全部动态

本页 40 条
10月9日 周五
arXiv cs.LG69

Brain foundation model-guided source-selective domain adaptation for cross-subject EEG decoding

arXiv:2507.21037v3 Announce Type: replace Abstract: Cross-subject motor-imagery electroencephalography (MI-EEG) decoding remains challenging because substantial inter-subject variability can cause both negative transfer from poorly matched source subjects and persistent distribution discrepancies between source and target domains. Existing multi-source domain adaptation methods often incorporate all available source domains or estimate source relevance using signal-level or task-specific representations, while distribution alignment is frequently performed only at the feature level. These limitations may introduce irrelevant source knowledge and fail to preserve class-discriminative structures across subjects. In this study, we propose a brain foundation model-guided multi-source domain adaptation framework (BFM-MSDA) for cross-subject MI-EEG decoding. The method first utilizes representations learned by a pretrained brain foundation model to estimate source--target compatibility and retrieve target-relevant sources. Subsequently, a relevance-weighted dual alignment strategy is applied to the selected sources and the unlabeled target domain. Specifically, Cauchy--Schwarz (CS) divergence is used to reduce discrepancies in marginal feature distributions, while conditional Cauchy--Schwarz (CCS) divergence further aligns class-dependent decision distributions. Source relevance is incorporated into both alignment terms so that more transferable source subjects contribute more strongly to adaptation. Experiments on two public MI-EEG benchmarks achieve average accuracies of 86.16% and 78.41%, outperforming representative cross-subject decoding and domain adaptation methods. Additional experiments with a large source pool further demonstrate that target-aware source retrieval improves scalability while mitigating negative transfer. These results highlight the importance of informed source selection under cross-subject distribution shift.

arXiv cs.LG79

迈向合理的概念瓶颈模型

arXiv:2506.05014v3 发布类型:替换 摘要:我们提出了一种新颖、灵活且高效的框架,用于设计概念瓶颈模型(CBM),使实践者能够在模型进行预测时,将其关于概念—概念($C-C$)关系和概念—任务($C \to Y$)关系的先验知识与信念显式编码并扩展到模型的推理过程中。由此得到的概念推理模型(Concept REAsoning Models,CREAMs)能够在架构层面编码任意类型的 $C-C$ 关系,例如互斥关系、层级关联和/或相关性,以及可能较为稀疏的 $C \to Y$ 关系。此外,CREAM 还可以选择性地纳入一个经过正则化的旁路通道,以补充可能不完整的概念集合,在保持具有竞争力的任务性能的同时,促使预测以概念为依据。为评估此类设置下的 CBM,我们提出了一种与 $C \to Y$ 无关的指标,用于量化预测部分依赖旁路通道时的可解释性。在实验中,我们表明,在不增加额外计算开销的情况下,CREAM 模型支持高效干预,能够避免概念泄漏,并在概念缺失时达到黑盒级别的性能。我们进一步分析了可选旁路通道对可解释性和可干预性的影响。重要的是,即使在仅有有限数量概念可用的场景中,旁路通道也能使 CBM 保持有效。

arXiv cs.LG77

InfiFPO:通过大语言模型中的偏好优化实现隐式模型融合

arXiv:2505.13878v4 公告类型:替换 摘要:模型融合通过轻量级训练方法,将多个具有不同优势的大语言模型(LLM)结合为一个更强大的集成模型。现有模型融合工作主要聚焦于监督微调(SFT),而对偏好对齐(PA)这一提升大语言模型性能的关键阶段探索甚少。目前少数针对 PA 阶段的融合方法(如 WRPO)仅利用源模型的响应输出,舍弃了其概率信息,从而简化了流程。为解决这一局限,我们提出了 InfiFPO,一种用于隐式模型融合的偏好优化方法。InfiFPO 将直接偏好优化(DPO)中的参考模型替换为一个融合后的源模型,该模型在序列层面综合多源概率,既规避了以往工作中复杂的词表对齐问题,又保留了概率信息。通过引入概率裁剪和最大间隔融合策略,InfiFPO 使枢轴模型能够与人类偏好对齐,同时有效地从源模型中蒸馏知识。在 11 个广泛使用的基准测试上的全面实验表明,InfiFPO 始终优于现有的模型融合和偏好优化方法。以 Phi-4 作为枢轴模型时,InfiFPO 将其在 11 个基准测试上的平均性能从 79.95 提升至 83.33,显著增强了其在数学、编程和推理任务中的能力。

arXiv cs.LG70

C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models

arXiv:2502.17920v2 Announce Type: replace Abstract: Pre-trained visual models have become fundamental in computer vision, but they face challenges in continual learning scenarios where data and tasks evolve over time. Low-Rank Adaptation (LoRA) offers efficient fine-tuning capabilities but remains limited for such dynamic environments. Standard LoRA cannot distinguish important subspaces, causing critical knowledge to be overwritten in sequential training. Existing approaches address this by dynamically expanding the set of LoRA adapters, either maintaining a growing pool of task-specific modules or merging new adapters into prior ones, at the cost of unbounded parameter growth or increasing inference complexity. We propose Continual Low-Rank Adaptation (C-LoRA), a method that enables a single, shared LoRA adapter to handle sequential tasks without catastrophic forgetting, without requiring any module selection or fusion at inference. The core of C-LoRA is a learnable routing matrix R that explicitly controls how each rank-one subspace contributes to the weight update. This matrix is decomposed into a stability component (R_base), which preserves knowledge from prior tasks, and a plasticity component (R_delta), which drives adaptation to the current task, providing direct control over the stability-plasticity trade-off. We analyze how R governs gradient flow during sequential training, and demonstrate competitive performance across multiple benchmarks.

