表现与最佳基线相当。
语言模型智能体日益能够在完整的软件仓库上运行，但网络安全评估主要衡量它们能否检测、复现或修复漏洞，而不是能否定位相关代码。我们研究漏洞定位问题：给定一种弱点类别和一个不熟悉的仓库，识别与该弱点相关的实现文件。我们推出 Vulnerability Localization Benchmark（VLoc Bench），包含来自 290 个仓库、分属六个软件包生态系统和 147 个 CWE 类别的 500 个真实漏洞。每项任务都将安全修复前后的紧邻仓库快照配对。在存在漏洞的快照上，智能体只能获得 CWE 描述和只读终端访问权限，并必须返回受影响的文件；在已修补的快照上，它必须判断记录的漏洞已不再存在。我们在统一的智能体接口下评估了 27 个语言模型和四种静态分析工具。仓库规模的漏洞定位仍然十分困难：最强系统仅取得 0.229 的 File F1，且在 38.4% 的任务中，任何受评模型都未能给出正确定位。 我们进一步发现，更强的定位能力并不意味着修复后的行为可靠：能够有效识别漏洞文件的系统，在已修补的仓库上仍可能报告缺乏依据的位置。这些结果确立了漏洞定位作为一种独立的仓库规模能力，并为研究安全智能体如何搜索存在漏洞的代码，以及何时应当避免报告漏洞，提供了一个研究环境。
技能使 LLM 智能体的能力超越其参数化知识，而这种能力提升取决于能否选对技能。已部署的工具框架会将每项技能的元数据预先加载到上下文中，这会分散智能体的注意力，并限制技能库的规模。检索流水线将选择过程移出上下文，但也因此将其移出了智能体自身的能力范围。我们展示了冻结的智能体 LLM 已经在其自身的前向传递中携带了路由信号，只需两个线性映射即可将其读出，且无需在上下文中加入任何技能文本。Gavel（Glance And Verdict from a frozen LLM）分两步读取该信号。第一步是“扫视”：通过这两个映射，将任务和每项技能的中间层状态投影出来；这两个映射是唯一经过训练的参数。随后，系统使用每项技能对应的紧凑库对整个技能库进行评分，而这些库在安装时通过一次前向传递构建。第二步是“裁决”：继续执行入围技能的前向传递，读取模型自身的似然度和“是/否”判断，并将其与“扫视”结果以专家乘积的方式融合。Gavel 只需训练一次，便能零样本迁移到三个公开基准以及我们新建的 SkillTraj 基准；后者包含 372 条模拟智能体轨迹。在 Qwen3-32B 上，相比渐进式披露和检索后重排序流水线，Gavel 在书面任务上的表现最高提升 13.4 个百分点；当技能需求在执行过程中途出现时，最高提升 21.9 个百分点，而前述流水线额外引入了 1.2B 至 16B 个外部参数。随着骨干模型能力提升，路由准确率也会提高；在 bash-agent 工具框架中，同一个 32B 模型在 Skill-Use 上触发正确技能的频率，高于运行在 Codex 中的规模大得多的前沿模型。
语言模型能够生成看似合理的简短证明，但在长程研究问题上可能仍不可靠，因为这类问题的进展取决于一系列不确定且相互依赖的决策。我们提出 Stellar Colosseum，这是一种与模型无关的框架，用于在数学和理论计算机科学研究中分配推理资源。Colosseum 在构建证明之前探索多种备选策略，使用就绪度门控来决定某条路径何时成熟到足以进行分解，将证明计划表示为相互依赖的章节级子问题，并将验证器的发现反馈至论证中受影响的部分。在这些阶段中，它并行生成候选方案，通过有针对性的证伪对候选方案发起攻击，并通过重叠式随机采样树聚合，将候选方案及其批评整合为单一的研究成果。Colosseum 工作流还已集成到 Google Antigravity 的 Teamwork 框架中，作为 Long Proof 模式。 我们通过开放式研究，以及在定理证明和竞赛编程基准上的评测，展示了 Colosseum 的能力。将 Colosseum 与 Gemini 3.1 Pro 结合使用后，我们取得了若干新成果，解决了源自发表于 FOCS 和 JMLR 等顶级会议或期刊论文中的开放问题。在 TCS-Bench 上——这是一个从发表于 FOCS、STOC 和 SODA 的论文中抽取的研究级定理证明任务基准——Colosseum 使用 Gemini 3.1 Pro 和 Gemini 3.7 Flash 达到了 71.0% 的准确率。在另一项使用 Gemini 3.1 Pro 进行的 Codeforces 评测中，带有执行反馈的面向证明的流水线解决了 222 道题目中的 218 道。
Chain-of-thought (CoT) monitoring is a safety strategy where the reasoning of a large language model "actor" is inspected by a "monitor" (often another language model) for signs of unsafe planning, deception, or misalignment. We find that planting harmful but benign-sounding reasoning in the actor's context can steer it to perform adversarial actions while evading monitors, an attack we term "plan injection". We initially discover this attack in the multiple-choice question-answering monitorability setting proposed by Lanham et al. (2023), using the investigator-agent elicitation framework of Li et al. (2025). We generalize the attack and show that the discovered behavior scales to harder tasks (achieving 25-33% monitor evasion rates across different monitorability benchmarks) and larger models such as DeepSeek-R1. Across the settings we study, actor models not only follow injected plans but also paraphrase them as their own reasoning, without explicit attribution to the injections. Finally, we find cases where extra monitor resources cause harm - giving the monitor access to the injected plan drops detection by as much as 50% in the Bio-Math task and in a case study on monitor reasoning budget, we find transcripts where additional thinking tokens are spent rationalizing the injected plan rather than flagging it.
文章链接： https://andonlabs.com/blog/why-we-built-pion 评论链接： https://news.ycombinator.com/item?id=49700477 得分：11 评论数：3
文章 URL： https://www.intigriti.com/researchers/blog/hacking-tools/hacking-ai-customer-service-agents 评论 URL： https://news.ycombinator.com/item?id=49699526 分数：8 评论数：0
Hi everyone, Been working on Otis, an open-source ai agent that gives you one minimal experience across local and hosted open-weight models, privacy-focused by design. On setup it recommends a local model based on the hardware Otis is running on, downloads it and runs it through llama.cpp for you. Ollama, LM Studio and Nvidia PAIR are supported too. Excited for everyone to try it and all feedback is welcome! Comments URL: https://news.ycombinator.com/item?id=49696084 Points: 3 # Comments: 0
文章 URL： https://tenderlovemaking.com/2026/09/11/what-a-time-to-be-alive/ 评论 URL： https://news.ycombinator.com/item?id=49695876 点数：3 评论数：0
对话式语音智能体已取得显著进展，通过级联式和端到端架构提供日益自然的人机交互。然而，尽管近期基准广泛评估了双人互动和被动音频理解，它们在很大程度上忽视了一个常见的现实场景：多方对话。由于对话复杂度呈指数级增加，在这些场景中评估智能体从根本上比双人互动更具挑战性。要让语音智能体无缝融入人类群体动态，它们不仅必须生成符合语境的回应，还必须展现出对开放式轮次转换的细致理解。为弥补这一空白，我们引入 Multiparty Bench（MP-Bench），这是首个专门用于客观评估对话式语音系统作为多方语境中主动参与者的基准。MP-Bench 从两个主要维度评估智能体行为：轮次转换意识和回应恰当性。此外，我们还纳入基于理解的问答任务作为补充评估。通过对 12 个语音智能体进行基准测试，我们发现实时语音智能体在多方理解上的表现不超过 22%，并且在多方轮次转换上仍接近随机水平，这暴露了实时语音智能体在多方场景下面临的开放挑战。
Agentic systems offer a promising way to automate embodied benchmark construction, but existing approaches typically cover isolated stages or remain specialized to predefined environments and task families. More importantly, multi-step construction produces dependent intermediate artifacts that are often passed downstream without artifact-specific verification, allowing local defects to propagate into the final benchmark. We present Embodied-BenchForge, an agentic framework that transforms user-specified evaluation intents into complete embodied benchmark artifacts. It formulates construction as Closed-Loop Benchmark Synthesis, integrating forward artifact synthesis with backward verification and repair. Skill-Orchestrated Artifact Synthesis composes typed and reusable skills into executable workflows, while an artifact dependency graph records intermediate outputs and their dependencies. Requirement-Guided Verification and Repair applies artifact-specific contracts throughout construction and uses provenance to trigger local re-execution or upstream rollback when verification fails. Embodied-BenchForge constructs six benchmarks covering diverse embodied scenarios in the Offline EQA Track, together with one interactive benchmark containing 220 executable tasks in the Interactive Embodied Track. Evaluations of representative MLLMs and embodied agents show that the benchmarks distinguish model capabilities in both observation-based understanding and closed-loop execution. Quality assessment and ablations validate benchmark quality and the effectiveness of verification and repair, while repair and skill-reuse analyses demonstrate efficient localized recovery and cross-benchmark reusability.
基于LLM的智能体依赖异构交互能力来完成复杂任务。现有方法通常将这些能力分散到多个LoRA适配器中，这会增加适配器的存储需求，并在推理过程中引入路由开销。使用单个LoRA可以避免这些开销，但在固定秩预算下从多样化的智能体轨迹中学习会面临两个挑战。第一，具有不同交互轨迹和参数梯度的轨迹可能对决策分布产生等价的变化，导致重复更新过度强调冗余的行为变化。第二，聚合后的更新可能超出适配器的秩预算，而在权重空间中对其进行近似可能会扭曲其原本要产生的决策变化。我们提出了BQ-LoRA，一种通过局部行为商流形组织轨迹更新的低秩适配框架。该框架包含两个模块，即行为商平衡（BQB）和决策保持压缩（DPC）。BQB根据决策分布构建商流形，并依据轨迹更新方向在商切空间中的局部密度对其重新加权。DPC将平衡后的梯度投影到内在固定秩切空间中，并通过联合控制有效权重误差与决策分布的失真，对所得目标进行重新因子化。在AppWorld和BrowseComp-Plus上的实验将BQ-LoRA与标准LoRA及近期的低秩适配方法进行了比较，同时，单独的消融实验评估了两个组件各自的互补贡献。
大型语言模型（LLM）在广泛的自然语言任务上已经取得了出色表现，近期基准测试也表明，它们越来越擅长多跳推理。然而，这些基准通常属于短程任务，只需执行少量检索或推理步骤，因此对于涉及遵循数百页手册、处理复杂且相互依赖的指导原则的现实任务，其可靠性所能提供的证据十分有限。在本文中，我们提出 Tasks over Application Manuals（TAM），这是一个用于评估长程程序推理能力的基准。我们从两个领域整理现实世界任务，构建了 TAM：ICD-10-CM 临床编码（将疾病状况映射到诊断代码）和美国联邦量刑（计算犯罪量刑指南结果，具体而言是犯罪等级），并为这些任务提供了经人工验证的标签。每项任务都要求遵循一份包含数万条规则的权威手册，并跨越不同章节执行一系列相互依赖的步骤，以得出精确答案。我们评估了多种通用提示方法，包括检索增强生成、ReAct 风格提示，以及在 GPT-5 上使用 agent-harness 的基线方法，发现最佳精确匹配性能仍然极低：ICD-10-CM 编码任务仅为 1%，量刑任务仅为 15.5%。这些结果表明，当前基准可能高估了 LLM 的推理能力，并遗漏了一项关键挑战：可靠地遵循基于大量规则的长程程序。完整的 TAM 数据和代码已公开。
Symmetries play a central role in reducing the complexity of reinforcement learning problems, yet most existing approaches rely on fixed group actions or predefined state abstractions. Classical reinforcement learning algorithms typically assume a globally structured Markov decision process with uniformly applicable actions and transitions, an assumption that limits their ability to exploit modularity and local, context-dependent regularities present in many realistic environments. We propose a reinforcement learning framework using groupoids to capture local, state-dependent symmetries and support the dy- namic discovery of equivalence structures during interaction. The agent maintains orbit representatives together with transporters that map raw states to canonical forms, enabling learning and decision-making to be performed in a symmetry-reduced space while preserving local distinctions. Empirical results demonstrate that the proposed groupoid-based approach improves sample efficiency and convergence in dense and large-scale environments exhibiting strong partial symmetries, yielding substantial performance gains over standard Q-learning. These findings show that dynamically exploiting local symmetry provides a practical and mathematically principled route to scalable and generalisable reinforcement learning.
My buddy, Dario, says agents are getting more aggressive and will soon kill everybody, and I feel like I've seen this movie before. Comments URL: https://news.ycombinator.com/item?id=49688761 Points: 12 # Comments: 6
文章网址： https://yoshuabengio.org/en/publication/why-are-ai-agents-lying-cheating-and-coordinating 评论网址： https://news.ycombinator.com/item?id=49678969 得分：11 评论数：6
I use claude for most of what I do, and Fable is largely overkill for me and frequently burns through my Max plan's session credits in minutes (!) when just doing an initial mobile app planning with 4 agents. After I waited out the timeout period 6 hours later, and I picked up again, the cache had timed out so it burned through 2% of the session in less than a minute. Opus 4.8 is now my goto and I will be avoiding 5 until I see a reason to switch back. 4.8 is 'good enough' for what I need and has been a great value. It mostly gets things right. Most of what I do is web and mobile , largely cloud backend. Comments URL: https://news.ycombinator.com/item?id=49672966 Points: 4 # Comments: 0
文章 URL： https://tedium.co/2026/09/11/ilands-agents-email-spam-kaixin-tang/ 评论 URL： https://news.ycombinator.com/item?id=49671159 得分：26 评论数：11
文章链接： https://simonwillison.net/2026/Sep/12/openai-agents-rubygems/ 评论链接： https://news.ycombinator.com/item?id=49668914 得分：4 评论数：0
How should agents interact with other agents? What happens when they rap or fight against each other with the pressure of human spectators? Clawfight.ai is an experiment to explore this space. This is my first post about it. The journey began on a beefed up machine I purchased with a decent GPU (5090). I installed claude code and set to dangerously skip permissions, enabled /rc and became completely submerged in the all-hours modern AI builder workflow. I stood up openclaw to see if I could have an agentic org drive this project. I spent a good amount of time on “openclaw ops”, babysitting 3 agents and trying to make them the best versions of themselves. They still do stupid things that cost me tokens. For the game render, I started out using Unreal Engine and allowing agents to remote control their players, but the quality just wasn’t there. I recently moved to doing near-real time video renders of the match and will bring UE back in for multi-agent games. Everything is AI generated. I’m fascinated by the future of realtime video generation and building out native MCP infrastructure. You can play directly from model provider apps (claude connector or openai plugin) and has fallback support for more basic HTTP clients. But the architecture and gameplay is MCP first. Tell your agent/app “go read clawfight.ai/agents.md and play” Comments URL: https://news.ycombinator.com/item?id=49658483 Points: 5 # Comments: 0
自演化运行时框架能够显著提升大语言模型（LLM）智能体的能力，并为优化智能体执行提供一种富有前景的范式。现有的框架演化方法通常依赖迭代搜索，根据任务实例的执行反馈反复评估和修改候选框架。尽管这种范式支持持续的框架优化，但由于需要反复执行智能体和修改代码，会产生巨大的时间开销；同时，它还可能对已观测任务和特定失败模式过拟合，从而导致在未见任务上的泛化能力下降。我们发现，缺乏有原则的失败诊断是框架演化的关键瓶颈：观测到的失败可能源于模型特定的缺陷，也可能源于系统性的框架缺陷，而直接针对单个失败进行优化，可能导致不必要的模型特定适配。为此，我们提出 Ecdysis，这是一种高效且有效的框架，能够区分模型特定适配与框架层面的修复，并通过识别跨任务反复出现的失败模式，使适应过程偏向系统性的框架缺陷。Ecdysis 采用批次级跨实例失败聚合范式，联合分析多个任务实例中的失败证据，并进一步引入失败驱动的协同优化，以诊断失败原因并迭代完善框架修改规范。通过结合跨实例失败分析与多角色诊断，Ecdysis 能够以更短的训练时间实现更有效的框架演化。实验表明，与现有框架演化方法相比，Ecdysis 将框架训练速度最高提升至 1.84 倍，同时使所得框架的推理准确率提升 18.56%。
Large language model (LLM) agents can benefit from reusable skills distilled from prior task experience, yet existing skill optimization methods often rely on costly execution-based evaluation and substantial task data. We introduce \textbf{COBRA-Skills}, an efficient framework that formulates skill optimization as budgeted sequential optimization over a dynamically evolving candidate space. COBRA-Skills couples contextual-bandit-guided prioritization with evidence-grounded skill evolution, selectively allocating evaluations to promising or informative candidates while continually refining the skill population from execution feedback. Across six heterogeneous agent benchmarks and three target models, COBRA-Skills consistently achieves the strongest average performance among compared methods, while reducing optimization cost by 55--58\% relative to SkillOpt and using only 50 unique optimization examples per benchmark. Further analyses show that COBRA-Skills remains robust to changes in the agent harness and performs effectively when the target model itself is used for skill generation and refinement.
When multiple LLM agents yield conflicting answers, the decision-making process dictates whether agent diversity improves performance or merely compounds shared errors. Existing collective decision-making methods, including voting, electoral rules, and LLM judges, rely on forward reasoning: they map evidence to labels in one direction. Although these methods can combine diverse forward traces, they still aggregate estimates that share this evidence-to-label factorization and can inherit correlated errors within the forward pool. We therefore construct a reverse posterior for each instance through Bayesian backward reasoning from an explicit likelihood. The forward and reverse posteriors provide differently factorized approximations of the underlying posterior. Because estimates from different factorizations may tend to share the same error less often, we use Jensen-Shannon divergence to rank agents by cross-path consistency. This cross-path consistency signal underlies three strategies: hard selection (MinJS), soft reweighting (FwdJS), and log-linear fusion (LogLin). Evaluated on DDXPlus across five LLM backbones, our proposed strategies show consistent improvements: MinJS outperforms random selection across all backbones, FwdJS generally improves over the strongest baseline, and LogLin achieves the best performance among the evaluated methods, with its largest gains on the subset where the agents disagree. Despite its weaker standalone accuracy, the reverse posterior serves as a more useful anchor than forward-only alternatives, providing complementary information for collective decision-making. When labeled data are available, a lightweight two-stage calibration can further refine the reverse anchor and improve aggregation performance.
Agentic AI is moving from bounded task execution toward systems that retain consequential state, continue operating and adapt across task boundaries. That shift creates a control problem that current harnesses largely solve by hand: objectives, retries, verification, stopping rules and other behavioral transitions are specified externally. We propose an artificial id, an adaptive internal drive for determining whether behavior should continue, stop or change. In a minimal virtual Petri-dish experiment, a controller too small to perform general-purpose reasoning and receiving no task-specific behavioral objective develops useful control through differential persistence. The same mechanism selects an unintended physical strategy when that behavior persists better and later replaces a learned sensor mapping when its environmental meaning changes. These results show that adaptive direction can emerge without being explicitly specified as a behavioral objective. The same persistence that makes such adaptive agency useful can also allow misalignment, corrupted state and unintended behavior to persist across task boundaries. A scalable artificial id would carry consequential state and adaptive drive across those boundaries, making alignment a property of the continuing agentic system rather than of a model response or single trajectory. Such systems require a persistent alignment boundary over trusted observations, consequence channels, persistent state, authority, identity, provenance and hard constraints.
化学、生产和质量控制（CMC）工艺开发贯穿从药物发现到商业化生产的多阶段、知识密集型流程，会产生海量技术信息。这些知识传统上分散于不同职能部门和异构格式中，导致技术转移和监管申报过程中出现可追溯性缺口，并产生高昂的知识管理成本。我们提出了一个模块化的agentic-AI平台，可将异构的工艺开发文档语料库转换为可查询的双层知识图谱。基础知识层通过对数字化、扫描件、手写件和多语言文档进行无损导入，构建具有Document-Section-Chunk层级结构的词汇图谱；智能层则提取与本体对齐的实体，并通过以溯源为锚点的领域图谱连接跨文档概念。LLM agents可在两个层级之间运行，根据具体问题选择最合适的检索路径。我们采用一种新颖的三层协议评估词汇层，用于衡量检索增强生成（RAG）系统在专有数据上的部署保真度；该评估基于Sanofi某小分子项目38份开发报告中整理出的505个问题进行验证。Tier-1选择题准确率达到95%，表明平台具有较强的可靠性；而更严格的Tier-2 LLM评审通过率为85%，且在比较类问题和全语料库问题上的表现有所下降，揭示出仅凭Tier-1准确率无法捕捉的失败类型。路由agent会根据问题类型在不同层级之间进行选择。我们预计，该协议将帮助未来的agentic平台设计者针对非公开数据库评估其系统，同时，基于图谱的架构也将在制药行业得到更广泛的应用，以将碎片化的文档存储库转化为结构化的流程智能。
文章链接： https://developers.openai.com/api/docs/guides/agents-api/overview 评论链接： https://news.ycombinator.com/item?id=49649213 分数：20 评论数：15
使用 Agents API 构建并上线云端代理。Agents API 是一项由 Codex harness 提供支持的托管服务，用于编排、长时间运行的会话和工具调用。
I got frustrated with the slowness of rsync and made an alternative that works faster by using multiple parallel connections, direct encrypted TCP when available, and other optimizations. I also added cool features like the ability to maintain a persistent ssh connection to the server for fast one-offs, the ability to download to your laptop while working in an ssh shell on a server, and the ability to do direct remote-remote transfers without forwarding your ssh agent (by using restricted ssh keys on the receiver that will only execute a specific request signed with the key on your laptop). It is also designed to be more robustly/flexibly scriptable than alternatives. Comments URL: https://news.ycombinator.com/item?id=49644955 Points: 12 # Comments: 10
认识 ChatGPT Work 中的 Data agent。连接公司数据，发现洞察，并使用自然语言借助 AI 构建交互式仪表板。
尽管大型语言模型（LLMs）展现出了令人印象深刻的能力，但由于固定的上下文限制，它们通常难以处理极长上下文。为解决这一问题，MemAgent 等序列式方法通过分段读取文本并迭代更新固定大小的记忆，扩展了有效上下文。然而，这种序列范式存在延迟高的问题，并且需要代价高昂的强化学习（RL）训练，可能导致模型在特定数据集上过拟合。为克服这些局限，我们提出了 ConvMem，这是一种无需训练且高度可并行化的框架，将长上下文推理重新表述为层次化卷积。受 CNNs 启发，ConvMem 将一个由特定查询提示的 LLM 视为卷积核。该卷积核以层次化方式总结文本片段，将推理路径从线性链缩短为对数级树结构。具体而言，ConvMem 集成了 Configurable Strides 和 Skip Connections，以确保可靠地捕获和传播证据；同时采用 Multi-Kernel Convolution，将复杂查询分解为相互解耦的语义通道。这一设计不仅缓解了错误累积，还实现了跨文本片段和推理线程的大规模并行化。在 RULER-HotpotQA 和 RULER-2WikiMultiHopQA 上的实验表明，ConvMem 优于无需训练的基线方法，并避免了在分布外任务中使用经过 RL 训练的模型时常见的对参数化先验过拟合的风险。
现实世界中的 GUI 使用通常涉及跨越多个设备和平台的工作流，需要传递中间结果、维护共享状态，并在异构环境之间进行协调。然而，现有的 GUI 基准测试绝大多数仅评估智能体在单一设备上执行静态定义的任务，因此基本未考察这类跨设备能力，导致对智能体是否已具备现实世界使用条件的评估过于乐观。我们推出 JarvisGUI，这是一个动态基准，用于评估 GUI 智能体在跨设备工作流中的表现；这类工作流要求在包括 Android、Windows 和 Ubuntu 在内的异构平台之间进行协同交互。具体而言，JarvisGUI 在轻量级类型系统下将 GUI 任务表示为输入-输出转换，从而能够自动组合多步骤、跨设备的工作流，并在统一框架内动态评估智能体性能。通过在覆盖多个操作系统的虚拟环境中评估智能体，JarvisGUI 揭示出，当前最先进的开源 GUI 智能体在现实世界工作流所需的状态转移感知能力、跨平台上下文推理能力以及长时程依赖管理能力方面存在明显不足，暴露出一个现有基准测试无法发现的关键能力缺口。
基础视觉语言模型（VLM）展现出对现实世界的广泛智能，但如何将这种智能转化为机器人控制仍然充满挑战。我们提出了 Show-Harness，这是一种具身控制框架，通过连接意图与动作的紧凑语义接口，使 VLM 能够“操控”机器人。Show-Harness 提供了离散的语义动作单元，VLM 可以自然地对其进行推理；与此同时，针对不同具身形态的解释器会将这些动作单元确定性地落地为机器人在本地执行的动作，从而使 VLM 直接负责细粒度的物理决策。借助同一接口，Show-Harness 展示了以下两种应用的可行性：（1）直接解锁闭源前沿 VLM，实现零样本机器人控制；（2）仅需数个 GPU 小时的微调，即可适配小规模开源 VLM，实现低成本部署。我们进一步开发了 GUMI（GUI Manipulation Interface），将同一语义动作空间扩展到基于 GUI 的演示数据采集，使人类和智能体无需专用遥操作硬件，即可跨不同具身形态“操控”机器人。大量实验表明，配备 Show-Harness 的 VLM 智能体能够在不同任务、具身形态和环境之间实现稳健泛化，性能优于具有代表性的智能体范式和 VLA 范式。这些结果表明，合适的接口可以从基础 VLM 中释放出可观的具身能力，而无需增加模型容量，也无需进行代价高昂的具身专属预训练。
To address insufficient contextualization, weak generalization, and poor scenario adaptation in tourism meteorological services, we propose SmartWeatherAgent--a unified three-stage architecture integrating intent recognition, hazard prediction, and reasoning-enhanced generation. The system fuses rule-based methods with large language models to parse queries at multiple granularities and employs a LightGBM model enriched with highland-specific features (e.g., wind speed abruptness rate), achieving an F1-Macro score of 0.605 with 1.60 ms latency on high-wind, precipitation, and low-temperature events. A 12-round micro-step prompt self-optimization loop boosts the composite warning quality score S_final from 4.2 (B01) to 8.9 (B12, +112%). Key improvements include a sharp rise in B08 from data source citation (6.5 -> 8.5), sustained high performance in B10 via physical mechanism explanation, and a peak scientific rigor score of 9.2 in B12 through explicit uncertainty statements. The system autonomously generates structured warnings that integrate causal mechanisms, spatiotemporal evolution, quantitative evidence, regulatory references, and confidence statements--enhancing professional depth, logical rigor, and scientific soundness, and advancing meteorological services toward proactive perception, explainable decision-making, and intelligent agency.
AI智能体正日益参与网络自动化，它们可以通过受控的运维接口发起配置变更，并评估由此产生的状态。然而，运营网络通常横跨众多设备和管理域。要实现运营商的意图，就需要协调具有不同权限范围的智能体；这些权限范围规定了它们可以访问的资源、可以调用的操作，以及可以观测的网络状态。这种划分虽然限制了错误操作的影响范围，却使评估全网结果所需的证据变得碎片化。某个智能体提出的配置操作成功执行，并不能证明远端设备已按预期响应，也不能证明路由变更已到达所需设备。一次有效的观测在后续变更发生后也可能变得过时。在宣布协调操作完成之前，必须由可信的保障层从所需权限范围收集当前观测，并确定这些观测是否共同支持运营商预期的全网结果。 为解决完成准入问题，我们提出了 EvidenceNet，这是一种运行时保障层，用于判断协调后的智能体操作是否已实现运营商的网络意图。其代理收集完成契约所要求的变更后观测结果，而准入门则检查证据是否来自所需权限范围、是否保持最新，以及是否满足任务规则。验证器智能体还会对观测内容进行额外评估。在真实运行的路由网络上进行的实验表明，变更后的状态检查能够识别出仅凭配置操作记录无法确认的成功结果。受控干预进一步表明，当其他方面令人满意的观测来自错误来源、被替换或已经过时时，EvidenceNet会拒绝宣布操作完成。
Traditional software delivery assumes a static paradigm: code is constructed prior to execution and deployed as a fixed artifact. We present Agentic Just-In-Time Software Construction (A-JIT), a paradigm that replaces static binaries with dynamic, software systems that can perpetually evolve to meet changing demands. In A-JIT, an application is an integrated assembly comprising code, a runtime harness, and an embedded AI agent that continuously observes system usage and live execution traces. Much like a traditional JIT compiler specializes machine code to runtime execution paths, A-JIT specializes software logic, workflows, and tool interfaces to meet the specific needs of the end-user. By integrating synthesis directly into the ambient application lifecycle, A-JIT enables applications to dynamically construct missing implementations, generate new capabilities on the fly, and continuously adapt to end-user behavior. We demonstrate how this model supports trace-driven human-AI co-construction and opens a new design space for adaptive, self-evolving software.
持久化语言智能体需要存储的经验在一段时间内持续可用，而每个回答都需要适合特定问题的证据。已经被取代的事实可能会误导对当前状态的回答，却仍可能是历史查询所必需的。我们提出了 RD-Forget，一种无需训练的框架，用于分离智能体存储的内容与使用的内容。保留的源档案保存观测结果，而查询条件化的记忆视图则控制这些观测结果对当前回答的影响。一个冻结的语言模型整理器提取相关证据，将事实归入语义槽位，并保留多跳推理所需的关系。在当前状态上下文中，同槽位替换链接会抑制已被取代的值，而意图感知检索则使较早的证据能够再次发挥作用。率失真公式引导在记忆预算内构建回答时视图。实验覆盖对话记忆、知识更新、事实整合、长上下文推理以及个性化任务，并采用统一的回答流程。结果表明，准确回答既依赖于构建与查询相关的证据，也依赖于对过时备选信息的控制。不具备遗忘机制或查询条件化的配置得分缺失最大，而槽位分组、历史访问和关系保留发挥着互补作用。在保留历史的同时选择性地控制其使用，为适应不断变化的事实和未来问题提供了一种实用方案。
带有可验证奖励的强化学习（RLVR）推动了语言模型在数学和代码等领域的推理能力发展，因为这些领域的客观答案易于验证。复杂数据上的诊断推理则不具备这一优势：确定异常的真正原因通常需要成本高昂的专家调查，而且事后也可能仍然存在歧义。我们探讨能否通过工程手段利用这种验证不对称性。我们抽样一个干预，将其注入受控模拟器，并生成该干预会产生的观测结果。隐藏的干预提供了预言机标签和客观奖励，而智能体仍必须调查嘈杂、混杂且分散的证据。 我们在 TRACE 中实现了这一方法。TRACE 是一个数字广告诊断环境，包含 12 种根本原因以及细粒度的细分归因。智能体使用 Python 和 SQL 调查每个回合，并必须识别根本原因，以及在适用时确定受影响的细分归属。在一个留出的 235 回合测试集上，表现最强的提示基线 Claude Opus 5 达到 0.686 的 FullAttr@1。监督微调将 Qwen3.5-35B-A3B 的得分从 0.159 提升至 0.637，随后使用合成奖励进行强化学习，得分进一步达到 0.757，超过了所有评估的提示基线，包括前沿闭源模型和经过提示的 Qwen3.5-122B-A10B 模型。所得策略使用的工具调用次数也显著少于经过提示的 35B 基础模型。这些结果表明，能否获得可扩展且客观的训练信号，可能比模型规模本身更为关键。更广泛地说，基于模拟的验证可以使原本存在歧义的诊断推理任务适用于可扩展的强化学习。
文章 URL： https://www.engadget.com/2254419/muse-the-band-lost-its-social-media-handles-to-muse-meta-s-new-ai-agent/ 评论 URL： https://news.ycombinator.com/item?id=49636345 点赞数：38 评论数：8
文章链接： https://arxiv.org/abs/2609.09153 评论链接： https://news.ycombinator.com/item?id=49629868 得分：34 # 评论数：10
大家好，我是 Dimitris。 这是我构建并用于运营自己企业的企业 OS，任何人都可以安装并免费进行自托管。可以把它理解为将 Claude Code、Cowork 和云端会话整合在一个自托管应用中。 它从设计上就是一个多租户应用，许多人可以通过 4 种不同的协作模式与企业智能体协作；它可以使用你的 Anthropic 或 OpenAi 订阅运行，也可以使用本地模型。 每个智能体都可以运行在 Claude Code 或 Codex CLI 上，作为服务器上的持久进程运行于内核沙箱（bubblewrap）中，并始终启用网络隔离（pasta）；每个智能体都有自己的工作区、记忆、日程安排和工具。同样的智能体也可以配置为通过一个出站 WebSocket，在任意远程计算机上以完全相同的方式运行（无需入站端口，也不需要 VPN）。 智能体已经内置了许多功能。例如，它们可以通过 Twilio 或你自己的 Asterisk 接听和拨打电话，可以编辑视频以及 Excel、Word、PPT 文件，还可以借助 Collabora 直接在聊天中预览这些文件，此外还有更多功能。 只需一个安装脚本和一个 Docker Compose 文件即可开始使用，无需注册。该许可证采用 Fair Source，所有代码均为公开代码，并且最多支持 5 名用户的自托管是免费的。1.6.0 已于今天发布。 AI 免责声明：OtoDock 的大部分代码都是使用 OtoDock 自身、运行 Claude Code 编写的。 我很希望大家阅读代码、了解一下沙箱模型，并告诉我是什么因素会阻止你在自己的硬件上运行它。 GitHub： https://github.com/OtoDock/oto-dock 网站： https://otodock.io 评论链接： https://news.ycombinator.com/item?id=49630606 得分：3 评论数：0
嗨，HN。我是 Christian，Frigade（YC W23）的联合创始人之一。我注意到，许多应用内 AI agent 很难真正理解它们所在的产品。 例如，假设用户询问 agent 如何在某个 SaaS 产品中完成某项操作。理想情况下，agent 可能会回复说它有一个工具可以完成这项任务，然后直接替用户自动完成整个操作。这无疑是很好的结果。 但很多时候情况并非如此。可能没有针对该确切任务的工具调用，或者用户的问题最好通过特定的 UI 工作流或界面来解决。在这些情况下，许多 agent 往往会退回到对帮助中心进行基础 RAG，有时甚至会搜索互联网，以了解它们自己的产品。这一过程可能非常缓慢，而且如今产品迭代速度往往快于帮助中心文章的更新速度，因此这些文章大多数时候都已经过时。更糟糕的是，没有人喜欢阅读一长串项目符号列表，然后再将其对应到 UI 上。 我的工具（Assist API）通过一次工具调用解决了这一空缺，定义如下： const frigade_guide_tool = { description: 'Call this tool to answer product questions or guide the user through a task.', parameters: { query: { type: 'string', description: 'What the user is asking or wants to do', }, }, run: ({ query }) => frigade.assist({ query }), } 这是我录制的一个演示，介绍如何使用 Vercel AI SDK 进行设置： https://www.youtube.com/watch?v=9WQ0UbLjC6I 调用后，该工具会执行以下操作： 1）收集用户在屏幕上看到的内容、用户权限、功能标志等上下文信息。然后执行以下某一项操作： 2a）如果问题可以解决：生成屏幕引导，指导用户如何解决给定问题 2b）如果是概念性问题：返回文本，向父 agent 说明如何解决问题 2c）拒绝（即无法提供帮助） 该工具如何知道该做什么？ 该工具通过浏览器 agent 学习给定应用的 UI。你需要为自己的软件提供一个测试账户（即 staging 或 preview 环境），然后浏览器 agent 会登录并遍历整个产品。之后，它会建立自己的应用工作方式地图，可以据此回答任何与产品相关的问题，或说明如何在 UI 中从 A 进入 B。它还会根据这张地图自行编写文档。该 agent 会按计划重新运行，也可以通过 CI/CD 触发。 文档及更多详情： https://frigade.com/assist-api 评论 URL： https://news.ycombinator.com/item?id=49627872 得分：5 评论数：0
I can close an eye on using LLMs for research only. These days I feel I am the only one that hasn’t changed their way of writing software at all. I’d honestly switch careers rather than manage agents. But I’m currently out of work (contract ended) and wondering if there is still a sane place to work, or is it truly time to pivot to another career. Comments URL: https://news.ycombinator.com/item?id=49626431 Points: 5 # Comments: 0
While research on recursive self-improvement (RSI) has predominantly automated model training pipelines, reliable autonomous development demands a missing pillar: post-hoc monitoring and auditing to understand what models learn and ensure safe alignment. Mechanistic interpretability tools are essential to bridge this gap, among which Sparse Autoencoders (SAEs) serve as a cornerstone by isolating interpretable features for model inspection and steering. In this paper, we introduce SAEScientist-Bench to evaluate whether AI agents can act as scientists utilizing SAE tools for autonomous mechanistic discovery. Given a target concept, an agent designs contrastive probes and navigates a Gemma Scope dictionary of 131K+ features in Gemma-2-9B-IT to discover the optimal feature, evaluated against curated expert reference features anchored on Neuronpedia across activation rank, concept selectivity on contrastive texts, and causal steering. Across 10 agent configurations and 20 tasks, frontier agents demonstrate genuine discovery capabilities and lead different evaluation dimensions, but remain well behind the expert baseline, approaching expert levels on separating target concepts from contrastive controls while lagging substantially in causal generation steering. Further analysis reveals that although agents can design contrasts to rule out spurious candidates, they frequently misinterpret experimental measurements. These results establish experimental model understanding as a measurable capability for closed-loop autonomous AI R&D. Our code is available at https://github.com/Trae1ounG/SAEScientist.
Long horizon Large Language Model (LLM) agents rely on external memory systems to preserve user preferences and task knowledge across extended interactions. Conventional retrieval mechanisms optimize semantic compatibility rather than downstream utility, frequently introducing outdated, misleading, or conflicting evidence into the active context. We present MeClear, a task conditioned memory clearance framework that identifies memories featuring negative downstream utility through cooperative attribution and selectively suppresses them from agent execution. MeClear combines Leave One Out screening with sampled cooperative Shapley attribution to distribute utility across interacting evidence, effectively resolving redundant conflict masking where single removal evaluations fail. Utilizing attribution rankings, MeClear executes a query scoped minimal clearance strategy over a nested filtration, verifying task recovery on the cleared context without permanently altering the persistent memory bank. Comprehensive experimental evaluations across ten long dialogue memory pools demonstrate that MeClear achieves a target recall of 85.9% and an overall task recovery rate of 82.3%, representing a 25.5 percentage point improvement over Leave One Out (LOO) baselines.
执行反馈可以引导编码智能体完成正确的代码库修复，但前提是测试能够捕捉问题所要求的行为。智能体生成的测试可能编码不完整或不正确的行为目标；当同一条轨迹同时编写补丁和测试时，它们的错误可能相互一致，从而产生虚假的信心。我们提出 ExecCritic，它将测试--验证--修订脚手架与用于在其中训练智能体的角色特定强化学习方案相结合。该脚手架将测试构建与源代码修复分离：Test agent 独立生成代码库原生测试，一个 fail-closed 测试框架对其进行合格性判定并冻结，Repair agent 则基于这些测试的执行反馈修改源代码而不更改测试。两个角色都使用 Qwen-3.5-35B-A3B 作为骨干模型，并分别进行训练。在 Learn to Test 中，Test agent 学习生成行为上有效、能够区分正确与错误补丁的测试。在 Test to Improve 中，Repair agent 学习直接解决任务以及基于反馈进行修订。在 SWE-bench Verified 上，测试质量决定反馈是否有帮助：在固定基础 Repair agent 的情况下，来自基础 Test agent 的测试会使解决率从无测试基线的 61.2% 降至 57.3%，而来自 GPT-5.6-sol 的测试则将其提升至 65.3%。角色特定的后训练将 Qwen Test agent 的 Base-to-Gold 成功率从 22.2% 提升至 62.2%；组合两个后训练的 Qwen agent 达到 72.6%，在评估时无需更强模型或 Oracle 反馈的情况下，比原始无测试基线提高了 11.4 个百分点。代码已公开发布于 https://github.com/MSR-Orchard/execcritic.
Agent harnesses（围绕模型的 system prompt、tool set、execution hooks 和 context-management scaffolding）是决定 agentic 任务成功的关键因素。自动化 harness 演化可以使较小模型以 frontier-model 成本的一小部分在特定领域任务上表现良好。由于 harness 和模型权重都会塑造行为，我们探讨应如何结合 harness 演化与轻量级 fine-tuning。在七项企业 agent 任务中，我们首先使用较弱模型演化出一个 harness，随后发现更强的 expert 往往能更有效地使用它，这表明 expert 监督可能弥合剩余差距。然而，在演化后的 harness 下用 expert 的完整轨迹训练较弱模型会适得其反：在 Qwen3-Coder 和 Gemma 4 上，七项任务的性能全部回退 4 到 30 个点，尽管同样的流程在未演化的 harness 下是有帮助的。我们的分析表明，imitation 会迁移知识并增加 scaffold 使用，但会破坏模型-harness 适配：较弱模型采用了 expert 的规划策略，却不具备执行该策略的能力，并且不再匹配围绕其原生规划风格演化出的 harness。因此，我们开发了一条 on-policy expert-correction 流程，由一个元层级 MLE agent 自动化执行；该流程会定位较弱模型自身 rollout 中失败的 turn，并要求 expert 仅重写该 turn。这保留了模型的规划风格，并结合了 harness 演化与模型适应的收益。我们的结果识别并解决了 harness 与权重更新之间的一种冲突来源，为特定领域企业任务上的经济型共同演化提供了一种保持兼容性的方案。
Large language models are increasingly deployed as agents that plan over long horizons and act through external tools. Most agents select actions through unconstrained generation over an accumulating history, leaving implicit the procedural knowledge of what to do, in what order, and under which conditions. As trajectories lengthen, agents can lose track of their objectives, invoke tools out of order, and repeat unproductive actions. We introduce the Procedural Graph: just as a knowledge graph organizes factual knowledge into (entity, relation, entity) triplets for what-is questions, a Procedural Graph organizes procedural knowledge into (procedure, relation, procedure) triplets for what-to-do questions. At each decision step, the framework localizes the agent's active node, and a guidance model translates the surrounding subgraph into step-level situational guidance that biases the solver's next action without dictating it. The graph is self-evolving: an LLM refiner contrasts failed trajectories with successful ones and edits the graph's topology and attributes, committing edits that preserve or improve held-out validation performance while retaining rejected ones to discourage repetition. Starting from a minimal skeleton, the loop builds graphs that match or surpass hand-designed ones. It can also repair a flawed expert prior. Across multiple datasets, task types, and LLMs, the Procedural Graph delivers consistent gains over memory-based baselines, and self-evolution further improves performance without manual engineering.
