全部动态

本页 40 条
10月9日 周五
arXiv cs.LG70

Refinement as a Service: Algorithmic Predictor Refinement

arXiv:2610.11415v1 Announce Type: cross Abstract: Prediction aggregation aims to combine information from multiple predictors into a more informative one. We study this question in the setting of calibrated predictors, where each prediction must equal the conditional expectation of the quantity being predicted given the predictor's signal. Given several calibrated input predictors and the feature distribution, but not the underlying Bayes probabilities, we ask when one can construct refined calibrated predictors that preserve the information in the original predictors and cannot be further refined using the available information. We formulate calibrated predictors as signaling schemes and define refinement through feature-independent garblings: a predictor refines another if its signal can simulate the other's signal. Constructibility is characterized through observable linear information: each signal corresponds to a vector over the feature space, and a new signal is constructible exactly when its vector lies in the linear span of the input signal vectors. Under this formulation, we establish a sharp algorithmic picture. For deterministic output predictors, bilateral refinement admits a polynomial-time algorithm based on a bipartite graph between the two input signal partitions, while refinement with an arbitrary number of input predictors is $\mathsf{NP}$-hard. In contrast, when randomized output predictors are allowed, we give a polynomial-time algorithm for any number of input predictors by decomposing constructible signal vectors into extreme rays of the associated polyhedral cone.

arXiv cs.LG75

超越分布保真度:用于合成表格数据的因果惩罚扩散模型

arXiv:2610.11407v1 公告类型:cross 摘要:合成表格数据生成器通常针对分布保真度进行优化,但仅有统计相似性并不能保证因果效应得到保留。本文研究了如何在完全生成式的表格模型中直接提升因果保真度。我们将因果保真度定义为:对于给定的目标估计量,从真实数据和合成数据中获得的推断分布之间的差异;理论结果表明,高统计保真度通常并不意味着高因果保真度。随后,我们提出一种因果保真度感知的训练框架,在生成目标中加入因果差异惩罚项。该框架通过因果惩罚的 TabDDPM 加以实现,并使用在策略得分函数估计量进行优化。我们进一步建立了因果正则化能够提升期望因果保真度的条件。在多种处理效应模拟实验和两个基准数据集上的实验,评估了我们的方法在保持具有竞争力的统计保真度的同时提升因果保真度的能力。

arXiv cs.LG78

WAM-Cache: Staleness-Bounded KV Reuse for Efficient World Action Models

arXiv:2610.11401v1 Announce Type: cross Abstract: World Action Models (WAMs) enable generalist robot manipulation by conditioning an action expert on representations from a pretrained video Diffusion Transformer (DiT). In closed-loop control, the video DiT runs at every chunk to encode the current observation into layerwise key-value (KV) pairs that the action expert queries. This prefill dominates the per-chunk computational cost, yet existing training-free accelerations leave it fully dense. We present WAM-Cache, a training-free framework that retains layerwise key-value representations across chunks and recomputes only a sparse refresh set of tokens. Crucially, we find that the intuitive heuristic of refreshing visually drifted tokens plateaus far below the dense baseline, even with an oracle predicting ground-truth KV drift. Downstream action accuracy is instead governed by where the action expert attends, not by what moved. WAM-Cache therefore selects the refresh set by uniting the action expert's cross-attention with visual latent surprise, complemented by a strict age bound that suppresses compounding error. On Fast-WAM, WAM-Cache cuts video DiT prefill FLOPs by 32-42% across RoboTwin 2.0, LIBERO, and real-world experiments, while staying within 0.7-1.8 percentage points of the dense policy in simulation and 2.5 points on a real robot.

arXiv cs.LG74

利用势函数估计自回归语言模型下的期望

arXiv:2610.11399v1 公告类型:cross 摘要:语言模型的许多应用关注的并非单个样本,而是模型下某个测试函数的期望。可靠地估计这类期望可能需要高昂的计算成本。本文展示了如何利用采样过程中作为副产品获得的下一词元条件概率,提高估计效率。具体而言,我们引入势函数:定义在前缀上的实值函数,可将测试函数分解为加性形式。我们构造了一种估计量,其方差取决于所选势函数,并推导出势函数能够降低方差的条件。随后,我们针对多个待估计量和应用开发了实用的势函数,并证明在相近计算成本下,这些势函数能够在多个待估计量上显著降低方差。

arXiv cs.LG73

PlanWAM:面向端到端自动驾驶的规划塑造未来表征

arXiv:2610.11382v1 公告类型:cross 摘要:端到端自动驾驶中的世界模型通过预测未来场景演化,为轨迹规划提供前瞻能力。现有方法主要研究如何预测未来以及如何利用未来信息,但较少探讨哪种未来表征对规划真正最有用。为此,我们提出 PlanWAM,一种 Planning-Shaped World Action Model。其核心思想是让规划任务塑造未来状态表征,使其保留对规划最有用的信息。随后,潜在世界模型根据历史观测预测这种经过规划塑造的未来潜在表征,并利用该表征进行规划,从而实现具备前瞻能力的规划。具体而言,我们首先使用 Temporal Register Pyramid,以关注信息时效性的方式压缩多帧历史信息,学习面向未来推理与规划的紧凑历史表征。随后,我们引入一个能够观测真实未来帧的特权未来后验分支,并利用轨迹规划目标塑造其未来潜在表征,从而获得规划塑造的未来潜在表征。Hindsight-to-Foresight Distillation 用于训练一个仅依赖历史信息的先验分支,以预测该未来潜在表征。预测得到的未来潜在表征充当规划上下文,并引导轨迹生成与选择。PlanWAM 在 NAVSIM-v1/v2 navtest 上取得 93.8 PDMS / 90.9 EPDMS 的成绩,并在零样本设置下的闭环 HUGSIM 中达到 38.7 HD-Score,表明其在开环和闭环评测中均具备领先的规划性能。大量实验进一步表明,规划塑造的未来表征为世界动作模型提供了一种有效且可部署的前瞻信息形式。

arXiv cs.LG83

RaReCache: Bridging the Gap in Cross-Model KV Cache Reuse via Rank disagreement-based Selective Recomputation