arXiv cs.LG72

BEAT: Balanced Frequency Adaptive Tuning for Long-Term Time-Series Forecasting

arXiv:2501.19065v3 Announce Type: replace Abstract: Long-term time-series forecasting supports a wide range of applications, including weather prediction and electricity demand planning. Frequency-domain methods address this task by decomposing observations into components that describe temporal variations at different scales. However, separate representations do not by themselves provide an explicit mechanism for adjusting the training emphasis across components. Under a shared forecasting objective, the frequency-specific networks can retain different levels of coefficient prediction error, motivating an error-dependent adjustment to their gradients. To this end, we propose BEAT (Balanced frEquency Adaptive Tuning), a framework that combines frequency-specific error monitoring with adaptive gradient modulation. We design a Frequency-Specific Monitor that compares predicted and target wavelet coefficients in a common normalized space and expresses each discrepancy relative to a reference error computed from the detail components. We further introduce a Dynamical Gradient Balancer that converts these ratios into positive, bounded coefficients. Components with higher relative errors receive larger gradient weights, whereas those with lower relative errors receive smaller weights. A shared modulation-strength parameter controls the departure from unmodulated training, and the monitoring and balancing operations are used only during training. Experiments on seven real-world datasets show that BEAT achieves competitive performance against state-of-the-art forecasting methods.

arXiv cs.LG76

通过概念推断与数据投毒实现的复杂数据类别机器遗忘

arXiv:2405.15662v2 发布类型:替换 摘要:机器遗忘旨在无需完整重新训练,即可从已训练模型中移除指定训练数据或知识的影响。对于现代图像分类器和大型语言模型(LLMs)而言,这一能力尤为重要,因为重新训练的计算成本可能十分高昂。然而,复杂数据上的机器遗忘仍然十分困难,因为目标信息通常分布在多个语义元素之中。现有方法主要通过移除样本、修改标签或编辑模型参数来减弱遗忘目标的影响。这些方法通常不会明确识别出哪些语义概念将遗忘目标与模型的预测结果或生成响应联系起来。因此,很难确定应修改哪些信息。这种不确定性可能导致目标信息残留,或不必要地影响本应保留的知识。为弥补这一空缺,我们提出了一种概念引导的数据投毒遗忘框架,能够明确识别出对目标类别或知识贡献最大的概念,并利用这些概念指导模型更新。对于图像分类,我们的方法首先通过 Post-hoc Concept Bottleneck Model 识别与类别相关的概念,定位表达这些概念的图像区域,并构造基于替换的投毒样本。对于 LLMs,我们通过多个问题引出目标知识,汇总所得响应中的 Integrated Gradients,以识别持续重要的内容,并对这些内容进行掩码处理,从而构造投毒训练目标。在多个数据集上的实验表明,所提出的框架能够在不同任务中实现有效的机器遗忘,同时在很大程度上保留模型在保留内容上的实用性。

arXiv cs.LG79

通过语言瓶颈进行策略学习

arXiv:2405.04118v4 公告类型:替换 摘要:现代 AI 系统(如自动驾驶汽车和游戏智能体)能够达到超越人类的性能,但通常缺乏类似人类的泛化能力、可解释性以及与人类用户的互操作性。受人类语言与决策之间丰富互动的启发,我们提出了通过语言瓶颈进行策略学习(Policy Learning with a Language Bottleneck,PLLB),这是一种使 AI 智能体能够生成语言规则的框架,这些规则用于捕捉产生高回报行为背后的高层次策略。PLLB 在由语言模型指导的*规则生成*步骤与更新步骤之间交替进行:在更新步骤中,即使一条规则不足以描述完整的复杂策略,智能体也能在规则的指导下学习新的策略。在包括双人信号博弈、迷宫导航、图像重建和机器人抓取规划在内的五项多样化任务中,我们展示了 PLLB 智能体不仅能够学习更具可解释性和泛化性的行为,还能与人类用户分享所学规则,从而实现更有效的人机协作。我们在 https://github.com/meghabyte/bottleneck 提供了实验源代码。

arXiv cs.LG75

CSF:面向运动生成器的上下文安全过滤

arXiv:2610.12467v1 公告类型:cross 摘要:文本条件运动生成器能够生成可跟踪的全身运动,但不具备场景依赖的安全概念:同一个动作可能是针对物体,也可能是针对人。现有的安全防护措施要么检查提示词,要么需要带标签的运动数据,要么施加几何约束;因此,它们无法直接考虑场景上下文如何改变运动的含义。我们提出上下文安全过滤(contextual safety filtering,CSF),这是一种无需训练的过滤器,将自然语言安全规则基于生成器生成的安全和不安全参考轨迹加以落实。对于每条激活的规则,安全和不安全参考轨迹定义一个仿射安全值,并由安全参考跟踪 CBF-QP 对其进行约束。在四个采用不同架构的预训练生成器上,CSF 在所有明确指定和由场景触发的不安全案例中均能激活预期规则,并将危险事件率最多降低 90%,同时保留 88%—100% 的无害运动。我们在真实世界的 Unitree G1 上展示了完整系统,该系统成功阻止了各种场景中的不安全运动,包括与人和物体的交互。

arXiv cs.LG78

均衡的数据饮食:解决机器人控制超大规模 RL 中的探索瓶颈

arXiv:2610.12465v1 公告类型:cross 摘要:通用机器人必须执行从敏捷运动到灵巧操作的广泛任务。尽管仿真到现实强化学习(RL)已被证明是实现这一目标的有用工具,但当前的 RL 流程依赖大量工程工作,并且需要针对每个任务设计结构先验,例如塑形奖励和示范。近期研究表明,在若干操作问题中,将多样化的模拟器重置与大规模并行仿真相结合,可以减轻大量工程负担。然而,我们发现,将这一范式直接扩展到更精确或更具动态性的任务上,仍然并非易事。尽管模拟器重置有助于探索,但在这一分布上进行均匀采样,会将越来越大比例的学习经验浪费在策略已经掌握或尚且无法尝试的任务配置上。这使得人们难以看到扩展 RL 并行环境所带来的预期收益,因为一个批次中的大量学习信号都会在训练过程中被浪费。为缓解这一问题,我们提出 Success Guided Sampling(SGS),这是一种简单的自适应采样器,将 RL 训练集中在策略能力边界附近的任务配置上。这样,基于大规模仿真的 RL 就能充分利用一个批次中的经验,从而更有效地扩展到大规模并行仿真。在使用多达 $2^{20}$(超过一百万个)并行环境的实验中,SGS 使 RL 能够解决具有挑战性的多地形四足机器人运动任务和接触密集型装配任务,而以往方法无法解决这些任务。最后,我们将学习到的操作策略蒸馏为基于 RGB 的策略,并展示了其在真实硬件上的若干项具有挑战性的装配任务中的零样本迁移。项目网站:https://sgs-rl.github.io/。