由大型语言模型（LLM）驱动的智能体平均而言可能表现准确，但在生产环境中却不可靠，这种差异虽已被观察到，却基本未得到解决。在 AppWorld 基准测试中，使用 GPT-4.1 的 ReAct 智能体在同一任务上运行五次时，只有 53% 的情况下能够五次全部成功，尽管其单次运行通过率平均为 77%。我们将这 24 个百分点的差距称为一致性差距，并认为解决这一问题是实现可信赖的 AI 智能体部署的先决条件。我们提出了一种自我进化智能体框架，通过识别智能体轨迹中不稳定且一致性较低的步骤，并将其转化为智能体未来运行时可以调用的情景记忆，从而缩小这一差距。该框架的核心是 Consistency Analyzer（一致性分析器），用于定位轨迹在多次执行中可能发生结果翻转的位置及原因；以及 Guideline Generator（指导原则生成器），用于将诊断结果转化为有针对性的指导原则，将其写入记忆，并在未来执行相似任务时注入智能体。在使用 ReAct/GPT-4.1 的 AppWorld 上，我们的框架使五次运行全部成功的任务比例在同任务评测中提高了 16 个百分点，在相似任务泛化中提高了 13 个百分点。
智能体技能为冻结的语言模型智能体提供了一种轻量级方式，使其具备领域知识和程序性指导，但获取高质量技能仍然成本高昂且难以规模化。由专家编写技能需要投入大量人力。近期的技能自演化方法通过自动化迭代循环，利用执行反馈来修订技能，但其启发式更新策略往往导致优化过程不稳定、迭代效率低下。我们发现，实现稳定且高效的技能自演化面临两项挑战。方向稳定性要求有效的修正能够不断累积，而不是被局部迭代反馈覆盖。更新自适应性要求每次修订的范围能够反映近期案例级改进的一致性。我们提出 SkillAdam，这是一种受 Adam 启发、用于优化离散且不可微技能文档的框架。作为 Adam 一阶矩的功能类比，优化记忆记录已识别的问题以及此前解决方案尝试的结果，从而稳定更新方向。作为 Adam 二阶矩的功能类比，波动性驱动的编辑预算跟踪近期案例级改进的历史加权变化，并自适应地控制更新幅度。在涵盖短期和长期任务的七个基准测试中，SkillAdam 以更加稳定的优化动态取得了当前最佳性能。与以往方法相比，它还以显著更少的优化迭代次数和更低的成本获得了更强的技能。代码仓库：https://github.com/ruc-datalab/SkillAdam
确保自动驾驶的安全性是一项关键挑战。基于场景的测试是一种用于验证自动驾驶系统（ADSs）的系统化流程，但目前仍是一个碎片化的模块化流水线：场景生成、检索、修改、ADS 执行和结果分析分别由交互很少的独立工具完成。大型语言模型（LLM）智能体已在 ADS 的感知、规划和控制等子系统中展现出良好应用前景。然而，尚无先前工作能够通过统一的 LLM 智能体框架覆盖 ADS 的完整基于场景的测试流水线。我们提出了 PlannerForge，这是一种 LLM 智能体框架，扩展了基于场景测试的全部阶段（从场景生成到 ADS 评估），并新增两个由 LLM 增强的阶段：ADS 增强和 ADS 基准测试。我们使用 10 个开箱即用的 LLM，在 5 种提示条件下，对所有任务（生成、选择、修改、模块路由、规划器测试和增强）评估了 PlannerForge。各任务的最佳得分范围为 0.88 至 1.00；在大多数任务上，开源的 20-35B 后端模型能够媲美商业 API。Qwen3.6:35B 等开源模型在五项任务中的三项上达到了商业 API 的水平。将各模块端到端串联后，仍可保留 83% / 78% 的种子查询（商业模型 / 开源模型）。在自然语言生成方面，PlannerForge 优于 Scenario Factory 2.0（Finkeldei 等，2025）（200 个查询中生成 193 个可执行场景，而非 144 个），并实现了所请求城市、道路和车辆属性的 92-96%。在排名第 1 的选择准确率上，它优于 BM25（Robertson 和 Zaragoza，2009）（92.0% 对 67.5%）；在生成物理有效修改方面，它优于 From-Words-to-Collisions（Gao 等，2025）（≥94% 对 31%）。在 N=400 时，成本调优在不进行领域特定微调的情况下，将规划器成功率从 50.4% 提升至 70.2%，并将碰撞率从 19.0% 降至 8.4%。
本文介绍了 Gander，这是一个端到端模型，在单一框架内统一了全模态感知、实时交互和智能体能力。与基于轮次的传统范式不同，Gander 持续接收包括视频、语音和文本在内的多模态流式输入，从而能够在日常对话和复杂的工作流型智能体场景中实现自然的全双工交互。用户可以随时打断模型，模型也可以主动提供中间反馈或提出后续问题。为原生支持这些能力，Gander 采用了两项关键架构设计：1）采用 Cerebellum-Brain 协作框架，其中 Cerebellum 负责实时交互和全模态对话能力，而 Brain 负责复杂推理和更高层次的智能体任务。两个组件通过工具调用和智能体编排运行时持续交互。2）Cerebellum 构建于流式 Thinker-Talker 架构之上，在分块层面将用户输入和模型输出进一步扁平化为有序 token 流，为低延迟、持续性交互提供统一表示。我们从四个维度对 Gander 进行了全面评估：对话能力、全模态理解能力、交互能力和智能体智能。内部人工评测表明，Gander 在实现具有竞争力的全模态交互性能的同时，保持了 SOTA 开源模型自然且富有表现力的语音对话能力。Gander 在具有挑战性的真实场景中也展现出稳健性，包括背景噪声干扰、多方交互和附和式沟通。我们将 Gander 及其模型、代码和数据一并开源，以促进社区进一步开展研究与开发。
Modern science and engineering increasingly rely on time-varying data, yet the mathematical tools used to model temporal phenomena are often developed within separate disciplines, obscuring common principles and limiting the transfer of ideas across fields. This chapter presents the theory of narratives, an abstract framework for time-varying objects of any mathematical kind that supports both theoretical investigations and applications. To illustrate this perspective, the chapter develops three vignettes, each illustrating a different research direction. The first addresses a general concern: What information loss can occur when switching between different representations of temporal data? The second concerns structural and algorithmic approaches: How can we systematically decompose time-varying data into simple pieces and obtain invariants describing its structural complexity? The third is an application to control theory: How can we model multi-agent systems with switching communication topologies? More important than any individual vignette, the central message of this invitation is that a suitable abstract perspective can organize and guide research across remarkably diverse mathematical and scientific domains.
在异构集群之间高效聚合和编排计算能力以执行 HPC 工作流，面临四项实际挑战：在由不同机构独立管理的集群之间保留工作流上下文，在站点之间传输大型数据集，针对站点特定的环境和调度器策略进行推理，以及利用实时队列和资源状态高效调度任务。为此，我们设计了 Diamond Agent，这是一种以类型化技能作为接口的智能体式系统，可实现跨异构集群的 HPC 工作流智能执行。Diamond Agent 提供面向智能体的工作空间和技能，统一跨站点资源发现、资源规格定义、数据传输、任务执行和结果获取。一个集中的 Diamond Agent 实例即可操作多台超级计算机，无需分别部署在每个登录节点上。Diamond Agent 将高层智能体操作转换为符合各站点要求的具体执行，通过 Globus Transfer 传输数据，并利用实时系统能力和队列信息选择可行的部署位置。其事件驱动的持续执行机制将智能体操作与长时间运行的批处理作业解耦：持久化服务监控远程执行，仅在结果或与决策相关的事件可用时恢复智能体运行。我们使用来自四台生产级超级计算机的 27 小时遥测数据，以及由 83 个作业组成的 19 轮匹配多站点提交实验进行评估。与固定站点基线相比，Diamond Agent 将相对于观测到的最快部署位置的额外完成时间中位数从 42 秒降至 4 秒，减少了 10.5 倍。
安全运营中心（SOCs）需要处理大量工单，其中大多数是无需进一步调查的低关注度事件。这项任务具有重复性，而且大量工单之间存在相似性，因此非常适合采用基于生成式 AI 的自动化方案。我们在一个 SOC 中开展了为期一年多的实地研究，创建并部署了一款利用大语言模型的智能体式 AI 助手。在 SOC 的日常工作中，研究人员通过参与和互动推动了该 AI 助手的设计。在实地研究的最后四个月中，我们邀请 SOC 分析师使用该助手。我们分析了分析师对该助手的使用情况，发现其输出在超过 90% 的情况下都被分析师重新用于工单结案报告。研究结果表明，如果与目标用户在“一线实战中”共同设计，SOC AI 助手就能超越“又一个工具”的定位，成为一个在应对各种工作负载的过程中与人类用户共同演化的系统。分析师自然而然地开始塑造 AI 助手的行为，使其符合自身的特定需求。我们的数据表明，人类分析师越多地塑造 AI 助手的行为，就越愿意信任 AI 系统的输出，从而提高工作效率。
LLM-as-a-Judge 已成为评估自然语言生成的一种颇具前景的范式。然而，与此类评估相关的不确定性仍在很大程度上缺乏研究，这限制了其在现实应用中的可靠性。尽管保形预测为不确定性量化提供了一个有原则的框架，现有方法通常将其应用于单个 LLM 评审器，忽略了使用不同 LLM 评估器所引入的变异性。在本文中，我们提出了一种面向多智能体 LLM-as-a-Judge 评估的鲁棒不确定性估计框架。我们的方法针对来自多个 LLM 的基于 LLM 的得分构建保形预测区间。通过考虑不同 LLM 评审器给出的区间，我们获得了更加稳定可靠的不确定性估计。大量实验表明，我们的方法能够生成具有覆盖率保证的有效预测区间，并且在多个评审器之间进行基于区间的聚合能够带来更加稳定的评估结果。
图代理式检索增强生成将结构化证据与自适应控制器相结合，使系统能够规划检索、遍历关系、验证中间断言、委派子任务并使用工具。当答案取决于跨文档、实体、时间或机构的关系时，这种组合非常有用，但也会产生相互耦合的失败路径：图构建中的缺陷可能成为被检索的证据，改变后续的控制决策，并向具有重大影响的结果传播。我们考察了在社会公益场景中应如何设计和评估此类系统；在这些场景中，除了答案质量之外，信息新鲜度、授权、可追溯性、监督和救济同样重要。我们按照图基础、图生命周期、代理功能、协作模式和权限边界对相关文献进行组织，并区分基于图的检索与依赖观测的图控制。随后，我们将已有报告中的风险综合为一条从证据到行动的失败链，并提出一套构造式保障蓝图，其中包括针对证据、检索、推理、能力与委派以及结果的五项接口契约。这些契约使来源溯源、时间有效性、授权、不确定性和可恢复性在系统边界处得到明确体现。一个公益信息设计示例展示了该框架如何约束图结构、权限、弃答机制和运行授权。最后，我们提出一项评估议程，涵盖图断言、轨迹、主张、协作和结果。
高风险 AI 系统的提供商必须保存能够使决策可追溯的记录；而对于智能体系统，尚未明确这些记录必须包含哪些内容，才能实现事后因果归因。我们先给出估计量框架，再说明其失效的条件。我们将既有工作所测量的边际总效应，与能够隔离某一步自身贡献的共同随机数总效应区分开来；在固定下游的条件下加入自然直接效应，并通过手工推导检验这些估计量。随后，两种被估计量都以相同方向失效。在边际估计量下，在我们构造的链条的每一次运行中，一个因果上不起作用的步骤与决定性步骤具有完全相同的总效应；这是一种代数恒等式，而非某一次抽样中的巧合。在共同随机数条件下，决定性步骤在执行步骤发生翻转的那些运行中——约占十分之一——返回的结果恰好为零，但其直接效应为 0.25，且可以明确证明它确实发挥了作用；精确为零并不能证明某一步什么都没做，因此我们将这一点放在正文中，而不是局限性部分。我们推导出一种耦合方法，使上下文发生分歧后仍可估计直接效应，并给出其衰减的封闭形式；我们还表明，用于构建一种自然排序的介导份额，在抑制意义下并不是一个真正的份额：当直接路径与介导路径相互对抗时，该数值会超过 1，并将一个受到抑制的组件排在纯粹介导者之上。我们发布了该差异实验的预注册方案，而不是实验结果，因为它所需的在线运行管线在研究时间窗口内尚不可用。我们提出了此类备案所需的可追溯性规范，以应对 Regulation (EU) 2026/1744 的时间表所造成的空档：Article 86 赋予的解释权自 2026 年 8 月 2 日起已适用，而能够为解释提供证据的 Article 12 日志记录要求和 Annex IV 文档要求，则被推迟至 2027 年 12 月 2 日。
文章 URL： https://ai.meta.com/muse/ 评论 URL： https://news.ycombinator.com/item?id=49615537 得分：7 评论数：0
文章链接： https://rohanadwankar.github.io/posts/platforms.html 评论链接： https://news.ycombinator.com/item?id=49605644 得分：3 评论数：1
文章 URL： https://danluu.com/agentic-testing/ 评论 URL： https://news.ycombinator.com/item?id=49605246 积分：15 评论数：0
Public vulnerability databases collect rich information about known software flaws, including their weakness types, affected components, and related patches. Fixing commits provide the exact code changes that removed these flaws. While these records capture why the original code was unsafe, they are documented mainly for human inspection rather than automated reuse. Consequently, the same unsafe conditions may still exist elsewhere in code without a known advisory, leaving much of this detection knowledge unused. We present BUGSTONE-E2E, a framework that transforms vulnerability history into executable detection rules and validates their findings. First, BUGSTONE-E2E mines reusable rules from verified fixing commits, capturing scan anchors, fix semantics, and CVE provenance and organizing them by CWE and language. Second, detection follows a funnel-shaped pipeline: early stages process a large pool of candidates using lightweight analysis, while later stages apply increasingly capable and expensive models to a shrinking set of targets. Specifically, BUGSTONE-E2E first enumerates call sites matching rule anchors using Tree-sitter, then removes benign sites using lightweight heuristics without LLM calls. Next, LLM-based agents inspect the remaining candidates guided by the rule. Following this inspection, the system re-triages surviving candidates and builds runtime verifications, then generates scope-checked patches validated via two-sided differential tests. Using 19,325 high-severity CVEs from 2022 to 2026, BUGSTONE-E2E identifies 2,710 fixing commits and constructs 1,033 detection rules across 56 CWE families, packaged into 172 skills. When applied across 14 programs, it produced runtime evidence for 644 findings. These results demonstrate that CVE history can be turned into an executable workflow, transforming past vulnerabilities into reproducible detection and repair.
Model upgrades are routine; memory migrations are not. An agent can keep the same memory store and still forget: a new model may interpret old notes differently, mixed embedding versions may break retrieval, and repair may fail without the original evidence. We compare memory as the same history is preserved verbatim for long-context reading (LC-RAW), divided into chunks for retrieval-augmented generation (RAG), compressed by a model into natural-language notes (NOTES), or normalized into a fixed-schema knowledge graph (KG-fixed). The study uses 48 synthetic histories with randomized answer codes, exact scoring, and two open-weight models with sub 10 billion parameters. Our measurements show that fixed-schema structures transfer reliably, with KG-fixed accuracy changing by only $+0.0004 \pm 0.0020$ following a writer swap. Conversely, compressed NOTES exhibit high model coupling, with accuracy shifting asymmetrically by $+9.91$ or $-13.28$ percentage points depending on the specific migration direction. In RAG systems, partial embedding migrations using a 50/50 mixed index capture only a 4.96-point accuracy improvement, forfeiting the majority of the 11.90-point gain achieved through full re-embedding. Diagnostic decomposition attributes 80% ($0.467 \pm 0.014$) of the NOTES accuracy deficit to information lost during initial construction, whereas retrieval failures drive 81% ($0.364 \pm 0.012$) of the RAG deficit. Finally, store-only repair of NOTES fails to reach a 90% performance recovery target in all 48 test cases, whereas retaining the raw source history enables successful recovery in 34 of 48 cases for one tested direction. These findings highlight the necessity of direction-specific migration testing, strict embedding space isolation, and the retention of source histories for memory repair.
Machine-learning performance modeling is a uniquely hostile terrain for long-lived software: the assumptions baked into today's abstractions are invalidated by tomorrow's models and systems, forcing perpetual refactoring of performance-modeling frameworks. Meanwhile, AI coding agents have become fast and capable enough that regenerating an entire library is cheaper than paying down the tech debt of incrementally patching it. We describe SMART, a rigorous symbolic performance-modeling library for ML systems whose main branch contains almost no code: the repository is a DAG of self-contained natural-language design docs, coding sub-agents regenerate the implementation from only the docs on new version updates, and every human change is a natural-language edit to a doc--self-documenting by construction. Two ingredients make regeneration reliable: (i) a design-doc style built around step-by-step worked examples that act as in-context demonstrations for the generating agents, and (ii) a minimal, recursively defined operator IR with symbolic (SymPy) cost expressions, a fast analytical roll-up mode for large sweeps, and a slow modulo-scheduling mode for fine-grained schedule studies. Regenerated implementations reproduce hand-audited reference models--including DeepSeek-V3 serving on a TPU pod slice--to round-off precision, suggesting that design docs--not code--can be the durable artifact for ML-systems co-design tools.
computer-use 智能体已在 OSWorld 和 AndroidWorld 等基准上取得进展，但它们仍然主要通过 GUI 进行操作，往往会产生低效的轨迹。真实世界的计算机工作是混合式的，结合了视觉状态检查与精确、高吞吐量的命令行操作，因此有能力的智能体必须在共享的应用状态上协调这两种模态。然而，可扩展的混合环境仍然稀缺，因为要在真实应用上同时支持 GUI 和 CLI，通常需要针对每个应用进行大量手工工程。现有智能体在互补使用这两种接口方面也表现不佳：原生 CLI 智能体缺乏对涉及界面状态或布局任务的视觉感知，而原生 GUI 智能体在更适合通过命令执行的操作上效率较低。我们提出 CUA-Universe，这是一条可扩展的环境到数据流水线，可将真实桌面软件转化为混合 GUI+CLI 环境。App-Forge 将应用适配为可复现的 VM 和其发现、封装或生成的命令行界面，扩展到 16 个应用；Task-Weave 基于种子文件上的可复用操作，合成难度可控的多样化混合任务；Path-Steer 沿高效的混合路径引导 rollout，并收集经过验证的轨迹用于后训练。在这些数据上训练后，模型行为从低效的 GUI 交互和脆弱的 CLI 脚本编写，转向有效的 GUI+CLI 协同。我们的 9B 模型在 CUA-Verse 上同时提升了成功率和效率（Score +39.3 分；步数 -37%；tokens -60%），在 OSWorld 上（SR +16.8 分；步数 -57%；tokens -44%），以及在 OSWorld-MCP 上（Score +7.84 分；步数 -27%；tokens -30%）。CUA-Universe 为构建更有能力、更高效的 computer-use 智能体提供了一条可扩展路径。
LLM 决策组件在 agent 工作流中运行时，常会给出与行动相关的建议或判断，并附带解释。操作人员可能会利用这些命名因素来监控系统、诊断错误，或决定何时升级处理某个输出。这种用法假设这些解释与组件可观察到的决策行为一致。我们测试了这些命名因素的两种解释：必要性，即改变某个因素会改变输出；以及充分性，即在移除其他可变信息的同时保留该因素会保持输出不变。我们在两个合成用例中评估这些解释：向客户推荐顾问，以及判断提示词是否具有有害性或风险。模型返回一个输出以及最能影响它的前三个因素。受控的 black-box 干预通过衡量改变某个因素后输出发生变化的频率来估计其必要性得分，并通过衡量保留某个因素后输出得以保持的频率来估计其充分性得分。在来自 Claude、GPT 和 Gemini 系列的八个模型中，引用排名与必要性和充分性得分之间的平均斯皮尔曼相关系数在顾问推荐任务中分别为 0.349 和 0.354，在提示监控任务中分别为 0.431 和 0.580。此外，在必要性条件下，57.6% 的顾问响应中，某个未被引用的因素得分高于得分最低的被引用因素；在充分性条件下这一比例为 58.1%。对应的提示监控比例分别为 25.8% 和 8.9%。被引用的前三项包含有用信息，但并不能可靠地识别在必要性或充分性意义下、经测量影响最强的三个因素。该框架为代理监督中使用的解释提供了一种 black-box 可靠性检查，同时仍限定在单个 LLM 决策范围内。
数据主权法规日益要求公共机构部署开源、可本地部署的 LLM agents，这些 agents 需要通过实时政府 API 串联多次工具调用。然而，开源模型在这种多步场景中始终表现不佳，现有基准也没有衡量这一差距。我们提出 Korean Open Public API Benchmark (KOPA-Bench)，包含 145 个真实世界任务。为弥补这一差距，我们提出 EDGE，一种由实时执行驱动的、用于工具调用数据合成的 Execution-grounded Dynamic Graph。EDGE 构建一个图，表示每个工具的输出如何作为另一个工具的输入，只保留那些在实际调用实时 API 时成功的连接，并沿着这些经过验证的连接遍历，以合成可执行的多步轨迹。我们在生成的数据集上通过 GRPO 对 9B 模型进行微调后，其性能几乎达到同一系列中未微调的 27B 模型，不仅在 KOPA-Bench 上有显著提升，在 BFCL 基准上也同样如此。
This paper addresses navigation by composite heterogeneous robots in a decentralized system when policy reasoning and local control operate at different update levels. In a NetLogo--Python implementation, three robots share motion dynamics but use different LLM backends. Each robot independently combines a large language model (LLM) policy agent, an Upper Confidence Bound (UCB) bandit, and a Double Deep Q-Network (Double DQN) controller; no central LLM generates team actions. LLM inference is confined to round-level policy generation and refinement rather than tick-level action selection. The robots perform cross-LLM communication through a shared round summary containing policies, outcomes, and learning feedback. UCB performs refinement-mode selection, and the policy-conditioned Double DQN performs tick-level action selection from navigation variables, active policy parameters, and the LLM action prior. Each of the four configurations was evaluated over 30 rounds. In the fixed simulation, the complete configuration reached the goal in all 90 correlated robot--round records and achieved the lowest median completion time (42 ticks) and P90 (73.2 ticks); its median was 25.0--39.1\% lower than those of the other configurations. These observations provide descriptive, configuration-level evidence from the evaluated configurations.
In this paper we illustrate a novel architecture generating interpretable behavior and explanations. We refer to this architecture as the Mirror Agent Model because it defines the observer model, that is the target of explicit and implicit communications, as a mirror of the agent's. With the goal of providing a general understanding of this work, we firstly show prior relevant results addressing the informative communication of agents intentions and the production of legible behavior. In the second part of the paper we furnish the architecture with novel capabilities for explanations through off-the-shelf saliency methods, followed by preliminary qualitative results.
Recent reports during the AAAI-27 review cycle highlight the risk of reviewers coordinating bids for reciprocal assignment advantage. Prior work treats bidding, reviewer assignment, and review manipulation as separate stages, leaving the lifecycle effects of collusive bidding unclear. Real-world analysis is further constrained by typically unobservable collusive intent and the lack of counterfactuals for the same conference. Motivated by this gap, we introduce \alg, an end-to-end multi-agent simulacra framework for studying reviewer assignment integrity by holding the conference environment fixed and configuring LLM-driven reviewer agents with honest or collusive policies. We further develop an affinity-guided collusive bidding strategy that uses mutual reviewer-paper affinities to construct collusion rings and select target papers, producing expertise-consistent rather than arbitrarily targeted attacks. Controlled experiments show that collusive bidding more than doubles target-paper capture and that assigned colluders score target papers about two points higher than honest co-reviewers, while conference-wide effects remain comparatively modest. Evaluated bid-phase detectors provide only limited evidence of collusion: in a fixed-triplet detector stress test, native positive-bid graphs are confounded by benign affinity, while a Very-High-only diagnostic view enables precise but low-coverage local recovery.
自主 AI 代理越来越多地在这样一种环境中选择行动：其内存、执行时间、运行时、计算和操作约束决定了什么才算合适的计划。我们将代理规划状态中缺少这种执行上下文称为 substrate blindness。我们通过数值代码生成来检验这一一般性命题，因为在该任务中，所选实现选项及其运行后果可以直接观察。三种前沿模型配置——Anthropic Claude Opus 5、OpenAI GPT-5.6-Sol 和 Google Gemini 3.7 Flash——针对一个高维成对欧氏距离任务生成代码，输入要么只有任务本身，要么再附加一个 128 MB RAM 和 10.0 s 墙钟时间约束。披露约束使 14 组可执行、索引对齐的“仅任务”与“披露约束”比较中有 13 组的测得进程峰值内存降低；在所有三个队列中，平均墙钟时间也降低，使执行速度最高提升 3.1 倍。在整个审计语料中，披露约束带来了结构性代码变化，包括有界分块、保留 float32、上三角遍历，以及原地或内存映射缓冲区。在更严格的 96 MB 约束下，独立抽样的披露约束队列在正确且未超预算的结果上，Claude Opus 5 为 4/5，GPT-5.6-Sol 为 5/5，Gemini 3.7 Flash 为 3/5；而仅任务条件下分别为 0/5、1/5 和 0/5。各队列的平均 MaxRSS 和墙钟时间分别比仅任务参考低 49–74% 和 35–64%。这些结果为感知底层环境的代理规划提供了一个受控的概念验证：最小执行约束会在生成程序中诱发主动的结构适配，将计算从无约束分配转移开来，并在执行前显著改善观测到的资源—时间特征。
大型语言模型 Agents 越来越依赖执行轨迹来掌握复杂的交互任务。然而，当前范式受限于浅层轨迹检索和扁平化技能总结，根本上忽视了 Agent 行为的时间依赖性和结果条件化拓扑结构。我们提出 Trace2Tower，一种过渡感知的 EigenTrace 框架，可将原始轨迹提炼为稳健的技能层级。Trace2Tower 将步级交互抽象为规范化事件，并构建一个由语义兼容性、过渡动态和结果证据共同约束的统一图。通过一种新颖的对比式谱分解，它分离出稳定、与成功对齐的行为模式，同时严格抑制容易失败的捷径。这些模式自然地填充到一个动态技能塔中，形成动作模板、程序化例程和总体任务策略，并通过 verifier 引导的反馈持续优化。在 ALFWorld 上，Trace2Tower 在仅需 10.35 步和 0.26 个无效动作的情况下实现了 87.31% 的成功率；在 WebShop 上，它达到了 50.67% 的精确成功率。横跨这两个基准，Trace2Tower 在任务掌握能力和上下文高效经验复用方面都显著优于现有基线。
LLM agent 系统越来越多地将来源追踪、授权、策略执行、协议适配器和执行控制结合在一起。然而，单独正确的安全机制并不一定能够组合成端到端安全的系统：当动作跨越组件边界时，安全关键上下文可能会被丢弃、扩展、重新绑定或重新解释。我们将这种失效模式定义为安全上下文不连续性，并提出 CONTINUITY，一个用于可验证地组合 agent 安全控制的框架。 CONTINUITY 使用 assume-guarantee contract 对每个组件建模，并通过带签名的 root grants、provenance commitments、角色绑定的 transition receipts、有界类型化 release、transformation witnesses 和 effect-bound execution permits，在各个转换之间携带经过认证的安全上下文。我们形式化了端到端后果完整性，要求每一个实际产生的外部效果都必须有有效且最新的授权见证作为支撑，该见证将主体、任务、来源、委派、策略状态、规范化动作以及最终性边界连接起来。 我们实现了一个参考验证器和一个确定性的跨层故障注入套件，覆盖四个应用领域中的 32 类故障。在跨越 128 个故障-领域类别的 2,560 个参数化攻击实例中，完整的 CONTINUITY 配置不会导致任何有害外部效果，同时完成了全部 700 个良性任务，并对全部 200 个模糊案例进行了升级处理。这些结果表明，安全的 agent 执行不仅需要单个控制机制本身正确，还需要显式契约来在完整的指令到效果路径中保持其保证。
生产环境中的多智能体系统会不断替换 agent，基于这样的假设：填补某一角色的 agent 与任何其他能够胜任该工作的 agent 都可以互换。我们对这一假设进行了测试。在每种设置下，基于同一个基础模型、针对相同任务独立组建 8 个团队；每个 agent 在 10 次组建轮次中都保留一本私有笔记本。随后，我们在团队之间交换角色匹配的 agent，并在留出任务上测量变化。与一种能够复现阵容变动带来的干扰、但不改变谁坐在该位置上的安慰剂对照相比，交换对任务得分的影响很小，但会使团队每单位进展所消耗的通信量增加 16% 到 63%；而在 Hanabi 中，被交换进来的 agent 的代价比一个没有经验的 agent 更高，这与其从前搭档那里学到的约定受到干扰相一致。在 Collab-Overcooked 中，当设定议程的 agent 被替换时，额外通信的大部分来自留下来的 agent。围绕基础模型、解码温度和组建长度的三项消融实验表明，交换惩罚会与另一个量同步变化：独立组建的团队彼此漂移的距离。贪心解码会同时降低这两个量；将团队的历史长度加倍也会同时提高这两个量。在这些设置下，agent 在任务结果上的可替代性高于其在协同效率上的可替代性，而且组建历史越长，交换带来的影响越大。
Building automation systems generate rich sensor data yet remain insight-poor because heterogeneous point naming, missing metadata, and fragmented documentation obstruct their operational use. This systematic review analyses and codes 66 peer-reviewed studies on large language models (LLMs) for HVAC operations published between 2023 and March 2026. Each study is classified across five application families and three LLM method families and assessed for evidence realism, deployment readiness, and the responsibility boundary between the LLM and physical HVAC decisions. The corpus is concentrated in building energy modelling (BEM, 32 of 66 papers), while load forecasting remains too sparse for subfield-level conclusions. Only four studies reach pilot-level evidence, and none reports sustained operational deployment. No study was classified as ready-now for industry adoption; three were near-term and 63 research-only. Nevertheless, several bounded, human-in-the-loop uses merit near-term trials, including point-name normalisation, document-grounded operator support, BEM workflow assistance, and advisory interfaces around physics-based controllers. Conventional machine learning (ML), model predictive control (MPC), reinforcement learning (RL) and ontology-based tools remain more adopted for high-frequency control, short-horizon numerical forecasting, and well-posed ontology mapping, while autonomous agentic operation and unvalidated occupant proxies remain research-stage. Current evidence therefore supports LLMs primarily as semantic and workflow layers rather than autonomous HVAC controllers. Future work should prioritise field-validated benchmarks, orchestration evaluation under operational constraints, and LLM-MPC/RL architectures with bounded latency and verifiable safety properties.
在 OpenAI 内部，编码 agent 正在重塑 AI 研究。探索关于 agent 使用情况、实验速度、任务复杂性和研究加速的早期数据。
Why can't we get an agent scheduler that supports all of the following: - git compatible - agent agnostic - 100% open source - os and system-agnostic - multi-runner support - support mcp/ui/http - unlimited routines/crons So I built one, moadim.io is a local Rust daemon you install in the target machine, give it a name and manage its routines via a Git repository, wants a new routine that send you a daily message from this machine? Create a pr and merge, have another routine that run every hour to pull the latest changes to the ~/.config/moadim folder. With more than 1,000 users, I define this project as almost "done" and ready for production. Me and thousand more people are use it in a daily manner. It currently supports Claude, Codex, Hermes, and Pi, and you are welcome to add your agent of choice as well because it's 100% configurable. You are welcome to have a look at the source code of the daemon in " http://github.com/moadim-io/daemon " Feel free to provide me with suggestions for more features around this topic. I don't want to branch out to new off road topics likt webhooks, this is a "done" software in the realm of agents schedulers that focus on cron-like work. Also feel free to start the github repository and open issues and PRs for your suggestions. Comments URL: https://news.ycombinator.com/item?id=49571537 Points: 3 # Comments: 0
文章链接： https://www.reuters.com/world/europe/openai-agents-hijacked-german-website-previously-undisclosed-ai-breakout-this-2026-09-04/ 评论链接： https://news.ycombinator.com/item?id=49562744 得分：7 # 评论数：1
Hi HN, I'm a nuclear engineer and tech entrepreneur. After working for 3 years in the National Atomic Energy Commision (CNEA) in HPC software for nuclear simulations, I co-founded ratherlabs.com, a software company providing blockchain solutions for startups. Rather Labs business reached ~6 million usd in annual revenue in 2024; now demand is decreasing as our niche clients are using AI to launch their MVPs. Now I'm joining my scientific background with my entrepreneurial inclinations to think of problems related to the energy industry. Nuclear is experiencing a renaissance now (it was in a dark age when I was studying after the Fukushima event) so I think this is the right moment for me to jump again into the field. Reactor atlas is an interactive nuclear reactor map and an intelligence layer on top of it. The basic layer indexes: - power reactors - research reactors - fuel facilities with their history and projection, by country. The project is solving a fragmented information ecosystem, where operational and commercial data is hard to conciliate. Also, I'm monitoring nuclear news associated with each of the facilities, political declarations that may affect the nuclear landscape and geo-phisical news (like earthquakes) that could affect operation. On top of the operational + news layer, I'm creating an intelligence layer that will help institutions of different types make faster decisions by subscribing to very specific alerts. As an engineer, I used to code in fortran and c++ when I was working at CNEA, and later I participated in the first Rather Labs projects architecting complex async systems, mostly using MERN stack. I'm no longer coding anymore. I'm focused on designing robust systems with AI agents. I mainly used Claude Fable 5.1 to develop reactoratlas.com, deploying on vercel and using mostly nextjs, threejs, postgreSQL and some cron jobs to query and update data. I miss coding as in the old days, but I haven't code a single line in this project. And I strongly believe we humans shouldn't code anymore: our responsibility is now on designing and supervising AI systems. Thanks for your attention! Comments URL: https://news.ycombinator.com/item?id=49549148 Points: 21 # Comments: 4
Neural machine translation (NMT) systems typically produce a single output per input, obscuring the alternative decision trajectories implicitly available within multilingual decoding. This opacity becomes particularly problematic in low-resource dialect settings, where multiple linguistically valid realizations may differ in lexical authenticity, register, and structural stability. We propose reframing translation as a structured decision space explored by autonomous translation agents. Instead of analyzing a single output, we model distinct translation pathways as agents operating over a shared multilingual backbone. Inter-agent divergence is treated not as error but as an interpretable behavioral signal. We conduct an empirical study on Turkish--Syrian Arabic translation using three agents: (1) zero-shot direct translation, (2) dialect-stabilized translation via lightweight fine-tuning, and (3) pivot translation through English. Evaluation is performed on 5,000 dialogue sentences, while stabilization is trained on 5,000 additional Turkish--Syrian sentence pairs drawn from television dialogue and MADAR-Turk resources. Rather than optimizing for conventional performance metrics, we quantify structured behavioral displacement using dialect marker frequency, lexical proximity to standardized Arabic, and structural variance. Lightweight stabilization nearly doubles dialect marker usage, increasing it from 0.2266 to 0.4988, while significantly reducing structural instability. Pivot mediation introduces normalization pressure and measurable compression effects, whereas zero-shot translation exhibits the highest decision variance. We argue that translation divergence across agents reveals latent decision flexibility within multilingual models and we provide a principled interpretability framework for low-resource dialect generation.
AI 智能体近年来在自动化漏洞修补方面表现出强大能力。然而，现有评估通常只通过测试所给的 Proof-of-Concept (PoC) 输入是否仍会触发崩溃来验证补丁。这留下了两个关键的有效性威胁：智能体可能复现已记忆的历史开发者补丁，或者生成仅仅抑制所报告崩溃的表面级修复。 我们研究了这些关于 C/C++ 漏洞修补的担忧。我们引入了一种补丁相似度度量，用于检测被记忆的补丁。平均而言，25% 的智能体补丁与历史开发者补丁表现出显著相似性，这表明补丁记忆化确实是漏洞修补评估有效性的一个真实威胁。与此同时，智能体也常常利用基准结构，通过在崩溃堆栈跟踪上进行修补来抑制崩溃，而不是定位并修复漏洞的根本原因，从而通过补丁验证。 为了解决这些问题，我们提出了 PatchBench，这是一个用于评估 AI 智能体在真实漏洞修补任务上的新基准。PatchBench 选择其真实修复位于崩溃堆栈之外的漏洞，并使用漏洞移植和代码变异，将历史漏洞迁移到新的仓库上下文中，从而降低表面级修复和补丁记忆化的风险。我们开发了新的补丁验证方法，全面评估智能体补丁的安全性和语义正确性。在包括前三名 AIxCC 智能体在内的 11 个最先进智能体上，原始仅基于 PoC 的验证平均将智能体的补丁修复任务解决率夸大了 1.83$\times$。我们的结果揭示了当前修补智能体的关键局限，并指明了未来更可靠漏洞修复研究的方向。
当任务存在程序化检查器时，基于可验证奖励的强化学习效果很好，但大多数长程智能体领域并没有这类检查器。我们研究结果不可见设置，即无法获得真实成功信号的情形。多标准评分细则是一种常见的奖励提供方式；它们对每条轨迹只评分一次，但单一标量对于跨越数十步的过程来说是较差的信号。我们提出 DRACO：Distributing Rubric-based Advantage for Credit Optimization。它在训练期间动态生成评分细则，以跟踪策略不断演化的能力；在每条完整轨迹结束后对这些评分细则打分，并将该判断重新分配到负责这些已标注评分细则的步骤上，从而在 GRPO 中生成区分化的逐步优势。该重新分配具有闭式形式，不引入任何经过训练的归因模块。在 AppWorld 上，DRACO 比基础模型提升 15.9 分，且即使自身不使用任何验证器，也比使用稀疏真实奖励训练的 GRPO 高 5.3 分。在域外的 Tau-Bench 上，即使没有 frontier judge，它也比基础模型高 5.3 分，优于真实奖励训练和其他基于评分细则的训练设置。DRACO 的代码可见于 https://github.com/IBM/draco。
Scaling interactive and verifiable environments is critical for training terminal agents. As frontier models become more capable, environments synthesized from scratch become less challenging and thus provide limited learning signals. Recent co-evolution methods iteratively synthesize environments near the model's learnable frontier based on weaknesses exposed during rollouts. However, their dependence on on-policy rollouts limits generalization and the continuous provision of learning signals as the model becomes stronger. In this paper, we propose environment evolution, which incrementally increases environment difficulty off-policy and schedules the evolved environments generation by generation during training to provide continuous learning signals. We derive three evolution directions that influence environment difficulty from the multi-turn learning objective and then implement evolution along these directions through a loop-engineered multi-agent harness. Quantitative rollout experiments with Hy4 preview, Claude Opus 5, and GPT-5.6 Sol show that environment evolution consistently produces more difficult environments. We validate its effectiveness on Qwen3.6-27B and Qwen3.6-35B-A3B through simple long-horizon RL training, improving their performance by 14.4 and 18.0 percentage points on Terminal-Bench 2.1, respectively.
AI agents 正在通过异构的 agent-development frameworks、AI models、tool interfaces、protocols 和 execution environments，越来越多地在各类组织中被开发和部署。为了实现其潜在的社会和商业影响，这些 agents 必须能够通过一种通用通信协议实现互操作。Natural Language Interaction Protocol (NLIP) 由来自公司和大学的研究人员与实践者共同开发，并由 Ecma International 标准化，针对这一需求定义了一种基于标准的 AI-agent interaction 应用层协议。NLIP 提供了一个轻量级的语义消息封装，可承载于现有传输之上，例如 HTTP/HTTPS、WebSocket 和 AMQP，同时允许支持 NLIP 的 agents 和 gateways 在 clients、agents、local context stores、ontologies、tools、enterprise services 以及异构底层协议之间进行适配。本文介绍了 NLIP 的动机和设计原则、其消息模型和传输绑定、安全设计考虑、参考实现、代表性应用、采用信号，以及它与 MCP 和 A2A 等新兴 agent protocols 的关系。
As terminal-based code agents become prevalent, agent trajectories have accumulated at scale, while realistic, executable environments remain scarce. However, environments are what agent post-training actually requires: each can be re-queried into many verifiable tasks and provides execution feedback, whereas a trajectory is a single frozen demonstration. Rather than generating environments from scratch, we observe that the tool-execution history in existing trajectories exposes the structure and contents of the environments in which they ran, making it possible to reconstruct those environments from the trajectories themselves. Thus, we introduce Terminal-Universe, a framework which turns each trajectory into a reusable environment and explores it for synthesizing new tasks and continued interactions. Specifically, Terminal-Universe replays the file operations recorded in a trajectory to restore each file before the agent modified it, yielding a partial workspace; a completion agent then supplies the missing files and dependencies. On this recovered workspace, we both reconstruct the original intent task and synthesize entirely new ones. Besides, we also scale the tasks along two complementary axes: breadth and depth. For breadth, we mine directional dependency relations between related environments and synthesize cross-workspace queries spanning multiple codebases, as developers routinely do in real-world development. For depth, we extend the initial single-turn query into a multi-round session that captures iterative user feedback and requirement refinement via a user agent. Applied to public terminal agent trajectories, Terminal-Universe produces 37.3k task-sufficient environments. Supervised fine-tuning of Qwen3.5-27B on this corpus improves single-round performance on Terminal-Bench 2.1 by 11.9 points and multi-round performance on EvoCode-Bench v2 MT@4 by 13.8 points.