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 cs.LG66

RL-ARC: Calibrating Large Reasoning Models via Reasoning-guided Uncertainty

arXiv:2610.11352v1 Announce Type: cross Abstract: Language models (LMs) are commonly trained with Reinforcement Learning with Verifiable Rewards (RLVR) to enhance their reasoning capabilities. However, since RLVR does not explicitly account for calibration during training, it can lead to severe calibration degradation, including overconfidence. Recent calibration-aware training methods for LMs, which incorporate objectives for uncertainty estimation into training, improve calibration but still exhibit overconfidence under distribution shift, while sacrificing reasoning performance. To this end, we propose RL-ARC, a calibration-aware training framework that jointly leverages reasoning confidence and answer confidence. Specifically, RL-ARC leverages reasoning confidence as an auxiliary signal for calibrating answer confidence, applying it as reasoning-guided regularization for correct cases and as an overconfidence penalty for incorrect cases. Comprehensive results across ID and OOD settings show that, beyond improving calibration, RL-ARC enables reasoning models to adaptively estimate confidence based on the given question without substantially sacrificing reasoning performance, thereby highlighting the importance of reasoning confidence for training reliable reasoning models.

arXiv cs.LG75

DivMoE:通过跨领域专家组合实现细粒度 MoE 增量升级

arXiv:2610.11317v1 公告类型:cross 摘要:混合专家(MoE)架构已成为扩展大型语言模型的关键技术,近期研究表明,细粒度专家设计具有显著优势。从头开始训练此类模型成本高昂,而利用预训练稠密模型进行稀疏化升级是一种颇具吸引力的替代方案。然而,我们发现了细粒度升级中的一种结构性缺陷:当细粒度专家源自同一个源模型时,朴素路由会发生坍缩,导致下游准确率降至接近随机水平(例如,在 Qwen3-1.7B 上,Drop-Upcycling-fine-grained 在 15 个基准测试中的平均准确率仅为 23.2%,与从头训练的 22.2% 基本相当,而同一方法的粗粒度版本则达到 50.2%)。我们提出 DivMoE,这是首个能够实现具备结构均衡路由的细粒度 MoE 增量升级框架。DivMoE 引入了领域专用的细粒度专家初始化方法:从经过领域自适应持续预训练的稠密模型中构建专家;同时引入多样性约束路由,通过硬性结构约束确保每个 token 激活来自不同领域组的专家。在两个基础模型和 15 个基准测试中,DivMoE 始终优于六种升级基线(在 Qwen3-1.7B 上达到 55.6%,而最强基线为 51.6%),并且在 Stage 2 持续预训练后,在每项基准测试中都严格优于稠密基础模型,从而弥合了困扰以往细粒度升级方法的性能回退差距。在公开推理混合数据上进行监督微调后,我们的 12B 参数 DivMoE 模型以 64.5% 的平均准确率达到了 Moonlight-MoE(16B)的水平,同时比受控的 NVIDIA-Upcycling 基线高出 6.1 个百分点。

arXiv cs.LG76

FloorSAV:利用2D平面地图阐明AV-LLMs的空间音频视觉上下文

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 cs.LG79

Being-M0.7: A Latent World-Action Model for Humanoid Robots

arXiv:2610.11283v1 Announce Type: cross Abstract: Humanoid loco-manipulation requires coordinated locomotion and manipulation informed by future scene evolution and whole-body motion, yet learning these capabilities is constrained by scarce robot demonstrations. Human video and motion datasets offer scalable supervision, but many contain only video or motion rather than paired video-motion data. Moreover, human motion does not directly specify executable robot actions. We present Being-M0.7, a latent world-action model that transfers visual-motion priors learned from mixed-modality human data to humanoid control through pre-training, robot mid-training, and action post-training. We curate a corpus from more than 10,000 hours of raw human-centric data, integrating video-only, motion-only, and paired video-motion streams to learn complementary visual dynamics and whole-body kinematic structure. Joint prediction of future latent visual states and motion encourages visual representations to encode future kinematics. Robot mid-training adapts this coarse-grained prior to robot viewpoints and body dynamics. During action post-training, an action expert combines visual predictive representations from the frozen, adapted prior with current images and proprioception through gated cross-attention, grounding predictive context in executable whole-body commands. Being-M0.7 achieves the highest aggregate success rate among the compared baselines on SIMPLE and matches the strongest baseline on real-world Unitree G1 loco-manipulation tasks.

arXiv cs.LG83

多语语音模型中的音系干扰

arXiv:2610.11275v1 公告类型:cross 摘要:音素级模型将语音转写或生成为音素序列,音素是能够区分词语的最小声音单位。这些模型能够实现细粒度的发音控制与理解,但对于不符合任何单一训练语言的输入往往表现不佳,例如两种语言交替出现的语音(称为语码转换),或训练数据中未包含的低资源语言。我们发现了其背后的一种系统性失效模式,即音系干扰:模型假设输入仅属于一种语言,并强行施加该语言的音系规则,覆盖与所假设语言相冲突的局部音素级决策。我们通过模型保留某种语言拥有而另一种语言缺少的音素的频率来衡量干扰程度。在语码转换输入上,两个音素识别器(语音到音素模型)和一个以音素为条件的文本到语音模型会丢失其中32%至79%的此类音素,但对于两种语言共有的音素,丢失情况少得多。在未见过的语言上,我们发现音素识别器会施加其判定语音所属的训练语言的音系规则;这种判定越有把握,它丢失未见语言拥有而被判定语言缺少的音素就越多。我们从模型的内部激活中探查其语言估计,并将干扰追溯到一个低维子空间。在单语语音上,将该子空间引导至另一种语言,会使模型丢失原语言独有的音素,并生成目标语言独有的音素。我们提出窗口化语言估计(WLE),这是一种推理时修复方法:用根据每个位置周围的短窗口计算出的语言估计,替换模型在该子空间中的语言估计。在语码转换输入上,WLE能够在全部三个模型中消除34%至69%的干扰;对于识别器而言,它基本不会改变单语语音上的性能。

