Daily Papers · 2026-10-15
📄 论文日报 | 来源：Hugging Face Daily Papers
🗓 Episode：2026.10.14 | VideoJev统一视频定位问答；Jeva低延迟工具决策
1. VideoJev：面向视频理解与交互的系统一模型
- 英文标题：VideoJev: System One Model for Video Understanding and Interaction
- arXiv：2610.18615
💡 一句话解读：视频AI要边看边做决定，但现有模型反应慢、每次都要重新算。VideoJev把看视频和做选择变成打分任务，让模型直接输出时间片段或选项，还能缓存和批量处理，速度提升百倍以上。
视觉智能体必须将视频观察转化为及时决策，以跟踪进展并选择下一步行动。基于新兴的“系统一”范式，我们提出VideoJev系列模型，将视频理解与交互视为结构化决策任务。我们改造预训练多模态大语言模型（MLLMs），用共享头替代自回归答案解码，对视频证据下的时间区间和文本选项进行评分，从而生成事件片段或答案选项，统一了时间定位与多项选择视频问答。这结合了MLLMs的语义灵活性与智能体可直接使用的高效决策。在交互过程中，新目标和中间结果会促使对同一视觉历史提出进一步查询，因此VideoJev通过缓存推理（针对陆续到达的查询）和联合推理（单次前向传播处理多个查询）来复用视觉计算。在三个时间定位基准上，VideoJev-4B和VideoJev-35B-A3B分别将骨干模型的平均mIoU提升了6.7和10.8个百分点，实测推理加速2.8-3.1倍。在四个视频问答基准上，两者在思考模式下性能与骨干模型相当，实测推理加速超过100倍。这些结果为将结构化视觉判断组合成复杂、开放式工作流提供了统一、可复用的决策接口。
📄 English Abstract
Visual agents must turn video observations into timely decisions to track progress and choose what to do next. Building on the emerging System One paradigm, we introduce VideoJev, a family of models that cast video understanding and interaction as structured decision tasks. We adapt pretrained multimodal large language models (MLLMs) by replacing autoregressive answer decoding with a shared head that scores temporal intervals and textual options against video evidence. These scores yield event spans or answer choices, unifying temporal grounding and multiple-choice video question answering. This combines the semantic flexibility of MLLMs with efficient decisions that agents can use directly. As an interaction unfolds, new goals and intermediate results prompt further queries about the same visual history. VideoJev therefore reuses visual computation through Cached inference for queries as they arrive and Joint inference for multiple queries in a single forward pass. Across three temporal grounding benchmarks, VideoJev-4B and VideoJev-35B-A3B improve mean mIoU over their backbones by 6.7 and 10.8 points, respectively, with measured inference speedups of $2.8$-$3.1\times$. On four video QA benchmarks, both achieve performance comparable to their backbones in Thinking mode, with measured inference speedups exceeding $100\times$. These results offer a unified, reusable decision interface for composing structured visual judgments into complex, open-ended workflows.
2. Jeva：从第一性原理构建的系统 1 模型
- 英文标题：Jeva: System One Models from First Principles
- arXiv：2610.18773
💡 一句话解读：当前AI做选择题时，常把问题和选项拼成一段文字来预测答案，这忽略了决策问题的结构。Jeva提出一种新架构，让模型能区分“没有答案”和“不确定”，选项顺序不影响结果，能处理大量候选，且重复选项的计算可缓存复用。
系统1模型能够实现直接、低延迟的决策。然而，近期从序列化文本预测答案的做法本身并未尊重决策问题的结构。我们认为当前设计低估了四个重要的结构要求：(i) 空值感知：模型应区分答案缺失与候选答案间的不确定性；(ii) 排列不变性：重新排序备选项不应改变决策；(iii) 选项可扩展性：候选集增大时不应需要重新定义输出空间；(iv) 可缓存性：对重复出现的备选项进行的与请求无关的计算应能在不同请求间复用。我们提出Jeva，一种决策原生架构，联合解决这些要求。共享编码器独立表示请求和每个选项，而紧凑的决策栈仅针对不可变的选项记忆精炼请求状态。对称聚合产生排列等变概率；共享评分适应可变大小的候选集。单独的覆盖估计区分答案缺失与候选竞争，从而支持显式的空结果。与请求无关的选项表示和层投影支持在新组装的集合中复用。这些机制统一了Choice、Noul和Score输出。在评估中，Jeva结合了强大的类型化决策性能与近乎完美的空值检测，并且在任何测试的排列下均无决策翻转。它能扩展到数千个候选，并通过跨请求复用显著降低延迟而不改变决策。
📄 English Abstract
System one models enable direct, low-latency decision-making. Yet recent practice of predicting an answer from serialized text does not, by itself, respect the structure of a decision problem. We argue that current designs underemphasize four important structural requirements. (i) null-aware: Models should distinguish answer absence from uncertainty among candidates; (ii) permutation-invariant: Reordering alternatives should not change the decision; (iii) option-scalable: Growing candidate sets should not require redefining the output space; (iv) cacheable: Request-independent computation over recurring