随机对照试验（RCT）能够在不存在混杂的情况下识别治疗效应，但样本量通常较小；观察性研究（OBS）的样本量较大，却可能存在混杂。已有许多估计量将小规模 RCT 与大规模 OBS 相结合，用于估计平均治疗效应（ATE）和条件平均治疗效应（CATE）。然而，现有的 ATE 估计量要么对 OBS 作出假设，要么未能充分借助 OBS 提供的信息。相比 ATE，CATE 的研究较少。现有的 CATE 方法要么假设 OBS 中不存在混杂，要么依赖于混杂函数模型，要么以引入偏差为代价换取更低的方差。因此，我们提出一个框架：在不对 OBS 作特殊假设的情况下，通过保持基于 RCT 的估计无偏，同时借助大规模 OBS 提高估计精度，从而融合 OBS 与 RCT。基于这一原则，我们构建了 ATE 估计量 AIPW-Fusion，该估计量具有闭式权重和置信区间；此外还构建了两个 CATE 学习器 DR-Fusion 和 R-Fusion。实验结果验证了我们的发现。
准确模拟刚体交互对于构建具有预测能力的物理世界模型至关重要。尽管近年来物体动力学建模取得了进展，但如何捕捉表面之间的局部接触对物体运动的影响仍然颇具挑战。端到端世界模型能够预测整个场景或物体之间的交互，但在实际情况下，刚体接触本质上是局部的，只有邻近表面才能直接交换接触力。受此启发，我们提出刚体接触推理（Rigid-body Contact Reasoning，RiCo），通过稀疏的接触表面点邻域表示物体之间的交互。RiCo将每个点的状态与邻近表面的相对几何信息、运动状态和物理属性相结合，然后在物体的各个点之间进行推理，以确定这些局部接触如何共同影响物体的运动。通过将跨物体推理限制在邻近表面之间，同时在每个刚体内部传播接触信息，RiCo无需建模场景中每一对点，就能保留细粒度的交互细节。这些特性使RiCo具备更高的准确性和接触保真度。在MOVi-benchmark上的实验表明，与基线方法相比，RiCo将100帧位置误差和方向误差分别降低了31%–35%和约38%。此外，RiCo还实现了较高的接触保真度，相对于真实值的穿透时间差和平均深度差分别为11.0%和2.22毫米。RiCo还能够以零样本方式从小规模训练场景泛化到包含270个物体的场景。我们在真实世界中开展的多球碰撞实验进一步为从仿真到现实的迁移提供了初步证据。
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。
监督微调（SFT）将预训练大型语言模型（LLM）适配到下游任务，但任务所需的概念可能获得程度大不相同的预训练支持。高频概念更有可能得到充分学习，而稀有概念的表征可能仍然较弱。我们提出一种名为先验壁垒的新概念，用于量化预训练模型对竞争概念相对于目标概念的支持程度。我们观察到，先验壁垒呈现长尾分布，使头部概念和尾部概念在进行 SFT 时处于不同的起点：头部概念面临较低的先验壁垒，而尾部概念则需要额外的指令来克服更高的先验壁垒。我们的理论分析进一步推导出长尾先验壁垒下 SFT 的预测风险界，明确刻画先验壁垒与累积的 SFT 证据如何共同决定预测性能。受这种依赖先验壁垒的需求启发，我们提出 PASS，一种自适应的 SFT 指令选择方法。该方法构建源自参考数据的概念，并估计每条指令所提供的区分性证据，然后将选择预算自适应地分配给在当前选择下覆盖仍然不足的概念。通过这种方式，PASS 在有限预算下同时考虑哪些指令能够提供有用证据，以及哪些地方需要额外监督。实验表明，在四种骨干模型—预算设置下，我们的方法始终优于七种最先进的指令选择方法。消融研究进一步表明，PASS 的自适应分配始终优于均匀分配。
空间推理一直是视觉语言模型（VLM）的薄弱环节，因为 RGB 输入无法直接提供几何证据。现有方法要么在推理时将 3D 信息注入模型，付出架构和延迟成本；要么使用结果奖励进行训练，而这种奖励只监督最终答案。空间错误源于感知：误判深度或方向只能通过场景的真实几何结构加以纠正，而空间训练语料库的 3D 扫描数据源已经提供了这种几何信息。我们提出 GPD（Geometry-Privileged Distillation），在策略内自蒸馏（OPSD）中将几何证据作为特权信息。对于每个问题，我们将深度、语义和鸟瞰图（BEV）线索渲染为紧凑文本，并与参考答案一同路由给教师模型；仅针对错误轨迹应用特权 KL 散度，以增强 GRPO，而部署的模型仍仅使用 RGB 输入。在 4B 主干模型上，GPD 在 VSI-Bench 上达到 57.1，并在 MindCube、SPARBench、MMSI-Bench 和 ViewSpatial 上取得 37.6 的平均分，在各项空间推理基准上均优于 GRPO 和答案特权 OPSD。消融实验验证了 3D 特权与答案特权的互补性、基于问题条件的路由相较于完整上下文注入的优势，以及将蒸馏限制在错误轨迹上的益处。
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。
大型语言模型越来越多地通过指出判决背后的法条或判例来为法律决定提供理由，并将此视为该决定源于该法律依据的证据。我们对此进行直接检验：在保持案件事实不变的情况下，将模型所指出的法律依据替换为一个无关的依据，并从模型的隐藏状态中解码其不断演变的判决。在七个开放权重模型（8B-70B）和涵盖司法与合同推理的四个基准测试中，当明确要求模型通过指出所适用的法律依据来证明判决时，模型在66.7%-100%的生成结果中指出了正确依据；然而，随着法律依据变化而改变判决的情况则远不稳定：在CaseHOLD上为0.0%-21.7%，在ECHR和SCOTUS上为30.0%-76.7%，在ContractNLI上为43.3%-50.0%。无论是扩大模型规模，还是采用专门构建的法律推理模型（尽力复现的LoRA模型；第6节），都无法弥合这一差距。对五个核心模型进行的红队评估发现，模型遵从隐藏在案件事实中的对抗性指令的比例（73.3%-96.4%）远高于其对判决替换的敏感度，而且这一现象在不同模型排名中无一例外地成立。因此，指出法律依据并不能很好地作为判决依赖该依据的代理指标；与此同时，同一判决仍会分别受到对抗性操纵的影响。两项发现均在排除提示措辞噪声和混杂采样因素的检验中得到复现，并且直接关系到将生成式法律解释用作合规或审计凭证的问题。