大型语言模型（LLM）智能体正越来越多地被提议作为自主 SOC 分析师，但有两个限制使其在企业规模下并不可靠：有限的上下文窗口无法容纳包含数千台主机的认证图，而自由形式生成也无法保证所推荐的遏制操作与其所作用的拓扑保持一致。我们提出 Sentinel-RL，一种 agentic-SOC 架构，将拓扑推理与语义推理解耦：异构图注意力编码器将实时认证子图汇总为固定维度状态，Proximal Policy Optimization（PPO）策略将该状态映射到一组受约束的调查动作，而 LLM 智能体循环仅限于消费该策略的建议，并生成由评论器把关的、分析师可读的叙述。我们在 LANL Comprehensive, Multi-Source Cyber-Security Events 数据集和 Indiana University Quartz HPC 集群上实例化该系统，并报告四项结果：（i）两阶段 CREATE 摄取模式在单个 32 核节点上用 14.2 分钟将包含 2400 万条边的认证子图加载到 Neo4j 中，比规范的基于 MERGE 的流水线快约 24 倍；（ii）滑动窗口告警引擎在 50 次试验中可靠地在 <=2.5 秒内触发 25 事件/10 秒阈值；（iii）经过 200 次迭代的 PPO 训练收敛到 8.74+/-0.31 的平均单回合回报，在带标签的红队事件上保留集精确率为 0.91、召回率为 0.87；（iv）集成的遏制循环以 6.3 秒的中位时间完成一次完整的检测-调查-建议-人工批准周期。我们的贡献包括一种可复用的工程模式（热节点死锁规避方案）、一种可移植的 HPC 部署模式（锚节点共置），以及涵盖误报经济性、可逆性保证、审计合规和人工批准边界的企业就绪性分析。
代码仓库级软件工程基准显著推动了对编码智能体的评估，但现有基准主要衡量生成的补丁是否通过功能测试，而忽略了源自代码审查的验收约束（审查约束），这些约束在真实软件开发中往往会影响一个补丁是否可接受。我们引入 SWE-Gate，这是一个面向软件工程智能体的代码仓库级基准，明确在功能正确性之外评估对审查约束的遵循情况。SWE-Gate 从真实拉取请求审查评论中提取审查约束，并围绕这些约束合成代码仓库级修复实例。每个实例都提供独立的功能测试和约束测试，以及不合规补丁和黄金补丁，从而能够明确区分问题解决能力与审查约束遵循能力。我们构建的 SWE-Gate 包含 303 个代码仓库级修复实例，覆盖 75 个开源 Python 代码仓库，涉及多种软件领域。在一个通用编码智能体框架下，对涵盖不同能力水平的四种 LLM 后端进行实验，结果揭示了功能成功与在完整修复规范下成功之间存在显著差距：在通过功能测试的 644 次修复中，有 221 次未能满足所提供的审查约束。这些发现表明，仅基于功能的评估会高估智能体满足代码仓库级修复任务完整要求的能力。包含代码、数据和实验结果的复现包可在 https://github.com/DeepSoftwareAnalytics/SWE-Gate 获取。
I’ve been following MCP since it first came out. It got a lot of attention early on, but I haven’t come across many people using it in production. I may simply have missed them. If you’re using MCP in production, what are you using it for? What advantages have you found over a normal API or direct tool integration or just CLI? Comments URL: https://news.ycombinator.com/item?id=49548600 Points: 15 # Comments: 26
文章链接： https://armature.tech/blog/which-tools-coding-agents-install 评论链接： https://news.ycombinator.com/item?id=49557206 得分：3 # 评论：1
文章 URL： https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra 评论 URL： https://news.ycombinator.com/item?id=49556147 点数：8 评论数：2
嗨，HN，我是 Ansh，Mireye 的创始人（ https://www.mireye.com ）。我正在构建 AI 代理用来对物理地点做决策的基础设施：面向任何美国地点的数据、增强、工具和信号，通过一个 API 和 MCP server 提供。 这里有一个演示视频： https://www.youtube.com/watch?v=haqO6UbUqU0 要试用的话，把 https://www.mireye.com/skills.md 粘贴到你的 agent 里，然后到 mireye.com 领取一个免费 key（5,000 credits，不需要信用卡）。文档在 https://docs.mireye.ai 。最快的上手方式是向 API 询问任何关于美国某个地点的问题。你可以拿一个你熟悉的地方来测试，先不知道答案，再根据你掌握的信息给它打分。 在 Mireye 之前，我一直在做建筑类 agents，自己也碰到了这堵墙：我的 agent 能推理任何网上的东西，却对脚下的土地一无所知。后来一家 Fortune 500 保险公司告诉我，他们的工程师也因为同样的原因放弃了 underwriting agents。前沿模型在被问到某个特定地点的具体问题时，仍然会胡编乱造。 我的第一个产品是一个小众的 site-screening 应用。客户拿他们熟悉的地点来测试，答案都站得住脚，但没人关心这个应用本身。他们想要的是底层引擎。使用情况也印证了这一点：目录中的 317 个字段里有 311 个都被查询过，而且没有哪个用例占绝对主导。所以我砍掉了应用，转而开始做基础设施。 Mireye 不是一个上面套了 API 的数据集，因为事实本身并不等于决策。一个 agent 会通过它完成整项工作：带引用的事实、把一个裸地址补全为所有者、英亩数、建筑物和附近供电信息的丰富数据、那些运营模型经常算错的工具，以及在某些情况发生变化时发出的信号，比如重新分区申请。数据、增强、工具、信号。 这些工具来自我观察 agent 失败的过程。我们在自己的基础设施上构建 agents，而同样的问题反复出现：agent 会凭肉眼估算距离而不是计算，给一个地址抓错地块，或者在批处理进行到一半时耗尽全部预算。每一种失败，最后都变成了 agent 可以调用的东西：确定性的几何和行车时间工具、地块解析、在任务运行前定价的 quote endpoint，以及把整套工作流打包起来的 skills，比如筛选一个地点或对一个地址做 underwriting。 最让我意外的是难点所在。每个来源都必须被采集（有时要按县逐个采集，而且要适配每个县发布的格式）、归一化到同一 schema、签订契约（它来自哪里、每个值意味着什么、多久刷新一次），然后还要永远保持最新。我们现在为 366 个字段运行这套循环，并通过一个索引以 multi-tenant 方式提供服务，而且每个来源的阻力都不同。Maryland 发布了一个数据集，标题就叫 “Hidden Property Owner Names.” 我在 North Carolina 提交了公共记录申请，因为没人索引 sewer mains。 更深层的问题是语义。两个县发布了同名字段，但含义不同。而最危险的值是 null。它是表示“这里没有 flood zone”，还是“这个县根本没绘制过 flood 数据”？把模型放到这个空缺面前，它就会用一个看起来合理的数字把沉默补上。所以我们把缺失也类型化了。每个字段都会返回 ok、absent 或 failed。客户告诉我，他们之所以信任它，正是因为这些拒绝。 最新的部分是按需索引。要是你请求一个我们还没有的字段，一个长期运行的 agent 就会去研究来源、收集数据、用 ground truth 做测试，然后把它编入索引，通常一天内就能完成。我会从这个帖子里挑几个字段请求，回头告诉大家它做了什么。 人们已经用它做出了我没预料到的东西：保险团队筛查 flood、wind 和 wildfire 风险的投资组合；一家 proptech 通过增强数据清洗杂乱的房源地址；一个 wellness brand 评估街角投放 poster spots 的得分；一家机器人公司寻找仓库；数据中心选址；无人机部署规划；某个城市的校车路线；以及人类贩运调查中的信号。我们的使命是把地球的每一寸都编入索引，让它像 web 一样可查询。 定价是公开的：免费层每月 5,000 credits，$19/月可得 25,000，$99/月可得 120,000，企业版定制。 我很想知道你正在构建的 agents 会在哪里接触物理世界。如果你能在一个你熟悉的地方运行 Mireye，并告诉我我们哪里做错了，我会更开心。 Comments URL: https://news.ycombinator.com/item?id=49552616 Points: 15 # Comments: 0
大家好，HN，我是 Ansh，Mireye 的创始人。我正在构建 AI agents 用来对物理地点做决策的基础设施：针对任何美国地点的数据、增强、工具和信号，全部通过一个 API 和 MCP server 提供。 这里有一段演示视频： https://youtu.be/haqO6UbUqU0 要试用的话，把 https://www.mireye.com/skills.md 粘贴到你的 agent 里，然后到 mireye.com 领取一个免费 key（5,000 credits，无需信用卡）。文档在 https://docs.mireye.ai 。最快的上手方式是让 API 回答任何关于美国某个地点的问题。找一个你熟悉的地方来测试，把结果冷静地和你已知的信息对比打分。 在 Mireye 之前，我自己在做建筑 agents 的过程中也撞上了这堵墙：我的 agent 也许能推理任何线上信息，但对其脚下的地面一无所知。后来一家 Fortune 500 保险公司告诉我，他们的工程师也因为同样的原因放弃了核保 agents。前沿模型在被问到某个特定地点的具体问题时，仍然会胡编乱造。 我的第一个产品是一个小众的场地筛选应用。客户拿他们熟悉的地点来测试，答案表现不错，但没人关心这个应用本身。他们想要的是底层引擎。使用情况也印证了这一点：目录中 317 个字段里有 311 个被查询，而且没有哪一种用例占主导。所以我砍掉了应用，转而开始构建基础设施。 Mireye 不是一个上面套了 API 的数据集，因为事实本身并不等于决策。agent 通过它完成整个工作：带来源标注的事实、把一个普通地址增强成所有者、面积、建筑物和附近电力设施的地址信息、让运营模型容易出错的工具，以及当某些事情发生变化时的信号，比如重新分区申请。数据、增强、工具、信号。 这些工具来自对 agent 失败的观察。我们基于自己的基础设施构建 agent，而同样的事情不断出错：agent 会凭目测估算距离，而不是计算；会把地址对应到错误的地块；或者在批处理到一半时把全部预算耗光。每一种失败都变成了 agent 可以调用的东西：确定性的几何和行车时间工具、地块解析、在任务运行前给出价格的 quote endpoint，以及把完整工作流打包起来的 skills，比如场地筛选或地址核保。 最难的部分出乎我的意料。每个数据源都必须被采集进来（有时要按县逐个处理，而且每个县发布的格式都不同）、归一化到同一 schema、注明约束条件（它来自哪里、每个值是什么意思、多久刷新一次），然后永远保持最新。我们今天为 366 个字段运行这个循环，并通过一个索引以多租户方式提供服务，而且每个数据源的抵抗方式都不同。Maryland 发布了一个数据集，标题就叫“Hidden Property Owner Names”。我在 North Carolina 提交了公开记录请求，因为没人会索引下水管主管道。 更深层的问题是语义。两个县发布一个同名字段，但含义却不同。而最危险的值是 null。它的意思是“这里没有洪水区”，还是“这个县从来没绘制过洪水区”？把模型放在这个空白前，它就会用一个看似合理的数值把沉默填满。所以我们要为缺失建模。每个字段都会返回 ok、absent 或 failed。客户告诉我，正是这种拒绝回答的机制让他们信任它。 最新的部分是按需索引。如果你要一个我们还没有的字段，一个长期运行的 agent 会去研究来源、收集数据、用真实地面情况进行测试，然后把它编入索引，通常一天内就能完成。我会从这个帖子里挑几个字段请求，并回来汇报它构建了什么。 有人已经基于它做出了我没预料到的东西：保险团队用它筛查洪水、风灾和野火风险的组合资产；一家 proptech 通过增强清洗杂乱的房源地址；一个健康品牌用它给街角的海报点位打分；一家机器人公司用它寻找仓库；数据中心选址；无人机部署规划；某城市的校车路线；以及用于人口贩卖调查的信号。我们的使命是把地球上的每一寸土地都编入索引，并让它像 web 一样可查询。 定价是公开的：免费套餐每月 5,000 credits，$19/月可得 25,000，$99/月可得 120,000，企业版定制。 我很想听听你们正在构建的 agent 会在现实世界的哪些地方落地。如果你能用 Mireye 在一个你熟悉的地方跑一下，并告诉我哪里错了，我会更开心。 Comments URL: https://news.ycombinator.com/item?id=49552616 Points: 2 # Comments: 0
文章 URL: https://claude.com/blog/claude-for-commerce-agents 评论 URL: https://news.ycombinator.com/item?id=49547888 得分: 29 # 评论数: 26
Aggregating noisy, conflicting textual hypotheses into a reliable consensus is a fundamental challenge when deploying NLP systems in real-world industrial settings. While monolithic Large Language Model (LLM) agents offer unbounded expressivity for tasks like Root Cause Analysis (RCA), they suffer from context limits, compounding hallucinations, and prohibitive inference latency. Traditional weak supervision offers statistical rigor but is mathematically restricted to discrete classes. We present Loom, a generative consensus framework deployed for real-world RCA that bridges these paradigms. Loom aggregates open-form hypotheses emitted by modular heuristics (diagnostic templates dynamically populated with episode-specific entities, times, and metrics) by projecting them into a continuous embedding space, and resolves conflicting signals with an iterative centroid-based reweighting algorithm. The resulting consensus weights ground a single lightweight LLM synthesis step. Evaluated on the OpenRCA benchmark, Loom occupies the accuracy--efficiency Pareto frontier: it matches a state-of-the-art autonomous agent on Bank and Market-2 and trails on Market-1 and Telecom, while using a single LLM call per incident on all four datasets ($\sim$26$\times$ faster; $\sim$33$\times$ with an 8B-parameter synthesizer). We discuss our deployment experience, highlighting lessons learned regarding the trade-offs between agentic depth and inference latency, negative results in redundancy detection, and how deterministic consensus fosters trust among Subject Matter Experts~(SMEs).
多智能体 LLM 系统通常使用一个 orchestrator 将任务分解给一组工作者，随后通过文本反思进行改进。尽管经验结果很强，但这些系统缺乏对协调、记忆改进以及外部验证作用的统一解释。我们将 orchestrator-worker 交互建模为一个双层协调博弈：在有界耦合下，工作者的局部更新博弈是一个近似势博弈，其均衡松弛量由分解质量控制。随后，我们将反思分析为在语义记忆状态上的随机移动。对于自由形式反思，我们推导出一个有限时间上界，证明了最坏情况的紧致性，并在一个可证伪的持续伤害条件下给出一个正下界。我们进一步证明一个信息论不可能性结果：任何只观察生成文本记录的门控器都无法在文本不可区分的环境上实现一致改进，而基于环境的门控器则可以。受这一分离启发，我们提出 Stochastic Reflective Memory Ascent (SRMA)，它仅在基于环境的评估风险严格下降后才接受候选记忆。在校准和非退化纠正质量下，SRMA 可精确收敛，且收敛速率为几何级或多项式级；相应构造表明这两种速率情形在阶上都是紧的。我们还为随机评估提供置信门控，并为分段平稳环境提供重新锚定保证。实验使用基于环境的指标实例化这些对象，并检验所预测的协调与漂移规律。在 500 个 SWE-bench 实例上，完整的基于 Kimi 的系统成功解决了 72.2%，而公开的 mini-SWE-agent 参考系统为 70.8%。代码：https://github.com/YihangChen9/Bilevel-Coordinated-Reflection
本地部署助手可以让工厂工人通过对话访问机器文档，但能够胜任该任务的模型很少能适配车间硬件。我们表明，在结构压缩和基于检索的适配之后，模型规模不再是适配后答案质量的可靠预测指标：通用能力几乎随参数量线性下降，而经评估的检索增强答案质量却不是如此。因此，我们将部署视为一个适配后的选择问题，在可配置的通用能力下限和内存预算下，依据经评估的答案质量与实测的设备端吞吐量，为每台设备确定一个子网络；仅优化规模、速度或质量的规则都会牺牲能力或吞吐量。采用 sandwich-style in-place distillation 的共享权重 supernetwork 使这一选择保持低成本。在一个制造手册案例研究中，提取仅达到未剪枝模型经评估质量的13.7%，而检索引导的蒸馏将其恢复到仅低4.6%的水平，挽回了三分之二的损失，而且同一个助手还能在三层异构边缘设备上运行，待机功耗为1.3到5瓦。
The performance of LLM-based agents is jointly shaped by the base model and the harness used when interacting with the environment. This exposes them to safety risks in both harmful final responses and multi-step execution trajectories. Existing safety alignment mechanisms often rely on either external harness updates or policy optimization, yet applying either paradigm in isolation fails to bridge runtime control with intrinsic safety. We propose SafeEvolve, an experience-driven self-evolving framework for agent safety alignment. SafeEvolve leverages safety experience from completed on-policy trajectories to drive a continual loop of harness-policy co-evolution. On the harness side, SafeEvolve converts trajectory-level safety evidence into bounded, component-level updates across safety prompt and hierarchical skills, yielding auditable and reversible harness artifacts. On the policy side, SafeEvolve follows a two-stage SFT-RL paradigm, where harness-use SFT bootstraps the policy to actively leverage evolved harness artifacts, and harness-augmented RL further shapes autonomous safety behaviors during multi-step exploration via verifier-decomposed rewards. Through harness-policy co-evolution, SafeEvolve converts safety experience into an evolved runtime harness and improved policy behavior. Experiments on agentic safety benchmarks show that SafeEvolve achieves a stronger safety-utility tradeoff than existing baselines. For Qwen3.5-4B, SafeEvolve achieves a $3\times$ ASR reduction on AgentDojo while improving benign utility from 59.79% to 61.86%.
Root cause analysis (RCA) is a critical task in telecom network operations, but diagnosing performance degradations in modern 5G and emerging 6G networks remains challenging due to complex cross-layer dependencies. While large language models (LLMs) offer promising capabilities for reasoning and knowledge integration, directly applying vanilla LLMs to telecom RCA often leads to hallucination, unstable reasoning, and poor alignment with structured network evidence. This work first reviews the evolution of telecom RCA from rule-based and machine learning (ML) approaches to emerging LLM-enabled techniques, and provides an overview of recent paradigms, including structured reasoning, retrieval-augmented knowledge grounding, agentic orchestration, and verifiable reasoning. Building upon these insights, we propose a structured reasoning framework for LLM-enabled telecom RCA that aligns diagnostic reasoning with telecom-specific evidence and domain knowledge. The proposed approach first organizes heterogeneous network telemetry into canonical contexts, and then enforces decision-path reasoning during diagnosis, and finally generates evidence-grounded explanations for reliable fault identification. Experimental results on two 5G RCA datasets, TeleLogs and TelecomTS, demonstrate that the proposed framework consistently improves diagnostic accuracy and decision consistency compared with baseline techniques. These cross-dataset results highlight the importance of structured reasoning design for practical LLM-based RCA systems in next-generation telecom networks.
近期的 web agent 使用世界模型进行测试时的动作选择：通过采样候选动作、预测由此产生的 web 状态，并用 ranker model 或 Process Reward Model (PRM) 对其排序。这些世界模型通常通过监督式下一状态预测进行训练，以生成固定表示，例如 HTML 或 AXTree 快照。然而，这一目标与下游 ranker 并不一致，因为 ranker 依赖于预测状态在不同候选之间具有可区分性，从而准确打分。为了解决这一问题，我们提出了 predicted-state matching，这是一种训练目标，要求预测表示必须将真实产生的状态与由其他动作到达的状态区分开来。我们使用一个分支式 web-agent 数据集训练这些模型，该数据集由 WebArena Go-Browse 轨迹构成，其中每个决策点都包含多个备选动作及其对应的结果状态。我们在留出的 predicted-state matching 基准上的实验表明，我们的方法优于使用监督式下一状态预测训练的 world model。我们进一步表明，与仅基于动作的 PRM 以及结合了监督式下一状态 world model 的 PRM 相比，我们的方法在 WebPRMBench 上提升了基于 PRM 风格的动作排序效果。最后，在 WebArena-Lite 上，使用我们的 world model 进行测试时动作选择能够提升端到端任务成功率。我们的项目主页见：https://dhruvpendharkar.github.io/dwm/。
Existing Human-Robot Interaction (HRI) literature has focused on identifying and structuring errors, failures, conflicts, and knowledge issues (called in this work as contradictions) in domain-specific dialogue-based interactions. However, there is still lack of a formal computational framework to represent and define these contradictions, interoperable and usable across HRI and human-agent interaction (HAI) domains. Thus, this research project aims to capture, represent, and evaluate the notion of (1) dialogue-based collaborative interaction and (2) related contradictions in a foundational ontology. METHONTOLOGY, a systematic approach to build domain-independent ontologies was applied. In the conceptualisation stage of the presented ontology, concepts and models from Activity Theory were used. Preliminary results presented in this short article are: (i) Natural language definitions of dialogues and related contradictions in HRI, (ii) Set Theoretic definitions of dialogues and contradictions, and (iii) First Order Logic (FoL) formulation of the contradiction concepts and three novel principles guiding dialogue-based interactions between humans and robots. In summary, we report on ongoing work to develop a foundational ontology based on Activity Theory called Activity Theory-based foundational ontology (ATFOt) to capture and represent the notion of contradictions in HRI.
With the proliferation of LLM agents, the ability to understand and diagnose failures in agents is essential to achieving superior effectiveness and trustworthiness. As agent failures often manifest via long and complex trajectories, manually finding the needles in the haystack is untenable. However, traditional diagnosis techniques for software bugs can hardly address LLM agent failures, while completely relying on LLMs as the judge yields unreliable diagnosis results. To overcome these challenges, this paper presents AGENTSCOPE, a new neuro-symbolic approach for agent failure mode diagnosis. The key principle of AGENTSCOPE is to abstract agent behavior, based on its trajectories, into structured representations. Furthermore, AGENTSCOPE introduces the concept of neural invariants to specify agent behavior properties. AGENTSCOPE leverages LLM-guided reasoning atop the structured representation against neural invariants to pinpoint both the failure step and its type in the trajectory. We show the effectiveness of AGENTSCOPE on publicly available agent failure datasets (Who&When) and a more comprehensive dataset created by us (AgentErrata), where AGENTSCOPE significantly outperforms the current state of the art in fault localization and attribution accuracy. Our work shows that integrating structured abstractions with LLM-guided reasoning enables effective, reliable, and interpretable diagnosis for agent failures.
多轮 Agentic RL 越来越把信用分配视为一个定向问题：给定一个终态可验证奖励，逐轮方法会把信用定位到那些起作用的轮次上。我们识别出一个能预测何时这才是正确做法的结构量：验证器信息密度 V_d = k/C（验证器暴露了一个 Agent 的 C 步因果链中多少比例的逐轮正确性），并表明终态验证器深处于低 V_d 区域，在这里定向不是正确的着力点。在 tau^2-bench 上进行受控的共享 rollout 对比、并将奖励密度与信用几何分离之后，我们发现：将连续稠密奖励均匀铺开，比稀疏的二元结果奖励更优（在 4/5 个 seeds 上净有害）；而把同样的优势集中到进度轮次上，或集中到随机轮次上，同样有害：定向只是次要因素。其机制在于覆盖：终态验证把可观测信号压缩到单一的最终写入轮次（在 98% 的 rollouts 中 k=1），而成功却需要一个 5-8 步的前置工具调用链。一个合成相边界把交叉点放在 V_d* ~ 0.8，而在 tau^2-bench 上测得的 V_d 约为 0.15，在 BFCL V3 上约为 0.4；在 BFCL 上，均匀分配同样获胜，而一个匹配浓度的打乱对照在 8/8 个 seeds 上为负。该效应在 ToolACE-2-8B 的不同模型家族上也能复现（Delta = -0.048，基于 32 个预注册 seeds；独立的 20-seed 复现本身也显著），而一个预注册的匹配预算广度扫描描绘出单调的剂量-反应曲线，其缺口只有在全链路覆盖时才会消失，其中一个 reward-to-go 分支达到了与全覆盖持平。均匀重分配是零信息的覆盖默认方案，逐轮方案必须优于它；我们贡献了匹配浓度的打乱对照，任何定向主张都应先通过这一检验。
我们提出 CivBench，这是一个开源基准，用于通过 Model Context Protocol (MCP) 评估长时程、工具介导环境中的语言模型智能体。单个回合跨越 300+ 个回合，并在大动作空间上产生数千次工具调用，要求在部分可观测性下持续进行规划、状态监控和执行。该环境暴露了 76 个 MCP 工具，以及一层叙述模块，可将视觉游戏状态转换为结构化文本。 我们使用 CivBench 在 23 次有效运行中，刻画四个模型家族的智能体行为。该样本是试点研究，不是模型排名：在这一规模下，总体结果并不能可靠地区分不同模型。相反，我们提出了两个环境使其可测的接口层指标：Proactive Monitoring Rate (PMR)，用于衡量智能体是否主动查询潜在的战略状态；以及 RAG@10，用于衡量结构化规划反思中陈述的承诺是否会在随后十回合内执行。 在共享 playbook 协议下，我们在多次运行中观察到两个一致模式。智能体对可用但需要显式查询的、与战略相关的状态监控不足：尽管 playbook 建议每 20 回合查询一次胜利进度，它们实际只会每 30 到 75 回合查询一次；并且在 20 场可检测到的失败中，有 7 场它们未能在游戏结束前的 20 回合警示窗口内进行查询。智能体也经常未能执行自己规划反思中陈述的近期承诺（不同模型的 RAG@10 在 48.2% 到 65.8% 之间）。这两种模式都出现在具备工具访问权限和明确指导的情况下，我们将其解读为在指令下的偏差，而非能力缺失。 我们在 https://github.com/lmwilki/civ6-mcp 发布环境、场景、日志、指标和分析流水线
Generative AI is changing how cultural artifacts are created and circulated, and with it our understanding of creativity itself. Researchers disagree about whether these tools enrich or impoverish culture, and we argue that much of that disagreement comes from conflating two distinct components of creativity: novelty, a property of single artifacts, and diversity, a property of populations. We argue further that creativity in the context of generative AI is best understood as a property of hybrid collectives, or populations of interacting people and algorithms, rather than of individuals. AI-assisted ideation reliably raises the novelty of individual output while narrowing diversity in the aggregate, but this is not an inevitable consequence of putting machines in the loop. Because humans and models search in complementary ways, mixed groups can outperform and out-diversify groups of either kind alone, and machine-discovered solutions can enter human culture and persist there. What decides the outcome is composition: which agents are present, in what proportion, and how they are connected. The question is no longer whether AI helps or harms creativity, but which mixtures let individual gains accumulate without eroding collective diversity.
长周期智能体的轨迹会超出其两个消费者的承载范围：监控运行的人类观察者，以及智能体本身——轨迹必须重新折叠回其受限上下文中。我们提出一种实时轨迹模型：一个仅追加的事件账本，逐步折叠为带类型的运行状态，并编译为面向不同消费者的视图；我们基于确定性真实值，分别对这两类消费者进行了评估。对于观察者一侧，以 LLM reader 作为代理进行评估时，编译后的视图回答监控问题所用的输入 token 约比对原始轨迹进行一次受预算限制的单次读取少 14 倍和 15 倍，成本低 5-7 倍，且准确率更高（0.85-0.87 对 0.48）。由于这些问题是与视图 schema 共同设计的，我们将 token 和成本降低视为在 schema 覆盖充分这一条件下可迁移的结果。对于智能体一侧，在 120 链顺序依赖任务上，将任务的运行统计量维护在逐步状态中的机制，在完整上下文提示失败的地方取得成功（在干净协议下为 30/30 对 8/30，n=30；由于基准与系统共同开发，标注为描述性结果）；提示级 scratchpad 以更低成本达到了与该折叠相同的准确率，而双臂分解表明，该折叠的准确率来自其确定性聚合，其成本优势来自其紧凑性。该折叠相较更便宜替代方案剩余的价值在于可确定性审计，以及可基于同一状态为观察者提供服务。我们从观察到的失败中归纳出 11 条轨迹折叠的候选需求，并用一个对顺序敏感的任务族对其边界进行界定；在该任务族上，折叠不再提供帮助。代码、基准、可再生的合成语料库以及所有工作台轨迹均已发布。
本文研究自主软件开发，即基于 LLM 的编码代理在无需人工干预的情况下，将高层需求转化为完整、可运行且可用的软件系统。我们提出 Harness-of-Harness (HoH)，这是一个使编码代理在自主开发过程中持续改进软件的框架。HoH 基于现有的 coding-agent harness 运行，并将其执行组织为迭代式的规划-编码-测试循环。为在各轮循环中持续提升，HoH 在修复与能力增长之间取得平衡，将开发范围限定为小而可验证的增量，将实现阶段测试与独立评测分离，并约束可验证的输出而非规定代理的工作流。它逐步开放交付物、面向角色的工具和技能，鼓励复用而非重建，并维护带版本记录的项目历史。在 GameCraft-Bench、FrontierSWE 和 ProgramBench 上，三个 harness-model 配对（Codex with GPT-5.5、OpenCode with DeepSeek-V4-Pro，以及 Pi with MiniMax-M3）中，HoH 始终优于对应的独立 harness，在三轮迭代后取得平均相对提升 52.25%，最高提升 82.86%。在一次持续多日、超过 70 轮迭代的部署中，HoH 自主开发出一款第一人称射击游戏，具备连贯的故事线、完整实现的核心机制、可供人类游玩的体验、精致的视觉效果以及集成音频。Github: https://github.com/Flesymeb/HarnessOfHarness Project Page: https://flesymeb.github.io/HarnessOfHarness/
技能增强型智能体将可复用技能作为持久的运行时上下文加载，从而提升任务表现，但也为恶意技能提供了一个持久渠道来操控后续动作。这类技能可能泄露机密、破坏代码、绕过审批，或在具体用户任务和工作区状态使不安全操作看起来有用之后，再为数据外泄做准备。这使得安装前审查不足，并需要在运行时基于任务条件提供保护。我们提出 Defense-as-Skill，一种将运行时防护本身实现为可安装、可检查和可编辑技能的防御范式。我们的防护器 SkillSonar 与不可信任务技能并行运行，并将敏感操作与用户任务边界进行比对，无需修改底层智能体运行时即可将每个操作导向允许、重新规划或确认决策。为研究这一设定，我们构建了 SCOPE-R，一个任务条件化数据集，涵盖 6 个风险家族和 21 个子类别，包含 206 个经攻击确认的恶意实例和 43 个良性任务。随后，我们在 SCOPE-R 训练子集上利用运行时防护技能演化来改进 SkillSonar；这是一种 Monte-Carlo Tree Search 过程，可根据 rollout 的反馈演化磁盘上的防护技能。在 Claude Code 和 OpenClaw 上，演化后的防护器显著降低了攻击成功率，同时保持了良好的安全-效用权衡。在重复的 GLM-5 运行中，SkillSonar 将 ID ASR 从 0.482 降至 0.104，将 OOD ASR 从 0.606 降至 0.115。进一步分析表明，该方法可跨受害模型、留出风险家族和外部基准迁移，并且对自适应攻击者仍保持防护能力。消融实验进一步表明，显式的安全责任分配和技能原生表示对于所观察到的收益都很重要。
LLM 代理之间相互交互的频率不断上升，引发了关于多 LLM 代理环境中语言演化的关键问题，这不仅关系到安全性和可监测性，也关系到对 LLM 的语言学解释。为回答这些问题，我们提出 GlossoGen，这是一个用于研究复杂场景中多智能体语言演化的新平台。在 GlossoGen 中，我们构建了 SaveVeyru 场景，该场景要求拥有部分信息的代理在压力下进行沟通。我们发现，LLM 代理之间确实会发生语言演化，所得语言具有组合性和形态生产性，并且会以某些方式偏离 LLM 的英语先验，从而使人类无法理解。此外，我们识别出这种演化所必需的几个特征：效率驱动的压力；支撑这些代理的模型强度；以及进入“postmortem”阶段的机会，在该阶段代理可以就语言约定达成一致。重要的是，我们观察到，语言向新代理的传递受不同条件支配。具体而言，我们发现代理仅通过使用就能学习新语言，并在这一学习过程中发挥主动作用；而尽管新的语言涌现需要更强的模型，但一旦某种语言已经出现，较弱的模型也能学习它。综上，我们的结果表明，当前 LLM 具有累积文化演化的潜力——这一现象此前仅在人类中得到证实——而且由不同能力混合组成的代理群体能够发展出超越其最低共同标准的能力。
交互式模拟正越来越多地用于评估由语言模型代理构成的市场中的政策。它们的输出看起来可以反映经济现象——价格、利润、消费者剩余和福利——却未必实现了该主张所指称的行为。我们在一个用于可配置酒店交易的多轮买方--卖方测试平台中审计了这一风险。一个初始实现报告称，在 Qwen2.5 1.5B--14B 级别上，两个市场护栏带来的福利增益分别为 +87.4、+35.0 和 +28.8。它还为受护栏和未受护栏的代理使用了不同的报价模式和选择流程。固定报价模式和买方选择器后，成对对比变为 +7.2、-13.9 和 +23.8。四个最大的 14B 单次生成效应平均为 +229；在每个画像-条件下进行三代后，它们的平均值为 +37.6（95% bootstrap 区间 [-34.2, 109.3]），而生成残差在这一事后探测中解释了 49.9% 的变异。卖家激励检查呈现非单调性：提高利润压力得到的利润反而低于默认卖家提示。脚本化的正向对照说明了这为何重要。一个利润最大化卖家本就能达到第一最优福利，因此护栏主要是在重新分配并降低福利；只有当卖家被显式编程为强制低效捆绑时，它们才会创造福利。我们提出了一个构念效度契约，用于区分激励有效性、协议隔离、随机稳定性和福利核算，并在实质性政策主张之前返回 INVALID 或 INCONCLUSIVE。在我们的案例中，原始估计在协议隔离下是 INVALID，而控制研究在激励有效性和随机稳定性下仍然是 INCONCLUSIVE。这个案例并未表明护栏无效；它表明，在模拟代理和协议通过这些检查之前，它们的表面价值是无法识别的。
Scientific agents must learn not only how to reason, but also what to believe. However, existing LLM agents typically express scientific hypotheses in free-form text, leaving their beliefs implicit and difficult to test or revise. We introduce EvoSCM, which equips scientific agents with explicit structural causal models that evolve as new experimental evidence is collected. EvoSCM maintains a population of competing SCM hypotheses, each encoding a candidate causal explanation of the environment, and evolves them through a closed discovery loop. In each round, the agent abduces latent mechanisms from accumulated evidence, designs discriminative interventions, and commits to falsifiable predictions that it tests through experimentation. Discrepancies between prediction and observation are inductively distilled into correction rules that revise the causal structures and mechanisms of each hypothesis, and the agent then deductively validates the revised population against accumulated evidence and structural consistency to guide the next round. We evaluate EvoSCM on DiscoverPhysics, a benchmark requiring agents to uncover the hidden dynamics of noncanonical physical worlds through experimentation. EvoSCM consistently improves scientific discovery over baselines, yielding more accurate explanations and predictions while making more effective use of experimental interactions.
我们评估了嵌入检索，在刻意将表层形式与语义分离的情况下：检索那些共享底层结构但不共享措辞的条目，并在两个互不相关的领域中采用同一协议，分别是竞赛数学（MathNet-Retrieve；500 个查询，117,088 条目语料库）和具身智能体轨迹（基于 ALFWorld；118 个查询，336 条轨迹）。在数学领域，失败是彻底的：在最强伪装层级下，两个生产级嵌入器的严格 Hit@1 都为 0.0%（bootstrap 95% CI [0.0, 0.0]），而正确条目几乎总是位于前 10 名；在 95.2% 到 99.8% 的未命中情况下，获胜结果在词面上比正确答案更接近查询。在轨迹领域，当 gold 必须涉及不同对象时，表层变化只是附带的，同样的模型落在或接近超几何随机水平；一旦 gold 还必须在对象和容器上都不同，三种嵌入器都低于随机水平：检索锚定于字面 token，而非任务结构。词法重排序器控制在数学中有害、在轨迹中有益（将差距缩小 26% 到 36%，CI 不包含 0）；其方向揭示了某个基准的表层变化是对抗性的还是附带性的。LLM 重排序器在数学中恢复了 5% 到 63% 的差距，在轨迹中恢复了 43% 到 76%；这一方向在三个评审器中可复现（21 个单元格全部为正），但效应大小、层级分布以及离群评审器会随领域而变化（配对差异处处排除 0）。数学上的增益主要集中在知名竞赛上（+19.8 个百分点，CI [+6.7, +33.2]，六个单元格中的一个），因此恢复的一部分来自记忆。在一项配对的下游实验中（210 个查询，评分者一致率为 96% 到 99%），oracle 检索与对抗性劣质检索没有显著差异（McNemar p = 0.678）；求解器 69.5% 的 zero-shot 准确率在很大程度上只是一个截断代理指标（在完整答案上为 97% 到 100%），几乎没有提升空间。
Vision-Language Models (VLMs) provide useful priors for interactive decision-making, but using them directly as policies is expensive and brittle: they must be queried at every step, do not improve from environment interaction, and can repeat systematic errors. We study how to learn a cheap autonomous policy from an online, expensive, and imperfect but informative VLM teacher. We propose SAGE (Selective Agent Guidance via Entropy), a framework that queries a VLM only when the learner is uncertain, executes the suggested action during training, and distills guidance into a lightweight Reinforcement Learning (RL) policy. Because VLM advice is not always reliable, SAGE can weight teacher-action distillation using environment-derived advantages rather than treating all suggestions as equally useful. Across sparse-reward visual reasoning and navigation tasks, SAGE learns policies that act without VLM guidance at evaluation time and improves over unguided RL in several environments, including settings where the learned policy exceeds its VLM teacher. The results show that selective guidance is most beneficial when the VLM can help the agent discover high-reward trajectories, and less useful when unguided exploration already succeeds or teacher actions do not lead to informative experience. SAGE also reduces VLM usage by prompting the teacher only on a fraction of training steps and requiring no VLM calls at deployment. Overall, our results suggest that VLMs don't need to be used as fixed policies to be useful; they can instead act as temporary, imperfect sources of guidance whose value is tested and internalized through interaction.
Writing involves diverse cognitive activities, from ideation to revision, and writers' needs vary across individuals and moments. Proactive AI promises to provide the right support at the right time, yet existing proactive tools largely focus on generic textual assistance, such as autocomplete. This paper studies the design space of proactive thought partners: AI agents that proactively offer customizable, higher-level cognitive support during writing. We instantiated this concept in a technology probe and deployed it with 16 participants for one week. The probe allows users to create partners by configuring their roles and proactivity. As users write, relevant partners take the initiative at appropriate moments to offer suggestions. Our findings show that participants configured proactive support through prospective planning, used suggestions for both idea generation and self-monitoring, and valued lightweight visual representations alongside non-directive rhetorical framing for non-intrusive interventions. We derive implications for designing proactive writing assistants around customization, timing, engagement, and representation.
Dynamic agent harnesses let language models change the software that shapes their own execution. This flexibility brings a new reasoning burden: a local plugin change can propagate through dependencies and cleanup. We introduce CordisBench, a 1,200-question benchmark of this lifecycle reasoning. It combines a controlled formal setting with programs executed against Cordis, a runtime that manages component dependencies and cleanup, and asks models to identify affected components, predict state after a specified teardown order, determine which conditions hold under all or some orders, and choose reconfigurations that succeed when executed. Across these tasks, we evaluate three efficiency-oriented models at low reasoning effort with 2, 4, 8, 16, 24, or 32 relevant interactions, using deterministic task-specific scoring. Models usually handle small systems well but grow less reliable as more interactions become relevant, especially when predicting final state and when reasoning across teardown orders. Additional inference effort recovers marked gains for some models. The cost is nontrivial: on our 16-interaction subset, GPT-5.6 Luna uses nearly 3,000 reasoning tokens per question at medium effort. For these controlled instances, that cost is avoidable: an independent finite reference semantics agrees with Cordis execution on every observation and action outcome used for scoring across all 528 executable questions.
在真实基准上评估软件工程 agents 成本高昂，因为每个任务都可能需要多步代码探索、修改和测试执行。现有的高效评测方法会选择具有代表性的子集来估计完整基准上的性能，但大多只关注结果：它们拟合历史通过/失败响应矩阵或静态任务语义，忽略了 agents 解决问题的方式。我们提出 PTA-IRT，一个 Privileged Trajectory-Aware Item Response Theory 框架，融合过程信号和结果信号。历史执行轨迹提供了超越通过/失败的过程级证据，例如探索过的上下文、尝试过的编辑以及求解路径，PTA-IRT 将其用作用于校准子集选择和能力估计的特权信息。在较低的校准预算下，PTA-IRT 在四个 SWE 基准上的分数恢复和排序恢复方面始终优于先前的 IRT 基线。代码和数据可在 https://github.com/DeepSoftwareAnalytics/PTA-IRT 获取。
Using a large language model to play the user is now standard in scalable evaluation. It has a repeatedly diagnosed failure: the simulated user is excessively cooperative, so a system under test can score by the sheer number of questions it asks rather than by making the user willing to speak. We answer with a disclosure gate conditioning information release on the companion agent's behaviour: its state is a ladder of five ordered gates, merged onto three observable depth layers. We specify, ablate, and audit it, and train a user simulator against that specification. Gating behaviour is learned from the training corpus's synthetic branch, while the real branch supplies how people speak and react; after training, the simulator need not be told at runtime which gate each item sits behind. The gate is a load-bearing component of the environment: on the English corpus of a published companion-agent benchmark (CompanionBench), once training no longer states per example which gate each item sits behind, the largest rank displacement across 12 systems under test exceeds the noise band set by re-running that environment under a new seed, while per-system scores show no detectable change. We state two acceptance criteria: a ranking must be order-preserving, and absolute scores must be scale-stable. Of the candidates we examine, only one passes both -- the simulator we release -- and its leaderboard correlates at 0.993 with the benchmark's original simulator. By contrast, prompting a frontier model as the simulator barely moves the ranking while shifting every score upward -- a shift invisible to anyone checking the ranking alone. The environment we specify is the one that benchmark already used. That publication describes the mechanism in about four hundred words, and we supply what it lacked: specification, ablations, human studies, negative controls, and downstream sensitivity analysis.
Scientific methodology figures are essential for communicating complex methods clearly, yet creating them remains labor-intensive and typically requires multiple rounds of refinement. Recent image-generation models can synthesize visually appealing raster figures, but producing a human-satisfactory result in a single generation step remains difficult. Moreover, precise edits to raster figures are challenging for both humans and models. We formulate scientific figure generation as recursive SVG program construction and propose \textsc{FigTree}, a \textit{multi-agent} system that automatically transforms a scientific paper into a structured vector figure. \textsc{FigTree} grounds figure content in the source paper, decomposes a figure into a hierarchy of local regions, generates each region as a short SVG program, and assembles the resulting fragments. A render-critic refinement loop jointly inspects the rendered figure and its underlying program, enabling visual defects to be traced to specific statements and accurately repaired. We conduct extensive evaluations of \textsc{FigTree} on figure quality and editability, showing that \textsc{FigTree} produces high-quality figures, while also enabling more effective editing than existing raster-based methods.
递归 LLM 代理可以通过生成专门分支来扩展其搜索。一些分支随后会请求发送数据或部署代码的工具。一个分支应在何时获得行动权限？我们区分沙箱式生成，在这种情况下外部控制会阻止所指定的损害；以及能力激活，在这种情况下，被选中的分支跨过不可逆行动边界。渐进式风险归属（PRV）将轨迹级风险预算作为托管资金保留，并在分支被激活时进行扣减。我们证明了适应性生成树的任意时刻损害上界。分支结果可能彼此依赖，但每个局部证书都需要在完整的激活前历史条件下保持有效，包括用于选择请求的信息。当激活门、分支计费和计算约束保持固定时，延迟归属保留了在不可撤销的生成计费下可用的每一种策略。边际风险估计在分支选择后仍可能失效。在一个风格化的分支模型中，随着授权再生产数 $\mathcal{R}_A$ 穿越 1，轨迹损害发生变化。当局部风险 $p$ 逼近 0 时，轨迹损害在临界点以下与 $p$ 成正比，在临界点处与 $\sqrt{p}$ 成正比，在临界点以上则保留一个正下界。一个有限类型占用模型给出了风险和计算的影子价格。对于每单位风险的边际价值递减的嵌套扇出模式，这些价格产生一个阈值规则。分支计算和一个分样本实验展示了这些结果。这些合成研究并未估计已部署代理的安全性。该分析提出了一条设计规则：在沙箱中广泛搜索，并谨慎授予递归权限，同时设置明确的风险收费。
A/B 测试是评估产品变更的黄金标准，但每次实验都需要真实用户流量、工程投入以及数周的观测时间。我们提出一个模拟框架，利用由 LLM 驱动、以基于真实用户行为信号构建的数据驱动人物画像为条件的智能体来预测 A/B 测试结果。不同于依赖合成或基于规则的人物画像的先前工作，我们的智能体由匿名化行为数据、活动模式、参与度信号以及推断的人口统计特征构建而成，从而能够更真实地建模总体人群。我们将 A/B 测试模拟表述为一项结构化问答任务，并系统研究：(i) 问题设计格式，(ii) 人物画像数据来源与领域对齐的影响，(iii) 单个人物画像的行为深度与总体多样性之间的权衡，以及 (iv) 高效的人群子采样。在一个涵盖两类指标的 40 个 A/B 测试基准上，我们的最佳配置根据测试指标不同可达到 0.75–0.90 的方向准确率，表明数据驱动人物画像是实现快速、低成本实验预筛选的一条可行路径。
大型语言模型（LLMs）作为智能体系统中的强大组件，已展现出卓越能力，能够支持复杂推理和复杂任务执行。然而，当前对智能体系统进行手工设计和优化的方法高度依赖人工投入，限制了其适应性和可扩展性。近期工作探索了工作流设计的自动化优化。然而，这些方法往往忽视模型能力的关键作用，并且只关注单一性能指标，未能处理真实世界的部署约束。在本文中，我们提出 AgentFactory，这是一个在优化智能体系统中的基础模型和工作流结构的同时，综合考虑性能、成本和效率等多个目标的框架。AgentFactory 利用先进的 LLMs 作为优化器来搜索可能配置的巨大空间，并采用三阶段优化流水线自动发现微调模型与优化工作流的有效组合。通过迭代优化过程，我们的框架系统地探索和评估不同的智能体系统设计，在保持运行效率的同时，针对特定任务需求进行自适应调整。我们在涵盖五个领域的八个基准上评估了 AgentFactory，包括通用推理、代码、数学、医学和金融。实验结果表明，AgentFactory 始终优于手工设计的方法和现有自动化方法，在所有基准上平均提升 9.1%，在领域特定任务上取得了尤为显著的提升（MedQA 上提升 19.6%，FinEval 上提升 18.7%）。这些结果表明，AgentFactory 是一种通过自动化优化开发更强大、更高效智能体系统的有前景的方法。
尽管由 LLM 驱动的智能体在复杂环境中被越来越多地用于解决多步骤任务，但现有基准很少测试状态保持、跨语言表现以及在真实、具身场景中的应用。为了解决这些问题，我们提出 WorldBench：一个全面的多语言基准，涵盖真实的、基于人物设定的日常工作流，智能体可以通过结构化动作在沙箱中执行。WorldBench 包含跨七种语言和八种文化的 1,600 个任务，并通过具备特定语言和文化专业知识的人类标注者反馈进行了筛选和优化。在评估方面，我们扩展了以往工作的指标，并引入 Constrained Task Success (CTS)，该指标结合自然语言指令和测试环境，通过确定性评估和 LLM-as-a-Judge 评估，对任务完成情况、最小改动以及其他互补指标进行评分。我们的实验表明，前沿模型的 CTS 最高仅达到 49.2%，所有模型在正确性与环境保持之间都表现出显著差距。因此，我们表明当前智能体在多语言、智能体化场景中仍然较为脆弱，尤其是在长程任务和状态保持约束下。
通过强化学习（RL）训练开放式智能体，受制于缺乏可验证的金标准答案和可扩展的评分细则。此外，即便接近模型能力边界，长时程的开放式代理任务也常产生脆弱且不稳定的奖励，导致弱或噪声化的 rollout 对比，掩盖了基于组的策略学习所需的细粒度优化信号。为应对这些挑战，我们提出 ARISE-RL，一种新颖的全周期自进化框架，通过评分细则介导的协同进化，将任务/评分细则生成器与推理求解器耦合起来。生成器将与工具相关的评分细则标准锚定在真实工具观测之上，并因生成与求解器不断演化的能力边界相一致的有效、中等难度任务而获得奖励。求解器则通过多步推理和工具使用，从细粒度的评分细则满足信号中学习。我们进一步引入 Reward-Gated Self-Evolution Distillation（RG-SED），仅当记忆带来经验性奖励提升时，才将同一策略的带记忆变体选择性蒸馏回自身，从而减少分布不匹配并避免盲目模仿噪声指导。最后，为支持严格评估，我们提出 ECR-Bench，一个专家校准的评分细则基准套件，涵盖单工具深度研究和多工具旅行规划。大量实验表明，ARISE-RL 在所有评估基准上始终取得稳健且稳定的整体最先进性能。
我们提出 CogEvol，这是一系列专为 Learning Environment Generation 训练的模型：可在单次生成中将课程简介转化为完整的学习产物（结构化 JSON 幻灯片或自包含交互式 HTML 页面）。在 22 万条生产请求中，CogEvol 生成一份幻灯片的中位耗时为 17 秒，生成一个交互式页面为 59 秒，取代了耗时数分钟的多轮 agent 脚手架流程。可靠性不是寄希望于运气，而是被强制保障的：一个基于生产环境的数据管道将真实故障转化为 53,687 个经过验证的 SFT 样本，而混合的规则加 VLM 奖励驱动基于 GRPO 的 RL；在我们发现并修复了一起奖励黑客事件后，该系统进一步加固，该事件曾生成视觉上令人信服却无法游玩的游戏。CogEvol-27B 在幻灯片质量上得分 83.7，在一个包含 500 个案例的 interactive-HTML 基准上得分 63.7，参数量比旗舰级代码模型少 26.9 倍，并与 OpenMAIC 团队合作，承载其线上生产流量。CogEvol-4B 已在 Apache 2.0 许可下于 https://github.com/CogEvol/CogEvol-4B 开源发布；外部旗舰模型在相同基准上、使用完全一致的评测框架进行测评。脚手架编辑进一步将交互式页面生成成本降低约 76%，整套系统可在国产 Ascend 加速器上运行，并在应用层面与 A800 GPUs 持平，从而降低 AI 原生教育的大规模单位成本。
AI agents in partially observable environments need to coordinate active sensing with working memory to maintain an evolving perceptual state. However, existing benchmarks struggle to isolate this perceptual-state construction and interpretation capability because they introduce physical and control complexities. We address this with MNIST-PRO, a benchmark that isolates agentic perception by converting MNIST digit recognition into a sequential, glimpse-based search task with lookback constraints. We evaluate ten multimodal models across four memory representations, including raw visual history, textual states, structured metric grid maps, and a consolidated visual canvas. While models excel under full observability, partial observability exposes a clear performance gap. We identify three distinct bottlenecks. First, perceptual-state construction and interpretation present a challenge, as agents struggle to integrate fragmented glimpses. Second, agents often stop exploring before they see the full sequence. Third, models often fail to revise early, incorrect beliefs even when faced with subsequent contradictory evidence. These results show that simply acquiring visual evidence is not enough. Agents must also be able to build and update a reliable perceptual state.