alternatives should be reusable across requests. We introduce Jeva, a decision-native architecture that addresses these requirements jointly. A shared encoder independently represents the request and each option, while a compact decision stack refines only the request state against immutable option memory. Symmetric aggregation yields permutation-equivariant probabilities; shared scoring accommodates variable-size candidate sets. A separate coverage estimate distinguishes answer absence from candidate competition, enabling explicit null outcomes. Request-independent option representations and layer projections support reuse across newly assembled sets. These mechanisms unify Choice, Noul, and Score outputs. Across evaluations, Jeva combines strong typed-decision performance with near-perfect null detection and no decision flips under any tested permutations. It scales to thousands of candidates and substantially reduces latency through cross-request reuse without changing decisions.
3. Agent-Reflex：教导语言智能体进行反思性行动
- 英文标题：Agent-Reflex: Teaching Language Agents to Act Reflectively
- arXiv：2610.17953
💡 一句话解读：现有语言智能体在真实环境中遇到工具返回的杂乱信息或用户纠错时，往往不会主动反思和恢复，容易一错再错。Agent-Reflex 通过三个阶段训练智能体：先让它在受干扰的环境中自己总结出可复用的反思技巧，再用这些技巧做监督微调，最后结合强化学习与蒸馏进一步激发主动反思，从而提升其抗干扰和纠错能力。
大型语言模型已从单轮聊天机器人演变为具备长程推理能力的交互式智能体。在实际部署中，MCP 工具服务可能返回带噪声的反馈，用户也可能介入纠正智能体，但现有方法对反思学习和失败恢复的监督有限。为弥补这一空白，我们提出 Agent-Reflex，一个用于激励智能体主动反思的通用训练框架。具体而言，Agent-Reflex 包含三个阶段：首先，环境反思技能发现（RSD）在环境扩展中注入基于工具反馈的扰动，采样能力多样的 LLM 池，挖掘匹配的成功-失败对，并总结可复用的反思元技能；基于该技能库，我们应用反思性监督微调（ReSFT），将每项技能转化为对自我反思、技能恢复和以用户为中心的纠正的监督，强化互补的反思行为；最后，我们提出在策略反思策略优化（ORPO），将噪声交互下的智能体强化学习与反思引导的在策略蒸馏相结合，进一步激发主动反思。基于 Qwen3-8B，Agent-Reflex 在五个长程工具使用基准上达到 58.6 的总体得分，在 AgentNoiseBench 上达到 29.1，超越了强大的环境扩展和反思基线。在六个具有挑战性的智能体基准套件上的实验证明了 Agent-Reflex 的有效性。进一步分析表明，Agent-Reflex 提高了智能体对噪声的鲁棒性，同时保持稳定的训练动态，支持可扩展的智能体训练。我们的代码可在 https://github.com/RUC-NLPIR/Agent-Reflex 获取。
📄 English Abstract
Large language models have evolved from single-turn chatbots into interactive agents capable of long-horizon reasoning. In real deployments, MCP tool services can return noisy feedback and users may intervene to correct the agent, but existing methods provide limited supervision for reflection learning and failure recovery. To bridge this gap, we propose Agent-Reflex, a general training framework for incentivizing proactive reflection in agents. Specifically, Agent-Reflex comprises three stages. First, Environment Reflection Skill Discovery (RSD) injects grounded tool-feedback perturbations into environment scaling, samples a capability-diverse LLM pool, mines matched success-failure pairs, and summarizes reusable reflection meta-skills. Building on this skill library, we then apply Reflective Supervised Fine-Tuning (ReSFT) that converts each skill into supervision for self-reflection, skill recovery, and user-centric correction, strengthening complementary reflective behaviors. Finally, we propose On-Policy Reflective Policy Optimization (ORPO) to combine agentic reinforcement learning under noisy interactions with reflection-guided on-policy distillation to further elicit proactive reflection. Built on Qwen3-8B, Agent-Reflex reaches 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, surpassing strong environment-scaling and reflection baselines. Experiments on six challenging agent benchmark suites demonstrate the effectiveness of Agent-Reflex. Further analyses show that Agent-Reflex improves agents' robustness to noise while maintaining stable training dynamics, supporting scalable agent training. Our code is available at https://github.com/RUC-NLPIR/Agent-Reflex.