arXiv cs.LG80

One Block, Multiple Depths: Recurrent Vision Transformers with Depth-Programmed Experts

arXiv:2610.12448v1 Announce Type: cross Abstract: In this work, we show that a single Transformer block, applied recurrently, can match the accuracy of a full-depth vision encoder at comparable inference FLOPs without intermediate feature distillation. reViT restores depth-specific transformations by representing the FFN at each recurrent depth as a convex combination of a small shared expert bank. A continuous normalized-depth coordinate programs this mixture, defining a resampleable trajectory through FFN parameter space. We evaluate this design in two regimes: supervised ImageNet-1k training and distillation from a DINOv2 teacher. Across both regimes, controlled adaptations identify weight-space merging as the strongest tested MoE family at a matching one-FFN budget, ahead of the token-dispatch and output-mixture alternatives. Trained from scratch, reViT-B/16 attains DeiT III accuracy with about 70\% fewer stored parameters. An 8-experts model distilled using only the teacher's output features retains nearly all of its DINOv2 teacher's linear-probe accuracy and transfers across classification, segmentation, and depth prediction. Elastic-depth training allows one checkpoint (trained model) to operate at multiple tested depths by resampling the same normalized coordinate interval. For fixed-depth deployment, the recurrent block can be materialized as a conventional dense graph, removing online routing and merging without changing the one-FFN-per-depth compute but expanding deployment storage.

arXiv cs.LG69

Density Ratio Estimation with Stein Displacement Fields

arXiv:2610.12437v1 Announce Type: cross Abstract: Density ratios quantify distribution shift from a probability-mass point of view, whereas displacement fields describe, from a dynamical point of view, how one distribution is transported onto another. Although both offer complementary insights, they are usually estimated separately, and converting one into the other requires post-processing. In this paper, we estimate the density ratio between a target and a base distribution by parametrizing it through a displacement field acting on the base: the log-ratio is modeled as minus the Stein operator of the base applied to the field, up to a normalizing constant. This gives both statistical and dynamical descriptions of the distribution shift through a single convex optimization problem. Iterating this estimate-and-move step gives two inference algorithms: push-forward moves the model and corrects a pretrained sampler without retraining it, whereas pull-back moves the data closer to the base and fits a transformation model one layer at a time. Applications to distribution shift in simulation-based inference and to nonlinear independent component analysis illustrate the benefits and limitations of the approach.

arXiv cs.LG81

VioLA:从人类数据中学习通用人形机器人控制策略

arXiv:2610.12435v1 发布类型:cross 摘要:教会人形机器人用全身遵循指令面临两大障碍。其动作空间庞大且高度耦合:机器人必须在保持平衡的同时协调腿部、手臂和手指的运动,这使得关节级动作难以学习。此外,人形机器人示范数据十分稀缺,因此当前的通用人形机器人策略无法开箱即用地遵循新指令,而是在部署前针对每项任务使用遥操作示范进行微调。人类示范的数量要多得多,但人的动作并不是机器人指令。我们通过改变通用策略的预测目标,消除了这两个障碍。我们提出了 VioLA,一种通用人形机器人策略,它预测身体和手部运动潜变量,而不是关节指令。预训练的身体控制器和手部控制器在机器人上执行这些潜变量。与之对应的运动编码器将人类和机器人的运动映射到相同的潜空间中。因此,人类记录可以在策略的动作空间中获得标注,训练示范池包含 1.406 亿帧,其中 93.2% 来自人类。由此,VioLA 无需针对特定任务进行微调,即可在真实机器人上以零样本方式遵循运动指令,成功率达到 100%;相比之下,GR00T N1.7 和 $\Psi_0$ 的成功率分别为 16.7% 和 0%。在无需针对特定任务进行微调的情况下,它在操作任务上的成功率也达到 88.6%。同样的方法适用于两个 VLA 和一个 world-action 模型骨干网络。仅使用人类示范训练的通用策略即可在真实机器人上以零样本方式完成运动任务。代码和检查点将会发布。

arXiv cs.LG78

FAITH: Feasibility-Aware Safety-Filtered RL for High-Dimensional Systems

arXiv:2610.12432v1 Announce Type: cross Abstract: Safe reinforcement learning commonly places safety and task performance in the same policy objective, where they can introduce competing updates. Safety filters separate them at action execution, but classical designs require an analytic safety function and dynamics model, and standard minimal-intervention filters are myopic to long-horizon task return because they minimize only instantaneous action deviation. Hard projections are also undefined when no safe action exists. We present FAITH, a feasibility-aware, model-free framework that approximates the optimal state-action safety value and amortizes minimal-intervention filtering with a feedforward network. The task policy optimizes the task return through the filtered dynamics, which recovers the feasible constrained problem without a competing safety term in the task-policy update. When no action satisfies the learned safety condition, the same filter approaches the action with minimum predicted peak harm. On a double integrator example and a Safety Gym environment, FAITH achieves the highest return among methods with no feasible-start violations and matches the lowest harm from infeasible starts. On a 29-DoF humanoid, it reaches a 99.95% safety rate while retaining 97% of the unfiltered return in Walking-Avoid, and obtains the highest measured safety rate in Push-Avoid by learning to sacrifice balancing and fall away from the protected region. The same policies are also demonstrated on a real-world Unitree G1 humanoid.

