论文

介绍 VisualPatchWorld，将世界动态表示为代码，先通过短时主动探查选择定性动力学形式，再通过最小化多步预测误差，从记录的状态-动作轨迹中拟合该形式的自由参数。VisualPatchWorld is introduced, which represents world dynamics as code and first selects a qualitative dynamical form with short active probes, then fits that form's free parameters from recorded state-action traces by minimizing multi-step prediction error.

Temporal-Distance-JEPA 通过发现离线日志中的时间进展结构，并将代价形式与规划时部署协同设计，缩小了 JEPA world-model 规划器的训练-规划差距。Temporal-Distance-JEPA narrows the train--plan gap for JEPA world-model planners by discovering temporal progress structure in offline logs and co-designing cost form with plan-time deployment.

本文提出基于双向 Mamba 块（Vim）的通用视觉 backbone，通过位置嵌入标记图像序列，并利用双向 state space model 压缩视觉表征，具有成为下一代视觉基础模型 backbone 的巨大潜力。This paper proposes a new generic vision backbone with bidirectional Mamba blocks (Vim), which marks the image sequences with position embeddings and compresses the visual representation with bidirectional state space models and has great potential to be the next-generation backbone for vision foundation models.

本文提出 embodied language model，将真实世界连续传感器模态直接融入语言模型，从而建立词语与感知之间的联系，实现真实世界中的通用推理。This work proposes embodied language models to directly incorporate real-world continuous sensor modalities into language models and thereby establish the link between words and percepts to enable general inference in the real world.

本文提出极简的预训练框架 SimVLM，在广泛的判别式与生成式视觉-语言基准上显著超越既往预训练方法并取得新 SOTA，包括 VQA、NLVR2 以及图像描述任务。This work presents a minimalist pretraining framework, named SimVLM, which significantly outperforms previous pretraining methods and achieves new state-of-the-art results on a wide range of discriminative and generative vision-language benchmarks, including VQA, NLVR2, and image captioning tasks.

基于 ONNX Runtime 的原生 C++ 推理引擎击败了所有 baseline，置信区间甚至不重叠；batch 策略的影响比语言或运行时的选择更关键，超出作者预期。A native C++ inference engine on ONNX Runtime that beat every baseline, confidence intervals didn't even overlap, and batching strategy mattered more than either the language or the runtime choice, more than the authors expected.

本文提出 Metis，首个 memory foundation model 原型，赋予 foundation model 原生记忆能力，并表明原生记忆在架构、端到端优化和效率方面具有优势。This paper proposes Metis, the first prototype of memory foundation models, which empower foundation models with native memory capabilities and shows that native memory offers advantages in architecture, end-to-end optimization, and efficiency.

跨任务的测试时扩展实验表明，SkillRise 跨任务复用可迁移的 skill，而非受益于对同一任务的重复采样，在保持强性能的同时显著降低多阶段 skill 学习流水线的运行时开销。Scaling at test time across tasks suggests that SkillRise reuses transferable skills across tasks rather than benefiting from repeated sampling of the same task, and retains strong performance while substantially reducing the runtime overhead of skill learning pipelines with multiple stages.

介绍 DecoEvo（Decoupled Co-Evolution），在解耦目标下协同进化一个求解器 skill 和一个 rubric 生成器 skill，优化过程中不使用 gold rubric。DecoEvo (Decoupled Co-Evolution), which co-evolves a solver skill and a rubric-generator skill under decoupled objectives without using gold rubrics during optimization, is introduced.

提出 CAST（Credit Assignment from Solver Teachers），将游戏求解器状态价值的变化转换为求解器优势，并将其作为 turn 级信号注入 RLVR，在 ALFWorld 和 WebShop 上取得最高的平均 zero-shot 性能。CAST (Credit Assignment from Solver Teachers), which converts value changes in a game solver's state value into solver advantages and injects them into RLVR as turn-level signals and achieves the highest average zero-shot performance on ALFWorld and WebShop.