4. Urbex：面向城市级三维场景的智能体空间定位
- 英文标题：Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes
- arXiv：2610.18356
💡 一句话解读：城市级三维场景中，要找到某个物体（比如“那栋红色屋顶的建筑”）通常需要先列出所有可能的候选区域再逐一打分，在大场景里又慢又麻烦。这篇论文提出 Urbex，让一个能看懂图像和文字的人工智能像人一样主动探索：先看地标、放大局部、旋转视角，最后直接框出目标，并用强化学习训练这种搜索策略，速度更快、定位更准。
城市级三维定位通常被形式化为候选排序问题，这需要预先计算实例提议，在大规模户外场景中效率低下。我们提出 Urbex，一个智能体式视觉-语言框架，将定位任务转化为主动空间搜索。Urbex 将城市级三维场景表示为一个交互式多视角环境，允许视觉-语言模型（VLM）智能体查询地标、放大局部区域、渲染斜视角视图，并提交一个定位边界框。我们在 CityRefer 数据集上使用强化学习优化这一工具使用策略，采用平滑定位奖励和轻量塑形奖励以实现高效探索。实验表明，Urbex 在 CityRefer 上取得了最佳的 Hit@1 和 IoU，支持无需候选提议且不依赖真实边界框的定位，并且比最强的城市级排序基线显著更快。在 STPLS3D-Refer 上的零样本结果进一步表明其跨场景鲁棒性有所提升。
📄 English Abstract
City-scale 3D grounding is commonly formulated as candidate ranking, which requires precomputed instance proposals and becomes inefficient in large outdoor scenes. We propose Urbex, an agentic vision-language framework that instead casts grounding as active spatial search. Urbex represents a city-scale 3D scene as an interactive multi-view environment, allowing a VLM agent to query landmarks, zoom into local regions, render oblique views, and commit a grounded bounding box. We optimize this tool-use policy with reinforcement learning on CityRefer, using smooth localization rewards and lightweight shaping rewards for efficient exploration. Experiments show that Urbex achieves the best Hit@1 and IoU on CityRefer, supports candidate-free grounding without ground-truth boxes, and is substantially faster than the strongest city-scale ranking baseline. Zero-shot results on STPLS3D-Refer further indicate improved cross-scene robustness.
5. GreenWAM：面向机器人学的世界基础模型
- 英文标题：GreenWAM: A World Foundation Model for Robotics
- arXiv：2610.18774
💡 一句话解读：机器人既要能预测自己动作会怎样改变环境，又要能据此选动作，但以往模型往往只干其中一件事，且各机器人动作格式不统一。GreenWAM 用统一的关节坐标描述20种机器人，联合训练出既能预测世界又能当策略的通用模型，还能适配真实人形机器人。
机器人需要一个能预测其动作将如何改变场景的世界模型，以及一个选择这些动作的策略。现有模型通常只针对其中一种角色进行构建和评估，而且每台机器人记录动作的格式各不相同。我们探究一个机器人世界模型能否在不同本体上同时承担这两种角色，以及其成功的关键是什么。我们提出 GreenWAM——一个面向机器人学的世界基础模型，以及产生它的方法配方：（1）一个共享的解剖学动作空间，用统一的关节坐标描述从双臂机械臂到灵巧人形机器人的20个机器人家族；（2）在这些家族上进行视觉、动作条件和世界-动作模式的联合训练；（3）从生成器中分离出动作专家（action expert）的策略。作为世界模型，GreenWAM 在 WorldArena 1.0 上取得了顶级的轨迹、深度和3D精度，在我们比较的所有世界-动作模型和视频生成器中均领先。作为策略，它在 RoboTwin 2.0 排行榜上名列前茅，在比较方法中取得了最高的零样本 LIBERO-plus 成功率，并在 CALVIN 和 LIBERO 上与专用的视觉-语言-动作（vision-language-action）和世界-动作策略相比具有竞争力。我们进一步将 GreenWAM 适配到真实人形机器人上，其成功率高于从 π_{0.5} 和 Cosmos 3 Edge 适配的策略，并将其蒸馏为单步策略以实现实时控制。受控实验得出三个发现：机器人专业化同时改善了世界预测和控制；在适配开始时分离出动作专家是有益的；在策略适配期间保留世界模型模式可保持对场景变化的鲁棒性。
📄 English Abstract