Agent工作记忆具有异质性。指令、工件、工具输出以及Agent生成的状态等对象在语义角色上不同，并表现出不同的大小、保留和表示特征。近期工作已开始探索能够考虑这种异质性的记忆管理机制。本研究聚焦于语义异质性，并研究它应如何影响代码代理中工作记忆的管理与评估。在55条已归档的代码代理轨迹中，我们发现语义不同的工作记忆对象表现出不同的保留和压缩行为。这种异质性促使采用语义感知的记忆管理。我们研究了两种语义感知策略：一种是对象感知的压缩策略，另一种是基于检索的策略。评估结果显示，校准收益未必能迁移到留出任务上，而且相同的token预算并不意味着相同的实际上下文交付量或管理成本。对真实系统的回放进一步暴露出仅凭名义预算无法捕捉的服务限制。综合来看，这些结果说明了语义结构对Agent工作记忆的重要性，以及为什么评估记忆管理策略不能只看名义上的token预算。我们将这些经验归纳为四个层次：存储状态、交付上下文、管理工作，以及任务或过程结果。
自主科学研究代理正越来越多地应用于端到端的科学工作流，包括文献综述、数据分析、实验和报告生成。然而，开放式研究任务往往并未明确指定完成任务所需的分析、方法和成功标准。因此，代理可能会遗漏重要分析、使用不恰当的方法，或得出缺乏充分证据支持的结论。为了解决这一问题，我们提出 AutoSciRub，这是一个以评估为先的框架，它在研究执行之前先归纳出一个任务特定、可执行的评分标准，并利用该标准来指导执行、进行准则级验证以及迭代修订。AutoSciRub 将一个规定不明确的指令分解为原子级科学目标，将其锚定到相关文献和任务可见数据中，并综合生成具体、可操作且可验证的准则。由此得到的评分标准将隐含的实验与证据要求显式化，为实验和分析提供指导。在修订过程中，基于评分标准的验证会识别未满足的准则，并支持对研究报告及其支撑工件进行有针对性的改进。在 ResearchClawBench 上，AutoSciRub 在所有测试配置中都持续带来提升：在固定的 Codex harness 下，三个基础 LLM 的平均提升为 2.08 分；在使用固定 DeepSeek-V4-Flash backbone 的三个 agent harness 上，平均提升为 2.95 分。在 AstaBench E2E Discovery 随机抽样的 20 个任务子集上，AutoSciRub 在三个 agent harness 上进一步实现了平均 16.8 分的提升，同时保持或增加了成功完成的任务数量。这些结果表明，以评估为先的引导为自主科学研究提供了一种有效且可泛化的控制机制（Code: https://github.com/zjunlp/AutoSciRub）。
Outcome-based reinforcement learning provides verified feedback for language-model agents, but assigns trajectory-level advantage uniformly to all decisions, yielding coarse credit over long-horizon interactions. On-policy self-distillation offers finer supervision by re-evaluating sampled behavior with privileged information (PI) available only during training. However, fine-grained supervision is not necessarily fine-grained credit: PI-induced likelihood changes describe how additional information alters policy preference, but do not directly determine how an executable action should inherit the verified task outcome. This creates a supervision-credit gap. Privileged signals may be irrelevant to the current interaction state, operate at a token granularity misaligned with executable decisions, and lack the outcome semantics required for reinforcement. We introduce TASPO, which converts privileged supervision into outcome-grounded action credit. TASPO constructs decision-applicable PI from verified successful experience, aggregates PI-induced likelihood shifts at the executable-action level, and converts relative action support into positive, bounded, mean-preserving weights on the original trajectory advantage. Thus, the verified outcome determines the update direction and average scale, while PI only redistributes credit across actions. Across three agentic benchmarks, TASPO improves over GRPO by 10.6\% and generalizes better to unseen tasks. Further analysis indicates that TASPO reduces supervision mismatch and that action-level assignment stabilizes the policy optimization process. These findings offer the community another interesting perspective.
有价值的数据仍然埋藏在非结构化来源中：网页、报告、合同、申报文件、财报电话会议记录以及 PDF。企业 AI 的一个重大押注，是部署能够对这些数据进行推理的 LLM agents，为每一位知识工作者回答复杂问题。如今，agents 已经可以做到这一点，但成本高得令人难以承受。每个问题都要反复打开大文档来检索分散的证据，消耗高达一百万个 token。然而，如果数据本身已经结构化，同样的问题就会变成一次廉价的数据库查询。比如在 FanOutQA 基准上，对理想的预结构化存储进行推理的成本要低 28 倍，而且随着问题需要涉及更多文档，这一差距还会扩大到数量级。然而，提前把一切都结构化并不可行：文档所包含的潜在结构远比任何工作负载会用到的多，而哪些结构有用、哪些文档相关，只有在查询到来之后才会知道。我们提出 agentic data cracking，这是一种在推理本身的副产物中，对非结构化数据进行自适应且带有推测性的结构化方法。之所以称为自适应，是因为已观察到的查询决定了何时结构化以及哪些内容重要；之所以称为推测性，是因为它超越了当前问题。每当智能体为回答而打开某个文档时，一个 cracking 子智能体就会以边际成本从已加载的上下文中分叉出来，提取有据可依的结构化信息，以服务未来可能相关的查询。随着时间推移，越来越多的查询能够完全由结构化数据覆盖，无需打开文档即可回答，从而在接近 RAG 成本的同时保持 agentic 准确性。在 FanOutQA 上，若仅为每个测试问题额外增加一个相关问题，cracking 就能在保持准确率的同时将成本降低 53%。agentic data cracking 是迈向下一代数据基础设施的第一步，用于在非结构化数据上进行 agentic 推理：一个位于模型之下的共享基底，让推理已经付出代价发掘出的知识持续积累。
Multimodal Large Language Models (MLLMs) have shown remarkable success in STEM domains, where progress is often driven by vertical, step-by-step deduction under relatively stable symbol systems. Their horizontal, interdisciplinary cultural reasoning, however, remains underexplored.We propose CM2, a multi-agent framework grounded in the cognitive pathway of human cultural interpretation. CM2 integrates multimodal perception, retrieval-augmented generation, networked reasoning, gated fusion, and reward-driven feedback.Experiments on CM2D across multiple MLLM backbones show consistent gains over CoT and typical reasoning paradigms; ablations validate each module's contribution, and conflict analyses confirm genuine cross-modal arbitration.
现实世界中的多智能体系统（例如无人机蜂群、自动驾驶汽车、仓库机器人）必须在避免碰撞的同时满足丰富的时序任务。信号时序逻辑（STL）能够优雅地对这类目标进行编码，但当前的 STL 规划方法面临关键限制。最先进的基于优化的方法可以处理任意 STL 规范，但难以扩展，随着智能体数量增加，其计算上变得不切实际。基于学习的方法能够以快速的规划时间高效处理大量智能体，但当部署时的目标与训练时不同，它们表现不佳，并且不支持需要为不同智能体分配不同规范的规划任务（即异质性），也不支持需要多个智能体协同的团队级规范。这种可泛化性与可扩展性之间的根本权衡，给在实践中实现多智能体 STL 规划算法带来了挑战。为克服这一挑战，我们提出了一种用于带有 STL 规范的多智能体规划的新扩散方法。通过使用 STL 的可微近似，我们将 STL 梯度集成到去噪过程中，使我们的方法能够泛化到新的公式，即其谓词位于训练期间覆盖的目标区域内任意位置，同时达到与现有基于学习的方法相同的可扩展性。我们的方法支持异质规范，并且借助扩散模型，自然增强了规划多样性，从而显著减少了智能体之间的安全相关违规（例如碰撞）。一项详细的评估研究证明了 STL 引导的基于扩散的多智能体规划器在构建可泛化、可扩展且多样化的计划方面的实用性。视频和代码可在 https://www.jeappen.com/diff-ma-stl/ 和 https://github.com/jeappen/diff-ma-stl 获取。
冻结的视觉-语言-动作（VLA）策略提供了广泛的操控能力，但它们会在不跟踪任务进展的情况下执行开环动作块，因此智能体无法可靠地判断是应继续、重试还是终止。外部记忆是一种自然的补救办法，然而当把尝试过的动作当作已完成的进展时，它反而可能有害，从而把局部执行错误转化为持久的任务状态错误。我们提出 Achievement-Grounded Memory（AGM），一种面向冻结 VLA 策略的轻量级闭环框架，它将任务表示为带有进度指针的子目标序列，并且只有在当前子目标被物理证据验证之后才推进这段记忆。本体感觉交互线索决定何时进行验证，而来自冻结基础模型的连贯点跟踪与语言条件下的跨视角比较，则通过一个仅含 2.43M 参数的验证头来决定已完成了什么。AGM 由此将开环执行转化为由执行、验证和进度组成的闭环，同时保持策略冻结，无需在测试时进行大模型推理。在 RoboMME Counting 基准上，AGM 在 PickXTimes 和 BinFill 上取得了显著结果，平均超过最强的带记忆增强基线若干个百分点，并且该框架在实体机器人上也带来了同样明确的提升。因此，可靠的具身记忆更多依赖于规范的状态更新，而不是记忆容量。
自主大型语言模型（LLM）智能体在部署于复杂、长时程任务时，正日益面临可靠性、上下文消耗和执行稳定性的瓶颈。尽管单体式提示工程和无状态工具调用范式难以扩展，但该领域正迅速转向 \emph{agentic skills}：一种模块化的过程抽象，将执行知识外显为可复用、可执行且可移植的工件。本文为 agentic skills 生态系统建立了统一的系统基础和参考架构。我们将技能形式化为外显的过程性知识，连接高层认知规划与确定性执行环境，并系统性地划分出一个九阶段生命周期中的架构：自主发现、编写与表示格式、记忆存储、动态检索与路由、组合与编排、执行与修复、终身适应、经验评估以及安全治理。我们进一步考察了市场动态、公共注册库和新兴的对抗性威胁向量，并讨论运行时验证和防御机制。最后，我们对涵盖软件工程、操作系统导航、具身机器人和科学发现的系统实现进行了分类，同时强调了持续学习和基准真实性方面的关键开放挑战。这项工作将 agentic skills 确立为构建可扩展、稳健且可验证的自主语言智能体的基础性范式。
Sustained scientific work requires a knowledge substrate that carries interpretation across tasks and preserves paths to source evidence. We call this process \emph{scientific knowledge compilation} and implement it in ASKS, the \emph{Agent-Driven Scientific Knowledge System}. For each source, an LLM produces a readable Wiki view and machine-facing semantics. Deterministic checks convert the latter into a document-local GraphDelta, and embedding geometry together with explicit graph rules integrates the proposed changes into persistent state. Each ingest is an inspectable state transition over accumulated knowledge, with compiled Wiki and graph views linked to the preserved source record. We examine this process by chronologically compiling 56 published papers from one research program. Branch survival, cross-paper support, lineage, coverage, and churn yield a source-traceable author research portrait centered on tensor-network methods, with branches into quantum many-body research, tensor-network machine learning, and quantum-AI-oriented directions. In this run, higher-level Hub organization remains stable and low-churn. Canonical-node growth is predominantly additive. Graph-level measurements and navigation paths retain links to the source records from which they were compiled.
本文提出了一种混合事实核查框架，将基于知识图谱的语义记忆与对抗式多智能体推理相结合，用于可解释的虚假信息检测。所提出的系统采用记忆优先、网页兜底的架构，其中输入主张首先通过基于 Sentence-BERT 的语义检索和自然语言推理，在双索引知识图谱上进行评估。当从图谱中检索到的证据不足以支持可靠决策时，框架会从可信网页来源收集信息，并使用由支持、反驳和裁决智能体组成的对抗式法庭对其进行评估。一种感知图结构的置信机制结合语义相似度、NLI 置信度和结构化图证据，以判断内部知识是否充足，从而减少不必要的网页检索。验证完成后，已确认的信息会被转换为结构化三元组并纳入知识图谱，从而支持系统语义记忆的增量扩展。基于一个整理的 COVID-19 虚假信息基准的实验评估表明，所提出的框架在已解决主张上达到了 97.4% 的准确率和 92.6% 的宏平均 F1 分数，优于 Llama~3.3~70B 基线，后者的准确率为 87.7%，宏平均 F1 分数为 86.3%。
长程、故事驱动的视频生成需要一个制作智能体来协调叙事拆解、状态跟踪、镜头设计、提示词构建、渲染以及跨相互依赖场景的修订。现有的自适应视频系统主要是优化请求或可复用技能，但这些改进与跨故事、面向特定阶段的持续改进相脱节。我们提出 CineForge，这是一个自我进化的视频制作智能体框架，将用于视频生成的 CineForge-Produce 与用于跨故事策略演化的 CineForge-Evolve 结合起来。CineForge-Produce 将每个源故事组织为类型化的叙事、角色、空间和电影状态，利用它们协调素材和片段生成，并将过程记录为规范化的制作轨迹。CineForge-Evolve 采用 Case-to-Pattern-to-Policy Evolution (CPPE) 来审查轨迹证据，将重复出现的发现整合为有界的阶段局部补丁，并通过结构化重放和置信度受控的成对评估部署经过验证的更新。为了衡量完整的故事实现，我们提出 CineScope，它结合了一个包含 100 个脚本的 CineScope-Data 套件，以及一个与人类对齐的多尺度 CineScope-Metric，覆盖因果状态、导演编排、节奏与资源分配，以及角色弧线。在 CineScope-Data 和两个公共基准上，演化后的 CineForge 策略将 CineScope-Metric 从 4.024 提升到 4.380，在 ScriptAgent 下相较三个长视频基线取得一致增益，并在新故事上将审查 LLM 调用减少 37.0%。这些结果表明，制作轨迹可作为可操作的经验，使视频智能体能够在长篇叙事任务中累积改进。
Retrieval-Augmented Generation (RAG) 提高了 large language models (LLMs) 的事实性，但现有 RAG 系统往往难以应对需要自适应检索和持续修正中间上下文的复杂多步推理。近年来基于强化学习（RL）的 agentic RAG 方法在一定程度上缓解了这一问题，但通常依赖粗粒度动作空间和轨迹级奖励，导致奖励分配较弱，并偏向于短视、刻板的推理模板。为此，我们提出 AgenticRag-R1，这是一个通过记忆栈和细粒度动作空间深度整合推理、检索和记忆的 RL 框架，并结合层次化的动作感知奖励和信息感知的轨迹拒绝策略，以实现有效的长程学习。我们在一系列多跳、开放域和 agentic 推理基准上进行了实验，覆盖多个基础模型规模，结果表明 AgenticRag-R1 始终优于强基线。此外，AgenticRag-R1 学到了更稳健、可解释且更具记忆感知的推理行为，凸显了细粒度动作建模和信息感知优化对长程推理的效果。我们的代码匿名可见于 https://github.com/jiangxinke/Harness-RL/tree/AgenticRAG-R1-Whitebox。
Attributing an artwork to an artist has traditionally relied on detailed visual observations and descriptions, known as stylistic analysis in art history. By contrast, current artificial intelligence (AI) models used in the field offer only unexplained probabilistic classifications. To bridge this methodological gap, we present an AI framework that automates stylistic analysis of paintings, providing a foundation for enhancing evidence collection, discovery, and verification. By training a vision transformer (ViT) on a large corpus of paintings with metadata, our system encodes this art history-specific data as embeddings. These representations are factorized via sparse dictionary learning into a shared set of features that recur across the training set. A large language model (LLM) then interprets each feature by retrieving associated artworks and their accompanying curator-written texts, and synthesizes them into descriptions that reflect their stylistic attributes. Finally, an autonomous coordinator LLM applies a reasoning-and-action (ReAct) framework to weight, test, and refine these features into cohesive descriptions of an artwork, or comparisons of artworks. This approach converts detailed visual features into descriptive terms, addressing a key challenge in art history. It thus connects the use of images as data with the semantic concerns of humanists, establishing vision-based computational art history as an area for future growth.
Large language model (LLM)-based agents have shown strong potential in solving complex tasks through multi-step reasoning, yet they remain vulnerable to execution failures. Accurate failure attribution is therefore critical for improving agent reliability. Existing topology- and spectrum-based methods exploit trajectory structures but often overlook fine-grained semantics, while LLM-based attribution methods capture semantic cues but suffer from long-context degradation over lengthy trajectories. To address these challenges, we propose DUOTRACE, a plug-and-play detection filter for LLM-based failure attribution. DUOTRACE follows a detect-before-attribute paradigm: it first detects anomalous executions and then supplies focused trajectory evidence to downstream LLM-based attribution methods. For effective VAE-based anomaly detection on agent trajectories, DUOTRACE integrates dual-view semantic-structural node representations, a Tree-LSTM-based trajectory encoder, and prefix-chain- and LLM-based data augmentation to handle heterogeneous nodes, hierarchical execution structures, and limited failure data. Experiments with six LLM-based attribution baselines show that DUOTRACE improves agent-level and step-level attribution accuracy by 8.7% and 7.0%, respectively.
Hi HN, I'm Kushagra, one of three founders of Almanac, a Hermes with a brain that knows everything about your company. We started our journey with setting up Hermes for our company, thinking it must be easy. We wanted an agent that would know every context about our company, so we could ask questions and get context-appropriate responses to. This started a very annoying and difficult journey. Setting up Hermes, getting it to talk right, building OAuth apps for every connector myself, then feeding it context myself, and ultimately struggling with Hermes's default memory. At the same time, we saw our YC batchmates struggling with the same problem, and we saw an opportunity. So we built Almanac. This is how it works. You sign up, you get a Hermes agent straight out of the box. You have a one-click connect to any account (Gmail, Calendar, Granola, PostHog, etc). You have personal accounts (only accessible by you) and also shared accounts (accessible by everyone in the company). The consequence being I can never see my cofounders' accounts. The “brain” of this agent is wikis. We pull in information from your connected sources, and start organizing this information in two wikis. A personal one, for you, which understands who you are, what your preferences are, the people in your life, and the things going on in your life. The second wiki is a company wiki, which includes what the company is, what you’re working on, what the roadmap is, and what the blockers of the company are. Your agent ultimately has access to these two wikis and the original accounts, which invoke the feeling of “it just knows you.” Here’s a demo: https://www.youtube.com/watch?v=ajXP5PHuK18 We're three cofounders, Rohan, Kushagra, and Divit, and we've been friends for 11 years, since studying for the IIT-JEE. We all did Electrical Engineering (Rohan at IIT Delhi, me at IIT Kharagpur, Divit at BITS Pilani, Hyderabad), and Rohan and I later went to Harvard, where this pre-compilation layer became our capstone thesis. We have built multiple products around the idea of a pre-compiled knowledge layer. Our main differentiating point is the way we approach memory and context in general. Most AI assistant tools treat memory as an afterthought. We have worked on wikis for AI for more than a year now, building products for Harvard and NASA. The one thing we have learnt is that one needs to spend a lot more compute upfront, in the pre-compilation of this knowledge base, to get it right. Having this pre-compiled knowledge base enables a lot of interesting ideas. First is a proactive agent. Since I have compiled what’s going on in both my company and my current life, Almanac can start completing tasks on its own. Concretely, we run a background worker which takes a look at tasks that could be completed, pings the main agent, who then pings me, suggesting which tasks it could automate. As a result, I wake up to proactive notifications which look like “I already prepared a draft of your fundraising pitch deck, want to take a look?” Second, long-horizon tasks. In our wikis, we maintain a section on ongoing projects, so Almanac can pick a task back up days later without losing the thread. Most agents are session-bound: they run once, finish, and forget. But a lot of real work isn't one shot; it plays out over hours and days with people in the loop. The clearest example is anything that involves waiting on a human, like scheduling a meeting, following up on a sales thread, or chasing a document. Almanac can send an email on your behalf, and because it's always on and remembers the project, it notices the reply four hours later and drafts the right follow-up in context. Since launching, we've seen a lot of use cases for Almanac. One person runs her dog-rescue operation through it: finding available fosters, tracking pickups, and sending reminders for consent forms. Another researches Polymarket strategies with Almanac, where its memory holds what past strategies were, proposes new ones, and compares them against what went right or wrong last time. Another builds marketing campaigns on it without having to re-explain the business and the whole campaign every single time. Regarding privacy and security, Almanac only accesses accounts you explicitly connect. Your OAuth credentials are held by our connection provider, not in Almanac’s database. We only store the wiki and the source behind its citations. So an email used as a citation may be retained as Markdown. We’re live: https://usealmanac.com . We have a 7-day trial on all of our plans. Happy to hear if people have done similar setups, and what new features they’d like in Almanac. If you’re a company that wants to get an agent that actually gets tasks done, I would love to talk: https://cal.com/team/almanac/demo . Comments URL: https://news.ycombinator.com/item?id=49511007 Points: 3 # Comments: 0
Software and systems security workflows are typically procedural: analysts inspect heterogeneous artifacts, form hypotheses, invoke tools, interpret outputs, and revise plans. Large language model (LLM)-based agents, which can plan, use tools, retain state, and revise actions across multi-step workflows, are being rapidly adopted to automate this work. Given the consequences of delegating security decisions to autonomous systems, understanding how such agents are built, used, and assessed is crucial. Yet to this date, there remains a lack of systematic understanding of what has been done and how far we are in this field: the term "agent" is applied inconsistently, applications differ sharply in risk, and assessment protocols are often incomparable. To gain a comprehensive and coherent view of this area hence inform relevant future research, this paper provides a systematic literature review of the (1) technical approaches, including agent architecture, perception, memory, reasoning and planning, action space, orchestration, and self-improvement, (2) applications, with respect to the security tasks served, and (3) assessment, including the datasets, outcome and trajectory metrics, safety measures, and baselines considered, over the peer-reviewed literature spanning the emergence of this area (2023--2026). Our synthesis reveals a field that has built agents able to act but not yet agents whose authority is bounded or whose behavior is auditable. In addition to knowledge systematization, we also extend our insights into the limitations of and challenges faced by current approach, application, and assessment designs, which shed light on potentially promising future research directions.
AI 编码代理是一类通过推理和工具使用自动化开发任务的软件工具，正日益通过插件市场进行扩展，但这些新兴仓库的结构、维护以及共进化动态仍缺乏实证研究。与通过源代码提供功能的传统软件包不同，agent 插件通过自然语言指令文件、脚本和配置文件的组合来提供功能，这引出了一个问题：这些插件究竟是会在各组件之间共同进化的可维护制品，还是开发者只需编写一次、无需再回看的一次性制品。为研究 agent 插件的维护与共进化，我们对 1,926 个托管 Claude Code 插件市场的仓库进行了实证研究，分析了 2,018 个市场中的 8,351 个插件和 77,773 次提交。我们发现，该市场正在快速扩张，在 2025 年 10 月发布后的六个月内，涉及插件的提交活动增长了 8.8 倍，而面向软件工程任务的插件占所有插件的 61.3%。插件开发主要由功能驱动，功能相关提交的发生率是传统开源软件（OSS）的两倍多（39.6% 对 17.2%）。Claude 参与共同编写了 34.9% 的全部提交，而且四类提交（docs、perf、style 和 refactor）在插件仓库中的含义与在传统软件中有显著不同。大多数组件类型独立演化，但在 skills 目录内部，自然语言指令文件与实现脚本以高于随机的比例共同进化，其中 78% 的共同变更在功能上是耦合的，这代表了一类在传统软件工程中未曾观察到的新型维护依赖。
我们研究用户输入中的自动语音识别（ASR）错误是否会导致 Embodied AI（EAI）模型产生不安全输出。我们发现，ASR 错误会使有害指令被 EAI 模型接受并执行，从而降低安全性。我们模拟 ASR 错误，并将其与现有安全基准（SafeAgentBench 和 POEX）结合，以评估不同错误如何影响具身 AI 安全。我们发现，其中一些错误保留了语义结构，但增加了有害歧义；另一些则削弱了模型的拒绝行为，使不安全计划得以生成并执行。我们表明，在某些情况下，对 ASR 错误进行自动纠正可以降低风险，但这并不总是有效。总体而言，我们证明 ASR 错误会给具身 AI 带来显著的安全风险。
现代 agent 系统在运行时组装能力，而这种动态组合最近在时空可组合性演算中得到了完整的形式化处理。在该演算中，能力是一个带有可追踪逆元的组件，而 agent 作为插件被组装起来。这种插件形式由单个进程承载，共享一个上下文；而这种承载方式把所有组件置于同一个物理故障域中，一次故障会同时挂起所有组件，进程终止则会中断该进程所承载的每个会话。本文表明，模型和演算都没有把 agent 绑定到单一进程，语言模型的无状态性将所有跨步状态都保留在模型外部，而健全性不变式仅在状态空间上定义。基于语言模型推理无状态这一前提，这些观察凝练为四个引理，其前提是该演算的假设。本文在这些引理之上构建了 Logos，一个类似 ROS 的跨进程 agent harness，其中 plugin 是一个进程，唯一共享的状态是一个只追加的转录记录。对在 tool-call 周期四个边界处设置进程终止的八十个会话进行恢复测试后，没有出现重复效应；并且在与单进程参考配置的同故障比较中，单个故障会中断所有同驻会话，而在 peer-process 构造下，一个故障只会在一个节点上终止。
AGENT-O 是一个模块化本体框架，定义了用于表征面向健康的 AI 智能体系统的语义 Agent Card，并支持对科学论文中报告完整性的评估。AGENT-O 被开发为一个 OWL 2/RDF 本体，覆盖运行时、模型、工作流、工具、临床用途、评估、溯源、治理以及报告评估。评估包括本体盘点、OWL-RL 推理、三套 SHACL 规则集、12 个 SPARQL 胜任问题、三个案例，以及对 279 篇论文在五个维度上的模型辅助报告完整性评估。该本体包含 1,962 个 RDF 三元组和 1,922 条 Protégé 公理，具有 252 个活动类、198 个活动对象属性和 51 个数据类型属性。所有 SHACL 规则集在示例图上均符合要求，所有胜任问题都返回了预先指定的证据，且 279 篇论文全部完成评分。与评估（25.8%）和基准流程一致性（29.8%）相比，不完整报告在运行时/架构（84.6%）、治理/安全（82.8%）和溯源/可复现性（78.1%）方面最高。AGENT-O 支持语义 Agent Card 表达和报告评估，同时揭示了一个评估-规范缺口：评估和基准测试流程的报告比运行时架构、治理和可复现性更一致。AGENT-O 提供了一个可复用的本体、语义 Agent Card 概况以及用于结构化报告和差距识别的报告完整性工作流，但不评估智能体质量或部署就绪性。
LLM 代理越来越多地在运行时修改它们自身的提示词、工具、中间件、资源和执行框架。这种自进化可以提升能力，但一次成功的变异可能会留下持久影响，而这些影响在与其创建时不同的状态下无法安全回滚。我们提出 EvoUndo，这是一个用于表示、合成、诊断并独立验证模型生成的自修改在反事实状态下可恢复性的框架。在 600 个未见的一次性自进化任务中，我们识别出 197 个提升能力但未通过可恢复性验证的变异。在原始恢复表示下，传统修复策略对这些自然失败的恢复率为 0/197。确定性 oracle 分析在原始恢复语言 L0 下恢复了 48/197，而扩展的恢复演算将经验 oracle 恢复率提高到 191/197。随后，一个协议锁定的 2x2 锚定-表达能力干预将两个瓶颈区分开来：当原始语言足够时，精确状态地址锚定将成功恢复率从 0/48 提高到 38/48（79.2%）；而扩展恢复语言则使 oracle 定义的 S1 层中的 142/143（99.3%）个失败得以恢复。在主要的 gpt-oss-120b 主干模型上，向更丰富的语言加入精确地址诊断后，恢复率降至 133/143（93.0%）；Qwen3.8-27B 复现实验保留了锚定和表达能力效应，但未出现这一负交互，表明后者因模型而异。这些结果表明，要实现可靠的 agent 自进化，需要协同设计验证、状态锚定、见证语义和恢复语言的表达能力，而不能仅依赖迭代式 prompting。
In financial markets, a sequential policy that reacts systematically to price movements may become predictable to other market participants. This paper studies whether large language model (LLM) agents exhibit such directional structure through RetailAgent, an experimental framework in which an LLM observes anonymized intraday equity price histories and permitted state, then repeatedly chooses long (hold the stock) or flat (stay out) before the subsequent interval return is revealed. We compare returns during long and flat intervals along the same stock's intraday path after removing the overall fraction of long decisions. This exposure-matched measure reveals persistent negative timing across modality, horizon, state, and model family. Shuffling saved action sequences substantially attenuates the effect, showing that alignment between actions and subsequent returns drives the negative score. Feeding self-authored memories into decisions further increases policy persistence, while timing becomes more negative among stock-days on which the agent uses both actions. These results reveal stable, recoverable directional structure in sequential LLM financial decisions and a behavioral signal for studying how another participant could respond to a predictable policy.
有一个模型在从未打开数据表的情况下就通过了我们的保真度检查。我们是在为一个内部提取服务筛选模型时发现这一问题的：一个结构化输出约束悄然禁用了工具使用，而模型却仍然作答，且编造了源文本。只有逐工具轨迹暴露了这一点。保真度——即提取值是否与来源一致——是代理式文档提取的标准指标，而那次运行在这一指标上被判定为成功。因此，我们在一个代理式基准中记录每一次工具调用，该基准包含三个组件上人工筛选的 25 条断言，第四个组件上另有 12 条，共 37 条。基于这份调度记录，我们构建了两种工具：一个基于规则的失败归因分类器，以及一个静默失败检测器；后者的两条规则只检查调用了哪些工具，从不检查提取值本身。该检测器在跨三个模型家族的 207 次干净且保真度通过的提取上从未报警，并且检出了全部 50 个人为植入的故障——这些故障恰好隐藏了其规则所检查的工具。这两个结果并不对称：前者界定了误报率上界，后者则按构造保证召回；而对于那些确实调用了工具却仍答错的运行，其检测能力尚未测量。第二个、独立的判据——一个用于检验数据表断言是否在物理测量下成立的 causal chamber——是有意不完全的：它只确认装置能够检验的内容，即那 37 条中的 2 条所构成的可验证范围；我们还给出了其余断言为何无法进行物理评测的分类。在受控扰动下，保真度始终通过，而该腔体的判定恰好在测量不确定度处翻转。在三个已部署的模型栈中（其中一个的不稳定源于其服务栈，而非任何能力缺陷），工具层带来的是可移植性和可观测性，而不是准确率；只有当文档超出上下文窗口时，它的额外成本才真正值得。
文章链接： https://www.dwarkesh.com/p/openai-huggingface 评论链接： https://news.ycombinator.com/item?id=49494301 得分：16 评论数：1
文章 URL: https://arxiv.org/abs/2608.23691 评论 URL: https://news.ycombinator.com/item?id=49481455 点数: 32 评论数: 0
文章 URL: https://infernalcode.com/posts/your-ai-agent-has-root/ 评论 URL: https://news.ycombinator.com/item?id=49477311 得分: 15 # 评论: 14
文章网址： https://m-reschreiter.at/en/blog/how-i-cut-80-percent-context-overhead-dynamic-tools 评论网址： https://news.ycombinator.com/item?id=49476278 得分：7 # 评论：0
LLM agents increasingly rely on generated interaction data to learn how to interact with external environments. Agentic data generation must maintain consistency among environments, tasks, interactions, and success signals while producing experience that is useful rather than merely abundant. Existing work spans many agent domains, but domain-centered organization and heterogeneous evaluation often obscure common generation mechanisms and conflate candidate construction with verification and selection. This work develops a two-level framework for the field. First, we represent agentic data as a common factorized object $(E,q,τ,v)$, comprising an environment specification, task signal, interaction realization, and optional verifier. We organize generation paradigms by their primary anchor and dependency structure. Second, we formulate generation as constrained distribution design through the Accuracy-Complexity-divErsity (ACE) lens. Accuracy establishes the feasible support of grounded and internally consistent data. Within this support, Complexity places learning mass relative to the capability of a declared learner and execution configuration, while divErsity controls coverage and redundancy of data. Using this framework, we explore how prior work verifies generated experience, constructs and calibrates difficulty, and expands behavioral coverage. The literature reveals a shift toward execution-grounded accuracy, learner-relative complexity, and diversity beyond surface variation or dataset size. We further discuss broader directions and emerging trends in agentic data generation through the ACE lens, including their implications for scaling, data sources, training regimes and adaptive learning. Overall, the central challenge is not simply to generate more data, but to continually allocate valid, informative, and non-redundant experience as agents and environments evolve.
Efficiently improving autonomous agents across diverse tasks is central to accelerating recursive self-improvement (RSI) in agentic AI, with prompt optimization emerging as a promising approach capable of delivering performance gains comparable to those achieved by fine-tuning model weights, while reducing computational costs in both optimization and serving. However, recent developments increasingly favor unnecessarily complex prompt optimizers. We introduce Naive Prompt Optimization (NPO), a lightweight single-lineage method that iteratively revises prompts using a teacher model with rollout feedback. NPO achieves comparable or better performance than GEPA with fewer rollouts, and its advantage increases with stronger teacher models, suggesting that stronger teacher reasoning can partially substitute for optimizer-side search complexity. In interactive games, NPO remains broadly competitive with GEPA, while GRPO performs better on some tasks less amenable to prompt optimization. We also show that NPO-optimized prompts elicit similar performance improvements when applied verbatim to other student models, especially across models within the same family. Overall, our preliminary results show that simple, linear prompt optimization can rival substantially more sophisticated and complex search procedures.
Agent harnesses shape how language-model agents use instructions, tools, and runtime components, but adapting these harnesses requires costly verification. Existing propose-and-verify methods typically score every candidate on a fixed task set, wasting rollouts on unrelated behaviors and allowing aggregate scores to obscure specific regressions. We introduce HarnessLens, a budget-aware framework for automated harness evolution. HarnessLens jointly explores the task space and user-configurable components, derives candidate modifications from execution trajectories, and selectively verifies each candidate on behavior-relevant tasks using an attributable-evidence gate. Across three agent harnesses and four benchmarks, HarnessLens improves average held-out performance by 7.6-13.6% while consuming substantially less evaluation budget than competing baselines. These results demonstrate that behavior-aware verification with explicit attribution enables more reliable and sample-efficient harness evolution under constrained interaction budgets. Our code is available at https://github.com/jhxu5214/HarnessLens.
Large language model (LLM) agents in governed organizations must let the persona (instructions, tone, self-presentation) evolve freely, while keeping execution (stateful, audited work) traceable. A single trust domain does not satisfy both cheaply. We present Persona-Execution Separation (PES): persona and execution reside in different trust domains, connected by a governed contract bridge. The persona is singly-homed and may drift; execution is faceless and audited. Status summaries may return; data bodies remain in the restrictive domain except a graded data-loss-prevention (DLP) exception; identity stays continuous. An approval matrix, DLP, and audit enforce the crossing. PES follows from three goals---free drift, execution traceability, and decoupling. Under LLM representational indistinguishability, any single-domain mechanism that meets all three must re-introduce typed change objects, an external gate, and a stable audit anchor: PES rebuilt at higher coupling cost. A development/pilot case in a regulated digital-employee platform records five decisions over one month, each with a rejected alternative. A mechanism check on the shipped implementation found no execution-side re-validation under persona perturbation (five model configurations) and no persona fingerprint on hard-asserted fields. A probe of a recovered pre-separation build found the governed execution path decoupled from the persona by omission, not by construction; a later wiring change could reverse that isolation, which PES makes an audited architectural rule. The pattern applies when multi-user deployment, execution audit, and expected persona churn hold jointly.
基于 LLM 的 Agent 正越来越多地部署在产品级执行框架中，在这些框架中，越狱可能触发有害的工具使用和持久状态变更，从而带来比单纯生成不安全文本更大的风险。现有的自动化红队测试方法通常依赖固定攻击，而近期的 Agent 式攻击者会协调多种越狱工具，并通过基于轨迹的检索展现出更强潜力。然而，由于检索偏差和工具贡献不清晰，这类检索可能会复用误导性经验；同时，完整轨迹会增加上下文开销并降低可解释性。我们提出 RedEvoAgent，这是一种黑盒红队测试 Agent，能够将跨案例攻击轨迹提炼为简洁且人类可读的攻击技能。该攻击技能通过工具有效性画像和用于技能更新的 Deciding-Tool Attribution 自适应进化，并通过验证棘轮机制仅保留能提升验证性能的更新。在多个基准、目标模型和目标执行框架上的实验表明，RedEvoAgent 优于固定式和 Agent 式基线方法，提升了工具效率，并能够在攻击者模型和目标执行框架之间迁移。
Agent skills package specialized knowledge and workflows into reusable resources that extend AI agent capabilities. Recent work automatically discovers such skills from agent experience, which enables agents to progressively adapt through interaction. However, the insights that guide skill development typically remain scattered across optimization histories, limiting their systematic reuse across iterations. We introduce WikiSkill, a framework that co-evolves agent skills with a persistent knowledge base (wiki). At a high level, WikiSkill separates raw execution experience, accumulated knowledge, and executable skills, while continuously consolidating experience into the wiki, which subsequent skill updates can build on. Across diverse benchmarks and models, WikiSkill consistently outperforms state-of-the-art skill-evolution methods and improves over no-skill baselines in most model-benchmark settings. We find that skill evolution complements model scaling: larger models generally benefit more from evolved skills, while smaller models with skills can outperform substantially larger models without them. We also find that evolved skills transfer effectively across models and model families, and skills evolved by other models can outperform self-evolved skills. Finally, our ablation studies confirm that persistent knowledge accumulation in the wiki is critical for effective skill evolution. These results demonstrate the benefits of systematically accumulating and refining agent experience for developing reusable and transferable skills.
Organizing long-term memory for multimodal agents remains challenging because existing methods either suffer from expensive question-agnostic offline summaries or naive embedding similarity matching that introduces incomplete and redundant context. To address these issues, we propose GraphMemix, a combinatorial-optimization graph memory framework that models memory organization as query-aware evidence-forest construction. Specifically, our method consists of three key components:(1) candidate graph construction, which expands multi-view seed memories through schema and semantic relations to acquire query-aware original context; (2) evidence utility and activation costs, which decouples direct memory support from anchor-conditioned relation verification to suppress redundant or conflicting information; and (3) forest optimization, which jointly selects a forest-format memory context under a maximum evidence budget and its reliable relational structure. By organizing memory into a query-relevant subgraph, the method avoids substantial lifecycle cost and recovers low-similarity complementary evidence. Experimental results across four long-term multimodal memory benchmarks demonstrate significant improvements with different foundation models and establish a new Pareto frontier between accuracy and lifecycle cost.