arXiv cs.LG63

Toward Joint Optimization of Circuit Depth and Training Data Size in Adaptively Grown Quantum Classifiers

arXiv:2610.12428v1 Announce Type: cross Abstract: Building a quantum model involves a tradeoff: how complex the circuit should be, and how much training data it needs. Caro et al. show that models with fewer trainable gates need less training data to generalize well. Q-FLAIR shows that a quantum feature-map circuit can be grown gate-by-gate, stopping once further growth stops improving the training loss. We ask whether these two results combine into a predictable scaling law. Does Q-FLAIR's own stopping rule pick larger or smaller circuits as training data grows? Does the resulting generalization behavior track Caro et al.'s bound? We reimplement Q-FLAIR's growth mechanism faithfully, including its analytic reconstruction and exact stopping rule. We run it on full-resolution (784-pixel) MNIST 3-vs-5 classification, at five training-set sizes from N = 2000 to 10000. We then fine-tune each resulting circuit, so we can measure Caro et al.'s notion of active gates, K. We find no predictable relationship between training-set size and the circuit size Q-FLAIR converges to. Circuit size and test accuracy both vary non-monotonically with N, and seed-to-seed variance is nearly as large as any trend across N. The empirical generalization gap never exceeds Caro et al.'s bound in 14 of 15 runs, so the bound holds as a valid guarantee in those runs. But the gap correlates only weakly with the bound's value (r = 0.12). This shows that K does not explain most of the variation we observe. Why a valid guarantee can coexist with such weak predictive power remains an open question, and answering it may be necessary before circuit depth and training data size can be jointly optimized in practice.

arXiv cs.LG78

超越时空先验:一种用于稠密对应匹配的可泛化方法

arXiv:2610.12421v1 公告类型:cross 摘要:稠密对应匹配长期以来受限于简化的时空先验,例如平滑运动和刚性几何。尽管这些假设对经典任务十分有效,但在图像编辑和参考引导生成(IEG)中却会失效,因为这类变换能够保持视觉身份,同时打破物理连续性。为了在此类变换中建立保持身份的对应关系,我们提出了 FreeMatching,这是一种可泛化的框架,将生成式和语义基础表征与来自经典数据集、跟踪视频和合成场景的异构监督相结合。教师引导的迭代细化进一步提升了 FreeMatching 在 IEG 中的对应关系效果,且无需稠密对应标注。在实验中,单个 FreeMatching 模型显著提升了具有挑战性的 IEG 图像对上的对应质量,同时在经典基准测试中保持了有竞争力的性能。此外,我们还展示了其作为评估身份保持程度的定量指标的实用性,其得分与人类判断具有较高相关性。代码可在 https://github.com/luping-liu/FreeMatching 获取。

arXiv cs.LG82

WOVEN:将视觉世界建模融入多模态LLM

arXiv:2610.12417v1 公告类型:cross 摘要:多模态大语言模型(MLLMs)在空间、具身、物理和时间推理方面表现不佳。我们假设,这些失败反映出一种共同的视觉状态转移推理能力缺陷,并检验该能力能否作为一种共享训练原语:不同模型可以从不同监督来源中学习这一能力,并将其复用于不同任务,同时采用系统化的训练方案。现有基准分别记录了这些缺陷,但不支持跨场景、动作和推理操作的受控比较。因此,我们提出WOVEN——一个用于视觉状态转移推理的训练数据源和基准。WOVEN按照场景、动作和推理类型组织转移监督,使用由视频预训练生成模型产生的多样且逼真的演化过程数据:涵盖20种场景类型、5种动作类型和8种推理类型,共36,076个样例。我们首先评估了38个前沿MLLMs(例如GPT-5.4和Qwen3-VL-235B-A22B),发现了显著且系统性的能力缺陷:即使是最强的模型,表现也远低于人类;这些失败在不同模型系列中反复出现,并且随着模型规模扩大仍然存在。随后,我们在多个规模上使用WOVEN训练MLLMs,发现它们学会了一种能够广泛迁移的共享能力:仅分别使用约2,000条样本训练的子集,就能使26个外部基准中的22个得到提升,最高提升27.3个百分点;此外,在保持相当准确率的情况下,WOVEN数据可以替代某项任务自身训练数据的30%至50%。进一步的受控比较得出了一套视觉世界建模训练方案,并在预留基准上进行了前瞻性验证:应根据监督所教授的推理操作来选择监督数据,而不是根据其呈现的动作、场景或领域来选择;同时,为增强鲁棒性,应优先选择视觉状态变化更大的数据。我们的工作确立了视觉状态转移推理作为一种可复用的基础能力,用于在MLLMs中开展系统性的视觉世界模型训练。

arXiv cs.LG79

使用价值表示预测对齐泛化

arXiv:2610.12410v1 公告类型:cross 摘要:LLM 开发者会对模型进行后训练,使其表现出亲社会价值观和行为特征,并将这些价值观和特征列举在对齐目标中。然而,尽管近期的后训练发展已经产生了在对齐评估中得分很高的模型,在狭窄行为集合上训练模型,仍会以意料之外的方式影响其在未见过的上下文和环境中的行为。本文提出对齐泛化预测这一任务,即预测将模型微调为遵循某一给定价值观后,其行为在广泛的留出价值观集合上会如何变化。我们针对现代对齐目标中发现的 66 种价值观,对对齐泛化效应进行了大规模分析,并对多种表示技术在对齐泛化预测任务上的表现进行了基准测试。我们发现,在上下文中应用价值观时基于模型激活的表示,显著优于基于价值观文本描述的方法。具体而言,表现最佳的基于激活的方法与我们的泛化矩阵达到了 0.45 的相关性,而基于描述的基线方法仅为 0.05。随后,我们通过使用这些表示来衡量多价值对齐目标中的价值观之间的相似程度,展示了能够预测对齐泛化的表示在下游任务中的适用性;我们发现,这种相似程度与模型鲁棒性存在显著相关性。最后,我们展示了关于共享的、与模型无关的价值空间的初步证据,并据此构建了首个以经验泛化动态为基础的 LLM 价值观分类体系。我们的工作证明了研究 LLM 中价值观泛化的重要性,以及将其应用于更加经验化的模型行为设计与训练的价值。