arXiv cs.LG75

Gated Memory: Admission-Controlled Memory Formation for Conversational AI

arXiv:2610.11270v1 Announce Type: cross Abstract: Personalized conversational AI relies on long-term memory systems that extract facts from user utterances and store them in persistent vector stores. Despite progress in retrieval, deduplication, and lifecycle management, the formation stage, the moment a fact is first written to storage has received almost no principled attention. We identify this as the binding constraint on memory quality in production systems. Critical contextual signals, such as the distinction between a permanent user attribute and a transient situation, exist only in the original utterance and are irreversibly lost the moment extraction produces a subject-relation-object triple. No downstream process can recover them. We propose Gated Memory, a lightweight, modular formation framework that interposes two decision checkpoints between conversation and storage: an admission gate that evaluates every candidate fact against the full utterance context before extraction runs, and a conditional enrichment stage that grounds admitted facts through an entity scope taxonomy with privacy constraints. The gate evaluates only the current exchange while using prior turns as read-only reference context, and produces a structured formation record. Admitted content is decomposed into atomic facts, each categorized, tagged with provenance (directly stated versus inferred), scoped to its condition of applicability, and grounded in resolved time and place, subject to a constraint that no entity absent from the context may be asserted. On the LoCoMo-10 benchmark with atypical emotional density in utterance data, Gated Memory achieves an overall +2.6% relative improvement in LLM-judge accuracy over a strong baseline with identical retrieval and generation, establishing formation quality as a measurable constraint on memory performance.

arXiv cs.LG73

Q-Capsule:一种用于缓解贫瘠高原问题的局部化胶囊式量子神经架构

arXiv:2610.11261v1 公告类型:cross 摘要:变分量子算法常受到贫瘠高原问题的限制:随着电路规模和深度增加,梯度会逐渐消失,使量子神经网络难以训练。我们提出了 Q-Capsule,这是一种局部化的胶囊式量子神经架构,通过寄存器划分、局部读出、稀疏胶囊间耦合、可训练的数据重上传,以及由量子 Fisher 信息矩阵(QFIM)引导的自适应深度增长来缓解这一问题。通过将每个可观测量的主要作用域限制在较小的胶囊内,并控制胶囊间纠缠,Q-Capsule 在保留局部量子表示之间通信能力的同时,维持了有用的梯度信号。随着寄存器宽度增加,Q-Capsule 始终保持稳定的梯度方差,而全局纠缠基线模型则表现出指数级衰减,其对数梯度方差的斜率接近每个量子比特 -ln 2。Q-Capsule 还能产生结构更清晰的优化景观、更高的参数效率、更强的去极化噪声鲁棒性,以及更低的测量需求。在二分类任务中,其自适应策略达到了 98.1% 的准确率;在四分类任务中达到了 97.7% 的准确率,同时在多分类任务中使用的双量子比特门数量比固定深度模型少约 73%。

arXiv cs.LG77

V-CoLA:采用线性注意力的视觉令牌压缩

arXiv:2610.11251v1 公告类型:交叉领域 摘要:视觉语言模型(VLM)展现出了令人印象深刻的能力,但由于视觉令牌占据输入序列的主要部分,因而面临显著的计算开销。这推动视觉令牌压缩成为减轻计算负担的关键方向。然而,随着融合线性注意力的混合架构(例如 Qwen3.5)的出现,针对 softmax 注意力设计的既有方法难以实现泛化。我们的分析表明,基于注意力和基于相似度的方法都会遭遇明显的性能下降,凸显了针对这一场景设计压缩方法的迫切需求。为此,我们提出了 \textbf{V-CoLA},一种专为线性注意力设计的高效、无需训练的令牌压缩框架。V-CoLA 引入了一种新颖的\textit{感知独特性的​​重要性准则},用于识别关键视觉令牌,并结合\textit{自适应令牌合并策略}执行压缩。所有组件均在实现层面进行了优化,以保持与线性注意力的分块并行机制兼容,从而确保较高的实际应用价值。在多个基准上的大量实验表明,V-CoLA 具有优越性能:仅使用 50.0\% 的视觉令牌即可达到原始性能的 99.5\%,在视觉令牌数量减少至 12.5\% 时仍能达到 88.0\% 以上的性能,同时实现 1.86$\times$ 至 6.15$\times$ 的预填充加速。

arXiv cs.LG82

当重构损失更低反而有害:面向低比特 LLM 量化的分布鲁棒细化

arXiv:2610.11226v1 公告类型:cross 摘要:仅权重量化的训练后量化(PTQ)高度依赖于最小化重构损失,以在低精度下保持模型质量。我们发现,通过最小化该损失所选出的权重未必能在新任务上带来更好的模型性能。事实上,我们发现,更低的重构损失甚至可能导致模型在相同校准数据上的性能下降。我们的进一步分析表明,在校准数据上具有更低重构损失的权重,在输入激活分布发生变化时,其损失可能高于其他权重。基于这些观察和分析,我们提出分布鲁棒量化(DRQ),一种事后细化过程,通过在受约束的输入激活分布集合上最小化最坏情况下的重构损失来优化量化结果。DRQ 在现有量化网格内细化表示量化权重的整数代码,同时保持量化参数和推理算子不变。大量实验表明,DRQ 能够改进由六种具有代表性的 PTQ 方法量化的模型,包括 AWQ、GPTQ 和 ParoQuant,并且在稠密大语言模型和混合专家大语言模型上均能带来收益。这些结果表明,DRQ 是一种适用于仅权重量化 PTQ 的通用事后细化框架,无需增加推理开销即可实现更好的下游性能。