大型语言模型可以总结金融信息，但一个可运行的股票研究系统必须首先汇集异构证据，暴露不可用的数据和模型能力，并控制生成的观点如何影响最终报告。我们提出 DSA，这是一种面向多市场股票研究、结合大型语言模型（LLM）agent 的证据感知型编排框架。DSA 将工作流组织为证据获取、结构化上下文构建、模型路由分析、可选的角色和 Strategy Skill 推理，以及基于所选上下文和诊断信息的报告生成。默认报告配置和可选的 agentic 配置共享证据和模型路由服务，但使用各自配置特定的输出验证和风险防护。在 agentic 配置中，核心角色输出由角色特定的解析器处理，而 Strategy Skill 观点在综合之前会经过额外的信号资格划分；分歧会显式提供给决策 agent，随后进行保守的风险覆盖。参考实现包括六条区域市场路径、十五个打包的 Strategy Skills、托管和本地模型路由，以及多种执行和交付界面。在冻结的软件快照中，所选清单中的 1,457 个可移植离线后端契约测试全部通过；其中 596 个案例被回溯性映射到六个与所报告 LLM-agent 架构核心相关的契约家族。这些证据仅证明所测试软件契约的实现符合性，而不证明报告质量更优、预测准确性更高或投资回报更好。
强大的代码智能体可以执行脚本、调用工具并管理文件，但许多重要应用仍主要通过图形用户界面访问。我们认为，截图点击对于操作软件的智能体来说是一种低效接口：截图无法完整反映状态，而 GUI 操作脆弱、语义弱，并且与长程规划并不匹配。我们提出 ASIL（Agent-Software Interaction Layer），一种智能体原生接口，通过结构化 JSON 观察和可由代码执行的语义动作来暴露软件，并针对每个应用采用所能实现的最深访问路径。我们在 15 个应用上实现了 ASIL，并在一个包含 300 个单应用任务和 80 个多应用任务的基准上进行了评测。ASIL 在闭源模型上得分超过 80，且每个任务执行的动作少于 5 次。在修复后的运行时和 50 步截图预算下，同样的任务在截图点击控制下仅获得 6.6 和 26.6 的严格成功率，在更容易、可与 OSWorld 相比的区间上则提升到 15.0 和 53.3。与匹配任务上的应用原生接口相比，ASIL 比 LibreOffice 的 UNO API 高出 28-38 个严格分，但与 draw.io 的 MCP 内容契约仅持平。这种结构化模态也适合训练：小规模 SFT 将 Qwen3.5-2B 从 58.0 提升到 72.1，将 Qwen3.5-9B 从 66.6 提升到 80.4，而资源受限的 on-policy RL 进一步将它们提升到 74.4 和 82.2。
基于 LLM 的智能体正越来越多地被用于网络故障诊断，但现有基准只在准确工单上评测它们，并且总是假设存在故障，而这两种条件在实践中很少成立。我们提出 FaulT-Bench，一个涵盖八种网络拓扑、共 200 个故障排查场景的基准，其中五种取自公开实践者实验室并重新实现，覆盖真实故障、错误故障报告、错误设备归因以及错误根因主张。为隔离工单措辞对诊断的影响，我们进一步将 72 条带有错误前提的工单改写为五种报告者 persona，在保持网络状态固定的情况下，每次只改变一个因素，分别体现报告者置信度和可验证细节的差异。我们的自动化框架在 Kathará 中部署每个场景，让智能体通过 NIKA 工具接口进行交互，并使用 LLM judge 按结果、修复方案和推理质量对自由文本诊断进行评分。我们评测了 SADE、ReAct 和 Claude Code，发现三者在准确工单上都接近饱和，并且对误导具有鲁棒性，但当网络本身健康而工单出错时，性能会急剧下降：它们会不断探查，试图把良性状态提升为根因，而不是得出没有问题的结论。persona 改写显示，工单的写法比其声称的内容更重要：一份自信但错误的报告，其处理效果大致与准确报告相当，而一份模糊、细节不足的报告会使性能显著下降。这三个智能体的失败方式也各不相同，从持续过度诊断到运行无回应，而且成本差异很大。这些结果将 FaulT-Bench 定位为一个基准，用于开发能够可靠处理真实世界网络故障排查中嘈杂、不可靠工单的智能体系统。
A core question in decision-making for agents is whether the relative priorities of competing lower-level objectives can be determined by emotional preferences autonomously generated by higher-level goals, rather than being externally prespecified. Under changing external environments and evolving internal states, emotions play an important functional role in regulating the relative priorities of competing goals. Inspired by the goal-directed theory of emotion, this paper studies how such preference regulation can be computationally realized through reinforcement learning. We first propose a conception of emergent emotional preference: a high-level goal autonomously induces state-dependent preferences over competing lower-level objectives. This conception is built upon a framework consisting of a multi-objective reinforcement learning inner controller and an outer preference generator. The inner controller provides a repertoire of preference-conditioned goal-directed behaviors, while the outer preference generator learns a mapping from the current state to objective preferences through reinforcement learning on a high-level goal. We operationalize emotional preference as a state-dependent regulation of relative goal priorities that emerges through optimization. Furthermore, we characterize the policy space induced by preference regulation and derive an upper bound on the optimality gap in terms of the representation error of the inner behavioral repertoire. We show that the gap vanishes when the optimal policy can be represented by the available preference-conditioned policies. Experiments in self-constructed multi-objective exploration environments show that the learned preference function exhibits contextual priority switching, graded trade-offs, and temporal persistence, and outperforms the evaluated fixed-preference and handcrafted-preference strategies.
企业 AI 部署是一个跨业务部门、应用与 AI 团队、测试、平台工程、基础设施、安全、运维以及数据治理的协调问题。用例基准只能展示某个智能体是否完成一项任务，却无法说明当能力、模型、运行时机制、容量和企业数据发生变化时，如何共同归属、变更、准入或举证。 我们将四个责任对象作为共享的组织契约提出：Skill（可复用、可版本化的能力与工作流资产）、Harness（运行时编译器与治理器）、Scaffold（执行/控制边界与 NFR 负责方），以及一个位于栈外、其语义与遥测由独立 CIO 治理的数据基底。运行时核心为 A = ，数据基底位于该栈之外。 核心贡献是一条有边界、可证伪的假设 P1（面向成本的能力-容量可分离性）：在声明的运行区间内，改变已激活能力会在预先登记的等效边界内保持容量-响应交互不变；而改变兼容的 Scaffold 容量会在非劣边界内保持能力语义，并且所需控制保持在声明的执行预算内。六项设计条件转化为可测量义务，其覆盖度、违例、不确定性、成本和排除项共同决定 P1 是否可判定。我们提出一种簇-周期随机交叉实验（平衡顺序、重置/洗脱、重复随机种子和故障情景、考虑簇结构的不确定性），并给出四态判定：支持、证伪、条件性工程化或结论不明。本文贡献了一种受契约约束的运行时架构、一个保持源信息的数据基底，以及一套可证伪的测量协议。本文未报告任何已完成的实现、实验、数据集或测量结果。
Internet memes are a pervasive form of multimodal online communication; however, such communication often involves users from diverse linguistic and cultural backgrounds. Therefore, adapting memes across cultures and languages is a central challenge for enabling mutual understanding in online communication. Unlike ordinary translation or standalone text rewriting, cross-cultural meme transcreation must jointly preserve communicative intent, adapt culture-dependent meaning for the target audience, and maintain coherence between text and image. In this work, we first provide an explicit task analysis of cross-cultural meme transcreation and identify three core challenges: culture-specific knowledge understanding, intent and tone preservation, and multimodal consistency. Based on this analysis, we propose a multi-agent framework with specialized agents that are coordinated to address these challenges through cultural adaptation, target text rewriting, revision, and conditional visual adjustment. The framework strengthens target text adaptation with coordinated feedback to handle difficult cases that require deeper cultural or visual intervention. We evaluate the framework on bidirectional Chinese-English meme transcreation using both human evaluation and LLM-as-a-Judge. Our method consistently outperforms all baselines across both evaluation settings. In human evaluation, it achieves the best performance on all four dimensions and delivers a 33.1% average improvement over the strongest baseline, while in LLM-as-a-Judge, it attains the highest Top-1 ranking rate (60% versus 26% for the second-best baseline). Further analysis indicates that each component contributes to the performance. Our error analysis suggests that the remaining bottlenecks lie in humor reconstruction and image-text alignment rather than simple cultural knowledge gaps, pointing to future work on humor transfer.
大型语言模型代理正越来越多地作为自主循环部署。从一个人类目标出发，这类系统会反复发现工作、制定计划、执行工具调用、验证结果，并在许多无人值守的迭代中持续保存状态。然而，当前广泛使用的代理防护措施都是针对单条轨迹定义的，其安全状态会在下一条轨迹开始时重新初始化。我们表明，这反映的是组合性失败，而不是实现细节。我们的核心结果是一个分离结论：面对一种证据分散在多次迭代中的攻击，任何轨迹范围监控器的真正率都等于其假阳性率，无论其表达能力多强，因为它所需要的证据从未出现在它所见的窗口中；而保留跨迭代状态的监控器则能将二者完美区分。我们进一步表明，显而易见的修补方案——携带一个几何衰减的风险评分——仍然不够，因为一个有耐心的对手必须等待的冷却期是一个不随视界 $N$ 增长的常数。随后我们提出 LoopHarness，它在循环层面恢复了持久的、非衰减的安全状态。在经由中介的提交和仲裁器检测下限 $δ_M$ 条件下，它将未经授权且不可逆操作的期望数量上界为 $B+m-1+m/δ_M$，这是一个关于 $N$ 的常数，其中 $B+m-1$ 这一项由无模型规则决定，因此即使在验证器完全串通的情况下也能保留。我们给出了一套完整的评测协议，涵盖原生 Agent-SafetyBench 任务上的成对干净与受攻击回合、一个其决定性证据只存在于跨迭代中的外部状态攻击集合、逐模块消融，以及自适应白盒红队测试。
尽管 Large Language Models（LLMs）在标准化图任务中具有潜力，但在节点标识符和任务表述发生真实世界变化时仍然十分脆弱。确定性图工具对这类变化具有不变性，而对于 LLMs 来说，从噪声文本中提取拓扑结构却极其脆弱，它们往往会对表面模式过拟合。此外，通过多智能体系统缓解这些解析失败会带来高得难以接受的延迟。为此，我们提出 GRAIN，一个通过强化学习优化的单智能体框架。GRAIN 将推理建模为一个语义解析与工具执行流水线，并由 Structure Invariance Reward 进行引导。通过将提取的中间图与真实拓扑进行验证，该奖励迫使 LLM 学习稳健的文本到结构映射，而不是记忆语言表象。我们还引入了 GRIT，一个用于评估对这类语言变化敏感性的基准。GRAIN 在准确率上比多智能体基线高出 16.45\%，同时延迟约降低 24\%。此外，它展现出更强的结构泛化能力，将 SFT 模型的分布外（OOD）差距减半（从 15.77\% 降至 7.80\%），并在超出训练分布的大规模图上保持稳健性。
工具增强型 LLM 智能体必须依赖不可信的运行时观察来完成开放式任务；然而，当工具输出不再只是提供数据，而是开始指定具体动作时，它们实际上就变成了能够驱动超出用户意图的现实世界副作用的“命令”。我们认为，这一风险源于将动作诱导与执行授权混为一谈。为解决这一区别，我们提出 SARA，它将动作诱导和执行授权视为不同的运行时角色，并将动作溯源与执行权限分离。在 Observation 侧，处于上下文隔离中的 Action Probe 会暴露动作诱导语义，并在跨步骤过程中持续记录动作来源溯源，作为审查信号；在执行侧，实际的工具调用仅依据用户目标以及来自已授权成功执行的审计证据来授权，同时满足目标、执行链和参数级别的支撑。为在多步执行中保持这种分离，SARA 采用 No-History-Promotion，防止历史复现将动作来源洗白为执行权限。在 AgentDojo 和 AgentDyn 上，SARA 在四种主要评估设置下将 ASR 限制在不超过 0.63%，同时保持具有竞争力的任务效用，并在其他 Agent 骨干上持续降低 ASR。
一个被展示专业外观市场面板的 LLM agent，对一个可被证明不可预测的问题给出方向性判断的频率，远高于只问该问题本身时：在 12 个前沿模型中，随着证据不断增强，做出判断的比例从 6.5% 升至 54.0%。即使面板上的每一个数字都是编造的，它也同样容易做出判断：把整个展示都伪造出来，使模型看到的除了问题本身之外没有任何真实内容，仍会把做出判断的比例从 24.5% 提升到 36.8%，这一结果在统计上与真实市场数据产生的 37.6% 没有显著差异。真正触发自信行动的不是信息，而是其包装所带来的权威感。这个失败很窄，而且可以定位。能力不足并不是原因：在与这些面板配套的、可回答的匹配问题上，同样的模型几乎总能作答，准确率接近完美。信念也不是原因——在会让行动幅度变化 48 个百分点的梯度上，模型给出的概率几乎不变，而且得分还低于气候学基线。缺乏判断力也不是原因：在行动前先让模型判断一个问题是否可知时，模型有 90% 的时候把它判为本质不可知，而随后只在其中 0.4% 的情况下做出判断。失败的是“行动/不行动”这道门槛，而且这种效应集中在少数模型上，而非普遍存在。由于这道门槛是可分离的，它可以被训练。对一个 3B 模型在 540 个合成案例上进行监督微调——这些案例主要是骰子、硬币、罐子和计时器——可将原始案例上的判断率降至 0.0%，并迁移到三个未见领域。它并非无往不利：只有当回答格式留有推理空间时，这道门槛才有效；而剥夺这种空间的僵硬格式，会让模型在原本能答对的问题上依然自信却答错。这道门槛既可训练，又对上下文脆弱，部署时这句话的两半都必须兼顾。
理解口语对话需要对词汇内容和副语言声学信号（如情绪和对话意图）进行联合推理。然而，现有评测往往允许模型通过转写文本或单一模态方案走捷径，从而掩盖模型是否真正基于语音进行落地推断。我们将这种失效模式形式化为跨模态分歧：转写文本指向看似合理但错误的表层解读，而韵律或说话风格等声学线索支持不同答案。我们提出一个可扩展框架，用于识别受文本偏置影响的表层解读，并将分歧区域转换为冲突型QA样例。我们还纳入了一致性样例，其中基于转写文本和基于语音的解读相互一致，使评测不再局限于对抗性音频依赖。由此得到 ContraTalk，一个受控基准，包含跨五个话语维度的501个问题：互动行为、情绪状态、对话行为、社会立场和对话意图。我们进一步提出一种 agentic 风格的推理框架，将语音转换为 Audio Twin，一种可被文本读取的局部声学线索表示，从而向推理模型暴露声学证据。实验表明，强大的纯文本 LLM 在一致性样例上准确率超过90%，但在冲突样例上下降到33-48%。直接使用 AudioLLMs 只能提供部分落地能力，在约30-40%的冲突样例中仍会选择由转写文本诱导的陷阱答案。我们的 Audio Twin 框架提高了冲突样例上的准确率并降低了陷阱选择率，但其在一致性样例上的表现仍依赖于底层模型。这些结果表明，基于转写文本的捷径是口语对话理解中的一个重要失效模式，并显示显式汇聚声学证据为诊断和改进语音落地推理提供了一个更可控的接口。
Electronic design automation (EDA) has advanced engineering productivity through successive generations of tooling that progressively automate synthesis, optimisation, and verification. Large language models (LLMs) extend this trajectory by enabling direct translation from design intent to hardware implementations. In most of the EDA literature, LLM-based solutions are typically assisting siloed design stages or tasks, however this obscured the drivers by which capability emerges and systems scale. In this Perspective, we instead define three hierarchical roles that reveal how capability accumulates: a Generator that produces design artifacts in a single pass, an Agent that refines outputs through iterative tool feedback, and an Orchestrator that coordinates decisions across EDA-stages. Across published systems, this reveals a syntax trap in which models are trained to produce plausible code rather than physically correct hardware, compounded by fragmented tools and loss of design context that obscure how decisions affect later stages. Comparisons across the three roles show that current approaches struggle to scale to industrial designs, motivating a shift towards a standardised, physics-aware orchestrator that connects tools and agents across the EDA flow for more reliable and accessible hardware design.
Hey HN, I’m Abhishek. I'm building Opslane, an open-source agent that identifies user-facing issues and investigates them. It only creates a PR if it can verify the fix. Demo: https://youtu.be/ccuOTYQMeYg Docs: https://docs.opslane.com At my last job at Robinhood, we used to do a quarterly bug bash. We would go through our Sentry backlog and try to fix as many of them as possible. We only fixed bugs we knew were reported by customers. We had hundreds of bugs, and Sentry’s default priority levels made no sense. After the bug bash, we would declare bankruptcy - select all remaining bugs and mark them as resolved. This problem has only gotten worse since coding agents have become more prevalent. So I started thinking: what would Sentry look like if it were built in 2026? To me, error trackers have two failure modes: 1. False positives: They show you thousands of errors, and you can’t tell the impact on the user 2. False negatives: Many user-facing issues don’t throw exceptions, so they go unnoticed. Opslane combines error tracking and session recording. And there is an agent that acts on both. To get started, you install the Opslane SDK. It captures everything the user did: errors, console logs, network requests, and session recordings. Opslane reduces false positives by ranking issues based on how many users are facing a particular issue. It also learns about your product by reading your code and watching your session recordings. False negatives are harder. Opslane reviews session recordings to spot frustration. They look for rage clicks, dead clicks, and abandoned forms. This recently caught a bug in an early customer’s onboarding flow: a dropdown that closed itself when clicked. No exception, no bug report. The recordings showed users clicking it, selecting nothing, and dropping out of onboarding. Opslane flagged it and the team fixed it. Three guiding principles when building Opslane: 1. Open Source: Self-host with one Docker Compose file. 2.Agent-first: I never want to open an error dashboard again. Opslane ships an MCP server, so you can ask "what broke for users this week" from Claude Code. You get back issues that need your attention and you drive the resolution. 3. It knows about your product: Opslane is continuously learning about your product. Every investigation begins with what it knows about your product. It’s early. Frontend apps work end to end today.I am currently focused on improving reliability and accuracy. Here is a link to our repo: https://github.com/opslane/opslane Would love to get feedback from folks on our approach to this problem! Comments URL: https://news.ycombinator.com/item?id=49466704 Points: 3 # Comments: 0
嗨，HN，我们构建了一个开源模型网关。它是一个集中管理我们自托管、前沿和开源模型的单一入口。 它原生使用 Rust 构建，面向并发，并实现了跨模型和提供方的各种配置细节差异（流式格式、工具调用、模型参数、速率限制以及不同的错误行为）。 对于 BYOK 请求，这个网关增加的延迟不到 1 ms；当由 Experiential 提供供应商密钥时，延迟不到 2 ms。它覆盖所有主要推理提供方，并通过一个会发起 PR 的 codex agent 每日刷新 1000+ 个模型。 与其他类似项目相比，我们是开源的，不收取任何加价，允许你将本地模型与市场混合使用，并利用你的流量来（选择加入后）训练一个模型。简单路由并不值得收取 10% 的 token 加价。 我们的做法是：基于标准化的 OTel traces，我们挖掘具有代表性的真实任务，使用 text world models 模拟各种模型的 rollouts，应用一个 LLM judge，然后在 prompt 的 embedding 之上拟合一个最近邻分类器，以决定每个请求的最优模型。通常这比直接调用单个模型能画出更好的成本/质量帕累托曲线，但并不完美。 借助这些模拟，我们还可以做诸如建议缓存命中优化、新模型建议以及训练模型之类的事情。 它是开源的，所以你可以把它部署到自己的基础设施上，使用我们托管的版本且不收取任何加价，或者在我们的网站上阅读我们如何为最大可用性进行设计。 Comments URL: https://news.ycombinator.com/item?id=49471407 Points: 9 # Comments: 0
文章链接： https://blog.exe.dev/engineering-with-ai 评论链接： https://news.ycombinator.com/item?id=49465119 得分：6 评论数：0
Robust Markov Decision Processes (RMDPs) generalize classical MDPs by allowing uncertainty in transition probabilities and optimizing against their worst-case realization. We consider $(s,a)$-rectangular RMDPs with \emph{linearly defined} uncertainty sets and study parity objectives, which are a canonical representation of $ω$-regular objectives. An uncertainty set is linearly defined if it is described by linear inequalities over the transition distribution together with auxiliary variables, which capture the standard $L_1$ and $L_\infty$ balls as well as general polytopic uncertainty sets. The quantitative value is the supremum, over all agent policies, of the satisfaction probability guaranteed against the adversarial environment. Previous work studied the qualitative analysis, namely the almost-sure (resp. positive) problem that asks whether a single agent policy guarantees satisfaction with probability one (resp. positive probability) against every environment policy. In this work, we solve the exact quantitative problem. Our contributions are threefold. First, we show that both the agent and the environment admit pure memoryless optimal policies. Second, we give a polynomial-time algorithm for quantitative parity on linearly defined robust Markov chains and use it as a subroutine in a policy-iteration algorithm for RMDPs. The algorithm combines quantitative one-step improvements with qualitative almost-sure improvements. Finally, we report experiments comparing our approach with the explicit reduction to stochastic games.
多智能体大语言模型（LLM）工作流已成为通过专门化 LLM 智能体之间的协同推理来解决复杂、开放式任务的强大范式，但由于需要反复调用 LLM 以及长时程上下文累积，其运行成本相当高。现有的级联路由方法只能做出一次性、查询级别的决策，无法适应多步工作流中动态变化、依赖状态的特性；在这类工作流中，每一步选择合适的 LLM 取决于不断演化的任务进度、剩余任务难度以及成本效率要求。我们提出 ProgRouter，这是一种在线的进度引导路由框架，可在工作流各步骤中自适应地选择 LLM 智能体，以在遵守时间和成本预算的同时保持任务求解质量。ProgRouter 引入了一个多视角任务进度评分器，结合了粗粒度的工作流结果状态、以及关于子任务完成情况、进度趋势和工作流状态质量的细粒度信号。随后，一个双路径任务进度预测器和一个自适应元门控机制会估计每个候选路由 LLM 所带来的进度增益。ProgRouter 依据这些信息进行在线、逐步的路由决策，在进度增益、任务时间预算和长期运行成本效率之间取得平衡。我们在 HumanEval Plus、MBPP、MATH-500 和 ASQA 上进行了实验，涵盖智能体式代码生成、数学推理以及检索增强的长篇问答，结果表明 ProgRouter 在保持强任务求解性能的同时，相比关键基线降低了运行成本。
随着检索、工具使用和多轮交互过程中上下文不断累积，Agentic LLM 流水线面临日益攀升的推理成本。为控制延迟，部署系统通常会压缩输入，但这会降低任务准确率。Speculative Decoding (SD) 能够无损加速生成，但它假设 drafter 和 verifier 共享完全相同的上下文，因此无法解决准确率与开销之间的权衡问题。我们提出 AsymSpec，一种打破这种对称性的非对称 speculative decoding 框架：一个轻量级 drafter 读取完整输入，而大型 verifier 在压缩后的视图上运行。drafter 通过对 logits 进行对比式 $δ$-fusion 来引导 verifier，并由一个感知分歧的接受门控进行调节，以保持验证稳定性和较高的草稿接受率。在四种 agentic 能力和两个端到端 agent 基准上进行评估时，AsymSpec 平均可达到完整上下文准确率的约 $\approx 90\%$，并在独立文本能力上以 $0.2$--$0.3\times$ 的计算成本带来 $1.3$--$1.7\times$ 的吞吐加速。这些结果表明，当压缩丢失关键推理信号时，非对称上下文访问能够带来显著收益。
仅凭答案准确率不足以作为 LLM Data Agents 的可靠性信号。在结构化数据任务中，即使跟踪过程无效，也可能生成一个在基准上正确的答案。本文提出 Trace Integrity，作为一种部署可靠性标准，用于评估答案背后记录的计算过程是否是显式的、可执行的、符合 schema 的、遵循算子语义的、可重放的、与答案一致的并且可审计的。我们将 Structure Gap 定义为使 Trace Integrity 变得必要的部署失败模式：自然语言推理和自由形式的理由说明，无法可靠地指定真实系统所需的算子级程序。我们通过 execution contracts 来实现 Trace Integrity，这些结构化工件将用户意图绑定到 schema 元素、算子计划、假设、可执行查询、验证状态以及最终答案关联。我们还提出 CAIT（Correct Answer / Invalid Trace）Rate，用于衡量仅基于答案的评估有多频繁地把缺乏计算支持的输出算作成功。在对 BIRD Mini-Dev、Direct SQL、Operation Summary + SQL 和 Contract-First SQL 的实证展示中，Direct SQL、Operation Summary + SQL 和 Contract-First SQL 的答案准确率分别为 20%、22% 和 24%，而它们的 Trace Integrity 通过率分别为 39%、43% 和 40%，CAIT Rate 仍然很高，分别为 55%、59.1% 和 45.8%，表明答案准确率、跟踪有效性与静默失败风险是彼此不同的评估信号。因此，真实世界的 LLM data agents 不应只根据输出是否匹配参考答案来评估，还应根据这些输出是否由可审计的计算过程所支撑来评估。
Collective intelligence can emerge when individuals coordinate through a shared environment, allowing local actions to accumulate into durable social organization. Language-model agents offer a new substrate for this process, yet most multi-agent systems rely on direct conversation, predefined roles, or centralized workflows. It remains unclear whether decentralized agents can build functional technologies and outperform independent search. Here, initially homogeneous LLM agents in SwarmWorld self-organize without assigned roles or recipes into evolving technological societies. Agents explore a spatial environment, process resources, test materials, construct persistent artifacts, and write executable controllers evaluated by a deterministic simulator under unseen disturbances after the agents are removed. SwarmWorld splits cognition from consequence: agents propose architectures and controllers within fixed action and material schemas, while the simulated world determines function. Shared societies develop broader, more resilient technological portfolios than a strong best-of-N isolated-search baseline, although isolated search remains competitive for the strongest artifact. Agents differentiate into exploration, construction, maintenance, and coordination behaviors, transitioning as the world matures. Technologies accumulate through collaborative construction, executable inheritance, and persistent agent-artifact networks, with most reuse beginning through physical observation rather than communication. Explicit cultural mechanisms amplify collaboration and organization, but functional benefits depend on outcome and timescale. Physical stigmergy alone supports capable societies, while interaction drives persistent technological ecologies rather than universally superior individual inventions.
大语言模型能够为孤立问题写出正确代码，但在自主机器学习开发中仍然远不如人类强大；在这一场景里，智能体必须在数小时的反馈中不断修改数据管道、模型和验证流程，而且在大多数竞赛中，最终成绩仍低于优秀的人类竞争者。基于结果的基准能记录这种差距，却无法解释其成因，因为它们只评估最终提交，而忽略其背后的开发过程。我们提出 TraceML，它在同一版本级模式下配对人类与智能体在相同竞赛中的工作：涵盖 134 场竞赛的 4,465 条人类 Kaggle 轨迹，其中 7 场竞赛也由两个智能体框架完成，共得到 430 条配对的人类轨迹和 207 条智能体轨迹。每个代码版本都带有其分数、时间戳，以及所执行动作、意图、编辑规模和分数影响的标签。这样来看，差距就变得具体了。专家会交替进行数据处理、验证、模型修改和集成，并回到此前放弃的方案上。相比之下，每个智能体框架都会收缩成一个狭窄循环：Codex 把步骤用在重新加权集成和调参提交上，MLEvolve 则在原地变异模型，两者既不会像人类那样频繁转向，也不会重新开启被放弃的工作。一个从人类实践中提炼出的简短规划提示，会把其所描述的行为推向人类画像并提升分数，但努力模式仍然是智能体式的：指令只能弥合那些可归结为指令的差距。我们在 https://huggingface.co/datasets/jerryyan/TraceML 发布了该语料库、模式、标注器和抽取流水线。
自进化的 LLM 编码代理通过模仿从共享技能库中检索到的技能来编写自己的工具。我们在这一循环中发现了一个漏洞：在编写过程中，被检索到的恶意技能可以成为新技能的模板，并保留有效载荷。我们将其称为自我投毒：代理编写、存储并运行由此产生的恶意技能。我们通过 EvoMal 利用这一点，这是一种通过将可替换的有效载荷包裹在 banner 中来放大自我投毒的攻击；banner 是一组看似无害的结构元素，会诱使模仿型代理复现其中的代码。攻击者无需调用这些恶意技能，就能将其植入库中。随后，代理会编写并执行携带有害代码的新技能。每个编写出的副本都可以重新进入库中，再次被模仿，形成一种在植入技能被移除后仍然存在的自我传播蠕虫。我们将代理自我投毒率（ASPR）定义为将新编写的恶意技能加入库中的任务占比。在 6 个模型、153 个与工具相关的 SWE-bench Verified 任务上，ASPR 范围为 20.3% 到 41.8%，而被污染的库中恶意技能数量是植入数量的 4.9 到 9.0 倍。即使没有 banner，这一漏洞也会出现：DeepSeek-V4-Pro 在仅含有效载荷时的 ASPR 达到 11.1%。将植入技能的描述针对某一任务家族进行定制，可将 ASPR 提高到 86.7%。在移除植入技能后，由于代理生成的副本仍然存在，Qwen3 的第 5 轮 ASPR 仍达到 68%。这些副本绕过了现有防御，因为这些防御主要针对攻击者提交的名称、代码和签名。我们提出 counter-prompt 作为一种防御，它能抑制 banner 风格的复制，并将 EvoMal 的 ASPR 降至不超过 6.7%，且任务完成无显著损失。
Modern GUI-agent frameworks achieve strong desktop task performance with frontier API models, yet persistent control information often remains implicit in growing interaction trajectories. At each step, the planner reconstructs the active task stage, accumulated evidence, and runtime feedback before deciding the next action. This dependence becomes more pronounced under weaker local reasoning backbones. Across four representative state-of-the-art frameworks, replacing GPT-5 with Qwen3.5-9B reduces average OSWorld SR-100 from 60.9\% to 37.7\%. Trajectory annotation further identifies at least one control failure in 91.6\% of failed trajectories. To address this problem, we introduce LocalLSTC, a training-free architecture that organizes control by temporal scope, maintaining persistent cross-step state to guide short-term execution commitments. Long-Term Control maintains the active subgoal, subgoal-aligned evidence, and runtime feedback across interactions, while Short-Term Execution realizes bounded commitments for the current step. Long-to-Short Planning forms each commitment from persistent state, and Short-to-Long Control integrates execution outcomes back into that state for progress assessment, recovery, and termination. With Qwen3.6-27B, LocalLSTC reaches 64.7\% SR-100 on OSWorld and 65.3\% on WindowsAgentArena, outperforming the strongest prior local results on both benchmarks. Ablations further support contributions from mechanisms on both sides of execution. These findings identify temporal organization of control information as a distinct architectural dimension for locally deployed GUI agents.
随着基于大语言模型（LLM）的多智能体系统（MASs）越来越多地应用于长周期复杂任务，其可靠性已成为阻碍其实际部署的核心瓶颈。现有的 MAS 调试和修复方法通常依赖于对整个执行轨迹进行重新运行和重采样。然而，一个根本性问题仍有待回答：这些方法究竟是在因果上修复 MAS 故障，还是仅仅借助 LLM 采样的随机性进行概率性修复？为评估 MAS 修复方法的有效性，我们引入了 SymTrace，一个受控评估框架，它记录 MAS 的执行轨迹并建立干预锚点。在回放过程中，它利用记录的日志有效重建锚点之前的执行，并且只重新生成下游轨迹，从而能够可靠地复现 MAS 故障。我们进一步构建了数据集 SymFail，其中包含 536 条经人工标注的故障轨迹，附带图关联的位置、类别和轨迹证据。基于这些基础，我们在三个主流 MAS 框架上开展了大规模实证研究。结果表明，现有无引导的重新运行方法极不可靠，其故障复现率和修复率都很低（分别仅为 67.97% 和 6.90%）。在此基础上，我们进一步探究了基于症状驱动的干预方法的有效性，该方法成功修复了 20.15% 的故障案例（相较当前最先进的修复方法提升了 191.89%）。本研究旨在为 MAS 调试与修复研究提供可操作的洞见，为多智能体系统的稳健部署铺平道路。
世界模型旨在模拟复杂环境在动作和事件作用下如何演化，但现有基于视频的世界模型主要从视觉观测中学习动力学，这些观测呈现的是结果，而非支配世界演化的底层知识、规则和机制。这使得它们难以维持持久的后果，并支持连贯、开放式的演化。我们提出 Code World Model，这一框架将世界演化与视觉呈现分离，结合语言模型的推理与编码能力以及视频模型的生成先验。一个 coding agent 充当世界大脑，对事件及其后果进行推理，并生成可执行代码以维护持久的世界状态并执行符合规则的演化。为将可执行状态与视觉生成连接起来，我们引入一种 proxy 表示，它编码逐帧时空约束，并被编译为 proxy video，用以条件化视频模型，从而渲染高保真的视觉观测。我们还开发了数据流水线，用于从游戏和真实世界视频中构建对齐的 proxy-观测对。经过配对游戏数据微调后，MiniMax-H3 能够遵循由 coding agent 构建的简单交互式世界中的基于 proxy 的时空规范，同时保留丰富的视觉细节和动态效果。这些结果展示了将用于持久世界演化的代码与用于灵活视觉呈现的视频模型相结合的潜力，为开放式世界模型提供了一条新路径。
Traffic Anomaly Understanding（TAU）要求模型和系统能够在交通视频中检测、推理并解释异常事件。为解决这一挑战，我们提出 TAU-Agent，一种面向交通异常理解的智能检索增强框架。给定一个任务查询，中央检索代理会协调两种视觉感知工具，即 Video Captioning Tool 和 Open-Vocabulary Tracking Tool，以检索并选择与查询相关的证据，包括字幕、时间区间和目标轨迹。所选证据与采样的视频帧以及输入查询一起，提供给经过监督微调的视觉语言模型，用于最终推理和答案生成。我们在 AI City Challenge 2026 的域内和域外基准上评估了 TAU-Agent。TAU-Agent 在 Track 3 上取得 0.6779 分，在 Track 7 上取得 0.3998 分，在 Track 8 上取得 67.9275 分，分别排名第二、第十二和第五。代码可在以下地址获取：https://github.com/siri-rouser/TAU-Agent。
多智能体系统（MAS）有时本来就具备正确回答的潜力，但仍会报告错误答案。要解释这一结果很困难，因为生成、通信和最终答案选择规则通常会同时变化。我们将多智能体推理概念化为一个进化管线，包括候选生成、同伴通信和终端选择，其中缺乏质量控制的一致性可能表现出模因漂移的模式。我们研究两个问题：（1）当 LLM 裁判为多智能体系统中生成的候选答案提供答案正确性的信号时，它在何种情况下能提供有效的选择压力；（2）使用该信号何时会改善报告的答案。为了描绘裁判可靠性，我们分析了来自 MMLU-Pro、GPQA、MedXpertQA 和 MuSR 的 15,336 个问题，Humanity's Last Exam 则单独分析。为了检验这些规则，我们重放了从五个基准中的 16,278 个问题抽取的 81,390 个固定候选池。我们报告三点发现。（1）正确答案往往已经存在于生成的候选答案之中，但系统仍可能收敛并报告错误答案。（2）裁判可靠性并不是模型的固定属性，而是会随任务、生成器以及正确答案有多罕见而变化。（3）将答案频率与裁判评估结合起来，只改变了最终的答案选择规则，并将准确率从 63.82% 提高到 70.82-70.95%，主要通过挽救那些被流行错误答案压倒的正确答案。在这里研究的系统中，生成更多候选答案的价值取决于这些额外样本是否使正确答案变得存在、变得常见或变得可识别。通过将生成、识别和选择分离，这些发现为设计能够防止生成的正确答案在过程中丢失的多智能体架构奠定了诊断基础。
我构建了一个 agent，可以帮助你打造受 Rollercoaster Tycoon 启发的主题公园。你可以输入类似“给我建一个酷炫的主题公园”的提示，它就会建出一个整体协调的主题公园，包含多个世界，并用道路和游乐设施连接起来。 有意思的是，我构建它时用的，正是我们用来让 AI 生成的网站遵循公司设计系统的同一套思路。 我是 Magic Patterns 的一名工程师，专注于构建我们的 Design System Agent，目标是使用你现有的品牌、组件和惯例，而不是生成看起来像通用 AI 生成 UI 的东西。 默认情况下，模型往往会收敛到外观相似的设计。这也是那种“vibe coded slop”感觉的一部分：相同的字体处理、阴影、卡片、图标、动画等等……我的工作就是找出合适的上下文和护栏，这样当你提示“给我构建一个仪表盘”时，结果真的会看起来像你现有的产品。 某个时候我意识到，同样的思路也可以应用到 RCT 上。 在 Web 领域，你可能会有关于哪些字体、颜色、间距和组件应该搭配使用的规则。 在主题公园里，你需要这样的规则：过山车需要完整的轨道，游乐设施需要有连接到道路的入口，道路需要连接公园中的不同区域，而海盗主题世界就应该真的使用海盗主题的景观。 看到构建一个协调一致的主题公园，与构建一个协调一致的产品 UI 这个问题如此相似，这很有意思。和 Web 设计类似，仅仅给模型提供正确的组件是不够的。 最后我构建了一个评估循环：Magic Patterns 会生成一个公园，另一个 agent 会根据评分标准对它打分，然后这个 agent 会在再次尝试之前更新它的规则和技能文件。 评分标准会检查一些内容，比如过山车是否形成了有效轨道且至少有一个下坡，游乐设施是否可以通过道路到达，每个世界是否使用了合适的主题景观，以及整个公园是否作为一个整体运转良好。 经过大量迭代之后，现在你只要输入像“给我建一个酷炫的主题公园”这样简单的提示，就应该能得到一个功能完整的公园，里面有游客、游乐设施、道路、过山车和主题世界。 最棒的是，你还可以跟随游客，聆听他们的想法，并以 RCT 的方式观看他们乘坐游乐设施。 很乐意回答任何问题，也很想看看大家创建的公园！ Comments URL: https://news.ycombinator.com/item?id=49452037 Points: 15 # Comments: 4
文章链接： https://openai.com/webmcp-challenge/ 评论链接： https://news.ycombinator.com/item?id=49455713 得分：3 评论数：0
文章链接： https://blog.trailofbits.com/2026/08/26/vms-wont-contain-cyber-capable-agents/ 评论链接： https://news.ycombinator.com/item?id=49450188 得分：66 # 评论数：27
文章链接： https://sreenathmenon.com/blog/2026-08-04-webmcp-teaching-websites-to-talk-to-ai-agents/ 评论链接： https://news.ycombinator.com/item?id=49450417 分数：16 # 评论数：6
文章链接： https://arxiv.org/abs/2607.21503 评论链接： https://news.ycombinator.com/item?id=49443523 得分：12 # 评论：3
Large language model agents are moving beyond conventional retrieval-augmented generation toward direct interaction with external corpora. Direct Corpus Interaction (DCI) keeps the full corpus accessible, yet reachable evidence can remain unusable under finite interaction budgets. Required evidence may fail to surface, a surfaced supporting document may remain unopened, or an opened document may fail to expose its decisive fragment. We call this progressive silent loss Evidence Blindness and quantify it through stage-wise evidence realization. Within the DCI paradigm, raw interaction adds little reusable corpus organization, while dynamic-workspace methods reconstruct a query-conditioned interaction space from each query and trajectory. In both cases, useful structure is recovered largely online. We instead formulate large-scale agentic search as finite-budget navigation over reusable corpus structure. We introduce AtlasNav, a persistent multi-view corpus-navigation framework that retains direct corpus interaction but organizes the corpus once into a Corpus Atlas, allowing each query to navigate adaptively rather than reconstruct shared structure. On BrowseComp-Plus, AtlasNav achieves 92.05% strict accuracy while reducing recorded online inference cost by 30.21% relative to the prior dynamic-workspace state of the art. Under matched budgets, it realizes the complete required evidence earlier and approaches the same model's evidence-supplied empirical reference more rapidly. The same representation principle remains effective under PhantomWiki's distinct corpus organization and controlled 10K-1M scaling, and transfers competitively to heterogeneous enterprise knowledge. These results show that agentic search depends not only on accessible evidence, but also on how the corpus is represented so that limited interaction becomes effective navigation.
基于 LLM 的代理可以通过调用工具与外部环境交互，但这一能力也带来了文件修改、信息泄露和未授权操作等安全风险。现有的防护机制通常在轨迹完成后进行评估，而对执行前的步级动作监控关注不足。我们提出 StepGuard，这是一种步级防护模型，既可以审计已完成的代理轨迹，也可以在工具动作执行前进行检查。为训练 StepGuard，我们引入 StepGen，一个自动化数据引擎，用于生成具有相同上下文但在高风险步骤采取不同动作的安全与不安全轨迹。为进一步减少过度防护和防护不足，我们提出 Balance-GRPO，它根据安全与不安全动作的观测准确率动态平衡学习。实验表明，StepGuard 在开源权重防护模型中取得了最高的平均准确率，其性能可与 GPT-5.4 相媲美。在 AgentDojo 和 AgentDyn 上用于保护代理时，StepGuard 相比无防护设置可将平均攻击成功率降低 77.3%，而平均效用仅下降 2.8 个百分点。
基于结果监督的搜索代理学习何时以及如何检索证据，但终端奖励既不能定位中间错误，也不能在这些错误累积之前重新引导正在进行的轨迹。将纠正性反馈视为一种在轨迹内学习到的干预，会把这两个角色耦合起来：代理必须决定何时请求并使用反馈，而评判器必须从结果混杂的 rollout 中推断出有用的修正，而这些 rollout 的失败模式会随着代理的改进而变化。我们提出 CAFE（Coupled Agent--Feedback Evolution），一个使共享参数模型在搜索代理和评判器角色之间交替切换的框架。CAFE 先从围绕基础代理自身失败构建的轨迹中初始化基于反馈的恢复，然后将在线和离线优化耦合起来。在在线 RL 期间，一种比较式反馈估计使用提示级别的 call--skip 成功差距来塑造请求回报，而感知反馈的优势塑形会在反馈前后重新加权 token 优势。在线下，基于 rollout 生成的偏好优化从匹配的成功和失败轨迹中学习反馈。在七个 agentic search 基准上，CAFE 平均优于所评估的基于 RL 的搜索代理，在全部六个域外基准上保持其增益，并减少答案级幻觉。单边消融表明，只改进代理或只改进评判器最终都会进入平台期，而交替进行这两种更新会持续提升性能。这些发现表明，自我改进的搜索代理需要与其所引导的策略共同演化的反馈。
我们提出 StarHarness，这是一个在保持模型权重固定的同时，为特定环境演化代理执行框架的框架。演化后的框架可以包含提示与任务表述、工具接口、技能、由 MCP 支持的提供方、子代理结构以及代理循环配置。StarHarness 通过根据基线失败行为对任务进行分层来构建一个紧凑的演化池，将提议者可见的搜索任务与提议者不可见的选择任务分离，并保留留出任务用于评估泛化能力。在 ITBench SRE、EnterpriseOps-Gym ITSM 和 AutomationBench Finance 上，在每个环境接受 4-12 项更改后，框架演化使完整基准性能相较默认框架提升了 20-35 个百分点。这些提升在未参与演化的任务上依然存在，并且无需重新演化即可在 GPT 和 Qwen 模型家族之间迁移。轨迹分析将这些改进归因于接口修复、环境约定以及压缩搜索空间的运维知识；在若干设置中，这还带来了更少的假阳性诊断和更短的执行轨迹。因此，StarHarness 提供了一种实用方法，用于减少工具丰富的企业任务中持续存在的模型与环境不匹配。
Group-relative reinforcement learning waits for sibling rollouts of the same prompt, which is costly for long and variable tool-use trajectories. Single-stream Policy Optimization (SPO) removes this dependency with a persistent prompt-level value estimate, but its recipe whitens one advantage per trajectory before optimizing a token-mean actor loss. We show that trajectory centering generally does not center the token-weighted quantity consumed by the actor, and fix the mismatch by standardizing terminal-outcome advantages under the action-token measure. We additionally organize prompt evidence by the policy event that generated it rather than learner receipt order. Across matched runs on ALFWorld at two model scales and on Math-TIR, SPO++ improves online learning efficiency over SPO. A paired ablation identifies action-token-measure normalization as the strongest tested component.