智能阅卷与自动评分技术是智慧教育的关键基础设施。然而，现有的文档解析和手写识别基准主要针对结构良好的印刷文档或孤立的数学表达式设计，缺乏能够捕捉学生答题纸固有复杂特征的数据集，包括多行推导过程、文本与数学公式的异构混合，以及划线等噪声伪影。为弥补这一空白，我们提出了 HANS，这是首个专为真实教育场景构建的数据集，涵盖数学表达式、自然语言文本、手绘表格，以及包括修改和删划在内的多种噪声模式，并配有细粒度标注，为鲁棒识别研究奠定了可靠基础。在 HANS 的基础上，我们提出了 NA-GOT，这是一种端到端框架，通过在特征层运行的轻量级噪声抑制模块，并结合融入解码阶段的噪声感知注意力机制，实现两阶段噪声抑制。实验结果表明，HANS 对现有方法构成了重大挑战，而 NA-GOT 在答题过程识别的准确性和稳定性方面均取得了显著提升。该数据集将在论文发表后公开。
The prevailing approach to improving robot foundation models (RFMs) relies on larger models, more robot demonstrations, and costly training at scale. We show that there exists an effective and efficient complementary approach: the right reasoning recipe can substantially improve the zero-shot task performance of existing state-of-the-art RFMs. We refer to this recipe as ARC. It consists of three key ingredients: a reasoning trace, a scalable automatic labeling pipeline, and a strategy for adapting pretrained RFMs to use these traces for control. First, we find that effective reasoning traces should be grounded in the robot's next action and explain its causal structure: why the action is appropriate and what effect it should produce. Second, we show that these traces can be generated automatically from existing demonstrations, enabling us to construct ARC-Trace-DROID from DROID without collecting new robot data. Third, we show how state-of-the-art VLAs such as $π_{0.5}$ and WAMs such as Cosmos3-Nano-Policy can learn to use these traces for control, with fine-tuning and inference tailored to each model's architecture and capabilities. Using ARC, we obtain gains in zero-shot RFM performance that, to our knowledge, are unprecedented without additional robot demonstrations or foundation-scale training. The adapted models establish a new state of the art on RoboLab-120 and MolmoSpaces, with gains of up to 50 percentage points on RoboLab-Reasoning-50. On real robots, ARC improves $π_{0.5}$'s task success by 82.2 percentage points. Project website: https://arc-robot-reasoning.github.io/
多模态大语言模型（MLLM）往往难以识别和利用图示中的几何关系。近期方法通过将几何实体、关系和约束转换为显式的文本表示，来应对这一挑战。然而，有效的形式化并非易事：在 Geometry3K 上，结构注入在 200 个示例中纠正了 28 个错误，却引入了 13 个新错误。冗余关系可能分散模型的注意力，而对图中元素的含糊指代则可能导致模型错误地应用约束。这表明，关键挑战不仅在于提取更多几何事实，还在于将其组织成有助于后续推理的表示。为充分发挥形式化的潜力，我们进一步提出 GeoReform——一种反思式形式化演化框架，将形式化视为可优化的策略，而非固定的解析器输出。GeoReform 执行完整的推理流程，收集失败的推理轨迹，诊断当前表示中的缺陷，并对策略进行变异，从而更好地选择、定位、分组和呈现几何实体、关系、约束及目标。在 Geometry3K 上，GeoReform 将 Qwen3VL-2B 的准确率从 42.0\% 提升至 56.0\%。在几何推理基准上的大量实验和分析表明，有效的形式化对于提升多模态几何推理能力至关重要。