arXiv cs.LG80

跨物种表征学习对齐小鼠和人类的神经动力学并追踪临床药物疗效

arXiv:2610.11222v1 公告类型:cross 摘要:临床前模型难以预测人类药物疗效,尤其是在神经系统疾病中。神经活动提供了极为丰富的转化研究信息来源,因为它捕捉了神经系统功能的高维变化,而这些变化可以在动物和人类中进行测量。然而,其高维特性使得研究人员难以区分保守的疾病相关特征与由物种、记录方式和实验环境引起的变化。在本研究中,我们检验了这样一种可能性:能否通过从电生理数据中学习按照生物学状态而非物种组织的表征,直接识别共享的神经动力学。我们开发了一种双规则对比学习框架,在保持不同表型之间分离的同时,对齐小鼠和人类中相对应的状态。该框架在不同物种间恢复了保守的感觉反应结构,并在癫痫研究中解析出三个小鼠模型与异质性人类患者群体之间不同的关系。当将接受治疗的动物投影到冻结的跨物种表征中时,其因药物诱导而向与人类对齐的健康状态移动的程度,回溯性地追踪了十种模型—药物组合中已知的临床疗效,其中包括一种疾病特异性的不利影响。尽管遗传病因和记录方式存在差异,该框架还识别出了Fmr1-knockout小鼠与人类16p11.2拷贝数变异携带者之间共享的疾病相关神经动力学。总之,这些发现表明,跨物种神经表征学习有望将异质性人类疾病映射到实验上易于研究的临床前状态,并评估干预措施是否能够恢复与人类相关的回路功能。

arXiv cs.LG65

BRACE: Differential Privacy for Dense Associative Memory with LSR Energy

arXiv:2610.11218v1 Announce Type: cross Abstract: Dense associative memory (DAM) provides an energy-based framework for memory retrieval with close connections to attention mechanisms in modern artificial intelligence. Despite growing interest in differential privacy for AI, the privacy of DAM retrieval dynamics remains relatively unexplored. In this paper, we develop a differential privacy framework for log-sum-ReLU (LSR) dense associative memory, whose finite-support retrieval dynamics pose distinctive challenges for privacy-preserving computation. We propose the Boundary-Responsive Adaptive Correction Evolution (BRACE) algorithm, a differentially private retrieval mechanism for LSR-DAM that adaptively corrects boundary-sensitive perturbations to control their cumulative effect over the retrieval trajectory. In theory, we prove that our method is minimax optimal by deriving dimension-independent terminal and full-trajectory retrieval error rates, with optimal dependence on the inverse temperature and, in the growing-horizon regime, the retrieval horizon. We further establish central limit theorems that enable uncertainty quantification for private retrieval by characterizing its asymptotic distribution and the additional variability introduced by privacy. Numerical experiments compare our proposed method with baseline differential privacy approaches and evaluate its retrieval accuracy. Together, our results provide a theoretical foundation for optimal privacy-preserving retrieval and uncertainty quantification in energy-based associative memory systems.

arXiv cs.LG74

剖析 Vision Transformers 中的表示结构:一项严谨的架构研究

arXiv:2610.11205v1 公告类型:cross 摘要:表示结构对于理解 Vision Transformer(ViT)架构及其泛化行为至关重要。然而,既有研究既未隔离和分析模块级特征,也未探究这些特征之间的交互如何影响性能估计。在本工作中,我们首次对不同架构规模下的特征信息进行了严谨分析,从实证角度揭示了 ViT 表示与泛化行为之间的关系,并利用这些洞见指导高效的 ViT 设计。我们的贡献主要有以下五点:在不同架构规模下,1)我们发现了初始化阶段的特征坍缩现象,该现象会导致冗余,并提出了一种缓解该问题的降维方案。2)我们利用熵和最小特征值量化特征信息,证明这些指标可作为预测泛化能力的可靠信号。3)我们表明,token 空间中的特征比嵌入空间中的特征更能忠实地表示信息。4)我们发现了一个出乎意料的现象:ViT 层内线性子模块产生的特征对于预测泛化性能至关重要。5)与以往基线方法相比,我们提出的代理指标将相关性排名提升了 18%–48%,并能有效识别出在计算成本更低或相当的情况下实现更高准确率的 ViT 架构。

arXiv cs.LG75

PageWeaver:KV 引导的稀疏注意力查询并集

arXiv:2610.11201v1 公告类型:cross 摘要:动态稀疏注意力会限制每个查询所选取的 KV 页面,但较小的支持集并不一定能带来高效的 GPU 工作。查询并集可以共享页面加载并填充 Tensor Core 瓦片;其开销取决于哪些查询被归为一组。我们提出 PageWeaver,这是一种执行设计,利用所选页面的亲和性来组装查询组,同时保留每个查询原有的支持集和完整的输出归属。受限的 GPU 搜索会生成查询 ID,随后由一个感知 ID 的双 CTA 内核使用这些 ID,而无需具体化重排后的 Q 张量或跨页面的部分输出。直接实现 KV 页面并集则提供了关于非局部复用和归约开销的补充性设计研究。在全程使用 FP8 KV 的情况下,H200 Union8 实现基于六组采集数据,相比实测的 FlashInfer 路径,实现了完整调用速度 1.70 倍的几何平均加速。在五组选定的 64K 上下文采集数据上,在线重新分组进一步将延迟降低了 3.26%–7.66%。整体模型预填充吞吐量比测试的原生路径高出 7.88%–14.36%;增量式重新分组带来的收益较小,在 32K/64K 下观测到的中位数增益为 0.47%–0.73%,而在 8K 下出现性能回退。B300 对比实验识别出了一些准备开销较高且原生内核更强、从而无法体现该优势的情况。这些结果将执行组复用与在线利用该复用的完整成本区分开来。

arXiv cs.LG74

选择性聆听:大规模音频-语言模型中音频影响的机制引导控制

arXiv:2610.11196v1 公告类型:cross 摘要:大规模音频-语言模型(LALMs)会利用多模态证据,但在无需聆听时,与任务无关的音频仍可能改变文本推理决策。总体准确率可能掩盖这种配对漂移,因为音频带来的修正与损害可能相互抵消。配对漂移分析与针对性干预确定了特定于架构且对干预敏感的后期音频路径,并将其识别为可操作的控制点。我们提出 ICAP-Gate,对每个模型的路径实施机制引导、任务条件化的控制。在四个 LALMs、两个推理基准,以及环境声音和自然语音干扰的评测中,ICAP-Gate 在全部 16 个完整划分的模型—条件评估中,其 Influence Rate 和 Answer Flip 的点估计均低于未门控推理。固定抑制会导致四个模型的自动语音识别(ASR)性能均下降,而 ICAP-Gate 通过在明确要求音频的指令下保留该路径,使 ASR 性能达到与未门控推理相当的水平。在全部四种评估设置中,ICAP-Gate 的配对漂移点估计均低于缓解式提示;相较于使用八个样本的 Self-Consistency,它在每次查询仅生成一次的情况下提供了具有竞争力的稳定化效果;在受控的 ARC 测量中,Self-Consistency 的延迟为未门控推理的 $7.0$--$9.2\times$。这些结果确立了选择性模态影响控制这一稳健多模态推理的设计原则。