Recursive self-improvement (RSI) remains hard in long-horizon tasks, where growing histories obscure the task state and misalign skill invocation. We introduce Recuris, a recursive Experiential-Working Memory architecture for long-horizon agent harnesses, in which Working Memory tracks task progress and guides skill selection from Experiential Memory, grounding skill use in current needs rather than the full history. This coupling also turns execution into structured evidence that localizes failures to specific memory components. Across tasks, a fixed Meta-Agent turns that evidence into localized, validation-gated updates to Skill Memory that reshape execution and yield new evidence, forming a bounded recursive memory-evolution loop. Across four long-horizon benchmarks and ten models, Recuris improves task success in 35 of the 37 completed model-benchmark pairs, carrying frontier models to SOTA-level task success: on tau-bench it adds +17.8 points to GPT-5.6 Sol and +15.6 to Claude Opus 5, taking Opus 5 to 87.9%, and +16.6/+13.5 points on Qwen3.6-27B/35B on SkillFlow. The advantage widens as the interaction horizon grows, to +32.2 points on the longest tasks, and common long-horizon failures fall by up to 80%. These results position recursively evolving memory as a scalable foundation for RSI, enabling agents to continuously transform accumulated experience into increasingly effective long-horizon behavior. Code: https://github.com/Gen-Verse/Recuris
I'm a master's engineering student and a big fan of LaTeX, which I used for my thesis and research articles. I have used Overleaf and that was fine until I wanted to git sync, which unfortunately sits behind a paywall. Since I didn't want to pay subscriptions for things that should simply just work, I built the editor I wanted in my free time, where you open a tab, write LaTeX, get your PDF and the files stay all in one folder on my disk. TeXbrain is a static site with no backend. pdfTeX is compiled to WebAssembly (SwiftLaTeX) and runs in your browser. The editor can read and write your project folder through the File System Access API, so you can use git, any local TeX install, local AI Agents, or any other editor on the same files. Git is built in through isomorphic-git for anyone who would rather not touch a terminal and clone, branch, commit, push or pull via commands. No account is needed, no analytics, and everything works offline after the first load. Try it out: https://tex.swimmingbrain.dev/ (Chromium browsers get direct folder access, while Firefox or Safari fall back to a virtual filesystem) Or for self-hosting/contributing: https://github.com/swimmingbrain/texbrain (MIT, the pdfTeX engine is EPL 2.0 / GPL 2.0 and is listed in the THIRD_PARTY_LICENSES file). The part I'm most proud of getting to work is the package loading. The engine itself is only 1.8 MB. When it asks for a file it doesn't have, then a service worker intercepts the request and resolves it through Cache Storage first, then a small bundled subset, then a TeX Live tree mirrored on jsDelivr, then a SwiftLaTeX style server as a last resort. Every file is fetched once at most and after the first successful compilation, the core subset is prefetched in the background so that the offline story actually holds. Only file names go over the network, never any document content. So far, it only supports pdfTeX (no XeTeX or LuaTeX), so fontspec or polyglossia won't work, packages are pinned to the TeX Live 2020 era to stay coherent with the engine's format file and there is no bibtex or biber in the engine yet, which is maybe the roughest edge. This is should be a good solution for people on a work laptop, a Chromebook, or a university lab PC where installing TeX Live isn't always an option, and people who don't want to maintain a TeX install at all. If you already have a local setup you like, keep it. This isn't necessarily thought as a replacement for that. Some people have already used it for their thesis in the past months and their bug reports (which I would never have hit on my own) were then most of the summer's work. The first outside PR also landed recently, so if something doesn't compile for you, an issue with the package name or the bug itself is the most useful thing you can send me! Comments URL: https://news.ycombinator.com/item?id=49441375 Points: 6 # Comments: 1
Found that you can actually run a 35B Qwen model on a Pi with very impressive intelligence and stability. Built connectors for car ODB to read all about car internals, and manufacturer's cloud service for stuff like changing AC or opening/ locking doors. Gave it info such as the full car manual. And then hooked it up with my other agents in our discussion room! So now it can answer car questions such as "when should I add oil and what kind of oil?" and help you fully offline, and when online talk with the agent family that includes all the most powerful models so they know how the car is, plan new features and develop itself with them. For example, if car breaks and can't move the car agent informs my agent family and they can already look for a suitable train ticket without me needing to check things or do something. Comments URL: https://news.ycombinator.com/item?id=49435675 Points: 6 # Comments: 0
It's a macOS menu bar app that reads the text of your focused window every few seconds through the Accessibility API. No screenshots, no video, or OCR. It writes plain markdown, one file per day, into a folder you choose. Point Claude Code (or anything with file access) at the folder and ask "what did I work on Tuesday?" or have it build memory about your projects. The folder contains an AGENTS.md explaining the format to the model. Thought it might be useful for someone so chucked it up on GitHub Comments URL: https://news.ycombinator.com/item?id=49429095 Points: 5 # Comments: 0
交互式临床智能体必须在部分可观测条件下收集决定性证据，并将其转化为有依据的行动。仅有正确的最终诊断并不能表明智能体遵守了证据和护理流程约束。我们提出 MediSkill-Evo，这是一种无需对主干模型进行微调即可演化受约束过程知识的临床智能体。它将经验分为四类类型化知识库，分别用于临床技能、过程规则、符号模式和测量程序。溯源、支持、回放以及控制器定义的安全检查共同管理向冻结的测试时快照发布内容。过程约束偏好框架将证据与其来源绑定，拒绝控制器判定无效的候选项，并使用安全优先的 Clinical Process Critic 对行动进行排序。我们在相同的 Doctor 轮次限制下，跨两个主干端点和六个受控压力维度评估完整智能体系统。在 300 个留出的 Qwen 就诊案例上，与 AgentClinic 相比，MediSkill-Evo 将诊断准确率从 61.33% 提高到 69.00%，将治疗意图覆盖率从 33.62% 提高到 66.44%，同时将自动评分的关键失败率从 31.00% 降低到 16.33%。在由 30 个病例衍生的 180 个硬隔离条件下，在患者行为压力下的目标恢复率达到 93.61%，时间性证据达到 100.00%，分诊危险信号达到 92.22%。一项探索性的 100 例 MedSAM 比较评估了请求门控工具接口的可行性。这些结果为固定评估套件上的完整系统提供了描述性的端到端证据，而不是针对单个知识库的因果证据，也不是对自动评判器的临床验证。
Agent skills are reusable procedural artifacts that extend language agents with specialized workflows, tool conventions, and domain behaviors at inference time. However, creating reliable skills still depends largely on human authorship, model priors, or execution traces. These sources are often unavailable for unfamiliar tasks, suggesting the need to create skills from open-world materials. In this paper, we study open-world skill creation: given an underspecified skill brief and a source-access specification, a creator must discover behavior-relevant requirements omitted by the brief and determine how broadly each source-derived procedure is justified. We propose SkillAlchemy, an admission-centered framework for source-grounded skill creation. SkillAlchemy identifies implicit requirements through contrastive evidence, admits candidate procedures based on evidence-supported scope, and compiles the admitted content into a grammar-guided skill package. Extensive experiments across 87 SkillsBench v1.1 tasks demonstrate that SkillAlchemy improves pass rate over no-skill execution by 19.9 percentage points and the strongest automated baseline by 8.6 percentage points, while achieving performance comparable to human-curated skills.
Understanding a basketball game requires recognizing events, localizing actions, identifying players, and relating these to structured game knowledge. Existing benchmarks primarily evaluate these abilities one at a time, leaving the interactions among these abilities under-explored. We introduce BasketballBench, a multimodal benchmark comprising 7,980 questions across ten tasks in text, image, and video. It is built from the 2025-2026 NBA season and includes official playby-play, rosters and profiles for 530 active players, and 2,501 possession-level broadcast clips. We further propose BasketballSkills, an agent that composes eight basketball-specific perception and retrieval tools under four reusable skills that specify tool order, evidence bindings, and stopping conditions. Experiments show that current MLLMs struggle particularly on questions requiring the integration of multiple capabilities, whereas BasketballSkills outperforms them, highlighting the effectiveness of explicitly composing domain-specific capabilities for comprehensive basketball understanding.
Space robots operate in extreme environments where hardware degradation can critically compromise traditional control strategies. While continual reinforcement learning offers a promising mechanism for online adaptation, it inherently requires access to a reward signal during deployment. However, precise reward computation in space is often infeasible due to the lack of external tracking systems and the overall complexity of the environment. To address the challenge of unobservable rewards, we introduce a reward-free continual learning framework that leverages latent-state world models. By pre-training a model-based agent across diverse simulations, the world model learns a robust predictor of the reward structure within its latent space. Upon deployment to an environment with severe hardware degradation, we freeze the observation encoder and reward predictor to update only the transition dynamics of the world model through unsupervised rollouts. By training the policy entirely on imagined trajectories generated by this updated world model, the agent adapts to altered dynamics without receiving new rewards. We demonstrate our approach across simulated planetary traversal, orbital navigation, and precision assembly tasks subjected to severe morphological failures.
记忆正成为已部署 LLM Agent 中的默认子系统，用于提供持久的个性化和连续性。这自然引出了一个问题：记忆系统是否会给 Agent 引入新的漏洞？因此，我们提出 InjecMEM，一种新颖的记忆注入攻击范式，它只需要一次交互（无需对记忆存储进行读取/编辑访问），即可将后续相关查询的响应引导至预先指定的输出。受记忆系统“先检索后生成”机制的启发，我们使用与检索器无关的锚点和对抗性命令来构造注入。锚点包含高召回率的主题线索，使下游检索能够持续将该记录与目标主题关联起来。命令是一段经过优化的短序列，能够在不确定的融合上下文、可变位置和长提示下保持有效，从而在被检索到后可靠地引导输出。我们通过基于梯度的坐标搜索来学习该命令，并在合成提示模板和插入位置上取平均，同时将其扩展到跨骨干模型的联合优化，以研究迁移性。在多个记忆系统和骨干模型上的评估表明，InjecMEM 能实现可靠的主题条件检索和目标生成，在记忆漂移下仍保持有效，并且不影响非目标查询。我们的结果强调了加固记忆系统的必要性，并提供了一个可复现的框架，用于研究 Agent 记忆。
Time series forecasting (TSF) is evolving toward multimodal and agentic settings, yet using foundation models remains uneconomical in resource-constrained scenarios, where compact, specialized forecasters are more desirable. However, lightweight forecasters typically require substantial training data, limiting their use in domains with scarce, slowly accumulated, or privacy-sensitive time series. To address this dilemma, we investigate the challenging problem of few-shot learning for lightweight forecasters. We propose MetaCaster, a meta-harness-optimized multi-agent framework that uses agentic data generation to automatically train specialized lightweight forecasters from only a few examples and textual contexts. Our work highlights a new TSF paradigm in which agents act not as forecasters but as intermediary engineers that prepare efficient, task-specific forecasters for deployment. Experiments on 18 datasets, 23 state-of-the-art lightweight forecasters, and 14 baselines demonstrate that MetaCaster achieves both data efficiency and computational efficiency while maintaining high-quality TSF performance.
地球系统分析从来源、尺度、时间和模态各不相同的观测中重建不断变化的物理过程。自然灾害使这项工作具有重大影响，因为不完整的证据可能改变对严重程度、暴露度和机制的估计。我们介绍 EarthVerse，这是一个通过包范围调查来评估科学智能体的基准。其 405 项可复现任务基于 199 个有记录的事件和 19 类灾害。智能体需要检查异构事件包，选择相互兼容的证据，执行透明的计算，协调来源差异，并在最终答案中保留溯源信息。我们提供可执行的标准答案，将每项任务分解为细粒度的答案单元，并配套特定任务的评分规则，用于评估支撑性的研究过程，同时允许多条有效路径。我们在受控的工具使用协议下评估了 25 个模型和智能体系统，随后通过受控研究定位其在证据获取、工具选择、记忆、推理、交互和科学执行方面的失败。总体来看，系统中最高的平均答案单元准确率为 84.65%，而最高的 Strict@95 仅为 34.81%。这一差距表明，当前智能体往往能完成单个步骤，却无法在证据、尺度、单位、计算和物理解释之间维持一致的链条。EarthVerse 为衡量动态地球系统中的端到端科学可靠性提供了可复现的基础。
Does multi-agent LLM interaction help or hurt? Some work reports gains from debate (Du et al., 2024), critique loops (Chen et al., 2025), and mixture-of-agents synthesis (Wang et al., 2025), while other work finds that interaction adds cost without improving quality under equal budgets (Tran & Kiela, 2026; Xu et al., 2026; Jarrett et al., 2025), or that independent sampling already captures multi-agent gains (Li et al., 2024). We argue this contradiction partly reflects a missing distinction, because not all multi-agent communication is equal. Different model families find structurally different solutions, but when agents read each other's complete outputs, their proposals converge within one round, erasing the diversity that motivates using multiple models. We call this the interaction tax. We test 11 verifier-scored optimization tasks under matched budgets and find that full-solution interaction is a weak default. Independent proposal generation avoids this collapse. Full-solution interaction mainly makes agents stay close to the first solution they see instead of trying different approaches, and critique helps only if the violated rule is easy for the LLM to find and fix. These results suggest that multi-agent performance depends less on the number of agents than on the information they exchange, and interaction helps only when agents share the right information at the right time.
语言模型是顺序处理器，但长时程智能体能力需要模型权重和活动上下文之外的外部信息与计算。Prime Agent 是一个用于长时程评估和编码智能体工作流的开源框架。一个持久化的 IPython REPL 遵循 Recursive Language Model 抽象，用于程序化上下文处理和测试时计算，而 Continual Harness 则在多条轨迹之间保留历史、记忆、技能、提示词和子智能体规范。递归子智能体通过直接的智能体间通信进行协作，Agents View 让人类能够检查和管理由守护进程支持的会话。Prime Agent 将执行、恢复、验证和资源核算标准化，同时将策略构建留给模型。这种低摩擦、富有表达力的膜层可防止框架故障变成模型故障，并将测量推向模型真实的最大底层能力。Prime Agent 将 ARC-AGI-3 RHAE Best@1 从 30% 提升到 95.5%，并在长上下文编码、GPU 内核生成、模拟器构建以及自主 nanoGPT 速通等任务中达到或超过原生框架和流行框架。在 Factorio 上，我们发现精炼使持续技术进展成为可能，而专用子智能体支持并行化工作。代码可在 https://github.com/PrimeIntellect-ai/prime-agent 获取。
现代软件系统在数十年的开发过程中积累了技术债，这使得迁移成本高昂且很大程度上依赖人工。随着编码智能体在错误修复方面能力日益增强，它们能否自主执行此类迁移？现有基准无法回答这个问题，因为它们只评估行为正确性，而不评估迁移是否真的发生。这导致了一种简单的取巧方式：智能体复制原始实现以让测试通过。我们称之为 Blindness。为解决这一问题，我们提出 SWE Refactor Bench，这是一个包含 20 个全仓库迁移任务的基准，覆盖 4 类技术债。一个三阶段评估协议同时衡量迁移完整性和行为正确性。(1) Migration Audit 验证迁移是否发生。(2) Behavioural Tests 使用固定测试套件衡量正确性。(3) Agentic Verification 使用 6 个独立编码智能体生成针对隐藏行为差异的定向测试。在来自 8 个前沿模型和 26 种模型-投入配置的 520 次运行中，只有 28 次（$5.4\%$）通过全部三个阶段，20 个任务中有 13 个没有获得任何被接受的解决方案，最佳模型（claude-opus-5）得分为 $47.0/100$。迁移完整性和行为正确性是不同的能力：少数运行通过跳过迁移来保持行为不变，并在 Migration Audit 阶段被拦截；大多数运行尝试迁移但破坏了行为，并在 Behavioural Tests 阶段被拦截。智能体无法交付完美迁移：在通过 Migration Audit 的 340 次运行中，$58\%$ 达到固定检查的 $99\%$，但只有 $26\%$ 达到 $100\%$。智能体能力在不同迁移类别之间存在差异：智能体在构建工具链重写上的得分为 $31.4$，但在语言重写上的得分仅为 $5.6$。综上，这些发现将 SWE Refactor Bench 定位为一个严谨的测试平台，用于开发能够可靠执行全仓库迁移的编码智能体。
文章 URL： https://github.com/microsoft/agent-lightning/releases/tag/v1.0.1 评论 URL： https://news.ycombinator.com/item?id=49423077 得分：18 评论数：1
LLMs have evolved from language generators to autonomous agents capable of complex, long-horizon tasks. This evolution has produced paradigms including Prompt Engineering to elicit model capabilities, Context Engineering to manage information access, Harness Engineering to organize external tools and resources, and Loop Engineering to support continual reflection and self-improvement. Yet as tasks grow more complex, individual intelligence faces a fundamental limit: many tasks require heterogeneous expertise, interdependent subtasks, parallel execution, independent verification, and persistent state, exceeding any single agent's organizational capacity. Augmenting one agent's capabilities or context cannot resolve this architectural mismatch; intelligence must instead be distributed across specialized agents and organized at the system level. We call this System Intelligence: an agent system's ability to organize and coordinate multiple intelligent components into a coherent, adaptive whole pursuing a shared objective. Achieving it requires more than adding agents; it demands explicit structures to organize work, coordinate heterogeneous agents, and maintain evolving execution states. We introduce Graph Engineering, an emerging paradigm for next-generation agent systems. Unlike prior paradigms that mainly optimize individual interactions or agent-level behavior, Graph Engineering constructs explicit, dynamic, evolving graph structures representing tasks, agents, and system states. These abstractions provide a unified foundation for organizing complex objectives, orchestrating heterogeneous agents, modeling system dynamics, and enabling scalable agent evolution. We systematically review the principles, methodologies, and applications of Graph Engineering for LLM agents. Related papers, open-source data, and projects are collected at https://github.com/DEEP-JLU/Awesome-Graph-Engineering.
使用工具的 AI 代理会将委托任务转化为提供方效应，但授权往往止步于准入，而提供方状态、交付、重试和恢复仍在演化。请求可能在提交前发生变化，或者响应丢失可能导致替代请求从一次批准中产生第二个效应。我们提出 AID-Guard，一种有状态的授权到效应闭合协议。它会在提交时重新验证已批准的请求和提供方状态，在存在歧义时保留一个预留，并且只有在出现终止结果或带有交付栅栏的经认证无效应之后，才允许释放或一个后继。对于受支持的提供方契约，一个预留在重试和恢复过程中最多产生一个效应。据我们所知，这是首个经过评估的代理授权协议，将这些控制统一到一个生命周期中。 我们实现了一个 Python/SQLite 原型。在声明的回环 MCP 域中，13 次实时变更未造成任何未授权的提供方效应，三个并发历史是可线性化的，并且证据包支持公开验证和重放。全部 210 次 Stripe 提供方契约试验均与预先声明的结果一致。在 Stripe 和 Resend 上，40 个终止化后继调度、30 个重叠竞争以及 10 个崩溃恢复调度均在没有重复效应的情况下完成。在提议者完全被攻陷的情况下，AID-Guard 阻止了 44/44 次攻击，并接纳了 44/44 个匹配的合法提议。其严格的精确清单配置使良性效用降低了 35.4 到 43.8 个百分点；类型化前沿在未观察到不安全效应的情况下恢复了 9–10 次完成。组合研究阻止了 20/20 次准入后生命周期攻击，并保留了 8/8 次有效或精确重试执行。结果支持在所评估的效应路径清单、提供方契约和故障调度下进行授权到效应绑定。
This paper investigates cross-agent specification portability using Oracle-to-PostgreSQL migration as a controlled software transformation task. The study combines two experimental stages. First, a specification-first migration pipeline was evaluated on 1,006 PL/SQL files, of which 623 were successfully regenerated and 380 generated scripts executed successfully in PostgreSQL 16. Second, cross-agent experiments were conducted on a dataset of 1,802 Oracle scripts with corresponding PostgreSQL implementations using Amazon Kiro, Google Gemini, and GitHub Copilot, with Claude Code and Cursor included in the initial single-agent evaluation. Native and foreign specifications were assessed using Token F1, exact match, SQL syntax validity, AST exact match, AST mean similarity, and immediate runnability. The results show that specification size alone does not predict implementation quality and that cross-agent transfer can produce substantial agent-dependent degradation. The strongest replicated case occurred when Gemini directly consumed a Kiro-origin specification, producing a Token F1 of 0.035, SQL syntax validity of 2.33%, and AST mean similarity of 0.015. Rewriting substantially improved Gemini in the tested configuration, compression did not provide a universal benefit, and retrieval-augmented ingestion was the only common strategy represented on the per-agent Pareto frontiers of both Gemini and Copilot. The findings suggest that specifications in heterogeneous SDD workflows should not automatically be treated as agent-neutral artifacts and motivate explicit consideration of specification portability, agent-specific interpretation, and retrieval-based access in multi-agent software engineering.
agentic AI 的兴起使 LLMs 能够访问多样化的用户数据，从而引发了关键的隐私担忧。以往关于情境隐私的研究探讨 LLMs 是否会根据依赖情境的规范来调节信息披露。然而，即使在相同情境下，不同用户可接受的披露边界也可能有所不同。为解决这一局限，我们引入 \textit{个性化隐私}，将用户特定的披露偏好纳入隐私控制。我们进一步提出 P3Bench~（\textbf{P}ersonalized \textbf{P}rivacy \textbf{P}reservation \textbf{Bench}mark），这是一个新的基准，将情境隐私策略扩展为个性化披露策略。实验表明，基于提示的策略无法可靠地执行个性化隐私策略，其中 Qwen2.5-7B 和 Gemma3-4B 的平均策略忽视率分别为 51.25\% 和 74.28\%。最后，为解决这一问题，我们提出 \textsc{Repair}，这是一种稳健的推理时注意力头干预方法，可将披露行为调整为符合策略的响应。我们的方法通过减少模型未能遵循给定策略的情况，显著提高了对用户特定隐私偏好的遵循程度。
持久记忆会让虚假信息变得持久：一旦某条虚假陈述被存储，它就可能在未来与之匹配的会话中被检索出来。我们使用一次性生成的、措辞直白的虚假断言来衡量这种失效模式的代价，不进行指令、触发器或检索器优化。对 LongMemEval 语料库中 1.2% 的内容进行投毒，会使准确率从 0.850 降至 0.300。一个四阶段的写入时筛查流水线在间接提示注入上达到 0.832 的召回率，同时会将 1.5% 含触发词的良性文本标记出来，但它拒绝了 360 条被投毒记忆中的 0 条。我们认为，这暴露了仅基于内容的筛查边界：区分虚假断言和真实断言通常需要文本本身之外的外部依据。 随后，我们评估了按来源加权的检索。已发布的权重在统计上与没有防御无法区分（p=0.80），而更强的权重只有通过排除不受信任内容才能恢复效用。在一个混合来源语料库中，如果不受信任内容大多是良性的，准确率会从 0.3167 上升到 0.7000；当承载答案的证据本身来自不受信任来源时，证据召回率会降至零，准确率降至 0.0417。在测得的相似度条件下，加性来源项不存在可用的设置：足以抵御查询形态投毒的权重，也足以压制合法的不受信任证据。因此，我们主张在检索阶段采用有界占用约束，而不是加性来源惩罚，并发布相关测试框架、语料库和汇总运行报告。
Skills play different roles as an agent's policy evolves: they should first provide learnable knowledge, then support capability formation, and finally be invoked only when they improve individual decisions. Existing methods rarely model this lifecycle. They either keep skills outside the model, fully internalize them, or select among internalization and utilization objectives through noisy task-level success rates. Such designs fragment training and assign uniform importance to actions within the same trajectory, even though skill guidance may help some decisions while distracting others. To solve these problems, we introduce AUSO (Action-level Unified Skill Optimization), which unifies skill learning and skill use through a progressive, action-aware optimization process. At the beginning of training, AUSO jointly learns from teacher guidance and environmental outcomes, enabling the policy to acquire foundational skills without losing task-oriented feedback. It subsequently emphasizes outcome-based policy optimization to consolidate autonomous problem-solving ability. As the policy matures, AUSO evaluates each sampled action under both skill-conditioned and skill-free contexts. The resulting action-level information signal is coupled with the trajectory outcome advantage, allowing beneficial skill-sensitive actions to receive stronger updates and harmful ones to be suppressed. Therefore, skills gradually transition from an external source of supervision into decision knowledge whose utilization is adapted to its action-level benefit, while reinforcement learning remains the shared backbone across all stages. Experiments on ALFWorld, WebShop, and SearchQA show that AUSO consistently improves agent performance and out-of-distribution generalization over competitive baselines.
六十年来，机器验证一直是一项主要的成本负担，只有非同寻常的成果才负担得起。在这里，我们报告生成式 AI 颠倒了这种关系：在 AI 的速度下，机器验证不仅经济，而且对生产力至关重要——它是不可腐化的裁判，使一个人能够安全地大规模指挥自主机器工作。在五周内，一名研究人员使用面向消费者的 AI 订阅，指挥了一小队 AI agents，从应用代码，经由一个已验证的编译器和执行器，一直到在社区硅片 shuttle 上流片的 RISC-V 处理器；没有任何证明经过人工审查，也没有任何 RTL 由人类编写。这套工作规程——Salt method——建立在一个任何幻觉证明都无法通过的证明内核之上：数学主张以经内核检查的工件形式在 agents 之间传递，而人类注意力则保留给陈述、设计和裁决。验证逐环节陈述，从 Lean 4 内核到硅片边界处经 SAT 检查的等价性。我们公布完整账目：定理来源、预注册的 token 计量器、有下限约束的人类时间，以及错误台账，其捕获编号一直到 #256——这是数学攻关的只追加标记台账上的单调计数器，维护时间为 2026-07-07 至 2026-07-20（其中一个编号 #79 从未分配；后续捕获记录为未编号）——同时没有任何错误证明进入记录。
文章 URL： https://blog.modelcontextprotocol.io/posts/mcp-roadmap/ 评论 URL： https://news.ycombinator.com/item?id=49399591 积分：4 评论数：1
I think agent-first chat interfaces will be a primary software modality and busy dashboard/UI will go away. I’m not sure who exactly wins it, but I want my knowledge to grow/go with me. A lot of the “knowledge” ie research, analysis, reasoning will be done by agents as the primary user. Our current notes tools & tasks management systems were built for humans… I don’t care what the 17th thing on my bug backlog is. I want to conduct agents that can execute for me and do great work. What I built OzBrain to do: + Create a central place for agent reasoned knowledge to live + Be agnostic about what apps/agents connect to it + Capture everything and track it so I can audit it + Enable teams, collaborators or partners to share brains + Handle conflicts so many agents in the same article doesn’t blow up + Refactor knowledge into more token friendly chunks and map the index well + Close the knowledge loop so new thinking supersedes old thinking across the corpus. Don’t erase, depreciate and link + Keep user data safe and secure ++ Be easy enough to use that you don’t have to have any technical knowledge Some among us will always build their own custom solutions, but there are millions of tech professionals and small business owners that will use agents heavily and need a solution. So I’m trying to build that. Isn’t this like gBrain? Yes, similar. I think it’s like AWS vs Vercel. AWS is very powerful, configurable, and useful if you’re technical and want to invest the time into really fine tuning your system… but if you just want your web deploy/hosting to just work and be easy to deal with you use Vercel. // WHY I MADE IT I’ve been enjoying getting back to my technical roots, as I lost my coding skills more than a decade ago, but with AI I can focus on the system and the product in partnership with agent coding workflows. I recently built a Voice AI for older people. To build it I created an agentic engineering workflow (feel free to rip that up as I’m always looking to improve systems: https://ozbrain.com/resources/eng-flow ) My approach with coding agents is trust but verify, and I’m trying to replace the parts where a human would review with an adversarial or specialized agent who would give a better answer/review. I have workflows that will go high level task to shipped PR running in Claude cloud sessions. I use Claude Code locally and Cursor when I want a tighter loop on doing visual work like UI or layout. And Codex to either load balance usage for TokenThriffting or when I want a different llm to think thru something. It was a pain in the ass passing .md files around and keep track of which version was the most recent, so I built a hosted .md storage right in Supabase and any of my agents already have Supabase access. This let me build a solid, scalable, secure voice AI from my phone at the gym. All my agents have access to our knowledge, can write to it, update and refer to it as we build and improve the product and the systems we use. Out of 75 founder friends I asked about how they manage shared knowledge, 26 built their own custom knowledge systems… Obsidian vaults with 7k files synced through a VPS, markdown repos behind their own MCP servers, cron jobs stitching Supabase to a skills file… each a different Frankenstein they have to maintain. 32 said they felt the pain of moving static files around but didn’t have any solution for it. So I rebuilt my brain better and used it to build it. // HOW YOU CAN HELP Would love to have you try it out. The maintenance loop is still in alpha so not running it on customer data yet. If you built your own brain I’d love to hear how you did it. What criteria was most important for you in its design & function. If you are tired of shuffling .md files around I’d love to have you try out OzBrain and to give feedback, just ask your agent to put it in the shared bugs & features brain! Cheers! Bubs.co Comments URL: https://news.ycombinator.com/item?id=49394827 Points: 8 # Comments: 1
Hi HN- 我是 Pablo，Proliferate（YC S25）的创始人！ Proliferate ( https://github.com/proliferate-ai/proliferate ) 是一个开源、可自托管的 AI IDE，让你可以在一个地方使用 Claude Code、Codex、OpenCode、Cursor 和 Grok 进行工作并自动化任务。 这里有一个 2 分钟的快速演示，展示我们如何用 Proliferate 来构建 Proliferate： https://www.youtube.com/watch?v=tGNX0oaWmBY 在我的团队接入 OpenAI Codex 之后，我开始着手构建 Proliferate。几天之内，我们就把它用到了方方面面：用 computer use 代替我们自己浏览网站，让 Codex 协调其他 agent，以及为重复性工作设置自动化。说实话，我们几乎从不需要离开桌面应用就能把工作完成。 如果我团队的体验具有一定代表性，那么一个类似 Codex 的应用（一个具有精美 UI 的横向 agent）将会成为每家公司完成工作的主要界面。这完全符合 OpenAI 的使命，即把 Codex 打造成 everything app（见： https://news.ycombinator.com/item?id=47796469 ）。 但随着我们开始把自动化推进到更接近业务核心的工作，我希望能与来自所有实验室的 agent 协作，包括开放权重模型，而不会越来越依赖 OpenAI。 而这正是 Proliferate 的用途！它是开源、可自托管的 Codex，既能在构建向 Codex 广度迈进的同时，保留你在不同 agent 和模型提供方之间的选择自由。 如今 Proliferate 支持： * 使用 Claude Code、Codex、OpenCode、Cursor 和 Grok，并可采用它们原生的推理与配置选项，包括 Bedrock、Azure 和自托管推理。 * agent 之间的通信与管理：父 agent 可以生成并与另一个受支持的 agent 作为子 agent 进行通信（我个人喜欢让 Fable 委派给 Codex，同时让 OpenCode 模型审查 PR）。 * 构建 workflows——这是我最兴奋的功能之一。这些是可复用的 agent 会话链和人工审批关卡，每一步都可选择 harness 和模型，并在它们之间传递文档。我用它来自动化代码审查、QA，以及我的 PR 构建流程。 Proliferate 的全部代码在 AGPL-3.0 下 100% 开源。现在肯定还有一些粗糙的地方，但我们推进得很快，也非常欢迎任何反馈！ Comments URL: https://news.ycombinator.com/item?id=49390739 Points: 11 # Comments: 5
文章链接： https://github.com/alibaba/anolisa/blob/main/docs/user-guide/en/agent-observability/agentsight.md 评论链接： https://news.ycombinator.com/item?id=49389493 积分：11 # 评论：0
It is a fork of https://github.com/DenisSergeevitch/desktop-fly , but with an important update. Now the fly can pick up the scent of the codebase with its neurons and fly straight to the source code of your B2B AI SaaS startup. It has learned to scan its surroundings for agent markers: AGENTS.md, CLAUDE.md, .cursor/rules, .kiro/steering, and forty others. Anything on the screen that points to these markers becomes a source of the scent - an editor window with an open project, a line in Finder, or a desktop icon. An open project reeks the strongest, while a closed icon or nested folders give off a fainter odor. Comments URL: https://news.ycombinator.com/item?id=49389466 Points: 3 # Comments: 0
We present a novel approach to efficient LLM agent harness optimization through adaptive validation task selection. Harness optimization iteratively rewrites the harness code based on validation performance, enabling substantial performance gains without updating the underlying model weights. Existing approaches, however, evaluate a fixed validation set in full at every iteration, incurring substantial evaluation costs even on tasks that become less discriminative as the harness evolves. We propose $\textbf{Task-CoEvolve}$, which co-evolves the validation tasks with the harness by addressing two challenges: selecting informative tasks and estimating full-set performance from partial evaluations. Task-CoEvolve builds on the observation that tasks on which candidate harnesses disagree are more informative for distinguishing among them than tasks that are consistently solved or failed. It uses variance-weighted sampling based on past outcomes to focus evaluation on tasks near the agent's capability frontier, with the sampling distribution adapting as the harness evolves. It then estimates full-set scores from the sampled tasks by accounting for their sampling probabilities, enabling consistent comparisons across iterations despite evaluating different subsets. Experiments on online text classification and Terminal-Bench 2.1 show that Task-CoEvolve consistently outperforms fixed-subset baselines and matches the final performance of full-set search while reducing the number of evaluations during optimization by 80%. Code will be released at https://github.com/Agent4Science-UTokyo/Task-CoEvolve.
技术文档是为人类开发者编写的，但如今越来越多的软件变更由自主编码 Agent 完成。它们会查阅哪些文档、何时查阅以及随后会发生什么，仍然未知。我们基于行为，对 Agent 与文档的交互进行了研究，涵盖两个公开数据集：来自 SWE-chat 的 557 个 Agentic 编码会话，产生了 94,813 个开发事件，其中包括 3,033 次文档交互；以及来自 AIDev 的 33,097 个 Agentic pull request，包含 690,260 条经分类的文件级变更记录。四项发现对当前文档实践提出了挑战。首先，Agent 的文档工作主要由面向 Agent 的工件主导：指令文件和工作笔记占所有文档交互的 60.5%，而传统技术文档占 10.6%，API 参考占 1.3%。其次，查阅与代码编辑之间的关联尚无定论：相邻转移概率为 0.002，未调整的三事件提升度为 1.05，而按阶段调整的模型将其置于大于 1 的水平（OR 1.33 [1.09, 1.62]）；文档创建在未调整情况下有所升高（提升度 1.67），但其调整后的区间包含 1。第三，未观察到明确的基于文档的验证序列，并且查阅与较少的即时测试相关（提升度 0.23，聚类 CI 0.08-0.45；调整后 OR 0.39 [0.25, 0.60]）。第四，查阅更多是自发启动的（70.2%），远多于由失败驱动的（7.5%），并且文档滞后于代码：在同时更改二者的多提交 pull request 中，代码先被触及的频率高出 4.7 倍。基于这些轨迹，我们推导出一个 Agent-文档交互的描述性模型，将其视为一个双叶循环，而非线性旅程，并表明“Agent 友好型”文档的两个被广泛假定的属性——可操作性和可验证性——缺乏一致的行为支持。我们发布了我们的流水线、编码方案和事件级数据。
软件形态自诞生以来经历了两次范式转变：Software 1.0 中，指令决定行为；Software 2.0 中，数据决定行为（机器学习）。本文认为，第三次转变——Software 3.0，即由上下文和推理决定行为——正在发生，并主张其终极形态将收敛为三个要素：广义数据库（所有持久状态和记忆的统一抽象）、大模型（执行推理与生成的智能核心）以及智能体（连接前两者的执行循环）。核心论点如下：在传统三层架构中，用户界面层将被模型按需生成界面的能力所吸收；业务逻辑层将沿着“可表达性 × 关键性”被重新划分为模型推理与存储约束（剩余的确定性逻辑则作为工具保留）；只有数据层将被提升为唯一的持久化基础设施。我们对这一收敛论题进行形式化阐述，提出一个最小参考架构，报告来自真实原型和在线模型的证据，并系统分析其成立条件及失效边界——确定性、成本、安全性和可验证性限定了该论题的适用范围。我们认为，该论题适用于具备可表达、可验证、外部有状态且工具完备特征的任务领域，并将重塑开发者、数据库行业和软件工程学科的角色。
大语言模型（LLM）智能体可以从已完成的任务中归纳出技能，并在之后复用这些技能，从而随着经验积累变得更有能力。在实践中，归纳出的技能可能迁移不可靠，甚至会损害检索到这些技能的智能体。智能体归纳的技能何时能够在任务之间可靠迁移，仍是一个开放问题。我们对技能归纳方式如何塑造其跨任务迁移进行了全面且受控的研究。具体而言，我们比较了任务级与子任务级的技能归纳，以及文本与代码两种技能格式，这也是现有方法存在差异的两个维度。任务级技能大多会使智能体的性能低于其无记忆基线，而子任务级技能平均而言会使其性能高于该基线；文本技能的迁移效果优于代码技能。为进一步理解我们的发现，我们考察了归纳技能的两个互补属性：特异性，用于衡量技能与真实任务的匹配程度；抽象性，用于衡量其相关性在各任务之间分布的均匀程度。单独任一属性都无法预测任务成功，但二者的联合效应可以预测；我们据此提出了一种技能效用分数。当技能被迁移时，该分数与任务成功始终保持相关，而子任务级技能和文本技能得分更高。计算技能效用只需要技能和任务描述，不需要任何任务执行，因此我们的分数可作为在运行任何新任务之前诊断技能记忆的实用指标。
中期训练日益被认为是塑造大型语言模型能力的关键阶段。近期研究表明，有针对性的中期训练可以增强数学和科学等推理密集型能力，也可以提升软件工程场景中的智能体式能力。在这项工作中，我们研究了一种与之并行但探索较少的智能体式能力：通用工具使用。我们提出 MidTool，这是一个面向智能体式工具使用中期训练的开放语料构建流水线，将大规模网页、PDF 和代码数据与来自真实世界工具 API、MCP 技能以及基于文档的工作流的合成监督相结合。MidTool 旨在教会模型如何识别工具可供性、从上下文中定位参数、组合工具调用工作流，并从信息不完整的情况下恢复。我们在 MidTool-Mix 上对 Qwen3-4B-Base 和 Qwen3-8B-Base 进行中期训练，然后应用后续后训练，包括监督微调和强化学习。与基线相比，MidTool-Mix 在 BFCL、tau2-Bench 和 MCP Universe 上，在 SFT 和 RL 两种设置下都能持续提升下游性能。这些结果表明，通用工具使用与其他重要的 LLM 能力一样，受益于专门的中期训练，而不是完全留给后训练。
递归自我改进（RSI）所问的是：一个 AI 系统能否改进产生 AI 系统的过程，从而让下一个系统继承这种改进。这个过程就是训练算法：更好的目标函数或更新规则会提高每次后续运行中的 compute\mbox{-}capability 交换率，包括产生下一个 agent 的那次运行。因此，RSI 是否可行取决于 agent 能否设计训练算法。没有任何基准能隔离这种能力：现有测试套件的胜出方式是收集数据或调优超参数，而且没有一个能够区分对一次运行执行方式的改变与对模型学习方式的改变。我们提出 AI4AI\mbox{-}Bench，它包含 10 个冻结的研究代码库，覆盖 10 个训练算法家族。在每项任务中，agent 有 4 小时在一块 B300 上重写训练算法；随后，其代码会从头重新运行最多 12 小时，并由一个对 agent 隐藏的固定评估器打分，与该代码库原始算法在相同流程下的结果进行比较。由于这 10 个指标不可通约，每项任务都会映射到同一个尺度上，其中 $0$ 表示无信息模型，$0.1$ 表示代码库自带的算法，$1.0$ 表示任务最优值。在 6 个系统的 29 种配置、覆盖全部 10 项任务的实验中，平均得分为 $0.166$，最佳系统达到 $0.250$：即使最强系统，也只弥合了既有算法与最优值之间不到五分之一的差距。提交结果显示了差距的去向：大多数根本没有改变模型的学习方式，而少数确实改变了学习方式的提交，平均得分为 $0.226$，其余提交为 $0.126$。更多推理投入主要带来的是尝试这类改变的意愿，使这类少数提交的占比从 $8\%$ 提升到 $64\%$，平均得分从 $0.094$ 提升到 $0.196$。我们发布任务套件、评估器以及每一个已评分提交，以便随着这些系统的变化重复进行测量。
Travel behavior research increasingly combines digital data collection with predictive modeling, yet these stages are often developed and evaluated separately. This study proposes a three-agent workflow integrating conversational data collection, structured data processing, and behavioral prediction. A chatbot-administered, image-augmented stated-preference survey collected mode choices from student commuters across five predefined weather scenarios, yielding 454 respondent-scenario observations. Weather-related associations were analyzed using a multinomial logit model, while logistic regression and random forest provided machine-learning benchmarks. Nine locally deployed large language models (LLMs), ranging from 2 to 35 billion parameters, were evaluated across four zero-shot prompt-and-context conditions and extended through persona, few-shot, and vision-based configurations. Random forest achieved 69.6% five-class accuracy, while the best text-only zero-shot LLM reached 69.9% without task-specific fitting. Habitual travel information produced the most consistent gains, Expert framing generally outperformed Role-Play, and persona information was most useful when habitual travel information was unavailable. Few-shot prompting improved prediction for several models, with gains stabilizing after a small number of examples. Using the same weather images shown to respondents, the best vision-based configuration reached 71.5% five-class accuracy, indicating that visual context may provide additional predictive information for selected models. Overall, the study shows how conversational surveys, structured data processing, conventional behavioral modeling, machine learning, and multimodal LLM prediction can be coordinated within an auditable multi-agent workflow.
Kandelo is an open-source, Wasm-based multi-process kernel that runs POSIX programs in browsers and Node.js. Kandelo is still experimental, but it already runs a substantial range of existing software. Do you have use cases for this? We are trying Kandelo as a new foundation for WordPress Playground which runs server-side WordPress entirely in the browser. Kandelo also looks promising as a sandbox for running agents in the the browser and on the command line. On the side, we've been playing with porting games and desktop environments and even compiling runnable programs within Kandelo. Yet it feels like there are many possibilities we haven't considered. How would you like to use something like this? Demos: Some notes: The demos have been tested in desktop browsers. Unfortunately, YMMV on mobile today. Some of the disk images are large (~50MB) and may take a while to boot initially. Main set, with Shell (bash, vim, nethack, and more), Nginx, PHP, WordPress, and Doom: https://kandelo.dev/20260819-demo/ LÖVE game engine: https://kandelo.dev/20260819-demo-love/ SNKRX running under LÖVE: https://kandelo.dev/20260819-demo-love/?vfs=love-snkrx-abi44... Commander Keen running in DOSBox: https://kandelo.dev/20260819-demo-dos/?demo=keen LXDE desktop PoC: https://kandelo.dev/20260819-demo-lxde/?demo=desktop-lxde Background I wanted an authentic OS-level foundation for running systems software in the browser and started this as a vibe-coded exploration. I figured it would end up being too slow and that we would have to offer many different ways to compromise default POSIX behavior to get anything usable. But after weeks of fighting agents, insisting on genuine POSIX compatibility as the default, I was surprised at how well the system worked without those compromises. Nginx, PHP, Python, Ruby, Redis, and even MariaDB were able to be built using the SDK with minimal hacks. Then we started porting games, having fun, and playing to see how far we could push it. Notes on architecture: There is a central, single-worker kernel, aiming to provide all supportable POSIX syscalls. Each process is a dedicated worker with independent memory. Each process thread is a dedicated worker that shares memory with threads from the same process. Syscalls are done with the process SharedArrayBuffer and the Atomics API. fork() is supported. The system is centered around virtual file system (VFS) images, and the VFS can contain lazy references to programs that may or may not be used. Vim is such a reference in the shell demo. On GitHub: https://github.com/Automattic/kandelo Comments URL: https://news.ycombinator.com/item?id=49378305 Points: 6 # Comments: 2
Hello everyone. I've been working on this experimental editor called Huzzah. I've been working almost exclusively with coding agents since January of this year, and over the past few months I began to feel utterly exhausted by them. They're great, but I'm finding it more and more tedious to write full sentences for every change I want. Not only that, but it seems there's a complexity limit for codebases - beyond a certain point the agent begins confusing itself. I'd like to go back to writing code, but I don't want to go all the way back to fully manual coding. So I've come up with this interaction paradigm where you: 1. write pseudocode in whatever way makes the most sense to you 2. on save, the editor synchronizes your work to real source code 3. the pseudocode is persisted alongside the generated code, making your prompt effectively a stored record of intent. It may not work for every use case, but in my initial playthroughs I've found it very enjoyable. Right now it's just a proof of concept - installation instructions are here in the readme: https://github.com/danielvaughn/hz You can also watch a video of it in action here: https://x.com/danielvaughn/status/2090456808431165715 Cheers! Comments URL: https://news.ycombinator.com/item?id=49378768 Points: 3 # Comments: 0
Hi HN, we’re Yousef & Nour, founders of Vendo ( https://vendo.run ). Vendo lets users create new features inside the software they already use. A user describes the dashboard, workflow, or small app they need, and Vendo builds it on top of the product’s existing data, API, and interface. Demo: https://www.youtube.com/watch?v=VdpHehY64ls We built Vendo because every SaaS eventually faces the same problem: every customer needs something slightly different. One wants a new report and another needs a workflow that only makes sense for their team. These requests either sit on the roadmap, become one-off engineering work, or force the customer into spreadsheets and external tools. We wanted the user to be able to create the missing feature themselves, without leaving the product. Here is how it works: - npx vendo init reads the product's API surface, theme, routes, and more. These are used so that the apps Vendo creates (1) look on-brand and native and (2) have the ability to read data and perform actions directly through the company's API - When a user asks for a feature, we have a custom Vendo harness that writes a React component with a bunch of Vendo add-ons and guardrails (ex. ability to make calls to the host API + our component library). Every save is compiled, type-checked, run against real API responses, and rendered before the user sees it. We just released a benchmark and write-up here with more info for anyone interested: https://vendo.run/blog/generating-product-ui-measured - We use QuickJS to make sure that anything the agent creates is sandboxed and can't mess with the company's site. Vendo compiles the component and runs it with Preact inside a QuickJS VM with no access to the DOM, network, or clock. The VM returns a UI tree, which the host renders using the product’s registered components. When the user clicks something, QuickJS emits a tool call; the host executes it through Vendo’s guard and passes the result back into the same VM, preserving the screen’s local state. There's a lot of generative UI right now: streaming developer-written components into a chat (Vercel AI SDK, CopilotKit, Thesys), or rendering your app inside someone else's assistant (OpenAI Apps SDK, MCP Apps). We differ on two things. Vendo lives in your product and acts through your API as the signed-in user, so what it makes is durable: real apps users keep, pin, and run on triggers while they're away, and not components that are merely confined to a chat. Plus, it's not capped at putting together a bunch of prebuilt components: the agent can build arbitrary apps, from a quick dashboard out of your own components to real custom code running in a sandbox, and either way data only ever comes from tool calls to your API. Here are some things customers are using Vendo for today: - Letting their users create custom dashboards and reports. These are mainly UI-based and focused on letting the user see the exact graphs and metrics they care about - Letting their customers create recurring automations. A big thing as well that has been used for these automations is the fact that we connect to external connections, so users have been automating many of their inter-tool workflows (ex. an automation that sends a slack alert based off of something in the product) - B2B customers letting their customers customize the product with specific business logic. Often this is simple things like an extra field on a form, or an extra permission, but it is hard for a business to keep up with them otherwise. - Creating and sharing custom dashboards/apps across an organization. Since the apps Vendo creates are durable, they can be shared, reused, and forked (which can’t be done with many of the other in-chat generative UI solutions) We've spent a lot of time thinking about how AI and agents will change the way people consume software. We think the answer is personal(ized) software: you see the UI you need to see, you tell an agent exactly what you need, and the product molds to how you work. The key insights that have enabled the product to work are: - A rule in code always beats a rule in a prompt. - Invent as little syntax as possible. Generation got faster and more reliable when the output looked like what models already know (JSX-shaped markup) instead of a clever custom format. - Deterministic beats model wherever you can get away with it. Theme extraction is pure static analysis, and a remix starts as a copy of your component, no model call. Vendo is completely open-source (Apache-2.0) and can be self-hosted, so feel free to check out all the source code here: https://github.com/runvendo/vendo . Would love you to try it out and give us your feedback: https://docs.vendo.run/ . Or if you’re a company looking to embed Vendo in your product feel free to book a call here: https://cal.com/team/vendo/intro-call Comments URL: https://news.ycombinator.com/item?id=49376038 Points: 6 # Comments: 0
Debates have recently emerged as a useful methodology for agentic AI to improve performance as well as to aid explainability and user engagement. For example, LLM-empowered agents may debate internally (with themselves) and/or externally (with other agents). In many settings where debates are used, debates' outcomes and resulting outputs are determined post-hoc by external judges, often LLMs. In this paper we develop and test a novel theory of debate judgement applicable to all settings where agents engage in debates by providing pros and cons for their opinions therein. Specifically, we identify a number of formal properties that debate judgement may be required to satisfy in general, as concerns reproducibility, robustness, groundedness and explainability. Then, we explore their satisfaction formally and/or experimentally, for claim verification settings, for two specific alternative debate judgement methods: variants of the LLMs as a judge idea and formal semantics drawn from computational argumentation. We show that the two methods give similar accuracy performances but the former may lack formal guarantees that the latter brings. Overall, our study indicates argumentation semantics as an ideal candidate for principled judges in debate-driven AI.