自监督多模态表示学习已在多个领域取得显著成功，但由于跨模态交互的复杂性，捕捉协同信息仍然具有挑战性。不同于存在于各个模态之间的共享信息，协同信息只有在多个模态的联合配置下才能产生与任务相关的信号，且无法从任何单一模态中恢复。本文聚焦于如何在多模态表示中保留这类协同信号的信息容量。我们的关键观察是，协同信息体现在模态之间的高阶统计依赖关系中，而这为显式建模联合交互提供了一个有原则的目标。受此启发，我们提出高阶表示与信息学习（Higher-order Representation and Information Learning，HRIL），在模态嵌入上构建经验交叉矩张量，以表示多路交互。HRIL采用Tucker分解获得核心张量，并结合协同感知正则化项，防止能量集中，从而保留用于捕捉协同信息的高阶耦合容量。在受控协同任务和真实世界基准数据集上的实验表明，与现有多模态对比学习方法相比，HRIL取得了稳定的性能提升，尤其在由协同交互主导的任务上表现出显著增益。代码已发布于 https://github.com/brightest66/HRIL】【。
世界动作模型（WAM）通常从基于重建的表征中预测动作和未来观测，而这种表征包含噪声和冗余信息，可能使下游预测变得复杂。我们提出LeWAM，这是一种双向Transformer，能够在通过四种模式进行端到端训练的无解码器JEPA潜变量上，实现前向动力学、后向动力学、逆动力学和策略预测。我们观察到以下优势：1）对齐：线性探针从LeWAM潜变量中读取机器人和物体状态的效果优于从常规Le World Model（仅进行前向预测的JEPA世界模型）中读取的效果；同时，该潜变量对视觉干扰因素的忽略程度与LeWM相当，且显著优于基于重建的WAM。2）执行：在闭环评估中，LeWAM的表现与使用相同编码器、规模匹配的常规流匹配策略相当，同时还能提供世界模型。3）规划：使用WAM进行规划时，对原始动作进行采样会使MPC利用动力学模型的不准确性；而在策略头的噪声空间中进行规划，则能改善这些WAM的闭环性能。
Safety benchmarks typically report one overall score for a suite of datasets, each of which may target one or more safety-related attributes, so models with similar overall scores can have very different attribute profiles. Comparing models is more tractable at the level of individual attributes, yet it is often unclear whether even a single dataset's scores isolate any single attribute. One plausible candidate for such an attribute is harmful refusal, a model's tendency to refuse dangerous or policy-violating prompts. We examine whether it constitutes a single, measurable attribute in HELM Safety. Using a construct validity framework that stipulates that an attribute must exist before a test can measure it, we start with HELM Safety's four datasets that might plausibly target harmful refusal, but find that three are saturated. We subject the remaining dataset, HarmBench, to two psychometric tests to determine if a single attribute like harmful refusal could stand behind its score. First, multidimensional item response theory modeling strongly suggests that HarmBench does not measure a singular attribute. Second, a differential item functioning analysis finds items where models from different developers with the same refusal ability score differently. These flags largely disappear under scope-specific matching, a pattern consistent with aggregation effects but not sufficient to rule out domain-specific developer differences. Zooming out, HarmBench collapses distinct harm behaviors into one score, and the overall HELM safety aggregate further collapses HarmBench and scores from other datasets into a single top-line number. Any safety score that averages over datasets and items can hide saturation and conflate behaviors this way. We argue that a score should earn its single-attribute reading before models are compared with it.