arXiv cs.LG73

LadderEdit:面向LLM高效终身编辑的编辑级残差压缩

arXiv:2610.11160v1 公告类型:cross 摘要:LLM的终身编辑需要在完成获取后存储数千次编辑。一类广泛使用的方法是为每次编辑附加一个LoRA适配器,这能够保持模型行为,但存储量会线性增长。为应对这一挑战,我们提出LadderEdit,一种在获取每个LoRA适配器后对其进行压缩的方法。每次编辑首先以低秩形式存储,作为低成本的草图。随后,我们在探测提示上检查该草图是否仍满足改写、泛化和局部性约束。通过检查的编辑保留草图;未通过的编辑则沿着秩梯级提升至更高秩,直至满足约束。由于每次编辑都保留了一定的表示,覆盖范围得以维持,只有困难编辑会消耗更多秩。在LLaMA-3-8B、Mistral-7B和Qwen2.5-7B上的ZsRE、CounterFact和WikiBigEdit基准测试中,LadderEdit在内存占用减少5.2倍的情况下达到与完整LoRA存储相当的效果,并且在连续进行50,000次编辑时仍保持有效。

arXiv cs.LG70

Accelerating Non-Smooth and Heavy-Tailed Sampling

arXiv:2610.11139v1 Announce Type: cross Abstract: Anchored Langevin dynamics (ALD) is useful for non-smooth sampling where the density of the target distribution is possibly non-differentiable and heavy-tailed; reflected anchored Langevin dynamics (RALD) can sample possibly non-differentiable target density on a constrained domain. In this paper, we propose and study non-reversible anchored Langevin dynamics (NALD) for sampling possibly non-differentiable and heavy-tailed target density in the Euclidean space and the non-reversible reflected anchored Langevin dynamics (NRALD) for sampling possibly non-differentiable target density in the constrained space. Our construction adds a circulation drift generated by a possibly state-dependent divergence-free skew-symmetric matrix field and a stream potential. It preserves the target distribution without requiring derivatives of target density, admits a random-time-change representation, and applies both on the whole Euclidean space and on bounded domains with normal reflection. By breaking reversibility, we show that NALD and NRALD can converge to their target distributions faster than their reversible counterparts via finite-time non-asymptotic convergence analysis, a large deviations analysis and asymptotic variance reduction. Numerical experiments demonstrate the efficiency of the proposed algorithms.

arXiv cs.LG78

Can a System-One LLM Perform Knowledge Tracing When Few or No Learners Are Logged?

arXiv:2610.11135v1 Announce Type: cross Abstract: Knowledge tracing (KT) models need many logged learners, so a new course or platform starts without a usable model. In LLM-based KT the LLM generates the answer, which we call System-Two; it is either fine-tuned on the target data or reasons and votes over ten samples, which is slow and gives coarse probabilities. We ask whether an off-the-shelf System-One LLM, which returns a probability for a typed question directly in a single pass, can perform KT when few or no learners are logged. On seven datasets, Jev without any data from the target platform reaches a mean AUC of .706, above the best of 28 deep KT models trained on 8 learners (.689) and above System-Two Thinking-KT on all seven datasets (.650) at about 1/100 of its API cost. Adding examples and a similar-learner statistic from the logged learners (JevKT) raises this to .722; JevKT stays significantly ahead of deep KT up to 16 learners and ahead on average up to 64, and supervised KT catches up between 64 and 128 learners. Among the readers we tested, the gain is specific to Jev, since three other LLMs queried with the byte-identical typed request through the official System-One adapter fall below it on all seven datasets, and reader swaps and contamination checks find no evidence that the input format or memorised data explain the gain. For new learners the advantage holds from their first interactions, whereas on unseen items with all learners logged, deep KT remains ahead.

arXiv cs.LG82

SFT-as-context 缓解监督式微调中的遗忘问题

arXiv:2610.11132v1 公告类型:cross 摘要:监督式微调(SFT)赋予大型语言模型(LLM)专门能力,但往往会以遗忘其父模型(即微调前的预训练模型)的一般能力为代价。对于同时需要专门能力和一般能力的查询,这种权衡尤其具有限制性。我们提出了 SFT-as-context,这是一种无需训练的方法,其中父模型利用 SFT 模型的回答作为上下文来回答查询。通过上下文学习,父模型可以从 SFT 回答中获得微调能力,同时保留自身的一般能力。在 19 对父模型与 SFT 模型以及 11 个基准测试中,SFT-as-context 在微调能力上始终接近 SFT 模型:在 AIME 2024 和 LiveCodeBench 上的差距仅为 2.2 和 2.1 个百分点,在 NutriBench-English 上的宏观 MAE 差距为 2.0;同时,在一般能力上,其平均表现与父模型的差距不超过 2.2 个百分点。值得注意的是,它能够解决同时需要微调能力和一般能力的查询,即使父模型和 SFT 模型单独使用时都无法成功。该方法还可以扩展到父模型与 SFT 模型配对之外:来自小型开源 SFT 模型的回答能够提升强大的闭源 LLM 的表现,优于单独使用其中任一模型。此外,我们使用贝叶斯框架推导出理论保证,分别界定 SFT-as-context 在微调能力上相对于 SFT 模型、在一般能力上相对于父模型的误差。除此之外,我们可视化了注意力权重,发现父模型会更多地关注有用的 SFT 回答,而较少关注无关回答,这表明选择性注意力有助于父模型通过上下文学习利用 SFT 回答。