本文提出 CoRT，一种用于 rubric 条件化 GRPO 的 token 级 credit 加权方法，通过反事实回放在原始 rubric 条件化 prompt 和匹配的免准则 prompt 下对同一样本响应重新打分，并表明策略内部反事实似然对比为响应内 credit 分配提供了有效的训练信号。CoRT is proposed, a token-level credit weighting method for rubric-conditioned GRPO that uses counterfactual replay to rescore the same sampled response under the original rubric-conditioned prompt and a matched criteria-free prompt, and suggests that policy-internal counterfactual likelihood contrasts provide an effective training signal for within-response credit allocation.

本文贡献包括：圣训学概念到多 agent 流水线的形式化映射、实现 claim 链和分级 narrator 注册表的关系模式、结合链等级与内容批评的决策矩阵，以及对真实物理教材中 20,000 条 claim 的评估。A formal mapping from hadith-science concepts to multi-agent pipelines, a relational schema implementing claim chains and a graded narrator registry, a decision matrix combining chain grade with content criticism, and an evaluation on 20,000 claims from real physics textbooks are contributed.

描述一个 memory store：agent 本地的 Neo4j 属性图，增强 HNSW 向量索引，并采用完整的双时态数据模型，支持时间点语义检索而无需物理覆盖历史。A memory store is described: an agent-local Neo4j property graph augmented with HNSW vector indexes and a full bitemporal data model that supports point-in-time semantic retrieval without physically overwriting history.

Voice Memory，一个面向 agentic 语音识别的纯推理方案：流式推理时，冻结 corrector 读取单一 per-domain memory，逐 utterance 决定是否作用于假设或弃权并保留 1-best，跨 corrector 族可迁移，推理路径不增加任何参数。Voice Memory, a inference-only scheme for agentic speech recognition: at stream time, a frozen corrector reads a single per-domain memory and decides per utterance whether to act on the hypothesis or abstain and keep the 1-best, and transfers across corrector families and adds zero parameters to the inference path.

提出 pi-r2-flow，使 action-chunking flow 策略具备实时响应能力，同时保留大型 backbone、表达力强的多模态策略与多动作预测能力，并给出延迟自适应 flow 调度，将 in-flight 动作作为 inpainting 条件，每次调用仅需一步去噪即可输出动作。This work presentspi-r2-flow, which makes action-chunking flow policies reactive and real-time while retaining large backbones, expressive multi-modal policies, and multi-action prediction, and a latency-adaptive flow schedule that treats in-flight actions as inpainting conditioning and emits actions in one denoising step per call.

提出 CADENCE，一个统一框架，对每种叠加式失败模式给出针对性修复，证明通过原则化的蒸馏即可在不依赖数据中心级硬件的条件下获得强推理质量。CADENCE, a unified framework with a targeted fix for each compounding failure modes, is presented, showing principled distillation reaches strong reasoning quality without datacenter-scale hardware.

提出 DualG-MRAG，一种面向多模态 RAG 的双层框架，解耦 Macro-reasoning 与 Micro-matching Graph 两类图结构，通过分离全局结构推理与细粒度证据匹配来抑制检索噪声，并引入动态规划解码机制，从 GNN 前向过程中直接提取显式推理路径。DualG-MRAG is proposed, a Dual-tier framework that introduces a decoupled architecture comprising Macro-reasoning and Micro-matching Graphs for Multimodal RAG to suppress retrieval noise by isolating global structural reasoning from fine-grained evidence matching, and introduces a dynamic programming decoding mechanism that extracts explicit reasoning paths directly from the GNN's forward pass.

引入一个基于 GLM 的 retriever，并在单跳与多跳 RAG 场景下对比分析 GLM-based、GNN-based 与传统向量检索 retriever 的相对优势，指出微调后的 GLM retriever 具有更好的跨域泛化能力。This work introduces a GLM-based retriever and investigates the comparative strengths of GLM-based, GNN-based, and traditional vector-search-based retrievers in single- and multi-hop RAG settings, and suggests that finetuned GLM retrievers generalize better out of domain.

提出 ConMem，一个面向 LLM 辅助设备巡检的贡献感知 memory 框架，支持人在环的早期风险筛查，并在受限 memory 预算下保留高价值证据。This work proposes ConMem, a contribution-aware memory framework for LLM-assisted equipment inspection, supporting a human-in-the-loop early-risk screening system and retaining high-value evidence under a constrained memory budget.