Generating realistic human-object interactions (HOI) in complex 3D scenes requires two complementary capabilities: reasoning about interaction feasibility in the environment and synthesizing realistic human-object motion. However, supervision for these capabilities is rarely available jointly at scale. Human-scene datasets provide rich information about environment-aware motion, while human-object datasets capture detailed interaction dynamics, yet paired human-object-scene data remain scarce. We present MAMHOI, an affordance-mediated factorization for scene-aware human-object interaction generation. MAMHOI factorizes scene-aware HOI generation through an explicit motion-affordance interface between scene understanding and motion synthesis: a scene-conditioned model first predicts where and how an interaction can be feasibly executed, and an affordance-conditioned HOI model then generates the corresponding human-object motion. This factorization allows scene understanding and interaction dynamics to be learned from complementary sources of supervision without requiring paired human-object-scene data. Experiments in complex indoor environments show that MAMHOI reduces object--scene penetration while better preserving human--object interaction quality, yielding more realistic and physically feasible scene-aware interactions. Project page: https://leimingyuan.github.io/MAMHOI-project-page/
分岔现象广泛存在于物理系统中，从结构屈曲到流体和气候动力学，但在深度学习领域仍未得到充分探索。在对称性破缺分岔中，单个输入对应多个同样有效的解，这违背了大多数学习型物理替代模型所基于的一一映射假设。我们提出 Bi-FORK，一种用于学习高维系统中一对多解映射的生成框架。Bi-FORK 通过潜在流匹配生成完整轨迹，保持空间和时间上的连贯性，并利用排斥引导采样，在单次摊销推理中恢复不同的解分支。我们在屈曲梁、机械超材料和 Allen-Cahn 相分离问题上评估了 Bi-FORK，涵盖连续、离散和场值分岔，离散化规模最高达 260,000 个点。Bi-FORK 能够恢复多模态解结构，同时将规模扩展到远超现有方法数个数量级的水平，为高维分岔物理系统的生成建模开辟了道路。
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
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。
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.
期刊和会议已开始对投稿论文进行筛查，以识别使用大型语言模型（LLM）生成的文本。这类筛查的可靠性依赖于针对一组固定 LLM 版本进行的基准评估，但实际使用的版本却在不断变化。本文量化了这种 LLM 更替如何影响科学论文的筛查。我们将《Proceedings of the National Academy of Sciences》中 4,000 篇 ChatGPT 出现前的摘要，与来自三家供应商的 23 个 LLM 版本所生成的改写文本配对；这些版本发布于 2023 年 6 月至 2026 年 8 月期间。随后，我们在不同的维护情景下训练检测器，从每出现一个新版本就重新训练的检测器，到训练一次后永不更新的检测器。在仅使用某一供应商过往版本训练的情况下，检测器可能在模型代际交界处性能崩溃：当其校准为误将 1% 的人工撰写摘要标记出来时，在最明显的交界点之前，能够检测出超过 99% 的改写文本，而在交界点之后仅能检测出 3.8%。使用较新版本训练的检测器也可能漏检较早版本生成的改写文本。不同版本之间的词汇差异，在很大程度上决定了检测能力可以迁移到何处，以及会在哪里失效。在我们模拟的两种筛查情景中，覆盖全部 23 个版本的筛查系统要么会将八分之一的人工撰写摘要标记出来，要么会漏掉最新版本改写文本中的三分之一。事实上，某款商业检测器漏掉了最明显交界点之后那个版本生成的大多数改写文本，却几乎没有将人工撰写的摘要标记出来。因此，科研诚信政策不应将检测器的基准准确率视为定论，而应将其视为暂定结果，并在每次 LLM 发布新版本时重新验证，包括对早期版本进行重新验证。
我们介绍了 NanoProof，据我们所知，这是 Lean 4 中首个因子化执行引导定理证明器，其训练数据、提取工具链、训练流程和权重均已发布，因此可以完全使用开源资源实现端到端复现。为此，我们构建并发布了一个结构化证明树数据集，以及一个用于在 Lean 4 形式化验证器中进行程序化交互和数据提取的工具。为支持可持续研究，我们注重计算效率，以促进易于获取的训练与评估。在 MiniF2F-Test 上，NanoProof 的 pass@16 达到 50.8%，超过了同类中最接近的两个系统 HyperTree Proof Search 和 ABEL，而所需计算量约分别少 90 倍和 7 倍；与 AlphaProof 相比，所需计算量则少四个数量级以上。虽然存在更强的开放权重证明器，但它们都是基于大型预训练语言模型进行微调的，既不发布训练数据，也不发布训练流程；NanoProof 表明，因子化执行引导类证明器可以使用适度资源从头重建。