arXiv cs.LG74

Lapras:时间序列语言模型的潜在推理

arXiv:2610.11111v1 公告类型:cross 摘要:时间序列语言模型(TSLMs)通过对时间信号进行推理并生成自然语言答案和解释,为理解时间序列提供了一条有前景的路径。一种常见方法是思维链(Chain-of-Thought,CoT),即生成将相关信号模式与最终答案联系起来的逐步推理过程。尽管这些模型在后训练阶段能够从参考 CoT 轨迹中学习,但在推理时生成对输入时间序列忠实的描述仍然具有挑战性。将高维、连续的时间表示转换为离散的语言标记,可能导致模型忽略与任务相关的模式,或对其进行不准确的描述。由于后续推理步骤建立在这些描述之上,早期错误会不断传播,最终导致答案错误,同时产生看似合理却与输入信号不一致的解释。我们提出 Lapras(Latent Post-trained Reasoning Across Series),一种赋予 TSLMs 潜在推理能力的后训练框架。经过 Lapras 训练的模型在联合的时间序列—语言空间中,通过一系列连续思维进行推理,仅针对最终答案生成文本。该能力通过教师—学生自蒸馏进行学习:教师模型基于 CoT 参考轨迹进行训练,并通过文本显式推理。学生模型在答案阶段使其隐藏状态与教师模型的隐藏状态对齐,从而将教师模型的推理能力迁移到自身的潜在计算中。我们在五个时间序列问答基准上,使用四种 TSLM 主干模型对 Lapras 进行了评估。与显式 CoT 相比,Lapras 将平均 F1 最高提升了 10.79%,同时生成的标记数量减少了 23.9 倍。Lapras 的连续思维还可以通过标准语言解码转换为可读的推理轨迹,从而保留文本解释。总体而言,这些结果表明,Lapras 是一种有前景的后训练范式,能够实现高效、有效且可解释的 TSLM 推理。

arXiv cs.LG66

A General $\widetilde{\Omega}(\sqrt{T \gamma_T})$ Lower Bound for Kernel Bandits

arXiv:2610.11082v1 Announce Type: cross Abstract: The kernel bandit problem consists of sequentially optimizing an unknown function with noisy feedback, where the function has bounded norm in a given Reproducing Kernel Hilbert Space (RKHS). A central quantity in the regret analysis of kernel bandits is the maximum information gain $\gamma_T$. In particular, the best existing upper bounds scale as $\sqrt{T\gamma_T}$ up to log factors, and nearly-matching lower bounds have been derived for specific kernels such as squared exponential and Mat\'ern. However, lower bounds for general kernels are lacking, thus making it unclear in what generality the upper bounds are near-optimal. In this paper, we establish a general $\Omega(\sqrt{T\gamma_T/\log T})$ minimax regret lower bound for non-constant continuous kernels on compact domains, establishing near-optimality (within log factors) in a very general sense. We show that the log factor appearing in this bound is unavoidable in general, but that it can be removed under certain conditions. Among other things, our findings imply that the minimax-optimal scaling is exactly $\Theta(\sqrt{T\gamma_T})$ (i.e., within constant factors) for the Mat\'ern-$\nu$ kernel with $\nu \in (0,2)$, $\gamma$-exponential kernel with $\gamma \in (0,2)$, and certain piecewise-polynomial kernels.

arXiv cs.LG72

在噪声监督下的多模态学习中学习可信信息

arXiv:2610.11057v1 公告类型:cross 摘要:多模态分类通过处理和关联来自多个模态的信息来实现更准确的预测。然而,现有方法通常依赖高质量的真实标签,而在现实场景中获取这类标签十分困难。尽管带噪标签学习中的样本选择方法旨在从噪声数据中识别标注正确的样本,但传统方法主要关注单模态设置,未能充分利用多模态信息。这促使我们构建一种更可靠的多模态学习噪声检测器。为此,我们从理论上分析了表示结构与噪声检测能力之间的关系。基于这一分析,我们提出了 REFINE,这是一种多模态标签噪声检测框架,联合使用融合表示和单模态表示进行标签噪声检测。具体而言,REFINE 通过对目标类和背景类进行判别分析来构建判别特征向量,并为每个类别选择具有更强噪声检测能力的可信表示空间。在每个可信空间内,REFINE 衡量每个实例表示与判别特征向量之间的对齐程度。随后,它将从这些空间中选出的子集进行合并。合并后的集合为更新多模态分类器提供了更干净的监督信号,从而减小训练过程中错误标注样本的影响,并提升模型的泛化能力。在多个不同任务上的大量实验表明,与基线方法相比,REFINE 具有更优的性能。源代码将公开发布。

arXiv cs.LG83

AgentHorizon:评估用于长时程计算机使用任务的智能体式评判器

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 cs.LG77

FedAlphaEdit:面向协作式知识编辑的零空间对齐合并

arXiv:2610.11033v1 公告类型:cross 摘要:多个机构可能各自持有私有知识编辑,并希望在不共享原始编辑请求的情况下,将这些编辑整合到一个大型语言模型中。AlphaEdit 等零空间约束编辑方法从数学上保证每次更新都不会改变无关知识,而 CollabEdit 等协作框架则在不共享数据的情况下聚合来自多个客户端的编辑。将二者结合起来看似简单。然而,我们表明,这种朴素的组合在结构上会失效,并找到了其根本原因。在此分析的指导下,我们提出了 FedAlphaEdit。据我们所知,这是首个在统一的零空间原则下,同时对本地编辑和服务器端合并规则进行对齐、以保护现有知识的协作式知识编辑框架。FedAlphaEdit 构建于零空间对齐合并之上:客户端共享投影后的统计量,而服务器在一次性理想化设定下,可以从理论上恢复将所有编辑集中在一处执行的结果。在两个架构系列上,实验结果表明,所提出的方法修复了性能崩溃问题,并使编辑成功率和知识保持能力同时接近集中式编辑的水平。因此,对于无法共享原始编辑数据的机构(如医院和金融公司),FedAlphaEdit 使它们能够共同维护一个共享模型,其效果能够近似于将所有事实集中在一处进行编辑。

arXiv cs.LG83

NOMOS: Compiling Written Policies into Statically Verified Tool-Call Gates for LLM Agents