arXiv cs.LG73

HRIL:通过高阶张量建模学习多模态协同

arXiv:2610.12393v1 公告类型:cross 摘要:自监督多模态表征学习已在各个领域取得显著成功,但由于跨模态交互的复杂性,捕获协同信息仍然具有挑战性。不同于各个模态之间分别存在的共享信息,协同信息产生于这样的情形:只有在多个模态的联合配置下,任务相关信号才会出现,而从任何单一模态中都无法恢复这些信号。本文聚焦于如何在多模态表征中保留这类协同信号的信息容量。关键观察是,协同信息体现在模态之间的高阶统计依赖中,这为显式建模联合交互提供了一个具有理论依据的目标。基于这一认识,我们提出高阶表征与信息学习(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,从而保留用于捕获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上的实验表明,相较于现有多模态对比学习方法,HRIL consistently improved,并在以协同交互为主导的任务上取得了显著提升。代码发布于 https://github.com/brightest66/HRIL。

arXiv cs.LG79

OnTrack:通过流式结构感知最优传输实时监控并干预 LLM Agent 轨迹

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 cs.LG78

Subspace Uncertainty and Sharp Sampling Thresholds on the Boolean Cube

arXiv:2610.12358v1 Announce Type: cross Abstract: We study Gaussian regression under squared population $L_2$ loss in a known $m$-dimensional subspace of degree-at-most-$k$ functions on the $d$-dimensional Boolean cube. Random inputs can undersample regions essential for prediction, delaying the parametric rate even when the model is known. For fixed $q_0<1/2$, $1\le k\le q_0d$, and sufficiently large fixed $A$, the worst-subspace sample threshold for minimax error $A\sigma^2(m+t)/n$ with confidence $1-e^{-t}$, $t\ge\log4$, is \[ N=(m+t)\exp\{E_{d,k}+O(k^{1/3})\}, \quad E_{d,k}=d\Psi(k/d), \] where $\Psi(q)=\log2-\mathsf H(\tfrac12-\sqrt{q(1-q)})$ and $\mathsf H$ is binary entropy with natural logarithms. The upper bound holds for every feasible $m$; the matching lower bound holds when $m\le\binom d{\lfloor k^{1/3}\rfloor}$ or $t\ge m$. We sharpen the Polyanskiy--Samorodnitsky uncertainty principle in two respects. First, for fixed leakage $\rho\in(0,1)$, the smallest set carrying a fraction $1-\rho$ of a nonzero degree-at-most-$k$ polynomial's energy has probability $\exp\{-E_{d,k}+O_{\rho,q_0}(k^{1/3})\}$. An Airy-kernel construction proves that the remainder cannot be $o(k^{1/3})$ in general. Second, we construct a subspace of dimension $\binom d{\lfloor k^{1/3}\rfloor}$ such that every function in the subspace has at least a fraction $1-\rho$ of its energy on the same set, whose probability is at most $\exp\{-E_{d,k}+C_{\rho,q_0}k^{1/3}\}$. For sufficiently large $k$, this set is a Hamming ball. A striking consequence is an exponential cost of noise: the parametric rate can require $(m+t)4^k\exp\{-O(k^{1/3})\}$ samples, whereas $O((m+t)2^k)$ suffice for noiseless identification. As $k\to\infty$ with $k/d\to0$, the noisy threshold is $(m+t)\exp\{2k+o(k)\}$.

arXiv cs.LG75

克服先验壁垒:长尾分布下的监督微调

arXiv:2610.12345v1 公告类型:交叉收录 摘要:监督微调(SFT)使预训练大语言模型(LLMs)适应下游任务,但预训练对任务所需概念的支持程度可能存在显著差异。高频概念更可能被充分学习,而稀有概念的表征可能仍然薄弱。我们提出了一个名为“先验壁垒”的新概念,用于量化预训练模型相较于目标概念对竞争概念的支持强度。我们观察到,先验壁垒呈长尾分布,使头部概念和尾部概念在 SFT 中处于不同的起点: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了长尾先验壁垒下 SFT 的预测风险界,明确刻画了先验壁垒与累积的 SFT 证据如何共同决定预测性能。针对这种依赖于先验壁垒的需求,我们提出 PASS,一种自适应 SFT 指令选择方法。该方法构建从参考内容中提取的概念,估计每条指令提供的区分性证据,并将选择预算自适应地分配给当前选择尚未充分覆盖的概念。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,以及哪些概念需要额外监督。实验表明,在四种骨干模型与预算组合设置下,我们的方法均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。

arXiv cs.LG69

VFold: Symmetry-Aware Cross-Layer Value Cache Compression

arXiv:2610.12338v1 Announce Type: cross Abstract: While caching key-value (KV) states accelerates Large Language Model (LLM) decoding, this cache can dominate memory usage at long context lengths. One solution is to compress this memory by exploiting inter-layer cache similarities. However, most existing techniques necessitate architectural changes to LLMs and incur substantial overhead. In this work, we propose a symmetry-aware value cache merging strategy that reduces cache memory while avoiding both harmful performance degradation and architectural overhead during decoding. Furthermore, we show that this approach can be exploited alongside existing cache compression techniques, composing with high-ratio quantization or key cache pruning to reach compression ratios that neither method reaches alone, with minimal additional cost. Ultimately, our findings reveal a major source of underutilized capacity in the value cache, offering a simple yet highly effective direction for scaling context windows under memory constraints.