OptGraph 是首个引入 GraphRAG 的优化 agentic workflow，首次将可复用经验构建为类型化 graph，刻画建模模式、问题形式化、实现细节与错误修正之间的关系。OptGraph is the first optimization agentic workflow that introduces graph retrieval-augmented generation (GraphRAG) and first constructs reusable experience as a typed graph, capturing the relationships among modeling patterns, problem formalization, implementation details, and error corrections.

本工作提出 CoMem，将拆分深度 j 显式化为可复用的上下文轴：为每个 token 写入一个深度 j 的残差，选择有界 chunk 集合，并仅恢复层 [j:L)。This work introduces CoMem, which makes split depth j an explicit reusable-context axis: write one depth-j residual per token, select a bounded chunk set, and resume only layers [j:L).

提出 LedgerMind，配套三层 Grounding Protocol、一个按问题复杂度匹配推理深度的 Adaptive Dual-Path Dispatcher，以及一个具备形式化 provenance 非放大保证的事件触发验证与修复引擎，同时提升答案准确率与轨迹级忠实度。LedgerMind is introduced, augmented by a Three-Layer Grounding Protocol, an Adaptive Dual-Path Dispatcher that matches reasoning depth to question complexity, and an Event-Triggered Verification-and-Repair engine with a formal provenance non-amplification guarantee that improves both answer accuracy and trajectory-level faithfulness.

本工作对有损验证方法所诱导的分布进行了原则性分析，并表明许多看似不同的方法仅在表面上有别，可统一归为两类：基于截断的验证与协作式验证。This work presents a principled analysis of the distributions induced by lossy verification methods, and shows that many seemingly distinct approaches differ only superficially and can be unified into two categories: truncation-based verification and collaborative verification.

ShadowDancer 引入 shadow pair，即在同一动力学下对外观做独立重采样的成对视频，并由 Shadow Library 大规模构建；一个 dynamics family 可控，当且仅当能为其构造出这样的 pair。ShadowDancer introduces shadow pairs, video pairs that replay the same dynamics under independently resampled appearance, constructed at scale by the Shadow Library, so that a dynamics family becomes controllable exactly when such pairs can be constructed for it.

记忆是长时程 LLM 智能体的核心，但现有记忆系统主要保存交互内容，而未建模哪些智能体在何种条件下可信。这一局限在多智能体系统中尤为关键，因为中心模型可能无法直接验证来自对等方、看似合理或相关的响应。我们提出 Σ-Mem，一种在线可靠性记忆，记录单个对等方的历史能力证据以及跨对等集的对等关系证据。两种证据均以实对称状态形式维护，并基于后Memory is central to long-horizon LLM agents, yet existing memory systems primarily preserve interaction content rather than modeling which agents can be trusted and under what conditions. This limitation is particularly important in multi-agent systems, where a central model may be unable to directly verify plausible or correlated peer responses. We introduce Σ-Mem, an online reliability memory that records historical competence evidence for individual peers and peer relationship evidence across the peer set. Both forms of evidence are maintained as real symmetric states and updated from post

对 Fairness Pruning 的实证评估表明，群体偏置处理与模型能力运行在可分离的电路上，奠定了从盲目零化向定向行为调制过渡的方法论基础。Empirical evaluation of Fairness Pruning empirically confirm that demographic bias processing and model capabilities operate on dissociable circuits, establishing the methodological foundations for transitioning from blind zeroing toward directional behavior modulation.

提出 MisKnow-Agent，一个受控评测框架，通过可控的权威线索与来源风格构造支撑人工审核结论的任务文档，并采用报告级 false-conclusion 采纳率（仅统计认可错误结论的报告），基于三种 backbone LLM 评估 DeerFlow 与 WebThinker。MisKnow-Agent is introduced, a controlled evaluation framework that constructs task-specific documents supporting manually audited false conclusions with controlled authority cues and source styles that evaluates DeerFlow and WebThinker with three backbone LLMs using a report-level false-conclusion adoption rate that counts only reports endorsing the false conclusion.

提出 OmniScope，一个无需训练的 token 压缩框架，以 query 作为跨模态共享的语义锚点，并对音频与视频分别估计相关性；由此给出 OmniLLM 推理的简单设计原则：跨模态共享 query，但不共享显著性估计。This work proposes OmniScope, a training-free token compression framework that uses the query as a shared semantic anchor while estimating relevance separately for audio and video, and suggests a simple design principle for OmniLLM inference: share the query across modalities, but not the salience estimates.