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 cs.LG81

一种用于中期提前时效全球日尺度火点辐射功率预测的图神经网络

arXiv:2610.11022v1 公告类型:cross 摘要:提前数天准确预测生物质燃烧活动对于空气质量预报和气溶胶预测十分重要。本研究由两个业务运行约束所推动。首先,用于初始化 NOAA 的 GEFS-Aerosols 的 GBBEPx 卫星火点辐射功率(FRP)产品存在约1.5天的延迟,因此每个预报循环都依赖于最新可用、但实际上已经过时的火灾观测。其次,在随后的5天业务预报中,这些火灾输入会保持不变;在 GSL 实验系统中则保持7天不变,这实际上假设火灾活动不会发生变化。我们开发了一种数据驱动模型,可根据最新可用观测,预测未来1至7天的全球 FRP。该模型基于时空图神经网络进行改进,使用再分析气象资料、土地覆盖和植被信息、近期火灾历史,并以 GBBEPx FRP 作为训练目标。模型使用2020—2022年的数据进行训练,并在2023—2024年进行评估。模型再现了全球季节变化周期,且性能显著优于持续性预报。在0.1$^\circ$分辨率下,2023年提前1天和7天的均方误差分别降低32%和43%,2024年分别降低24%和40%。在1$^\circ$分辨率下,临界成功指数介于0.32和0.60之间。随着提前时效增加,探测能力仅小幅下降,而强度预报能力下降得更快。模型能够可靠地探测大型火灾,但会系统性低估其辐射功率。这些结果表明,火灾活动在提前数天的时间尺度上具有较好的可预报性,并指出在预测的 FRP 能够支持业务气溶胶预报之前,强度校准和小火点定位仍是主要挑战。

arXiv cs.LG79

在原始视频上对语言模型进行中期训练

arXiv:2610.11019v1 公告类型:cross 摘要:多模态大语言模型主要从图文配对数据或带注释的视频中学习,而原始网络视频很少被用于进一步训练现有语言模型。我们研究了没有字幕且不使用文本损失的原始视频,能否作为预训练语言模型的中期训练数据。我们将视频帧编码为连续视觉标记,并让语言模型学习预测下一个视觉标记。我们使用来自 YT-Temporal-1B 的原始视频片段对 Qwen3-1.7B 进行中期训练,随后对该模型和未经中期训练的模型采用相同的图像-文本指令微调,使两者仅在中期训练环节存在差异。经过中期训练的模型在四项视频基准测试中的平均得分高出 2.9 分,在十项图像基准测试中的平均得分高出 5.1 分,涵盖感知、文档和图表任务。尽管中期训练不包含文本,文本性能仍得以保持:在 14 项文本基准测试中,平均得分为 48.9,而未经中期训练的模型为 48.0。训练过程分析表明,图像和视频方面的提升在训练进度达到 30% 以内时便已显现,此后趋于平台期,变化幅度小于 0.5 分。预测字幕并未优于预测下一个视觉标记,这表明视频中期训练可以完全保持自监督形式,无需承担自动字幕生成带来的计算开销或标注噪声。

arXiv cs.LG81

Curating Always-Loaded Context for LLM Agents: A Capacitated Assortment Model with Censored Feedback

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 cs.LG73

跨运行工况的多尺度等离子体动力学重建

arXiv:2610.11004v1 发布类型:cross 摘要:利用少量传感器重建具有空间分辨率的等离子体动力学,对于诊断、降阶建模和控制至关重要,但这仍然十分困难,因为稀疏测量无法充分约束多尺度且依赖工况的自由度。Shallow Recurrent Decoder (SHRED) 通过利用测量历史,在一定程度上解决了空间稀疏性问题;然而,其全连接解码器没有显式机制来解析跨尺度空间结构,也没有显式的参数依赖关系。我们提出 Recurrent Multiscale Affine-modulated Inference Network (ReMAIN),保留 SHRED 的循环时间编码方式,但将其解码器替换为 U-Net,并通过逐特征线性调制,使 U-Net 的特征层级受到循环状态的条件控制。时间表征在整个 U-Net 中同时提供稠密先验和特定尺度的调制。我们进一步提出参数化扩展,将运行条件与传感器历史联合嵌入,使重建能够随着控制动力学随运行工况变化而自适应。首先,我们在代表多样动力学特征的六个一维非线性偏微分方程上,将 ReMAIN 与 SHRED 进行了基准比较。在所有基准测试中,ReMAIN 都降低了对未见轨迹的重建误差,并更准确地解析了陡峭过渡、局部极值和细尺度变化。随后,我们在一个无碰撞 $E \times B$ 等离子体上验证了参数条件模型;该等离子体受到垂直于轴向的电场和径向磁场作用,其中电场强度作为运行参数。对于训练中未包含的电场强度,ReMAIN 重建了高维多尺度等离子体状态,并恢复了其依赖工况的时空动力学。总体而言,ReMAIN 改进了基于稀疏传感器的全状态重建,并通过参数条件机制实现了跨等离子体运行工况的泛化。

arXiv cs.LG79

面向区域的 CLS Token 增强用于细粒度图像检索

arXiv:2610.10991v1 公告类型:cross 摘要:图像检索方法通常依赖从图像中提取的单一全局语义描述符,例如视觉 Transformer 中的 [CLS] token。然而,试图将图像的全部语义信息压缩到单个描述符中,可能会损害下游检索性能,尤其是在细粒度检索任务中。在本文中,我们使用精心选择的空间 token 集合,对新型视觉 Transformer 中的语义 token(全局 [CLS] token 和四个寄存器 token)进行增强,旨在捕获刻画每个语义 token 所包含内容的空间区域表示。我们采用 DINOv2-reg 模型,该模型包含能够自发学习物体及其部件表示的寄存器 token。对于每个“提示”token([CLS] 以及每个寄存器 token),我们找到一个对应的图像“伙伴”图像块 token,并提取一个 N x N 的图像块区域,从而生成一组局部化 ROI token。我们的方法无需任何外部边界框或显著性模块,仅通过将语义 token 与其空间表示区域进行匹配,即可自动捕获重要的感兴趣区域。此外,我们将这些 token 融入受 ColBERT 启发的多向量检索框架,通过逐 token 对齐机制实现细粒度匹配,同时避免存储所有图像块嵌入所带来的巨大存储开销。通过大量实验,我们发现:(1)寄存器 token 编码了有用的细粒度细节,可以补充 [CLS] token;(2)自动池化得到的 ROI token 能进一步提升细粒度区分能力;(3)使用少量 token 的多向量检索性能优于 DINOv2-reg 单向量基线,同时对于大规模搜索仍具备可行性。代码可在 https://github.com/IdhcbIan/Augmenting_CLS_with_ROI_tokens 获取。