arXiv cs.LG74

asdex:JAX 中的自动稀疏微分

arXiv:2610.12336v1 公告类型:交叉收录 摘要:科学计算和机器学习中的许多任务都需要计算函数的雅可比矩阵或海森矩阵。自动微分(AD)能够以机器精度计算这些导数,但显式构建一个稠密的 $m \times n$ 雅可比矩阵需要执行 $n$ 次前向模式或 $m$ 次反向模式的 AD,每次计算一列或一行。对于一大类函数,每个输出仅依赖少量输入,因此其导数矩阵是稀疏的。自动稀疏微分(ASD)通过四个步骤利用这一结构:检测与输入无关的稀疏模式;通过图着色,将可以共享一次 AD 计算的列或行分组;执行压缩微分,每种颜色进行一次 AD 计算,得到压缩的导数矩阵;最后将其解压缩,恢复原始稀疏模式。颜色数量以及相应的 AD 计算次数通常与问题维度无关:例如,具有 $b$ 条连续带的带状雅可比矩阵,无论规模多大,都只需要 $b$ 种颜色。asdex 是广受欢迎的 JAX 生态系统中首个独立的 ASD 工具包。它通过 asdex.jacobian 和 asdex.hessian,为 jax.jacobian 和 jax.hessian 提供可直接替换的稀疏版本。

arXiv cs.LG75

RiCo:通过局部接触推理进行刚体交互的神经模拟

arXiv:2610.12333v1 公告类型:交叉发布 摘要:准确模拟刚体交互对于预测性物理世界模型至关重要。尽管物体动力学建模近期取得了进展,但捕捉表面之间的局部接触如何影响物体运动仍然具有挑战性。端到端世界模型在整个场景或物体层面预测交互,而在实际中,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。基于这一观察,我们提出 Rigid-body Contact Reasoning(RiCo),通过接触表面点的稀疏邻域来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关系、运动和物理属性相结合,随后在物体各点之间进行推理,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 既保留了细粒度的交互细节,又避免了对场景中每一对点进行建模的开销。这些特性使 RiCo 能够实现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 帧的位置误差和姿态误差分别降低了 31–35% 和约 38%。此外,RiCo 实现了较高的接触保真度,其穿透时间和平均穿透深度相较于真值的差异分别为 11.0% 和 2.22 毫米。RiCo 还能够从小规模训练场景零样本泛化到包含 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提供了从仿真迁移到现实的初步证据。

arXiv cs.LG77

用于治疗效应估计的预测赋能数据融合

arXiv:2610.12332v1 公告类型:交叉收录 摘要:随机对照试验(RCT)能够在不受混杂影响的情况下识别治疗效应,但其规模通常较小;相比之下,观察性研究(OBS)规模较大,却可能受到混杂影响。针对平均治疗效应(ATE)和条件平均治疗效应(CATE),已有许多将小规模 RCT 与大规模 OBS 相结合的估计方法。然而,现有的 ATE 估计方法要么对 OBS 作出假设,要么未能充分借助 OBS 提升统计效能。与 ATE 相比,针对 CATE 的研究较少。现有的 CATE 方法要么假设 OBS 不存在混杂,要么依赖混杂函数模型,要么接受偏差以换取更低的方差。因此,我们提出了一个框架,无需对 OBS 作出特殊假设,即可融合 OBS 和 RCT,在保持基于 RCT 的估计无偏性的同时,借助大规模 OBS 提升估计精度。基于这一原则,我们构建了具有闭式权重和置信区间的 ATE 估计器 AIPW-Fusion,以及两个 CATE 学习器 DR-Fusion 和 R-Fusion。实验验证了我们的研究结果。

arXiv cs.LG70

Prior or Feedback? What an LLM Uses When Adapting Neural Operators

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 cs.LG66

Spatial Pattern Formation from Multi-Agent Learning in Public Goods Dilemmas

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 cs.LG79

借助 ARGUS 解锁调控基因组:用于解释单核苷酸变异的证据约束智能体框架

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 cs.LG70

ISBO: Scalable Spatio-Temporal Bayesian Optimization with Log Gaussian Cox Process Models via the INLA-SPDE Approach

arXiv:2610.12213v1 Announce Type: cross Abstract: Bayesian Optimization (BO) is a popular method for efficiently optimizing expensive black-box objectives. However, BO utilizing standard Gaussian Processes is ill-suited for doubly stochastic Cox Processes that are often used in spatio-temporal problem spaces. We introduce INLA-SPDE Spatio-Temporal Bayesian Optimization (ISBO): the first scalable BO framework for spatio-temporal data, that models the log-intensity with a Log-Gaussian Cox Process(LGCP) and performs inference via Integrated Nested Laplace Approximation and Stochastic Partial Differential Equations (INLA-SPDE) approach. Using a Matern field on meshes yields a sparse Gaussian Markov Random Field, where INLA provides fast and accurate posterior inference throughout sequential optimization. ISBO stably locates high-intensity regions and the peak of the latent intensity with minimal evaluations. A time-varying Upper Confidence Bound acquisition with masking avoids revisits, while penalized-complexity priors regularize early rounds. Experiments on synthetic and real-world spatio-temporal datasets show accurate peak discovery, intensity recovery, and substantial speedups over an RKHS-based baseline, positioning ISBO as a practical choice for BO with point-process data.

arXiv cs.LG77

SciTBERT:面向科学与技术语言处理的时间一致性语言模型家族

arXiv:2610.12207v1 公告类型:交叉发布 摘要:预训练 Transformer 模型正越来越多地用于研究科学与技术进步。针对论文或专利文本调优的编码器,在科学与技术领域的下游分类、回归和相近性任务中,表现优于通用模型。然而,由于这些预训练模型固有的前视偏差和领域偏差,它们在研究科学、技术及其交叉领域的时间依赖特征或历史档案特征方面的适用性受到限制。这些局限源于训练语料在时间分布和文本来源分布上缺乏约束。我们提出 SciTBERT:一个基于 BERT 衍生架构的时间一致性语言模型家族,使用科学论文、专利和高质量教育类网页文本进行训练,训练数据截止日期覆盖 2013 年至 2025 年间的每一年。我们还利用论文和专利引用,以保持时间一致性的方式对这些模型进行后训练,构建了 SciTBERT-CI 模型家族。我们发现,即使训练数据受到语料中较早年份截止日期的限制,这些模型的表现通常仍优于此前的领域专用编码器模型。为进一步探究此类模型在多大程度上能够学习连接科学与技术的表征,我们提出了 PatRepEval 基准,其中包含一组面向科学与技术交叉领域的专利相关文本嵌入任务。模型在涵盖论文和专利的多种分类、回归和检索任务中的表现,凸显了使编码器模型表征与其下游任务的领域分布相匹配的重要性;同时,保持时间一致性的编码器能够达到或超过在无时间约束条件下训练的模型的表现。