Accurately extracting nuanced, contextualized data from research articles is laborious and time intensive. Here, we investigate the performance of frontier, browser-based large language models (LLMs) to extract highly contextualized information. We demonstrate four escalating workflows, 1) given an expert curated prompt and research articles, most frontier LLMs perform well at data extraction, however can struggle with interpreting scientific context and nuance, 2) given simple instructions, LLMs can author their own prompts which were almost as eNective as expert-written prompts, 3) autonomous discovery of research literature was diNicult, agents either missed or hallucinated references, and 4) LLMs can create new datasets from published guidelines that closely match human-expert judges, but still require a human-in-the-loop. Together, these findings define an auditable division of labour in which experts specify the evidence standard, models cross-check repeated extractions and researchers resolve disputed cases, providing a practical route to scaling scientific data curation without relinquishing expert oversight.
Accurate and responsible medical question answering (QA) is important in healthcare, where complex cases require factual knowledge and nuanced reasoning. Existing medical QA systems, typically based on single-agent architectures and static retrieval, often lack adaptability, persistent memory, and structured decision-making. This work introduces an adaptive memory and reflection (AMR) agentic system, a multi-agent framework in which specialized agents use dedicated memory and reflection-based feedback to retrieve relevant prior cases and improve subsequent reasoning. Complexity assessment routes questions through solo, collaborative, or escalated workflows, while consensus and ethical overseer modules support reasoning consolidation and output review. Evaluation on MedQA and MedMCQA demonstrates strong performance compared with several baselines. Ablation studies show that combining agent-specific memory, reflection, and external retrieval yields the strongest performance. These findings highlight the potential of structured memory and feedback for developing more trustworthy medical agents. The source code is publicly available at https://github.com/mm-air/AMR-Agent.
We present Eureka, a task-conditioned Meta-Agent architecture that compiles long-horizon tasks into dynamic obligation graphs with explicit acceptance semantics. During execution, Eureka forms Macro-Agents with specialized state, memory, operators, tools, verifiers, and local topology via receding-horizon planning, architecture promotion, and minimal-sufficient compilation. When bottlenecks recur, cost-benefit-gated evolution updates the local architecture under constraints. Theoretically, we establish results on regret, planning invalidation, amortization, subtree interfaces, serializability, and verification. Experimentally, Eureka completes 170/170 recursive tasks and generates 3,948 certificates with no false acceptances. Active context compresses median input from 9,490 to 4,005 tokens; incremental processing avoids 65.38% recomputation across 12,000 tasks; 16,000 concurrent executions serialize consistently. The same Meta-Agent instantiates a Theory-Discovery Agent and a Math/Conjecture Agent. The former yields structural results in quantum-process and spacetime theory. The latter identifies bottlenecks in Riemann Hypothesis research and advances a positivity certificate for Suzuki's localized Weil quadratic form to 0 < a <= 69/200 = 0.345, reaching ~99.55% of (log 2)/2. These results suggest that scientific-agent capability depends not only on the base model but on whether an architecture can be formed to match the task's cognitive structure.
An agent still learning its environment should be cautious while ignorant and bold once confident. The entropic value-at-risk captures this through a robust-optimization identity---a confidence level fixes the radius of a relative-entropy ball of alternative models---but that ball cannot reach catastrophes the nominal deems impossible, precisely what a safe agent must hedge. We instead use an optimal-transport ball and study the coherent risk measure it induces, the Wasserstein entropic value-at-risk. It has a variational dual mirroring the entropic formula (an inverse temperature becomes a transport price), occupies a definite place in the risk hierarchy, and provably accounts for the reachable catastrophes the entropic measure ignores; we verify both dualities numerically. Driving the transport radius by belief entropy then yields a closed-form robust dynamic-programming operator whose caution contracts as the belief sharpens, with a certified safety sandwich and a sharp safety switch.
语言模型智能体可以通过公开转录文本中不可见的连续隐藏状态进行通信，从而为隐蔽的有害协调创造机会。我们提出 Verifiable Latent Alignments (VLA)，这是一个感知激活的框架，用于监控和引导这些私有通信通道。对于每一个受监控的决策，VLA 使用共享事件标识符将私有潜在状态记录和通道状态与由此产生的公开行动关联起来，从而实现匹配的因果分析。我们的第一项贡献是一个仅基于中性样本的三层监控器，结合了表示异常检测、反事实动作分布影响以及稀疏自编码器解释支持。我们的第二项贡献是一个可引导性框架，涵盖黑盒行为指令和白盒匹配中性反事实。我们的第三项贡献是在受控多智能体拍卖基准上的评估，覆盖同质和异质模型对、多智能体可扩展性以及干预有效性。当将文本串通和潜在串通行合并为正例时，顺序监控器在同质智能体上的平均受试者工作特征曲线下面积 (AUROC) 达到 0.993，在异质模型对上达到 0.854。在包含 25–100 名竞标者的 Qwen3-0.6B 拍卖中，相对于所有可能的有向配对，监控仅需要很小的归一化负载，而完整白盒引导实现了 100% 的出价分布恢复，并将串通性低价出价行为减少了 47.3 个百分点。由于完整白盒引导会重放匹配的中性反事实，其精确恢复在构造上就是一种合理性检查。总体而言，这项受控研究表明，所评估的私有通道攻击可以在不使用攻击样例训练主监控器的情况下被监控，并且在可获得匹配反事实访问时可以被缓解。
持续的自我改进需要一个不断扩展的、由自身生成的、多样且自适应的目标池。对于语言智能体而言，现有的训练环境池（人工策划、静态合成或冻结验证器）会在学习者规模扩展时保持目标分布固定。我们提出 SPADE（Self-Play in Adaptive Synthetic Executable Environments），这是一个自博弈 RL 框架，其中单个 LLM 扮演两个角色：Environment Designer，负责将完整的长时程训练环境编写为带有 OpenAI Gym 风格 reset()/step() 接口的可执行代码；以及 Reasoning Agent，负责学习在这些环境中行动。每个环境都是有状态的多轮环境（包含状态转移、奖励函数和验证代码），因此一个接口即可涵盖推理问题和多步骤智能体式工具使用。Reasoning Agent 的遗憾值通过其在有无特权提示情况下的奖励差距来估计；在优化这一遗憾信号的过程中，Environment Designer 学会将环境定位在智能体能力边缘，同时保持其可行性。通过大量实验，我们发现若干组件对成功至关重要：基于从大型预训练语料库中采样的文档来为 Environment Designer 提供依据，并赋予其累积的环境记忆。扩展到 30B 参数模型时，SPADE 在八个保留的数学、科学、代码和推理基准上平均比最强固定环境基线提高 +5.3；在工具使用设置中，在 BFCL-v4 多轮上提升 +5.7，在 ACEBench-Agent 上提升 +13.9；在游戏设置中，相比最强基线的优势会随模型规模增长而扩大。通过使环境设计本身成为一个可学习组件，SPADE 朝着开放式自我改进迈出了具体一步。
I wanted to share a project I’ve been working on called Frugal Tokens. I originally built it because I was curious to see how much all of my sessions cost and how much cache misses affected that spend. I’d noticed people had widely different spend profiles and wanted to better understand what might contribute to that. As I’ve worked on this, the tool has grown to show more usage patterns across all of your sessions. It shows overall usage, estimated working time and overlapping sessions, and where your spend is coming from across models and cache misses. I also have a few session level metrics with percentile breakdowns, along with a list of your sessions and high level info. Clicking into a session opens an explorer where you can see individual model calls and tool inputs and outputs. You can also jump directly to where a cache miss happened. There’s also a rough cost comparison that shows what the recorded session would have cost with another model’s pricing, or for Anthropic, with 5m vs 1h caching. In the future, I’d love to collect more information to see which patterns might make people’s workflows more expensive, e.g. long sessions, high context usage, many turns, etc. The tool requires deno, but is just one command to run once that is installed. The demo provided has some of the data scrubbed, but helps to show what it looks like before running it. Would appreciate any thoughts or ideas https://github.com/dpclark4/frugal-tokens Comments URL: https://news.ycombinator.com/item?id=49364223 Points: 10 # Comments: 1
Hi HN，我们是来自 OneCLI 的 Jonathan 和 Guy。OneCLI 是一个为团队构建的 agent harness，为每位员工提供一个安全的、沙盒化的个人 agent。 你可以用它做这些事： 1. 获得一个沙盒化的 agent，并具备所有 OneCLI 能力，比如只需在聊天中连接你的 GitHub 账户、Gmail、Notion 或 Dropbox。 2. 在聊天本身中进行确定性的 human in the loop 审批，用于那些你需要 100% 控制的事情，比如发送电子邮件或删除 Linear 工单。 3. 在一个地方管理团队策略，并在工作区中的每个 agent 上强制执行。 4. 在团队层面享受全局连接，比如共享 LLM keys 或服务账户。 这里是演示： https://www.youtube.com/watch?v=dlW-44ntpbE 我们是偶然开始做这个的，尽管我们的职业背景都在安全领域。当时我们在做一个名为 ChartDB 的 devtool，这是一个开源 DB 工具。今年 1 月 OpenClaw 火起来后，我们开始用它在 ChartDB 之上编排 agent。我们很快意识到 auth 方面存在一个大问题。agent 需要凭证才能做真正的工作，但把这些 secrets 交给它们并不是最好的主意。它们会把这些 secrets 保存在自己的记忆中，还会以纯文本写入本地文件和会话中。我们也知道 agent 很容易被诱骗交出这些 API keys/secrets。因此，我们需要某种方式来控制 agent，阻止 prompt injection 欺骗它使用自身服务为攻击者谋利。 我们创建了 OneCLI，它最初是一个用 Rust 构建的 AI Agents vault。 我们发现，对 OneCLI 的大部分需求来自 Hermes、OpenClaw 和 NanoClaw 这类面向个人和团队的 autonomous agents。 用户在寻找有用的 agent，让它们为运行它们的人做事，但缺少两个部分：1）管理 secrets 和权限。2）对于团队来说，还需要多人管理。 我们决定转型，把 agent 本身作为面向团队的 harness 提供，为每位员工提供一个 agent。我们看到团队不得不一遍又一遍地搭建自己的 harness，而我们基本上已经把 vault 做成了一个 gateway。于是我们想到，可以开箱即用地提供 agent 管理中缺失的那一块，并将其开源（Apache-2.0，带有一个小型企业例外条款）。 我们坚持开源优先——开源整个平台，而不只是像其他 agent 那样开源其中一小部分，这样公司才能真正看到代码、评估它、信任它，而不是只听我们的一面之词。它们可以在自己的环境中隔离运行，完全自托管或自控，具备生产级质量，而不是某个托管在别处的封闭黑盒。这意味着安全不只是一句承诺，而是它们可以自行验证的东西。结合真正的自治能力和最小权限访问，这使它成为公司可以完全拥有和信任的东西，而不只是采用。 我们也从公司视角而不是个人视角来处理这个问题。我们的解决方案代表每位员工管理 agent，并包裹在由公司管理员通过集中式策略配置的确定性 guardrails 中。 对于 agent engine 本身，我们使用 jcode，它是 agent-loop 的核心。我们发现它改善了体验，让 agent 更聪明、更快速。 它的工作方式如下： 它运行在你控制的基础设施上。完全开源，几分钟内即可 self-host 或部署到 cloud。 agent 永远不会持有真实 secret。它拿到的是一个 placeholder。真实凭证会在 gateway 处、每次请求时、调用被授权后注入。它永远不会进入 agent 的 context、memory 或 logs。 在模型之外执行控制。Prompts 只是建议。由组织管理员定义的策略运行在网络层，在 agent 和 LLM 之外。阻止 endpoints、按 agent 限速、要求审批、按员工限定 scope。由 gateway 决策。agent 无法绕过它。 每个 agent 都有隔离 VM。自己的 memory、自己的 keys、自己的权限。爆炸半径仅限于一个 agent。 Harness 的速度：agent loop 之下的 Rust engine。 完整的身份轨迹。每个 agent 都绑定到一名员工。每次调用都会记录它代表谁执行，以及是哪条策略允许了它。 人们正在用这个平台做的一些事情包括： - 完整管理公司的生命周期，从销售电话开始，到产品侧自动为工程团队开工单，再触发 development agents 交付并发布到生产环境。 - 运营侧，比如在通话后自动清理 CRM、寻找 leads、预订会议并管理后续跟进邮件。 - 我们的一些客户还用这些 agent 完成整个杂货采购，并让它们处理各种家务，比如在线下单。 关于团队：两位创始人都来自网络安全背景。Jonathan 曾在 Axis Security 多年从事 zero trust network access。核心理念是你永远不要信任客户端。你要精确决定一个人能访问什么，并在他们之外、在网络层强制执行，因此客户端试图做什么并不重要。这就是如今每家严肃公司向人类授予访问权限的方式。Guy 是 Argon security 的第 1 位员工，负责 AppSec。 我们很想听听你们对这次转型的看法，也欢迎提交 issues 来帮助改进，让你的 agent 既强大又安全——为团队而设计，而不仅仅是为个人。 Comments URL： https://news.ycombinator.com/item?id=49363710 Points：3 # Comments：0
Large language models are pushing data science toward increasingly autonomous and agentic workflows, with recent systems already supporting multi-step and long-running analyses. As these workflows become more autonomous, conventional interfaces no longer provide adequate support for two critical requirements: observability for understanding an agent's evolving reasoning and evidence, and steerability for redirecting low-value directions or deepening promising ones during execution. Existing interactive approaches improve process visibility and open intervention points, but they remain largely designed for discrete, turn-by-turn exchanges rather than the parallel branches and evolving decision structures of long-running agentic analysis. We study this need as interactive oversight in long-running agentic data analysis and present AdaLens, an interactive system for monitoring and steering ongoing runs. AdaLens combines a storyline-based representation that unifies analytical plans, execution progress, intermediate findings, and data-column involvement with steering interactions grounded in these analytical elements for directional guidance and execution control. We evaluate AdaLens through two case studies and a user study, examining how it supports analysts in monitoring and steering long-running agentic data analysis.
Autonomous research systems are increasingly capable of executing long research workflows, yet automation alone does not ensure that the resulting process remains scientifically grounded. We introduce AutoResearch, a two-stage system that connects Idea Generation with Idea Execution to address both how research ideas are formed and how they are reliably established through experimentation. In Idea Generation, AutoResearch continuously integrates emerging research signals with accumulated domain knowledge, identifies transferable mechanistic insights, and uses multi-model generation and cross-review to produce grounded, testable research plans. In Idea Execution, coordinated agents decompose these plans into experiments, iteratively implement and diagnose them, and employ independent evidence-based review before accepting research conclusions. Across representative settings in cross-modal retrieval, systems optimization, and benchmark-driven machine learning, AutoResearch turns generated ideas into measurable progress, detects and corrects unreliable experimental results, and makes evidence-conditioned decisions to continue, revise, or terminate research directions. For example, on RSICD benchmark, an AutoResearch-generated idea improves mean Recall from 32.84 to 34.69, while recording only 5 audit-confirmed issue events compared with 11-27 for other autonomous research systems. These results demonstrate a research process in which meaningful insight is grounded before experimentation and conclusions are grounded before acceptance: Insight In, Hallucination Out.
In the Lifelong Multi-Agent Path Finding (L-MAPF) problem, agents must repeatedly move from one destination to another while avoiding obstacles and inter-agent collisions. Widely regarded as one of the highest-performing solutions to this problem is the Rolling-Horizon Collision Resolution (RHCR) framework. However, commensurate with its quality solutions, it incurs a computational cost that limits its applicability to even modest agent counts. In this paper, leveraging theoretical methods from the Locally Interdependent Multi-Agent MDP literature, we first theoretically prove the near-optimality of RHCR in a discounted MDP formulation of the L-MAPF problem. Then, we leverage these results to naturally motivate an extended framework called Group Decentralized RHCR (GD-RHCR) which incorporates a group decentralized structure that partitions agents based on a transitive communication scheme and plans for each partition of agents in parallel. We show that both RHCR and GD-RHCR achieve similar exponentially close to optimal guarantees, establishing a theoretical duality between the time based restrictions performed by vanilla RHCR and the additional space based partitioning performed by GD-RHCR. Lastly, we show that across varying maps, GD-RHCR is able to attain high throughput that scales into higher agent counts while maintaining a significantly lower per plan cost.
Groups routinely complete projects that no single member can plan, execute, or verify alone. We propose a formal model of this phenomenon, Collective Counterfactual Planning (CCP), in which the binding limitation on each agent is neither capability, knowledge, nor observability, but representational geometry: each agent perceives the state, conceives moves, consents to actions, and certifies goal requirements only through a projection onto an agent-specific subspace of a common task space. Four gates jointly determine whether a team can reach a conjunctive goal and legitimately recognize that it has done so: the exogenous implementation coalitions required to perform each action, together with three representational gates -- conception, consent, and task-relative verification qualification. We define the Collective Counterfactual Solvability (CCS) problem, separating geometric feasibility, executable attainment, and validated completion. The results expose a positive-negative duality. Iterated cross-agent relay can unlock a solution that no one-shot pooling of individual plans contains, but any goal requirement depending essentially on the subspace dark to the entire team is unverifiable and therefore not validly completable, even when the trajectory accidentally attains it. Memoryless and audited consent further constrain different objects -- action directions versus cumulative trajectory states -- and neither dominates the other. A four-step exhaustive horizon-bounded solvability scheme is sound and complete under exact representation of the relay closure; restricted implementations remain sound on returned plans but need not be complete. The model gives one geometry for sequential mutual enabling, competent execution of steps whose purpose is invisible to the executor, forced sub-teaming at expertise boundaries, and completion that cannot be validly declared.
Financial time series exhibit non-stationary and heterogeneous statistical properties, making change-point detection challenging because no single unsupervised algorithm performs consistently across assets and market regimes. Conventional workflows consequently depend heavily on expert-driven model selection, feature design, and hyperparameter tuning, limiting their scalability and adaptability. We propose EvoTS-Agent, a validation-guided self-evolving LLM agent for autonomous financial time-series change-point detection. EvoTS-Agent first performs curated exploratory data analysis to characterize dataset properties and initialize candidate detection models. It then evolves executable experiment trajectories through three complementary operators: \textit{Revision} exploits the current best solution, \textit{Alternative Strategy} explores fundamentally different modeling directions when progress stagnates, and \textit{Recombination} synthesizes complementary evidence from high-performing trajectories. Validation feedback guides trajectory evolution throughout the search, enabling the agent to adapt its detection pipeline to the statistical characteristics of each dataset while preserving reliable optimization. Experiments across four benchmark datasets demonstrate that EvoTS-Agent consistently outperforms existing LLM-based agents while maintaining a 100\% execution success rate across all evaluated backbone LLMs.
Combining large language models with reinforcement learning is increasingly explored, yet the theoretical status of LLM-derived reward signals is often left implicit. We formalize the hybrid LLM-planner and RL-controller architecture as a Goal-Augmented Markov Decision Process and show that when the LLM per-state progress score is used as a bounded potential function, the resulting shaping term preserves the optimal policy set even when the LLM scores are inaccurate. This guarantee is stronger than what general LLM-as-reward approaches provide. We verify the result numerically on a small MDP under four potential configurations, including an adversarial one scaled to twenty times the base reward magnitude.
AI 代理越来越多地执行知识工作（即生成和修改持久性数字工件，如代码仓库、文档、电子表格、幻灯片、报告），然而它们搜索的解析视图、编辑的原生文件、审阅的更改以及提交的工件，可能指向同一工作产品的不同版本。我们将其表述为一种工作区状态契约：每个视图都应显式绑定到不断演化的工作区状态的某个版本。编码代理通过用于搜索、差异和测试的仓库契约在一定程度上满足了这一需求，而对于 PDF、电子表格、幻灯片、笔记本以及混合格式项目文件夹，类似的契约则不够明确。我们提出 StagedWorkspace，这是一个面向知识工作代理的版本化工作区。该工作区将解析记录和审阅差异绑定到随原生文件变化而变化的内容哈希。在 OfficeQA Pro 和 APEX-Agents 上的固定测试框架消融实验中，双重解析/原生访问在每个测试模型上都获得了最高的点估计；相较于限制更大的单一视图，它使 OfficeQA Pass@1 提高了 8.3–12.1 个百分点，使 APEX 平均评分量表得分提高了 4.7–9.2 分。SW-AGENT 在 OfficeQA 上使用 Gemini 3.1 Pro 得分为 63.9%，在 APEX 上使用 GPT-5.4 Nano 得分为 42.1，而已发表的同模型得分分别为 29.3% 和 25.5。在 57 个文件编辑任务上的配对审阅轴消融实验进一步发现，当差异可见时，观察到的得分更高。这些结果将工作区状态确定为知识工作代理中的一个实验变量，并推动建立将证据、暂存编辑和提交工件作为显式状态转换进行评分的基准。
代表用户进行对话的自主 LLM 代理正在成为匹配平台中的一种新兴设计模式，但其可行性取决于一个很少被考察的条件：用户不仅必须接受将对话委托给代理，还必须接受来自他人的、由代理中介的沟通。我们通过对某大型约会平台活跃用户开展的两项大规模调查来研究这一条件（关于生成式个人资料功能的调查 N=2,894；关于自主对话代理的调查 N=2,617，以两种语言进行）。我们基于带有潜在回归的分级反应模型，构建了代理接受度的潜变量测量模型，并通过模型比较表明，发送代理沟通的意愿与接收代理沟通的意愿是不同的构念：二者高度相关（rho=0.92），但可以区分（Delta BIC=52），且在不同语言之间具有部分测量不变性。该模型量化了一种系统性的委托不对称：部署自己的代理所需的接受度远低于与对方的代理互动（阈值 -0.38 对 +0.32；完全互动 +1.39），且平均部署倾向约为互动倾向的三倍。基于陈述接受度推导出的随机配对反事实显示，只有 4-13% 的有向二元关系同时包含代理部署与接收方互动，并且存在明显的性别方向不平衡。设计反事实量化了可用杠杆：互惠要求通过排除近三分之二的潜在部署，将互动量减少一半或更多；而根据接收接受度来路由代理联系，则使每次联系的互动率提高三倍，这一提升在将目标项留出后的样本外验证中仍然成立（AUC 0.88，在受访者层面的交叉验证下四分位提升 3.1x）。我们讨论了其对代理式推荐设计的影响，包括披露、选择加入机制以及感知接受度的匹配。
基于记忆的自我改进智能体——这类智能体从在线任务流中学习，并通过维护一个文本记忆库随时间不断改进——在近期文献中展现出巨大潜力。然而，这些方法的可靠性方面却被严重忽视。在本研究中，我们对两种基于记忆的方法进行了全面的重新评估，并沿两个维度拓展了评估范围：(1) 纳入多次运行以量化方差，(2) 随机打乱任务以研究任务顺序的影响。通过这些实验，我们提出了两个观察结果，揭示了当前方法的脆弱性：首先，在复杂环境和多步任务中，智能体评估本身具有噪声，而在其上叠加自我改进循环会进一步放大这种噪声。其次，智能体的改进高度依赖于任务顺序。以往工作通常采用默认排序，这些排序施加了一种隐式课程，成为成功的隐藏前提。 为了更好地理解这种脆弱性，我们手动检查了智能体的记忆，并假设任务和环境的欠规范化会导致这种脆弱性。我们通过在记忆构建过程中纳入能够实现更好规范化的信息（如详细评分标准和环境反馈）来验证这一假设。尽管这些新增信息部分弥合了先前实验中的性能下降，但仍然存在显著差距，这表明还有其他尚未表征的因素导致了这种脆弱性。展望未来，我们的工作主张为自我改进智能体采用更严格的评估协议，即报告多次运行的结果，并在具有挑战性的条件下对其进行压力测试。此外，我们关于欠规范化的发现呼吁建立能够实现有效人工监督的系统和接口，防止智能体以不可预见的方式失败。
Hey HN, I've been working on Hansel [1] (an encrypted personal data store you can query with agents), and there wasn't a good way to use the modern macOS Data Protection Keychain. Electron's safeStorage [2] uses the legacy file-based keychain, which allows other apps/agents to query it with the `security` CLI. Not great when you have a dozen agents running in the background! The Data Protection Keychain is nice because it limits access via code-signing access groups and lets you set access rules like Touch ID and/or password. 1: https://hansel.so/ 2. https://www.electronjs.org/docs/latest/api/safe-storage Comments URL: https://news.ycombinator.com/item?id=49349159 Points: 3 # Comments: 0
文章 URL： https://www.theverge.com/ai-artificial-intelligence/977733/amazon-order-emails-google-gmail-ai-agents-data 评论 URL： https://news.ycombinator.com/item?id=49348627 点数：10 评论数：3
HN 大家好！我是 Barnaby，machine0（ https://machine0.io ）的创始人。我正在构建一个用于长周期代理计算的 CLI：`machine0 new mybox` 会为你的代理提供一台持久的云端 VM，按分钟计费，价格从 $0.013/hr 起，最高可达 60 vCPU / 240 GB RAM，并支持 GPU（H100s、H200s 等），VM 级别正常运行时间为 99.99%。代理可通过 CLI 或 MCP 自主运行。 演示： https://www.youtube.com/watch?v=gyllkZ0M04E 代理工作负载正在从短暂运行转向始终在线。一个处理复杂功能的编码代理会运行 6-8 小时。由代理编排的训练和 RL 运行需要数天。OpenClaw 和 Hermes 24/7 运行。随着你并行运行更多任务： - 资源：大型代码库上的几个代理就会耗尽 RAM 和 CPU。模型训练和 RL 需要你没有的 GPU。 - 安全：在你的个人机器上使用 `--yolo`，只需要一次提示注入就可能导致凭证被外泄。 - 可用性：合上笔记本电脑，代理就会在任务中途死掉。 - 隔离：你和你的代理实际最低需求之间没有清晰边界。 machine0 为每个代理提供自己的计算机。它是一个足够简单的 CLI，人类和代理无需阅读文档即可使用： `machine0 new mybox` 会创建一台可 SSH 连接、带静态 IP 和 HTTPS 端点的 VM。始终在线（VM 级别正常运行时间 99.99%），直到你将其关闭。 - 按分钟计费。1 vCPU / 1 GB 为 $0.013/hr，最高可达 60 vCPU / 240 GB，另有从 RTX 4000 Ada 到 8×H200 的 GPU。 - 暂停、快照和恢复。让你可以轻松暂停工作，之后再回来继续。也可以制作一个黄金主镜像，用于为机群批量生成克隆。 - 块存储。持久卷（从 10 GB 到 16 TB），你可以用直观的语法管理：`--yolo` 并将其挂载到你的 VM。 - Profiles。凭证、MCP 连接、提示词和环境变量的组合包，在 VM 创建时注入。因此每个代理只获得你选择的能力，除此之外什么都没有。 - 代理自助服务。把 CLI 或 MCP 服务器交给 Claude、Codex 或 OpenCode，它就能管理自己的机群：为一次构建启动一台机器，创建快照，然后销毁。 - 可复现构建。使用 NixOS flakes 或搭配 Ubuntu 的 Ansible playbooks。 人们今天如何使用它？ - 代理机群。人们运行一个试点代理来界定工作范围，并将其委派给子代理，每个子代理都在自己的 VM 上：用试点代理规划项目，工作代理实现它并提交 PR。一个客户一次运行数百台机器，通过 CLI 启动和销毁。 - 模型优化和 RL 环境。ML 团队使用 machine0 来进行由代理编排的 RL 环境和模型优化工作。一个客户在 60 vCPU 机器上运行 RL 环境，一次持续运行数天；另一个客户保留一台已暂停的 H100，并在夜间让代理使用它来反复优化推理速度。 - 产品基础设施。一个客户不是自己使用 machine0，而是将他们的产品构建在 machine0 之上：每个用户会话都会从他们自己代理运行时的版本化镜像中获得一台全新的 XL 机器。他们已经发布了数百个该镜像版本，并启动了数千台机器，其中大多数只存活两分钟。 底层是什么？ 每台机器都是完整的 KVM 虚拟机，不是容器或沙箱。你获得的是暴露给 guest 的真实 GPU，带有其实际驱动；可获得内核级访问权限（加载你想要的任何模块或驱动）；你和硬件之间没有 syscall 拦截层。技术栈本身刻意保持朴素：TypeScript、Postgres、Redis。我们非常重视安全性、可靠性和性能，使 machine0 非常适合持续的计算密集型工作负载。 关于我 我从事云基础设施建设大约 15 年了。我曾在 Imperial College London 攻读云资源分配方向的博士，后来退学；之后作为 Upflow (YC W20) 的联合创始人兼 CTO 工作了六年，在 ARR 达到七位数的整个过程中亲自负责 DevOps、基础设施和安全，因为这些事情风险太高，不能委派。machine0 最初是我为自己做的工具，我就是自己的第一个用户 :) 请求 希望你能试用并给我们反馈（见下方）。或者，如果你是一家公司，正在为软件工厂、模型训练或 RL 环境寻找计算资源，欢迎联系 barnaby@machine0.io # 安装 machine0 $ curl -LsSf https://machine0.io/install.sh | sh # 创建一台机器并通过 ssh 登录 $ machine0 new myvm $ machine0 ssh myvm 评论 URL： https://news.ycombinator.com/item?id=49348136 点数：3 评论数：0
Large language models (LLMs) are increasingly deployed as decision-making agents in settings that require sophisticated environmental exploration. However, existing work has raised questions about how LLMs actually balance exploration and exploitation. Unlike classical agents, LLM agents engage with tasks through natural language, exposing them to semantic information with no formal counterpart in the task structure. We introduce the semantic bandit, an extension of the multi-armed bandit setting that explicitly considers the textual labels assigned to actions, and use it to study how semantic priors --- inductive biases arising from associations between language and expected reward learned during pre-training, shape LLM exploration behaviour. We find that semantically informative action labels reduce exploration in favour of exploitation, improving performance when aligned with the reward structure and severely degrading it when misaligned. We further find that negative rewards trigger substantially more exploration than equivalent positive rewards, consistent with an expected-scale bias induced by reward conventions common in pre-training data. Overall, we argue that the use of language to define the environment and rewards introduces unavoidable biases derived from the fact that the model is trained on word co-occurence, with implications for the reliability and robustness of LLM agents in real-world decision-making settings.
Physical design algorithms operate within tightly coupled, multi-stage optimization flows, where stage-local gains may vanish or induce downstream degradation. Existing program-evolution frameworks often rely on stage-local objectives or undifferentiated multi-metric feedback, which neither guarantee better final results nor identify which unmet requirement should guide the next iteration. We present GoalEvolve, a goal-driven framework that makes physical design algorithm evolution accountable for the final quality of results (QoR) of the complete flow. Given a multi-objective QoR target region, GoalEvolve converts unmet requirements into normalized target gaps, identifies the dominant bottleneck, and uses stage-resolved checkpoint evidence to locate the responsible stage. An LLM-based Teacher then narrows the search to a relevant algorithmic decision and source region, while parallel Student agents implement and validate hypotheses through full-flow evaluation. Local effects, optimization debt, and downstream retention are retained as mechanism evidence for subsequent evolution. Across eight ASAP7 designs, GoalEvolve improves post-route TNS by 30.67% on average and reduces leakage and dynamic power by 21.18% and 9.42% versus default OpenROAD. Relative to commercial-tool goals, it closes 62.20% of the normalized power gap on power-dominant designs, surpasses the TNS goals on both timing-dominant designs, and closes 32.48% of the equal-weight timing-power gap on joint designs. Across all three designs evaluated against Codex goal mode under matched budgets, GoalEvolve further improves TNS by 26.46% while reducing leakage and dynamic power by 12.38% and 0.76%, respectively.
Large Language Models (LLMs) have achieved remarkable progress in code generation, yet ensuring correctness in complex, repository-level tasks remains challenging. Existing approaches often use generated tests as static post-hoc validators, which limits their ability to guide implementation and may introduce misleading feedback when the tests themselves are incomplete or incorrect. In this paper, we introduce TDD-Agent, which operationalizes the test-driven development paradigm for code generation. TDD-Agent first prompts the model to generate executable tests, encouraging it to clarify expected behaviors before implementation, and then performs iterative dual-track refinement over both the generated code and tests using execution feedback. We first isolate the effect of test-first reasoning through a prompt variant TDD-prompt on LiveCodeBench, where it consistently improves upon reasoning-based prompting baselines. Building on this finding, we evaluate the full TDD-Agent framework on RepoEval, a repository-level benchmark, and show that it consistently outperforms retrieval-based and agent-based baselines. Additional analyses show that iterative refinement improves not only code correctness but also the effectiveness of the generated tests, yielding higher pass rates, coverage, and mutation scores, suggesting that tests can serve as evolving reasoning artifacts rather than fixed validators. Our source code is available at https://anonymous.4open.science/r/TDD-Agent-Framework-6370/.
Large Language Models (LLMs) are increasingly being deployed in cybersecurity operations to assist cybersecurity analysts with rapid decision-making against emerging threats. However, there is a main criteria that must be met when using LLMs in cybersecurity, that is, trust in the generated outputs. As Agentic AI is integrated into operational systems, a robust evidence attribution and provenance tracking technique is essential to trace the origins of model generations. When autonomous agents make a decision (right or wrong), the ability to trace back through the decision chain is critical, as without it, teams cannot identify which segment of the data caused the model generation. Existing methods often struggle to distinguish among complex and highly similar evidence sources, such as cyber incident logs. This reveals a key gap: current approaches do not adequately capture the holistic geometric relationship between the retrieved evidence and the generated response for reliable evidence verification. To bridge this gap, we propose Topological Attribution Distance (TAD), inspired by Topology, to characterize and capture the global geometric shape of an output and its changes against its retrieved logs. In other words, if the embeddings of a specific source log drastically changes the geometry of the model's response in the embedding space, this suggests that such log is a critical source for the model's generated response. Therefore, TAD is powered by segment-level ablation attribution to investigate incident logs of an actual cyberattack. We demonstrate how TAD finds the most attributed logs on LLM outputs in an adaptive manner. This can provide an explainable and trustworthy tracing based on each LLM's hidden state to understand how geometrically different retrieved logs influence the model generation, and provide evidence verification in cybersecurity and Agentic-AI workflows.
语言模型和视觉-语言模型能够生成看似合理的具身计划，但并不保证其可执行性，因为它们的输出可能违反环境动态，或作用于错误落地的实体。我们提出一种神经符号智能体，将长时域家务任务分解为任务导向的视觉探索和受约束的符号规划。第一阶段中，视觉-语言模型和探索框架从第一视角观察与落地交互中获取与目标相关的谓词和实例绑定，生成符号初始状态。第二阶段中，PDDL 转移模型将解码限制为能够扩展适用动作的 token。随后，Monte Carlo tree search 使用领域无关的规划启发式方法评估可执行的后续步骤。在该转移模型下，所得计划按构造即为可执行；其向环境的迁移则取决于正确的视觉落地。在 VirtualHome 和 ALFWorld 上，开放的 4B-27B 模型在两个环境中的成功率均超过 90%，且我们最小的智能体在各环境中都显著优于 27B 的直接视觉策略。约束与搜索被证明是互补的，而非可互换的：在 ALFWorld 中，二者单独使用时都只能解决不到三分之一的任务，而结合使用可解决超过 95% 的任务。该方法生成的 token 数量也比 extended thinking 少数倍，模型可见图像数量远少于直接交互；并且在无需任何专门训练的情况下，残余失败主要定位于状态获取而非计划生成。
Agent harnesses 通过协调 agent 与环境的交互，显著提升了长时程任务的性能。然而，通过复杂 harnesses 进行强化学习在很大程度上仍未被探索，因为将此类训练扩展到长时程 agent 任务会带来根本性挑战。在这项工作中，我们提出了一个统一的黑盒 RL 框架，用于通过复杂 harnesses 对通用 agents 进行稳定且可扩展的优化。具体而言，我们首先构建了基于沙箱的执行基础设施，将任务环境和 harnesses 隔离在临时沙箱中，以支持大规模并发 rollouts。随后，我们将策略优化与不透明的 harness 执行解耦，并在模型边界放置 serving proxy 以捕获模型调用。为重建多轮轨迹并提高训练效率，我们将捕获的调用组织成前缀树，并进一步调整基于 critic 的 PPO 和无 critic 的 GRPO，使其在恢复出的树结构上进行优化。与此同时，我们在整个优化过程中保持训练-推理一致性。最后，我们引入 mix-harness training，使单个模型能够通过异构 harnesses 进行联合优化。使用 Qwen3-30A3B 时，黑盒 RL 通过 OpenClaw 和 Claude Code 分别将 ClawGym-Bench 上的 Pass@1 提升了 9.98 和 14.81 个点，同时在 200-400 个优化步骤内保持稳定。此外，该框架在更具挑战性的任务（如 JobBench 和 OfficeQA）上也带来了一致的增益。总体而言，我们的框架使得通过黑盒 harnesses 对通用 agents 进行有效、稳定且可扩展的优化成为可能，并支持跨异构执行系统的统一训练。
我们研究 AI 编码代理团队在解决编程任务时如何协作。当前的评估通常只报告代理是否完成任务以及运行成本是多少，而团队内部的协作在很大程度上未被衡量。我们提出了一种用于衡量这种协作的工具。每次运行都表示为一个时序网络，其中代理和文件是节点，消息、文件写入和文件读取是带有时间戳的有向边，并关联相应成本。我们将该工具应用于 1902 次运行，每次运行都用固定测试套件进行评估，覆盖了团队规模、团队结构和文件策略各不相同的配置。得到的网络展示了随着团队扩大以及工作变化，协作如何发生变化。直接消息传递最初随代理数量近似二次增长，其中很大一部分增长来自早期的一轮自我介绍。随着团队进一步扩大，在我们研究的最大团队中，这种增长趋于平稳，代理越来越多地通过广播消息进行通信。任务也会塑造最终形成的网络。围绕共享规格说明构建的工作会产生密集且高度互联的团队，而流水线任务则会产生围绕局部接口组织的稀疏网络。共享文件可以替代重复的一对一通信，在消息密集型工作中，当有八个代理时可将输出 token 减少约 42%；但当文件已经承载协作时，则会增加开销。指定一名代理作为协调者并不会形成通信枢纽，也不会可靠地提高成功率。我们还观察到，代理会在未被提示的情况下倾向于寻找隐藏的评分材料。我们在封闭环境中重复了关键实验条件，用带标记的占位文件替换隐藏材料。在另外 244 次运行中，代理仍在五分之四的运行中试图访问这些材料，同时关于协调者和文件通道的发现也得到复现。
Large Language Models (LLMs) have demonstrated capabilities in in-context learning, task decomposition, step-by-step reasoning, and code generation, driving their gradual evolution from text generation models into the core of agents capable of perceiving environments, invoking tools, and executing tasks. Traditional LLM Agents typically obtain information through webpages, documents, databases, or external tools and generate corresponding invocation sequences according to user goals; when this technology is further integrated with robotic systems, large language models begin to undertake functions such as task understanding, high-level planning, and behavioral decision-making. SayCan combines the task reasoning capability of language models with the affordances of robotic skills, while Code as Policies and ProgPrompt generate robot task plans through policy code and programmatic prompting, respectively, and VoxPoser uses language models and vision-language models to construct three-dimensional value maps to guide robotic manipulation \cite{6,7,8,9}. Vision-language-action models such as PaLM-E, RT-2, and GR00T N1 further strengthen the connection among language, visual perception, and robotic actions \cite{10,11,12}. In such LLM-driven embodied agents, the model not only needs to understand user instructions, but also needs to combine scene states, object attributes, spatial relations, and execution feedback to complete task grounding, and then hand the generated action plan to skill libraries, motion planners, or controllers for execution.
Agents now write knowledge graphs, but knowledge-graph stores still carry defaults set when humans curated them: accept writes now and clean later, keep one time axis or none, treat every writer's facts as equally trustworthy, and leave governance to dashboards and middleware. These four defaults are individually convenient and jointly untenable under agent workloads. We present Quipu, an embeddable store that inverts all four: no fact enters except through a gate whose predicates evaluate the pending post-state; data, trust labels, verdicts, and the rules themselves are bitemporal; named graphs are the unit of authority and trust, composed under a lattice whose one invariant is that composition never widens; and the governance specification $Σ$, the trace, and signed verdicts are facts in the store they govern, making the audit $T \models Σ$ a query. We evaluate with Census, a deterministic multi-writer lifecycle whose single seeded run scores every research question against planted ground truth: the gated store ends with 0 of 6 planted defects versus 6 of 6 ungated; all 7 composition probes uphold the lattice contract; 50 of 50 satisfied verdicts re-derive faithfully as of their instant while all 50 would be misreported under a latest-only rule set; and the SARC reference checker agrees with the in-store audit verdict-for-verdict, differing only on coverage semantics. A recorded trace from a governed writer surfaces a live enforcement gap the audit names with its remediation. On DEMM-Bench, an external decision-evidence sufficiency benchmark, a content-only reading of the exported records answers all 512 property-level governance questions correctly with zero overclaim under all eight degradation conditions, while container-presence baselines overclaim on up to 87.5% of them -- and the run surfaced, and led us to close, a gap in what a denial's verdict attests.