arXiv cs.LG77

Omni-Diffusion-Distill:统一多模态扩散大语言模型的少步蒸馏

arXiv:2610.10990v1 发布类型:cross 摘要:统一多模态扩散大语言模型(dLLMs)通过单一架构同时支持图像生成和多模态理解,但其迭代解码需要进行数十至数百次前向传播。现有的少步蒸馏方法大多聚焦于图像生成或文本生成中的一者,因此,如何在保留生成与理解能力的同时,将完全离散的多模态 dLLM 压缩为单一高效的学生模型,仍不明确。我们提出 Omni-Diffusion-Distill,一种统一的两阶段蒸馏框架,在大幅降低统一多模态 dLLM 推理成本的同时,保留强大的生成与理解能力。Omni-Diffusion-Distill 在离散 token 空间中,对图像和文本的生成与理解蒸馏进行统一。第一阶段中,学生模型通过重放缓存的教师模型轨迹进行训练,以跳过部分解码步骤;第二阶段中,学生模型沿自身 rollout 过程中的中间状态进行进一步优化。我们还通过一对碰撞惩罚和熵匹配引导,解决统一蒸馏中的两类性能退化问题:前者减少并行文本解码中的重复,后者防止图像生成中因拟合锐化后的教师分布而导致的熵坍缩。Omni-Diffusion-Distill 在多模态 dLLM 的解码效率与生成、理解性能之间实现了业界领先的权衡,将图像生成的解码步数从 128 步降至 8 步,并将多模态理解的解码步数从 512 步降至 64 步,实现 18.2 倍和 21.2 倍的实际运行时间加速。在这些步数预算下,其在文生图生成任务的 GenEval 和 DPG-Bench 上分别取得 0.828 和 83.0 的分数;在多模态理解任务中,其在 MM-Vet 和 COCO captioning 上的 GPT 评判分数分别达到 20.0 和 57.2(在相同步数下,教师模型为 28.4 的两倍)。

arXiv cs.LG83

AI4Fire: Evaluating Large Language Models on Wildfire Tasks

arXiv:2610.10946v1 Announce Type: cross Abstract: Large language models (LLMs) are entering wildfire management, where overstated evaluations can cost property and lives. How do they perform on wildfire tasks, with and without grounding? Bare means a model receives the task input alone. Grounded means it also receives one task-specific addition: for smoke detection, a smoke-free reference frame from the same camera. AI4Fire runs six core models bare and grounded on five wildfire tasks, zero-shot; a sweep adds 29 more. Our literature search on fire tasks found 138 works; none combines this roster, task coverage, and paired bare and grounded runs. We report three findings. (1) Grounding helped most where the addition carried the answer: a read-only SQL tool lifted every core model's database accuracy from at most 16 to at least 88 percent. (2) Simple rules were hard to beat: no core model outperformed repeating today's staffing count, and two open-weight models mostly copied the median of similar earlier fire-days, a worse forecast. (3) Public releases carry hazards: a fire-danger column separates the holdout perfectly, and 67 aerial fire frames carry smoldering or fire-free labels read from a clipped thermal maximum. We release prompts, responses, scores, code, and the survey record.

arXiv cs.LG76

GHARP: Real-time Gaussian Head Animation from Large-scale Reconstruction Prior

arXiv:2610.10945v1 Announce Type: cross Abstract: We present GHARP (Real-time Gaussian Head Animation from Large-scale Reconstruction Prior), a method that animates 3D human heads in real time from a few input images of a subject and a driving expression signal. We decouple the problem into an identity stage that builds a representation of the subject's geometry and appearance offline, and an animation stage that predicts expression-dependent residuals on top of it at runtime. This separation offers a favorable trade-off with respect to fidelity, quality and runtime: the identity stage can be expensive while the animation stage runs a lightweight network, optimized for mobile devices. Our method performs animation in a semantically structured latent space of a pretrained reconstruction model, where expression changes remain spatially contained, making residual prediction efficient. This reconstruction prior provides a consistent spatial layout, allowing fusion of multiple input views into a compact, fixed-size canonical Gaussian representation. While this two-stage design improves the runtime-quality trade-off, it still inherits a problem common to all expression-driven avatar methods: expression codes describe only the face and thus omit body pose and clothing position, making these regions underspecified in the input. The animation network faces an ill-posed mapping and resorts to averaging over conflicting body appearances, producing blur and temporal flicker. We address this with a body alignment network that learns to align the person's body in the target image with the input reference images, removing the ambiguity from the training signal. Our method achieves state-of-the-art quality on the Ava-256 benchmark while running up to 13x faster on an A100 GPU with 8x fewer Gaussians.

arXiv cs.LG80

StoreBench:用于评估和训练自主运营代理的实时电商环境

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 cs.LG60

Higher-Order Morphology Priors for Quadruped Reinforcement Learning Under Actuator Degradation

arXiv:2610.10934v1 Announce Type: cross Abstract: Actuator degradation turns quadruped locomotion into a coordination problem requiring joints to compensate for lost actuation. Prior work suggests that morphology-aware graph policies improve learning and generalization under body perturbations. We ask whether these benefits can be strengthened by explicitly modeling higher-order mechanical structure. We represent the Unitree Go1 as a cell complex with limb- and body-level rank-2 cells and apply Hodge-based message passing. Under degradation training, the node-edge-face Hodge actor achieves the highest return on unseen actuator degradations, with higher survival and lower velocity-tracking error. These results support higher-order morphology as a useful inductive bias for whole-body compensation under actuator degradation.