arXiv cs.LG74

基于扩散积分分数的最快变化检测

arXiv:2610.12200v1 公告类型:交叉收录 摘要:经典 CUSUM 依赖于底层分布的对数似然比,而通常无法仅凭有限的变化前后样本计算该比值。我们提出扩散积分分数 CUSUM(DI-SCUSUM),这是一种无需训练的检测器。我们向样本添加高斯噪声,构造两个平滑密度估计,并精确计算其 Hyvärinen 分数,无需训练分数网络。对于每个新到达的观测,我们采样一个扩散时间,对观测施加扰动,并将重要性加权的分数差用作 DI-SCUSUM 递推中的增量。在观测服从固定经验分布的假设下,变化后的平均增量与从平滑后的变化后经验分布到平滑后的变化前经验分布的 Kullback-Leibler(KL)散度成正比。我们确立了误报的指数缩放关系和一阶延迟界;在阈值和增量缩放固定时,该延迟界与 KL 散度成反比。在经过校准的各向异性高斯仿真中,DI-SCUSUM 的表现几乎与似然比 CUSUM 相当,相较于基于分数的 CUSUM,测得的检测延迟降低了约 91%。在 MNIST 和 Oxford-IIIT Pet 上,DI-SCUSUM 在相近的误报水平下,其经验条件检测延迟也低于 SCUSUM。

arXiv cs.LG71

Learning to Plan by Looking Back: Hindsight Hierarchies for Training Reasoning Models

arXiv:2610.12168v1 Announce Type: cross Abstract: We introduce a self-improvement loop for reasoning models based on the following observation: Even when the difficulty of a problem exceeds the model's current solving abilities, an additionally supplied solution might enable the model to extract useful solution ideas in hindsight. We operationalize this by jointly training the same model to exhibit the following three capabilities: predicting solution ideas from problems alone, reverse-engineering ideas from problems and known solutions, and solving problems using provided ideas. The loop alternates between reverse engineering such ideas from problems with supplied solutions and using these ideas as additional supervision for joint training of all three capabilities. We give a formal specification of our method and a concrete instantiation for interactive theorem proving in the Lean theorem prover; empirical evaluation remains future work.

arXiv cs.LG79

海报:LLM 蒸馏推断的初步研究

arXiv:2610.12137v1 发布类型:cross 摘要:未经授权的模型蒸馏,即利用专有大型语言模型(LLM)的输出训练模型,正日益成为模型提供商面临的威胁。我们研究蒸馏推断:判断一个可疑模型是从另一个模型蒸馏得到的,还是独立训练得到的。我们将这一问题表述为假设检验,并通过训练影子模型来估计每个假设下的预期行为:蒸馏影子模型从教师模型的推理轨迹中学习,而独立影子模型仅从参考答案中学习。审计者衡量每个模型对教师模型推理输出的预测接近程度,然后利用影子模型将可疑模型的得分转换为经过校准的 p 值。在一项以 Qwen2.5-7B 为教师模型、以 Llama-3.2-3B 为可疑模型的初步研究中,我们的检验在显著性水平为 0.02 时达到了 1.0 的真阳性率。这些结果表明,利用蒸馏推断检测蒸馏攻击是可行的。

arXiv cs.LG82

复习全部,什么也记不住:Attic-KV 复习将被读取的内容

arXiv:2610.12133v1 公告类型:交叉发布 摘要:许多键值(KV)缓存在人们尚不知道会向其提出什么问题之前就已被压缩,例如为检索而缓存的文档、跨请求共享的提示前缀,以及长对话的记忆。当前主流方法通过复习为 KV 条目评分:模型重新阅读上下文,并保留其关注的条目,假设缓存对上下文的复习越完整,记忆效果就越好。我们表明,在预算紧张时,这一假设会适得其反:复习全部,什么也记不住。在 3% 的保留率下,重读全部上下文只能在 RULER 上保住原有 96.5 分中的 31.5 分,而在 LongBench 的自然文本任务上,其表现甚至不如完全不复习的方法。原因在于,缓存会保留它所复习的内容:重读将预算分散到整个上下文,使答案本身对应条目的留存概率仅略高于随机水平。如同考前的学生,缓存通过自测能比通过重读记住更多内容。由此得出两项原则:复习将被读取的内容,并根据内容量决定复习量。我们将这两项原则实现为 Attic-KV(简称 Attic),这是一种无需训练的复习方法,模型使用引用上下文的问答对进行自测,同时辅以数量随内容自适应调整的锚点词元。仅改变复习方式,就能提升三种采用不同评分方式的基础方法:在我们于 RULER 和 LongBench 自然文本任务上测试的全部八种设置中,单独使用 Attic 均是表现最佳的无需训练方法;将其接入基于梯度的 KVgrad 和经过训练的 RestoreKV+ 后,二者的得分分别最高提升 17.1 分和 28.1 分。随着预算缩减,其优势不断扩大,在 3% 的保留率下,相比完整重读的优势达到 41.9 分,而且其压缩速度也快于重读全部上下文。

arXiv cs.LG79

用于聚合物电解质离子的保结构神经密度泛函

arXiv:2610.12132v1 公告类型:交叉收录 摘要:预测非均匀聚合物电解质的结构与响应,需要一种既能保持分子尺度精度,又能跨空间尺度和几何构型迁移的离子关联描述。我们开发了一种用于电解质的神经密度泛函,它保持空间对称性、热力学可积性和 Noether 恒等式,并在稳定、非临界的体相状态下恢复完全屏蔽。其非线性密度依赖性能够捕捉成对闭合近似所遗漏的浓度依赖关联,包括强耦合条件下长波长粒子数涨落从增强到受抑制的交叉转变。该泛函能够描述未用于训练的盐浓度下的密度分布,并预测体相结构因子和长波长粒子数响应。该泛函仅使用分子动力学模拟得到的平面密度和内力分布进行训练,即可预测更大空间区域及二维外场中的离子结构。在相同的离子数据上,其精度高于另外三种神经密度泛函架构;当训练模拟次数减少至四分之一时,它仍能保持精度,而表现最佳的其他架构的误差则增加约三分之二。这种空间可迁移性为将分子关联与更大尺度的连续介质预测相连接提供了必要基础。

arXiv cs.LG76

LLM 水印检测能否公开?