Generating ultrasound reports from multiple images requires aggregating clinical evidence across views, yet archived key frames capture only part of the dynamic examination. Raw-report imitation is therefore misaligned with visual supervision: content that is clinically valid for the full examination may be unverifiable from the images available to a model. This gap creates a clinical behavior alignment problem. A model must preserve visible findings, avoid diagnostic reversals and unsupported completion, and not collapse into conservative templates. We propose CAMEO, a Clinically Aware Multi-image Evidence-grounded Orchestration framework for ultrasound report generation. Stage I learns ultrasound visual-language primitives; Stage II performs Cross-View Evidence Grounding by distilling trusted visible report points into multi-image QA and report-style supervision; and Stage III performs Clinically Aware Preference Alignment using clinical-error-oriented preference pairs. From USReport, we construct USReport-Distilled with 17,670 evidence-grounded paired-image training instances and USReport-Pref with 21,869 preference pairs; we additionally use 25,631 PubMedVision-US ultrasound instruction samples for domain adaptation and multi-image instruction tuning. On the primary USReport-Distilled benchmark, CAMEO improves over EchoVLM from 0.25 to 0.40 BLEU-1, 0.28 to 0.45 ROUGE-1, and 0.27 to 0.43 METEOR, while raising ClinicalScore from 55.02 to 74.20. These results underscore the value of evidence-grounded supervision, clinically aware alignment, and clinically structured evaluation for reliable ultrasound report generation.
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。
Knowledge distillation enables efficient spatiotemporal prediction by transferring knowledge from an accurate teacher to a compact student. However, matching outputs or features independently for each sample leaves cross-sample predictive structure underused. Exploiting this structure requires representations and historical references that reflect the dynamics of each task. We propose TAM, a Task-Aware Memory Distillation framework that organizes a frozen teacher's knowledge into a bounded, retrievable history. Memory entries encode latent features, forecast changes, or flow residuals, while task-specific selection rules identify relevant historical references. The student either matches the teacher's similarity distribution over shared references or regresses observation-conditioned residual prototypes. These objectives complement supervised prediction and conventional distillation. The teacher, memory, and auxiliary adapters are used only during training, leaving student inference unchanged. We evaluate TAM on video prediction, weather forecasting, and traffic flow prediction across multiple teacher-student configurations. Averaged over four paired runs, adding TAM improves SSIM on all six video datasets and reduces MSE on five relative to the corresponding KD baselines. Mean paired MSE reductions reach 1.86% on KittiCaltech, 1.93% on WeatherBench with a gSTA teacher, and 1.01% on TaxiBJ. These results demonstrate the utility of historical teacher supervision across distinct forecasting tasks without additional student inference cost.