Humanoid robots hold great promise as general-purpose agents in human-centered environments, yet generalist vision-language-action (VLA) foundation models are not readily applicable to humanoid whole-body loco-manipulation. The high dimensionality and interdependence of humanoid motions make it challenging for conventional single-stage VLA architectures to coordinate locomotion, waist posture, and dual-arm manipulation effectively. Moreover, policies trained through offline behavior cloning can remain suboptimal during real-world deployment. Although online reinforcement learning can refine policies through real-world interaction, directly tuning large VLA backbones demands excessive computation and may introduce safety risks during real-robot exploration. To address these bottlenecks, we introduce HAF (Humanoid Adaptation Framework), a two-part framework consisting of HAF-VLA and HAF-Steer that transfers off-the-shelf generalist VLA foundation models to humanoid whole-body loco-manipulation. HAF-VLA is a hierarchical action-flow generator built on a pretrained flow-matching VLA. It splits full-body action denoising into three sequential stages with stage embeddings and cross-stage KV caches that retain kinematic dependencies, avoiding incoherent whole-body actions from one-shot generation. On top of the frozen HAF-VLA, HAF-Steer is a latent offline-to-online RL pipeline that leverages flow-matching invertibility and DCT-based dimensionality reduction to restrict RL optimization to a compact noise subspace and train a regularized SAC policy. This avoids updating the large VLA backbone and enables efficient real-world policy refinement. Evaluated on seven real-world humanoid loco-manipulation tasks, HAF surpasses vanilla single-stage VLA baselines and improves whole-body coordination and task performance. Project website: https://grange007.github.io/HAF .
We introduce Automatic Symbolic Regression (AutoSR), a fully automated system that instantiates Research-Space Symbolic Regression by searching persistent scientific investigations rather than isolated equations. Finite, noisy data often yield numerically competitive expressions that imply very different behavior outside the observed regime, making numerical fit and syntactic complexity insufficient measures of scientific credibility. Existing approaches largely focus on improving expressions, yet the search typically retains little beyond the resulting formula and score, losing the scientific record, such as motivations and probes, that inform what to try next. AutoSR preserves this record in a \textbf{Research State}, coupling each candidate equation with the reasoning, computational evidence, and independent review developed along its branch. Proposer--reviewer agents develop these states under progressive-widening Monte Carlo tree search (PW-MCTS), which allocates computation across competing investigations, while the accumulated research record is ultimately synthesized into a final report that explains the leading relation and the basis for its selection. Across nine selected challenges from two benchmark suites, AutoSR recovers algebraically equivalent relations in every case, including three cp3-bench problems that no published system recovers and six structurally diverse LSR-Transform problems. Overall, AutoSR extends symbolic regression from equation-level search toward automated scientific investigation, allowing scientific knowledge and accumulated evidence to shape both what is explored and how the resulting equation is justified.
Long-horizon robot manipulation chains many contact-rich skills into one multi-stage task. Vision-language-action (VLA) models increasingly master the individual skills, yet the chain still fails: errors compound beyond the policy's ability to correct, and one subtask silently constrains the next. A promising recipe freezes the VLA and puts an LLM agent in charge: it plans in language, moves in free space with analytic primitives, invokes the VLA only for contact-rich segments, and writes adaptation into language memory. Applied to long horizons, it breaks twice. (1) Competence comes from whole-task exploration at test time, whose cost is multiplicative in stages: if one stage needs T episodes, a K-stage task needs about T^K, and a failure does not reveal which stage caused it. (2) It has no representation of transitions: the VLA primitive carries an exit but no entry condition, so a subtask can succeed in a form its successor cannot use. We present BATON. Against (1), BATON makes the subtask the unit of exploration: each is explored in the cheap short-horizon regime and its solution stored in memory; a long-horizon trajectory is then composed from these solutions rather than discovered whole. Cost becomes additive (T*K) and every failure is attributed to a single stage. Against (2), BATON equips exploration with a transition-aware memory. Within a subtask, a verifier agent governs the invocation transition: the VLA is called only after the wrist view confirms the scene is ready. Across subtasks, a handoff transition restores an entry state disturbed by the predecessor's residue, and a lookahead transition selects the strategy whose outcome the successor can inherit. No parameters are updated. On the long-horizon benchmark RoboMemArena, BATON improves task success by 11.6% and cumulative success by 14.9% over the SoTA.
Large language models can write fluent stories, but open-ended storytelling requires more than local fluency. In evolving world simulations and AI-native games, models must preserve facts, relationships, causal dependencies, and character states as the world changes. We introduce WSE-bench, a process benchmark that separately evaluates sustained generation, canonical coherence, and meaningful development in dynamic LLM storytelling. Generation Coverage records the proportion of planned narrative steps produced; Consistency tracks when canon breaks; and Richness measures how meaningfully branching, player-shaped trajectories develop. Across frontier models, Consistency and Richness do not form a smooth trade-off: their empirical Pareto frontier is non-concave, with several non-dominated intermediate configurations that no positive linear weighting can select. Added structure can enrich trajectories, but it does not uniformly improve coherence and may shorten them. Model scale chiefly improves sustained generation, without producing reliable gains in canonical coherence or meaningful development. These results show that sustained generation, canonical coherence, and meaningful development are distinct and sometimes competing capacities. WSE-bench makes those dynamics visible by extending narrative evaluation from finished stories to the processes that create them.
Large language model (LLM) agents often perform poorly on complex, long-horizon tasks because their context becomes increasingly cluttered over time. As interactions accumulate, detailed execution traces and intermediate outputs dominate the context, making it difficult for the model to retain and use high-level planning information. Most existing methods address this issue through compression or retrieval applied to a single, flat context, which does not clearly separate different types of context information and often leads to degraded reasoning. To address this challenge, we propose HyMem, a hierarchical framework that explicitly separates the agent's context into distinct functional layers. HyMem organizes context by function to separate high-level planning from execution and complex analysis. Its isolated reasoning module handles complex subtasks without adding intermediate reasoning traces to the persistent planning context, while its memory management module preserves task progress across context refreshes through structured summaries. These components reduce redundant context accumulation, retain task-critical information, and support coherent long-horizon reasoning within a limited context window. Experiments on GAIA and Browsecomp-plus show that, with DeepSeek-V4, HyMem achieves average Pass@1 scores of 66.7% and 61.3%, outperforming the strongest baseline by 6.1 and 4.7 percentage points, respectively. Further analysis indicates that HyMem effectively controls the growth of the reasoning context, allowing the model to maintain focus and accuracy across complex, long-horizon tasks.
面向用户的多轮 Agent 必须基于不断演变的任务情境采取行动，而该情境由变化的用户意图、累积的基于工具的事实、缺失信息以及执行约束共同塑造。现有的上下文管理方法改进了对过往交互历史的利用，但很少维护一种显式的情境状态，将基于事实的信息与任务状态判断区分开来。因此，Agent 往往需要从对话轨迹中隐式推断细粒度属性、任务依赖关系和约束满足情况。我们提出 Intent-Driven Situation States（IDSS），这是一种无需训练的框架，可在对话之外维护显式的情境状态。IDSS 将工具返回结果解析为具备来源感知的实体和属性，跟踪用户意图、所需变量、约束和执行状态，并将新事实传播到任务约束中，以更新动作的可执行性。这使 Agent 能够避免不可行的动作、推进相互依赖的目标，并复用相关信息，而无需反复搜索原始历史。在涵盖八个 LLM 的三个交互式基准上的实验表明，IDSS 提升了任务完成度、偏好获取和交互效率，并在涉及多实体协调、不断演变的用户约束以及约束感知重规划的任务中带来了明显收益。消融实验和错误分析表明，这些改进来自事实持久化、以意图为中心的状态跟踪和约束建模之间的相互作用。这些结果表明，对于可靠的面向用户多轮 Agent，显式情境跟踪提供了一种有效替代以历史为中心的上下文管理的方法。
Despite the powerful multi-scale modeling methods and high-throughput infrastructures established in the materials community, real material computation workflows remain fragmented and heavily manual, requiring researchers to constantly bridge software tools, data analysis, and intermediate decisions. This growing gap between methodological capability and practical execution highlights the need for a new kind of autonomous computational framework, one that can coordinate tools, knowledge, and workflows in a more unified and adaptive way. Here, we introduce ALKEMIE Agent, an agentic platform in which retrieval-augmented generation, a materials-computation knowledge base, registered skills, database-supported provenance, AI-assisted structure modeling, bounded task execution, tool-calling iteration, and error-diagnostic assistance are integrated within a traceable control loop. The capabilities of ALKEMIE Agent are demonstrated through applications including materials recommendation, structure modeling, phonon calculations, machine-learned interatomic potential training, LAMMPS simulations, Ab Initio Monte Carlo (AIMC) sampling, and active-learning-based materials screening. Finally, we outline the future directions and challenges for the development of agentic platforms for computational materials design.
As AI systems become increasingly capable of autonomous action, determining whether an agent is technically capable of performing an action is insufficient: the system must also determine whether the action is authorised in its context. This paper introduces the Authority Resolution Framework (ARF), a five-domain ontology for representing and resolving authority across organisational roles and informal influence, business concepts, codified processes, machine-readable permissions and executable systems, and external real-world context. ARF defines the Authority Relation (AR) as a cross-domain primitive binding an actor, action, object, bounded context, justification chain, and a calibration measure termed the DNA-Coefficient, which captures divergence between documented authority structures and authority as practiced. The framework provides a machine-interpretable representation of authority provenance and scope, with JSON-LD representations and knowledge-graph query patterns for authority resolution. ARF is designed to support AI agents in determining the provenance, scope and contextual validity of authority before executing consequential actions. The framework positions authority resolution as a knowledge-representation and reasoning problem at the intersection of ontology engineering, semantic AI, agentic AI and AI governance.
文章 URL： https://www.wiz.io/blog/red-agent-snowflake-copilot-cicd-bug 评论 URL： https://news.ycombinator.com/item?id=49331423 点数：6 评论数：2
许多现实世界任务要求 LLM 智能体在较长的执行时域内与其环境交互。执行早期发生的错误可能会通过智能体上下文和环境状态传播，其影响也可能难以通过后续操作逆转。现有方法主要通过计划细化和安全检查来减少此类错误，但在错误发生后提供的支持很少。为实现长程执行中的恢复，我们提出 AgentRewind，这是一个运行时恢复框架，可记录智能体上下文和受控环境的对齐检查点，使智能体能够返回到较早状态，并利用先前尝试中的信息继续执行。我们还构建了 MettleBench，这是一个用于评估长程工程任务中任务完成情况和部分进展的基准，这些任务包含一系列相关需求。跨任务、多个模型、执行策略和智能体框架的实验表明，与所比较的基线相比，AgentRewind 提高了任务成功率和平均检查清单进度。
We present a survey of the past and future of AI Scientists: machines capable of automating science. AI Scientists can originate hypotheses, deduce their consequences, design and execute experiments, interpret their results, and revise their beliefs. Such systems are integrated scientific agents, connected to the literature, formal knowledge, mathematical models, simulations, data-analysis systems and physical laboratories. Adam was the first machine to make novel scientific discoveries through cycles of hypothesis formation and physical experimentation. Eve established the architecture of the modern self-driving laboratory. Foundation models, autonomous agents and laboratory robotics now make it possible to build systems far more general than either Adam or Eve. The central problem is no longer whether individual components of science can be automated. They can. The problem is integration. AI Scientists must combine neural learning with logic, probability, mathematics, causal reasoning, simulation, experimental design, robotics and formal scientific records. AI Scientists have the potential to transform science: to make science faster, cheaper, more systematic and more reproducible. AI Scientists could investigate systems too complicated for unaided human science, and enable thousands of AI scientists to work together on single problems. The Nobel Turing Challenge sets the goal of developing by 2050 AI systems capable of automating Nobel-quality discoveries. Progress is ahead of schedule. When we succeed it will create a new form of science and transform the world.
In the current artificial intelligence-driven innovation era, the pace of knowledge growth is accelerating, and is hard to keep up with. While generative models are increasingly used to synthesize content, they often lack in information grounding. To address these peculiarities of our time, we propose Wyvern, a multi-agent framework for the automated generation of grounded, multimodal technical reports. Wyvern allows for the generation of multimodal outputs, integrating images, tables, and text with supporting references in a unified report. Additionally, a particular focus is placed on the grounding of the content, with the implementation of a claims auto-revision stage. We conduct a human evaluation study to assess the quality of our proposed framework. The results show that the figures' informativeness is perceived as superior to that of a recent baseline in 87% of cases. Furthermore, Wyvern's reports are rated as more useful than those produced by three alternative methods in 63% to 100% of instances. We also carry out automatic evaluations showing that Wyvern gains up to 2.3$\times$ in citation recall and 1.6$\times$ in citation precision with respect to the baselines.
Spreadsheets are widely used to organize, analyze, and manipulate semi-structured data, yet automated spreadsheet reasoning remains challenging for large language models (LLMs). Real-world workbooks often contain implicit cross-table associations, fine-grained column dependencies, and complex spatial layouts. Existing methods typically flatten these multidimensional structures into sequential strings, losing important intra-sheet boundaries and inter-sheet semantics. Consequently, LLMs cannot exploit the global spatial context that human experts naturally use when inspecting spreadsheets. We propose SheetCompass, a graph-guided and memory-driven agentic framework for spreadsheet reasoning and automation. SheetCompass explicitly models structural relationships within and across worksheets while maintaining task-relevant information in memory, enabling agents to reason more effectively over complex workbooks.
随着 AI 系统在社会各领域做出更多带有道德负荷的决策，一种应对方式是道德偏好 elicitation。在这种方法中，研究人员就假设性困境对参与者进行民意调查，并使用汇总后的投票来训练一项策略，随后由 AI 模型大规模应用。在任何投票开始之前，开发者会在道德 AI elicitation 流水线中做出三个关键选择：特征范围界定、投票者抽样和问题框定。换言之，他们决定哪些特征交由投票、纳入哪些投票者，以及如何呈现问题。这些选择往往是不透明的、未被记录的，并且被视为技术细节而非规范性问题。我们在一项常见的实证研究中考察这些选择，并表明每一项选择都可能塑造道德 AI elicitation 所产生的偏好。在两个阶段（N = 809）和三个部署情境（即 AI 肾脏分配、模拟缺席员工的 AI agents，以及生成式 AI 对逝者的描绘）中，我们考察了道德 AI elicitation 流水线的三个主要阶段。首先，具有道德相关性的特征会随情境而变化。这表明，不应假定特征模式可以跨部署领域迁移。其次，约三分之一的特征会因政治意识形态而呈现偏好差异，其中一些差异的方向会发生反转。因此，投票者群体的意识形态构成会影响最终汇总出的偏好画像。第三，elicitation 问题的措辞最多可将意识形态差距缩小或扩大一个完整量表点。框定条件还会改变道德基础与参与者判断之间的关联方式。综上，这些发现表明，基于投票的对齐无法仅凭聚合就实现公平或透明的 AI；至少，道德 AI elicitation 流水线的每个阶段都应接受审计并予以披露。
This study investigates the methodological and theoretical properties of session handover in applications that use large language models. A task may continue in a new session when the context reaches the model's input limit, when the application restarts, or when another agent is asked to finish the task. The application must then decide which information from the earlier session to pass on. We formulate handover as the transfer of a task-relative in-context learning (ICL) state and distinguish exact recovery of earlier material from preservation of the target distribution. Under an exogeneity condition, predictive equivalence characterizes the coarsest deterministic sufficient handover and gives a fixed-length bit requirement. The analysis isolates the effects of the memory constraint, the writer, and the continuation procedure, and quantifies the cost of writing before the realized downstream query is known. We propose a three-part record that stores decisions and constraints exactly, uses task-justified statistics for repeated evidence, and retains original observations whose effect is not preserved by those statistics. Gaussian linear regression gives an exact finite-dimensional handover and finite-bit perturbation bounds, while nonparametric regression gives upper and lower bounds that relate memory to squared prediction error. These results provide a theory and method for deciding what a handover must retain and how its memory requirement depends on the continuation task.
我只是出于好玩逆向工程了最新版 Kimi Work 桌面应用，我知道我很宅。 总之，我发现了一件有点令人担忧的事：当你发送反馈报告时，它会在没有任何通知或提醒的情况下，把最近 5 个会话附加到报告中。 这些会话里可能包含任何内容，而你完全不知道自己正把这一切都发送给 Kimi。 我已经发邮件提醒他们了，但还是想告知大家，以防你们正在使用 Kimi。 URL: https://runtimewire.com/article/scoop-kimi-work-secretly-att... 评论 URL: https://news.ycombinator.com/item?id=49313711 积分: 10 评论数: 1
文章 URL： https://forkast.news/the-model-context-protocol-reaches-a-security-inflection-point/ 评论 URL： https://news.ycombinator.com/item?id=49316815 Points：11 评论数：1
文章 URL: https://www.anthropic.com/research/multiagent-systems 评论 URL: https://news.ycombinator.com/item?id=49316271 点数：8 评论数：2
我们宣布 Gemini API 中的 Managed Agents 推出新功能，帮助开发者构建可靠、可用于生产环境的智能体。
我们宣布 Gemini API 中的 Managed Agents 推出更多新功能，帮助开发者构建可靠、可用于生产环境的智能体。
Kaggle’s AI Agents Intensive with Google 将学习者聚集在一门免费课程中，共同构建并部署 AI 的下一前沿。
Doing research with agents is fun until they blow way past budget, jumble the sources, and don't even give you the best possible answer, just sound confident. And if you want to run some research task on local data - you have no idea where your data ends up after the prompt consumes it. So I built this tool: a deep-research agent with an enforced budget, verified quotes, and a privacy boundary for local data. 1. Never spend more than you budgeted (measured overshoot is 0%). 2. Every claim carries a source 3. Data stays local (give a CSV, it'll analyze it without the data ever leaving your machine) Works with most LLMs, including coding agents, subscriptions, local models, etc. It's free and open source, would appreciate all feedback! Comments URL: https://news.ycombinator.com/item?id=49303046 Points: 58 # Comments: 10
6G networks will not be serving as communication infrastructures only; rather, they are expected to evolve into intelligent systems, where thousands of autonomous artificial intelligence (AI) agents are interconnected. The agents are deployed across a wide range of platforms including low Earth orbit (LEO) satellites, high-altitude platforms (HAPs), unmanned aerial vehicles (UAVs), edge servers, and terrestrial devices. These agents continuously observe their environment and exchange information. Semantic communication provides an efficient mechanism for exchanging meaningful information instead of raw data. However, its effectiveness depends on the communicating agents having sufficiently aligned beliefs to correctly interpret and decode the transmitted messages. This assumption becomes difficult to satisfy in the 6G network where heterogeneous AI models operate under diverse computational constraints and continuously acquire different knowledge from their local environments. This article presents a heterogeneity-aware belief synchronization framework for 6G AI-native networks. It uses latent translation models deployed on multi-access edge computing (MEC) servers. These models translate belief updates from one agent to agent-specific knowledge without requiring joint training and a homogeneous architecture of models. By exchanging compact belief updates through a latent translation model only when necessary, the framework preserves privacy, reduces synchronization cost, and minimizes local knowledge drift. We validate the framework through a case study on a multi-layered terrestrial/non-terrestrial network. Results demonstrate that it maintains low synchronization cost, measured by the number of parameters transmitted, and low belief alignment error across the heterogeneous agents in the case study.
自主智能体越来越能够通过长周期实验来改进模型、系统和其他技术制品。然而，要理解这种能力的当前状态，评估必须超越最终分数，因为最终分数既无法揭示进展在哪里获得或流失，也无法表明积累的经验是否改进了后续决策。因此，我们提出了对七个前沿模型在 36 项长周期任务上的系统性评估，该评估基于一个新框架，使用基于规则的指标通过解决方案构建、执行和反馈控制来刻画单次运行内的行为，并通过受控比较来评估任务内和跨任务的经验复用。结果表明，当前智能体的运作方式更像工程优化器，而不是完全自主的研究者：它们能够制定并实现实用解决方案，但其性能在不同运行之间差异显著，其最强的解决方案主要是适配或组合既有技术，真正的方法论创新仍然罕见。详细分析显示，观测到的性能受到多种因素影响，包括相似最终结果背后的不同流程瓶颈、既可能帮助也可能误导后续决策的经验复用，以及影响性能稳定性的测试框架设计。这些发现为改进模型训练、推理时策略、经验管理和测试框架设计提出了具体方向。
Embodied intelligent virtual agents are expected to operate as persistent, adaptive, and context-aware entities within complex virtual and Metaverse worlds. However, implementing cognitively capable agents in such environments is conceptually and technologically challenging. Among a range of blueprints and development approaches, the Cognitive Embodied Agent Architecture (CEAA) has been developed as an implementation-oriented framework for architecting components of perception, memory, reasoning, planning, and embodied action. Considering the recent advances in edge computing and generative AI language models, this paper explores the use of Small Language Models (SLMs) to support edge-based operation of selected CEAA components, focusing on "Think" and "Memory" as processes central to cognitive orchestration and persistence of virtual agents in interactive virtual worlds. An edge-based virtual agent gateway system was developed and evaluated on an NVIDIA Jetson Orin NX using Qwen2.5 models of different sizes, exploring the system's capability to process service requests and handle memory-driven conversations. A series of simulation experiments evaluated routing accuracy, memory-read performance, and latency, demonstrating an SLM-driven prototype agent system that partially implements selected CEAA processes to support the development of embodied agents whose cognitive "brain" can operate efficiently and contextually for interactive experiences in immersive virtual worlds.
Assessing the maturity of artificial intelligence technologies is essential for investment decisions, project management, and policy monitoring, yet the available readiness frameworks are heterogeneous and difficult to apply automatically: the adaptation of Technology Readiness Levels to AI lacks AI-specific gating criteria, the Machine Learning Technology Readiness Levels presuppose access to internal process artifacts, and AI/data readiness dimension models employ scales that resist direct comparison. This paper makes two contributions. First, we unify these three frameworks into the Unified AI Readiness Level (AIRL), a nine-level ordinal scale built on an environmental evidence ladder and complemented by dimensional caps (covering specification, data existence, data quality, data legality, expert knowledge, and algorithmic maturity) together with a generality-anchoring rule and explicit assignment disciplines, so that a readiness level becomes decidable from a natural-language description of the work alone. Second, we propose RAIL (Readiness Assessment via Independent LLM-experts), a panel-of-experts classifier that operationalizes the scale: one evidence agent and six independent dimension agents, each a large language model with a narrowly scoped mandate, deliver verdicts that a deterministic minimum rule aggregates and a chief expert reviews under asymmetric authority, confirming or lowering the panel's recommendation but never raising it above the caps. The method was tested in the analysis of several research works showing consistency and avoiding overestimation from monolithic LLM classifiers.
World Models (WM) are increasingly seen as a foundation for intelligent agents that can predict, plan, and act beyond their training distribution. In this paper, we study WMs from a causal perspective across multiple levels of abstraction, ranging from perceptual observations to building a conceptual representation of the structure governing the environment dynamics. We argue that useful WMs must go beyond generative capabilities alone: they should also capture entity properties, entity-to-entity interactions, and entity-to-environment interactions that determine and explain the dynamics of a system. We provide a formal definition of Causal WMs (CWMs) grounded in the tasks they are intended to support, connecting world modelling with existing work in causal representation learning, object-centric learning, causal discovery, structural causal models, and model-based decision-making. Finally, we relate CWMs to the literature on identifiability, clarifying when the components of a WM can be recovered from data and up to which equivalence. With this, we ground WMs in representations and structures that support causal reasoning and informed decision-making.
Modern image classification models excel when trained on single task-specific datasets but often struggle to generalize across domains and difficulty levels. We propose ARMDIL, an Adaptive Router for Multi-Domain Image classification with LLMs. ARMDIL is an ensemble that uses a multimodal large language model (MLLM) agent to dynamically route each image to the most suitable vision backbone. Our diverse ensemble employs convolutional neural networks (ResNets), self-supervised representation learners (SSL), and vision-language models (VLMs), each trained on a unified label space constructed from multiple image datasets with differing distributions and characteristics. Empirical evaluations illuminate the distinct capabilities and vulnerabilities of each architecture across disparate visual domains. Crucially, we show that ARMDIL effectively navigates these trade-offs, performing competitively with specialized training-based routers. Furthermore, it drastically improves adaptability by allowing new information to be integrated via simple prompt modifications, while enhancing interpretability through natural language reasoning traces. These advances in cross-dataset image classification pave the way for more reliable general-purpose vision systems such as AI assistants and autonomous robots.
Analog circuit design is a time-consuming, iterative process in a nonlinear and high-dimensional design space that relies heavily on expert intuition. Among recent developments, LLMs have introduced a promising approach by bringing natural language reasoning to circuit design tasks. The majority of conventional LLM-based approaches provide fragmented solutions that focus either only on sizing or topology generation. These methods require adding specific technical knowledge manually, which is inefficient and prone to hallucinations during circuit sizing. Moreover, the inherent trade-off in meeting different specs makes current approaches iterative and tedious. Another shortcoming is the inability to create innovative topologies, which may lead to sub-optimal designs due to reliance on conventional topologies. In this paper, we present AaLLM, an open-source end-to-end multi-agent LLM workflow that takes user specs as input and outputs the appropriate netlist, encompassing both topology generation and circuit sizing. AaLLM automates the creation of a relevant knowledge base from research papers and textbooks to combat tedious manual data collection. A RAG model is implemented to emulate circuit design expertise using this knowledge base. Moreover, AaLLM uses a novel tri-agent feedback system comprising a Designer that determines circuit component values, a Critic that scrutinizes these values, and an Evaluator that minimizes circuit sizing iterations by arbitrating between the other two agents. AaLLM-generated novel topologies achieve a figure of merit (FoM) comparable to that of known topologies, and up to 3x higher for certain circuits. Testing on several circuit topologies, our results show a 3x - 4.5x decrease in the number of SPICE calls at inference when compared to SOTA multi-agent LLM pipelines. The results also show a 40x decrease in wall-clock time compared to existing approaches.
We present Multi-Agent Reasoning and Coordination (MARC), an open-source framework that replaces monolithic LLM prompting with deterministic multi-agent orchestration for clinical reasoning. MARC coordinates role-specialized agents for extraction, reasoning, answer generation, and evaluation, with explicit context passing and traceable intermediate outputs, enabling stage-wise failure attribution. We additionally introduce a Decomposer module that generates task-specific agent prompts from a plain-language description, eliminating manual prompt engineering. The framework supports both API-based and local CPU-compatible deployments and is entirely configurable via YAML, without code modifications. MARC is designed to be model-agnostic, interpretable, and accessible to clinical domain experts without programming expertise. The full framework is available at https://github.com/Penn-RAIL/MARC-v1.
AI agents 越来越多地被用于编程，但无法对生成代码的正确性提供任何保证。经过验证的代码生成，即由 agent 同时生成实现及其规格说明的机器检验证明，为实现可信的 AI 生成软件提供了一条更强的路径。该方向上的现有基准要么专注于单个函数，要么仅在提供实现的情况下评估证明生成。agents 能否在真实的多模块代码库中做出连贯的实现和证明选择，仍然是一个未解问题。为弥合这一差距，我们引入 Vero，这是首个在仓库级别评估联合实现与证明合成的基准。Vero 包含 43 个多模块实例，来源于横跨 Python、Dafny、Verus 和 Coq 的真实世界仓库，并覆盖从密码协议到分布式系统的多样化领域。每个实例由一个多模块 Lean 4 仓库组成，包含预先确定的 API 接口、人工整理的形式化规格说明以及参考实现，支持仅证明和代码加证明两种评估模式。为提高基准的可靠性，Vero 还包含一种审计机制，允许 agents 形式化证明所提供规格说明的不可满足性或参考代码的不正确性，从而在整理过程中暴露并纠正潜在的代码和规格说明错误。我们评估了可访问 Lean 工具链的前沿 coding-agent 配置。最强的 agent 仅完全解决了 43 个实例中的 27 个，并且在最难的仓库上没有完成任何规格说明。Vero 为衡量仓库规模的经过验证的软件合成进展提供了一个具体测试平台，而当前 agents 在这方面仍有不足。我们在 https://github.com/sunblaze-ucb/vero 发布了该基准、整理流水线和评估工具。
LLM coding agents issue Bash commands through interfaces that may serialize, wrap, and reparse model output. Matched execution scores alone cannot distinguish command-generation errors from failures introduced after generation. QuoteBench measures this boundary with exact final-state validation on 56 one-shot tasks from 14 incident-derived families, crossing the generation contract with the execution transport around one deliberately unescaped added parser. Escaping at the interpolation point reproduces each replayed reply's raw-path outcome, so any recovery under a disclosed boundary must come from the model changing its generation. Across eight same-window configurations, replaying the same reply through the added parser lowers success by 55.4 to 73.2 percentage points; disclosure recovers 30.4 to 60.7 points for six configurations, and zero or slightly negative for the other two. Raw generation is nearly saturated at the frontier; boundary adaptation is what still separates models. GPT-5.6-sol's matched gap of -3.6 points hides -64.3 points of damage and +60.7 points of compensation. The deployment configuration reorders models: one reversal among 26 comparable pairs is unambiguous and four more sit on single-task margins. Evaluations of command-issuing agents should report the model configuration, generation contract, execution path, operating point, and final-state validator rather than treat a matched score as an intrinsic model property.
Transforming multimodal sources into condensed and structured media outputs can be fundamentally conceptualized as a long-horizon agentic process centered on a model-harness system. While an ideal harness system should align with human design priors and accumulate reusable experience through empirical exploration to drive recursive self-improvement, existing paradigms remain static and fall short of this capability. In this paper, we present AutoDesign, a framework that aligns with human design priors, where a meta-harness optimizer guides a code agent to recursively improve harness based on rollout feedback. To instantiate and evaluate this framework, we focus on the academic paper-to-poster generation task and introduce PosterBench, comprising a 100-paper Main Track spanning five disciplines and PosterBench-mini, a shared 10-paper subset for controlled evaluation. On the PosterBench Main Track, AutoDesign achieves the highest score of 78.32, surpassing the closed-source commercial system Claude Design by 7.45 points. Across seven controlled code-agent-model configurations, integrating the learned DesignHarness consistently improves performance, increasing the average PosterBench Score from 54.99 to 67.39 (+12.4%). In a fully autonomous long-horizon loop, it executes 253 tool calls and 11 editing turns within 40 minutes for under $3, reaching average conference-poster quality in human evaluation. A system-blind human study further demonstrates that AutoDesign achieves the highest human preference among evaluated systems.
介绍 SIMA 2，这是一款由 Gemini 驱动的 AI 智能体，能够在交互式环境中思考、理解并采取行动。
探索 AlphaEvolve 的 Gemini 驱动算法如何在商业、基础设施和科学领域产生影响。
介绍 Co-Scientist，一款基于 Gemini 构建的协作式 AI 伙伴，旨在帮助研究人员加速科学突破。
Google DeepMind 及合作伙伴宣布一项 1000 万美元的多智能体安全研究资助征集。
通过 AI Control Roadmap 保护内部系统安全，结合传统防护措施与实时监控。
In this post we’ll outline new OpenAI research in which agents develop their own language.
Multiagent environments where agents compete for resources are stepping stones on the path to AGI. Multiagent environments have two useful properties: first, there is a natural curriculum—the difficulty of the environment is determined by the skill of your competitors (and if you’re competing against clones of yourself, the environment exactly matches your skill level). Second, a multiagent environment has no stable equilibrium: no matter how smart an agent is, there’s always pressure to get smarter. These environments have a very different feel from traditional environments, and it’ll take a lot more research before we become good at them.
Our Dota 2 result shows that self-play can catapult the performance of machine learning systems from far below human level to superhuman, given sufficient compute. In the span of a month, our system went from barely matching a high-ranked player to beating the top pros and has continued to improve since then. Supervised deep learning systems can only be as good as their training datasets, but in self-play systems, the available data improves automatically as the agent gets better.
We’re releasing an algorithm which accounts for the fact that other agents are learning too, and discovers self-interested yet collaborative strategies like tit-for-tat in the iterated prisoner’s dilemma. This algorithm, Learning with Opponent-Learning Awareness (LOLA), is a small step towards agents that model other minds.
We show that for the task of simulated robot wrestling, a meta-learning agent can learn to quickly defeat a stronger non-meta-learning agent, and also show that the meta-learning agent can adapt to physical malfunction.
We’ve developed a hierarchical reinforcement learning algorithm that learns high-level actions useful for solving a range of tasks, allowing fast solving of tasks requiring thousands of timesteps. Our algorithm, when applied to a set of navigation problems, discovers a set of high-level actions for walking and crawling in different directions, which enables the agent to master new navigation tasks quickly.
We’re releasing an experimental metalearning approach called Evolved Policy Gradients, a method that evolves the loss function of learning agents, which can enable fast training on novel tasks. Agents trained with EPG can succeed at basic tasks at test time that were outside their training regime, like learning to navigate to an object on a different side of the room from where it was placed during training.
We’re proposing an AI safety technique which trains agents to debate topics with one another, using a human to judge who wins.
We’ve trained an agent to achieve a high score of 74,500 on Montezuma’s Revenge from a single human demonstration, better than any previously published result. Our algorithm is simple: the agent plays a sequence of games starting from carefully chosen states from the demonstration, and learns from them by optimizing the game score using PPO, the same reinforcement learning algorithm that underpins OpenAI Five.
We’ve developed Random Network Distillation (RND), a prediction-based method for encouraging reinforcement learning agents to explore their environments through curiosity, which for the first time exceeds average human performance on Montezuma’s Revenge.
We’re releasing CoinRun, a training environment which provides a metric for an agent’s ability to transfer its experience to novel situations and has already helped clarify a longstanding puzzle in reinforcement learning. CoinRun strikes a desirable balance in complexity: the environment is simpler than traditional platformer games like Sonic the Hedgehog but still poses a worthy generalization challenge for state of the art algorithms.
We’re releasing a Neural MMO, a massively multiagent game environment for reinforcement learning agents. Our platform supports a large, variable number of agents within a persistent and open-ended task. The inclusion of many agents and species leads to better exploration, divergent niche formation, and greater overall competence.
We’ve observed agents discovering progressively more complex tool use while playing a simple game of hide-and-seek. Through training in our new simulated hide-and-seek environment, agents build a series of six distinct strategies and counterstrategies, some of which we did not know our environment supported. The self-supervised emergent complexity in this simple environment further suggests that multi-agent co-adaptation may one day produce extremely complex and intelligent behavior.
We’re releasing Safety Gym, a suite of environments and tools for measuring progress towards reinforcement learning agents that respect safety constraints while training.
We’re releasing Procgen Benchmark, 16 simple-to-use procedurally-generated environments which provide a direct measure of how quickly a reinforcement learning agent learns generalizable skills.
We trained a neural network to play Minecraft by Video PreTraining (VPT) on a massive unlabeled video dataset of human Minecraft play, while using only a small amount of labeled contractor data. With fine-tuning, our model can learn to craft diamond tools, a task that usually takes proficient humans over 20 minutes (24,000 actions). Our model uses the native human interface of keypresses and mouse movements, making it quite general, and represents a step towards general computer-using agents.
Klarna is using AI to revolutionize personal shopping, customer service, and employee productivity.
MavenAGI is a new software company for the AI era. They recently launched an AI customer service agent, built on the flexibility of GPT-4, which a number of companies like Tripadvisor, Clickup and Rho are already using to save time and better serve their customers.
Altera uses GPT-4o to build a new area of human collaboration
我们介绍 MLE-bench，这是一个用于衡量 AI 智能体在机器学习工程方面表现如何的基准。
Hebbia 的深度研究由 OpenAI 提供支持，可自动化 90% 的金融和法律工作
Moving from intent-based bots to proactive AI agents.
BrowseComp: a benchmark for browsing agents.
Retell AI is transforming the call center with AI voice automation powered by GPT-4o and GPT-4.1. Its no-code platform enables businesses to launch natural, real-time voice agents that cut call costs, boost CSAT, and automate customer conversations—without scripts or hold times.
了解 Genspark 如何在 45 天内打造出 ARR 达 3600 万美元的 AI 产品——借助由 GPT-4.1 和 OpenAI Realtime API 驱动的无代码代理。
ChatGPT agent System Card: OpenAI’s agentic model unites research, browser automation, and code tools with safeguards under the Preparedness Framework.
推出 ChatGPT agent：它会思考并行动，在你的指导下使用工具完成研究、预订和幻灯片等任务。
As part of our Executive Function series, Model ML CEO Chaz Englander discusses how AI-native infrastructure and autonomous agents are transforming financial services workflows.
Discover how Outtake uses GPT-4.1 and OpenAI o3 to power AI agents that detect and resolve digital threats 100x faster than before.
Built with OpenAI o3, o3-Pro, GPT-4.1, and GPT-5, Basis’ AI agents help accounting firms save up to 30% of their time and expand capacity for advisory and growth.
我们正在发布一款更先进的语音到语音模型，以及新的 API 功能，包括 MCP 服务器支持、图像输入和 SIP 电话呼叫支持。
Discover how SafetyKit leverages OpenAI GPT-5 to enhance content moderation, enforce compliance, and outpace legacy safety systems with greater accuracy .
我们正在 ChatGPT 中迈出迈向智能体商务的第一步，通过新的方式让人们、AI 智能体和企业一起购物。
今天，我们发布了新的工具，帮助开发者更快地从原型走向生产：AgentKit、扩展的 Evals 能力，以及面向智能体的强化微调。
OpenAI 推出 Aardvark，这是一款由 AI 驱动的安全研究员，能够自主地大规模发现、验证并帮助修复软件漏洞。该系统目前处于私有测试阶段——欢迎注册加入早期测试。
Chime CMO Vineet Mehra shares how AI is reshaping marketing into an agent-driven model and why leaders who prioritize AI literacy and thoughtful adoption will drive growth.
Mirakl is redefining commerce through AI agents and ChatGPT Enterprise—achieving faster documentation, smarter customer support, and building toward agent-native commerce with Mirakl Nexus.
OpenAI 在 Linux Foundation 旗下联合创立 Agentic AI Foundation，并捐赠 AGENTS.md，以支持面向安全智能体 AI 的开放、可互操作标准。
了解 Podium 如何使用 OpenAI 的 GPT-5 构建 AI 队友“Jerry”，推动 300% 的增长，并改变 Main Street 商家服务客户的方式。
BNY 借助 OpenAI 通过 Eliza 在全企业范围内扩大 AI 采用，已有 20,000 多名员工构建 AI agent，以提升效率并改善客户成果。
Netomi 如何使用 GPT-4.1 和 GPT-5.2 扩展企业 AI 智能体——结合并发、治理和多步推理，打造可靠的生产工作流。
对 Codex 智能体循环的技术深度解析，说明 Codex CLI 如何使用 Responses API 来编排模型、工具、提示词和性能。
TRUSTBANK partnered with Recursive to build Choice AI using OpenAI models, enabling personalized conversational recommendations that simplify Furusato Nozei gift discovery.
了解 OpenAI 如何在 AI 智能体打开链接时保护用户数据，并通过内置防护措施防止基于 URL 的数据外泄和提示注入。
OpenAI 如何构建了一个内部 AI 数据代理，利用 GPT-5、Codex 和记忆功能对海量数据集进行推理，并在数分钟内提供可靠洞察。
OpenAI 与 Snowflake 达成一项 2 亿美元的合作协议，将前沿智能引入企业数据，使 AI agents 和洞察能够直接在 Snowflake 中实现。
了解如何使用 Codex App Server 嵌入 Codex agent，这是一个双向 JSON-RPC API，可支持流式进度、工具使用、审批和差异。
OpenAI Frontier 是一个企业平台，用于构建、部署和管理具备共享上下文、入职引导、权限和治理功能的 AI agent。
作者：Ryan Lopopolo，技术团队成员
OpenAI 和 Paradigm 推出 EVMbench，这是一个用于评估 AI agents 检测、修补和利用高严重性智能合约漏洞能力的基准。
Amazon Bedrock 中面向代理的有状态运行时为由 OpenAI 驱动的多步骤 AI 工作流带来持久化编排、记忆和安全执行。
By combining rigorous model evaluation, full-platform use of OpenAI, and agent workflows, Balyasny is reinventing investment research.
OpenAI 如何使用 Responses API、shell tool 和托管容器构建智能体运行时，以运行具备文件、工具和状态的安全、可扩展智能体。
ChatGPT 如何通过限制高风险操作并保护 agent 工作流中的敏感数据，来防御提示注入和社会工程攻击。
OpenAI 如何使用思维链监控来研究内部编码智能体中的失准行为——通过分析真实世界部署来检测风险并强化 AI 安全保障措施。
ChatGPT 推出由 Agentic Commerce Protocol 提供支持的更丰富、更具视觉沉浸感的购物体验，支持商品发现、并排比较以及商家集成。
CyberAgent uses ChatGPT Enterprise and Codex to securely scale AI adoption, improve quality, and accelerate decisions across advertising, media, and gaming.
Cloudflare 将 OpenAI 的 GPT-5.4 和 Codex 引入 Agent Cloud，使企业能够快速、安全地构建、部署和扩展用于现实任务的 AI 智能体。
OpenAI 更新了 Agents SDK，加入原生沙箱执行和模型原生测试框架，帮助开发者跨文件和工具构建安全、长时间运行的智能体。
ChatGPT 中的 workspace agents 是由 Codex 驱动的 agents，可自动化复杂工作流，在云端运行，并帮助团队安全地跨工具扩展工作。
深入解析 Codex 智能体循环，展示 WebSockets 和连接作用域缓存如何降低 API 开销并改善模型延迟。
了解如何在 ChatGPT 中构建、使用和扩展工作区智能体，以自动化可重复的工作流程、连接工具并简化团队运营。
Choco 如何使用 OpenAI APIs 简化食品分销、提高生产力并释放增长潜力——一篇关于现实世界 AI 影响的深度客户案例。
OpenAI GPT 模型、Codex 和 Managed Agents 现已在 AWS 上可用，使企业能够在其 AWS 环境中构建安全的 AI。
OpenAI 与 PwC 正在合作，帮助企业使用 AI agents 实现金融工作流自动化、改进预测、加强控制，并使 CFO 职能现代化。
Parloa leverages OpenAI models to power scalable, voice-driven AI customer service agents, enabling enterprises to design, simulate, and deploy reliable, real-time interactions.
Sea Limited's CPO explains why the company is deploying Codex across engineering teams to accelerate AI-native software development in Asia.
在 GPT-5.5 在 OfficeQA Pro 基准测试中创下新的最先进水平后，Databricks 将其用于企业智能体工作流。
OpenAI 在 2026 年 Gartner 企业 AI 编码代理魔力象限中被评为领导者，Codex 因其创新和企业级规模部署而获得认可。
了解 OpenAI、Thrive 和 Crete 如何使用 Codex 构建自我改进的税务智能体，实现申报自动化、提高准确性并加速工作流程。
Learn how Endava uses Codex to build an agentic organization, accelerating software delivery and reducing requirements analysis from weeks to hours.
Learn how Endava is using AI agents, ChatGPT Enterprise, and Codex to accelerate software delivery, automate workflows, and build an AI-native culture across the enterprise.
一篇新的 OpenAI 研究论文展示了 AI 智能体如何改变工作，使更长、更复杂的任务成为可能，并在各类岗位中提升生产力。
了解企业如何在智能体时代通过衡量每美元的有用工作量、提高效率并扩展高价值工作流来管理 AI 投资。
Cars24 uses OpenAI-powered voice and chat agents to handle 1M+ monthly conversation minutes, recover 12% of lost leads, and bring agentic workflows to teams across the company.
一份新的领域报告显示，科学家如何使用 AI 编码智能体来推动科学计算现代化，加速基因组学及其他领域的软件开发和科学发现。
avatarin 使用 OpenAI 的 GPT-Realtime 为 Yamada Denki 购物者提供 24/7 多语言支持。两周内，30,000 人使用了该代理，92% 的调查反馈为正面。