arXiv:2610.12106v1 公告类型:交叉发布 摘要:为大型语言模型添加水印是追溯聊天机器人和智能体输出的常用手段,但检测器仍未公开,因为公开检测器可能让攻击者利用其反馈进行针对性修改。然而,水印已经容易受到无先验信息的篡改攻击。因此,我们首先量化了在部署场景中,公开检测器在从词元级分数到二元判定的不同访问权限下,是否会带来额外风险。其次,我们提出了一种拆分密钥的公私结合水印方法,通过公开检测器暴露其中一个密钥,同时保留另一个密钥用于完整验证和取证。知情攻击者只能改变公开信号,从而造成公开分数与私有分数之间的不平衡。我们为这种不平衡引入了一项统计检验,并将其与完整密钥判定相结合,构建了一个两阶段机制。第三,我们在广泛的水印移除和伪造攻击下评估了该拆分密钥方法,对比了无先验信息与利用检测器信息两种设置。公开检测仅在修改预算较小时提高了水印移除效果,因为普通改写就已能以更低的质量代价去除水印;但公开检测确实使伪造成为可能,而私有处理流程能够识别此类伪造。总体而言,公开一半水印能够实现透明性和互操作性,而且针对已公开部分的篡改仍可被检测到。这限制了提供方的风险,并对将检测器完全保密的必要性提出了质疑。

arXiv cs.LG71

DVLA-RL++:面向小样本学习的基于强化学习门控的双层次视觉语言对齐

arXiv:2610.12095v1 公告类型:交叉发布 摘要:小样本学习旨在利用有限的标注样本识别新类别。近期研究引入文本语义,以弥补视觉观测不足并改善类别表示。然而,较高的图像文本一致性可能同时反映物体的固有属性和偶然出现的上下文,使支持集原型容易受到上下文污染。为解决这一问题,我们提出 DVLA-RL++,在 DVLA-RL 的基础上引入互补语义净化(CSP)和反事实强化学习门控(CRG)。具体而言,CSP 根据带标注的支持样本生成固有属性描述与干扰因素描述,并比较这些描述与每个支持样本词元的一致性。随歧义程度变化的拒绝间隔引导稀疏证据分配,而固有语义锚点则填补未分配的权重,在视觉证据不可靠时提供兜底。CRG 使用兼顾识别性能与干扰暴露程度的奖励,学习逐层语义融合强度。在同一小样本学习任务中独立执行的参考轨迹提供配对学习信号。理论分析将保留证据及锚点质量与原型稳定性联系起来,并确立了同策略梯度无偏估计的条件。在标准、细粒度和跨域基准上的实验表明,该方法达到最先进的准确率,较 DVLA-RL 平均提升 1.4%。项目页面:https://peacelwh.github.io/TPAMI27-DVLA-RLpp/。

arXiv cs.LG76

用于变分序贯蒙特卡罗的可微系统重采样

arXiv:2610.12094v1 公告类型:cross 摘要:粒子滤波器是非线性状态估计的标准工具,但其重采样步骤是离散的,从而阻碍了在变分序贯蒙特卡罗中进行基于梯度的学习。我们提出可微系统重采样(Differentiable Systematic Resampling,DSR),这是一种由温度控制的系统重采样松弛方法,在实现完整梯度流的同时,保留了系统重采样基于 CDF 排序的带状结构。当温度趋近于零时,DSR 收敛于精确的系统重采样;我们还证明了其所引入偏差的逐点指数收敛速率。与基于最优传输的可微重采样相比,DSR 无需迭代求解器,计算开销显著更低。在随机动力系统和真实世界手写数据上的实验表明,DSR 在滤波和动力学学习性能方面达到相当或更优的效果。

arXiv cs.LG75

Perception Test 2026:挑战总结及向城市级视听推理的扩展

arXiv:2610.12081v1 发布类型:cross 摘要:延续 Perception Test 挑战系列,我们在瑞典 Malm\"o 举办的 2026 年欧洲计算机视觉会议(European Conference on Computer Vision,ECCV)上以研讨会形式组织了第四届挑战赛。本届赛事聚焦空间智能,设有四个不同赛道:源自 Perception Test 基准的统一多选题视频问答和基于视频的定位问答,以及基于城市规模步行导览视频的两个新赛道(KilometerAudio 和 KilometerVision)。本文介绍了城市规模赛道所采用的新基准,并总结了所有赛道的获胜方案,其中包括一个在所有赛道中参赛且表现令人满意的通用模型。新增城市规模赛道的获胜方案表明,复杂的空间推理和多模态推理可以通过高成本的智能体流水线来解决,但对于单独使用的多模态模型而言,这仍然很困难。

arXiv cs.LG68

Diffusion Removes Langevin's Conditioning Dependence: A Sharp Gaussian Analysis

arXiv:2610.12052v1 Announce Type: cross Abstract: Despite their empirical success, why diffusion models overcome the bottlenecks of classical score-based samplers remains unclear. In this work, we leverage Gaussian distributions to isolate this phenomenon. We establish 2-Wasserstein convergence bounds for optimized hyperparameters, showing that diffusion processes achieve a sampling error of $O(\sqrt{d\lambda_{\max}}\log N/N)$, where $d$ is the dimension, $N$ the number of sampling steps, and $\lambda_{\max}$ the largest eigenvalue of the target covariance matrix. Unadjusted and underdamped Langevin dynamics suffer from an additional $\sqrt\kappa$ factor, where $\kappa$ is the condition number. These rates follow from spectral bounds which are sharp: we confirm them via matching first-order asymptotics as $N\rightarrow\infty$. Our analysis provides a rigorous characterization, in the Gaussian setting, of how time-dependent score trajectories remove condition-number dependence during sampling. By contrast, in the learning phase, we show that estimating the unnoised score by gradient descent leads to essentially the same estimator as estimating a noisy score, which suggests that the benefits of noising do not come from the learning phase.