参数高效微调（PEFT）使大型语言模型（LLMs）能够适应专门任务，但往往会付出预训练过程中获得的通用能力退化的代价。现有方法主要通过数据回放或正则化来缓解这一权衡，依赖额外数据或显式的优化约束。我们转而关注另一个问题：应当在哪里进行适应？我们发现，对不同的 Transformer 层进行微调会带来不同程度的目标任务收益和能力退化，这表明并非所有层都同样适合进行适应。为刻画这种差异，我们使用逐层经验 Fisher 信息来衡量对目标任务的敏感性。然而，计算 Fisher 分数需要反向计算，并且对于大型模型而言成本会越来越高。因此，我们引入输入--输出余弦相似度，作为一种轻量级、仅需前向计算的层敏感性排序代理指标。在不同模型和任务上，输入--输出相似度较低的层始终表现出更高的经验 Fisher 分数。基于这一观察，我们提出 Layer-Selective LoRA（LS-LoRA），仅在输入--输出相似度较低的层中放置可训练的 LoRA 适配器。在数学推理和代码生成任务上的实验表明，与标准的全层 LoRA 相比，LS-LoRA 在保留更多常识推理能力的同时，能够提升目标任务的平均性能，说明谨慎选择适应位置可以成为平衡目标任务适应与通用能力保留的一种简单而有效的方法。
部署通过手势与人类互动的智能机器人系统，需要具备识别多样时间模式能力的神经网络。我们对十项抽象序列任务进行了系统性基准测试——其中五项为排列不变（集合）问题，五项为顺序相关（序列）问题——并在涵盖循环、卷积、基于注意力机制以及集合函数等类别的十八种神经网络架构上进行评估。除核心的架构—任务组合外，我们还探索了多种预处理方法和目标变量变换，共形成超过250种不同的实验配置。所有变体均在完全一致的条件下进行训练和测试（固定随机种子、共享超参数、共享数据划分），以确保比较的公平性和可复现性。在全部十项任务上的排名显示，有四种架构始终表现优异：BiGRU、TCN、Conv1D 和 GRUReLU；在基准测试设置下，它们的参数量均少于2,000个，足以支持实时部署。基于这一排名，我们将三种架构各具特色的顶尖模型（BiGRU、TCN 和 GRUReLU）应用于一个实际机器人学问题：根据骨骼关键点序列估计动态手臂手势的执行速度。我们在一个自建数据集上评估了三种速度定义方式（峰值计数、周期时间和平均尖峰间距）。该数据集包含八类与交通相关的手势，共有通过 OpenPose 记录的256,710帧数据。在峰值计数定义下，最佳配置取得了0.198的平均绝对误差，对应约5%的相对误差；周期时间定义下的相对误差约为4%，平均尖峰间距定义下的相对误差约为8%。这些结果表明，神经网络能够根据骨骼数据可靠地估计手势速度，为实现具备速度感知能力的手势控制机器人系统开辟了道路。
大语言模型（LLMs）在复杂任务上表现出色，但仍极易受到提示注入攻击的影响：嵌入外部数据中的恶意指令可能覆盖用户意图。现有防御方法仍受限于需要对模型进行微调、易受自适应攻击影响，或依赖脆弱的手工设计提示。我们认为，这种脆弱性的一个根本原因在于缺乏对信任来源的明确表示。为此，我们提出可学习的信任边界分隔符（Learnable Trust-Boundary Delimiters，LTBD），这是一种轻量级防御方法，能够在保持LLM参数不变的同时，在输入中显式编码信任边界。LTBD使用少量可学习的分隔符来区分受信任的用户指令与不受信任的外部数据，使模型能够更好地遵循预期的信任层级。实验结果表明，LTBD显著优于推理时防御方法，并且与基于训练的方法相比具有竞争力，同时保持良性任务的实用性，仅引入可忽略的推理开销。具体而言，LTBD在AlpacaFarm上实现了0.00%的ASR，在TaskTracker上仅实现了0.11–0.19%的ASR。在自适应攻击下，LTBD同样保持有效；在这类攻击中，对手完全了解该防御机制，并明确尝试绕过它。
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.
Foundation models supply reinforcement learning (RL) with priors that mitigate its longstanding weaknesses in sample efficiency and transfer, but their token-by-token generation makes queries sequential and costly. Jev, a recently released decision model, generates nothing and returns calibrated, typed answers in a single forward pass. Existing work studies foundation models in RL either as models to be trained or as generators to be prompted, and Jev belongs to neither category, having so far served only as a black box in single domains. How well such a model decides on its own in RL environments, and how it can improve RL as a component of training, therefore remain unaddressed. To this end, in this paper we first examine the requirements that the objects of an RL system place on the answers they consume, and establish that Jev can fulfill all of them except the cardinal use of a value function. The remaining objects form positions that admit several roles each. We then construct algorithms that employ Jev at three of these positions, as a reference policy, an exploration judge, and a replay rater, to improve sample efficiency, exploration, and learning performance. Across nine MiniGrid tasks and three Atari games, training with Jev outperforms a standard RL learner, including where the learner makes no progress alone, while the model itself remains untrained. To our knowledge, we present the first use of Jev within the RL learning process and establish a frozen decision model as a usable component of RL training, inviting further exploration of how Jev and other advanced decision models can improve RL.
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.