研究表明，领域特定的 orchestration 可使缺乏专业团队的实验室也能实现可审计的预处理，并为其他科学领域的 AI agent 提供了可借鉴的设计原则。The results indicate that domain-specific orchestration can bring auditable preprocessing within reach of laboratories lacking dedicated expertise, illustrating design principles applicable to AI agents in other scientific domains.

本工作识别出记忆溯源洗白现象：基于LLM的记忆整合过程中，外部观察可能被改写为看似用户历史或工作流支持的内容，在保留动作触发的同时，抹去本应限制其权威性的低可信度来源。This work identifies memory provenance laundering: during LLM-based memory consolidation, an external observation may be rewritten as apparent user history or workflow support, preserving an action trigger while erasing the low-trust source that should limit its authority.

多维评估-验证奖励(EVR)将评估分解为独立视觉准则；针对每个准则，MLLM Evaluator生成多个候选假设，Verifier在具体视觉证据中grounding每个claim以接受或拒绝，产生可靠且细粒度的奖励信号。A Multi-dimensional Evaluation-Verification Reward (EVR) decomposes evaluation into distinct visual criteria; for each criterion, an MLLM Evaluator generates multiple candidate hypotheses, and a Verifier grounds each claim in concrete visual evidence to accept or reject it, producing reliable and fine-grained reward signals.

提出QWorld，用分位数-分位数匹配目标替代EP，直接将投影后的潜在样本与秩匹配的高斯分位数对齐，从而在尾部保持有效的修正梯度。QWorld is proposed, which replaces EP with a quantile-quantile matching objective that directly aligns projected latent samples with rank-matched Gaussian quantiles, thereby maintaining effective corrective gradients in the tails.

Criterion-Distilled Policy Optimization (CriPO)，通过 on-policy 自蒸馏增强基于评分标准的 RL，以同时应对未探索和被压制的标准。Criterion-Distilled Policy Optimization (CriPO), which enhances rubric-based RL via on-policy self-distillation via on-policy self-distillation, to simultaneously address both unexplored and suppressed criteria.

提出SA，一种Stable Advantage Fusion框架，避免熵崩塌，在所有六个模型-领域设定上持续优于固定系数的GRPO+OPD融合，聚合得分提升0.70%，训练更稳定。SA is proposed, a Stable Advantage Fusion framework that avoids entropy collapse and consistently outperforms fixed-coefficient GRPO+OPD fusion, improving the aggregate score by 0.70% across all six model-domain settings while achieving more stable training.

研究发现privileged shifts无法给出可靠的答案对齐方向，其幅度主要反映反事实敏感性而非token级学习价值；提出Counterfactual Sensitivity Credit Reallocation (CSCR)，作为GRPO的简单扩展，降低高敏感token的credit并对token级优势重新归一化，同时保留原始credit预算与verifier确定的方向。These findings show that privileged shifts fail to provide reliable answer-aligned directions, while their magnitudes primarily reflect counterfactual sensitivity rather than token-level learning value, and propose Counterfactual Sensitivity Credit Reallocation (CSCR), a simple extension of GRPO that reduces credit for highly sensitive tokens and renormalizes token-level advantages to preserve both the original credit budget and verifier-determined direction.

世界模型为规划和行动提供了预测性基础，但现有建模方式仅回答物理层面的问题：它是什么/在哪里，以及将如何演变。然而，人类行为由隐藏的心理状态驱动（一个人相信什么、想要什么、意图做什么、感受如何，以及认为在社会上何为可接受），因此仅追踪物理场景而忽略每个智能体所知与所信内容的模型，会对看起来正确的场景预测出错误的行动。我们将心理世界建模（MWM）形式化为一个通用理论框架，将心理变量作为世界模型的核心组成部分。World models enable a predictive substrate for planning and action, yet existing formulations merely answer a physical question: what/where it is, and how will it evolve. Human behavior, however, is driven by hidden mental state (what a person believes, wants, intends, feels, and considers socially permissible), so a model that tracks the physical scene but not what each agent knows and believes about it predicts the wrong action for the right-looking scene. We formulate Mental World Modeling (MWM), a generic theoretical framework that makes mental variables core components of a world model ra