A robot needs a world model that predicts how its actions will change the scene and a policy that chooses those actions. Models are usually built and evaluated for only one of these roles, and every robot records its actions in its own format. We ask whether one robot world model can serve both roles across different bodies, and what makes this work. We present GreenWAM, a world foundation model for robotics, together with the recipe that produces it: (1) a shared anatomical action space that describes 20 robot families, from bimanual arms to dexterous humanoids, in common joint coordinates; (2) joint training on these families in visual, action-conditioned and world-action modes; and (3) policies that split an action expert off the generator. As a world model, GreenWAM attains top-tier trajectory, depth and 3D accuracy on WorldArena 1.0, ahead of every world-action model and video generator in our comparison. As a policy, it scores among the top entries of the RoboTwin 2.0 leaderboard, reaches the highest zero-shot LIBERO-plus success among the compared methods and is competitive with dedicated vision-language-action and world-action policies on CALVIN and LIBERO. We further adapt GreenWAM to a real humanoid, where it succeeds more often than policies adapted from $\pi_{0.5}$ and from Cosmos 3 Edge, and distill it into a single-step policy for real-time control. Controlled experiments yield three findings: robot specialization improves world prediction and control together; an action expert helps when split off at the start of adaptation; and keeping world-model modes during policy adaptation preserves robustness to scene changes.
6. AesCode：基于解耦跨模态奖励的美学代码生成
- 英文标题：AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards
- arXiv：2610.18860
💡 一句话解读：让AI写幻灯片、海报这类代码时，它看不见自己排出来的画面，所以布局和配色常常很糟。这篇论文让AI一边写代码、一边参考同一需求生成的图片来指导审美，把画面拆成可核查的部件分别打分再训练，生成效果接近顶级闭源模型。
以代码形式生成幻灯片、海报、仪表盘等信息丰富的视觉作品很有吸引力，因为其输出结构化、可编辑且可验证。然而代码模型无法看到自己对布局和颜色的选择在画布上如何组合，这往往限制了空间规划与美学质量。图像生成器擅长视觉设计，却可能错误渲染所要求的内容，使其输出难以作为直接复现的可靠目标。为此我们提出 AesCode：它在内容提示的基础上生成代码，同时以同一提示生成的图像作为美学引导。具体而言，AesCode 将目标表示为一个描述整个画布的设计图（design graph），使每个属性都可归因。基于该图，我们导出互补的跨模态奖励：确定性验证器对可从代码和渲染中解析的属性打分，而 VLM 评判器则借助与图元素和关系绑定的、样本特定的视觉图评分标准（Visual Graph Rubric, VGR）评估不可解析的属性。我们通过带逐通道归一化的强化学习优化这些解耦奖励，以共同提升空间规划、美学质量和可验证正确性。为评估该设定，我们构建了一个分别对可验证正确性和视觉质量打分的基准，其视觉得分与人类判断高度一致。我们发现参考图像能显著提升现有模型所生成代码的视觉质量，消融实验表明 AesCode 既受益于将参考作为模型输入，也受益于将其作为视觉奖励来源。在给定相同参考的情况下，与同规模 Qwen3-VL 基线相比，AesCode-8B 和 AesCode-32B 的视觉质量分别提升 31.1 和 22.4 个百分点，达到与 GPT-5.5、Claude Opus 4.8 等闭源前沿模型相当的水平。
📄 English Abstract
Generating information-rich visual artifacts such as slides, posters, and dashboards as code is appealing because the outputs are structured, editable, and verifiable. Yet code models cannot see how their choices of layout and color come together on the canvas, often limiting spatial planning and aesthetic quality. Image generators excel at visual design but can misrender requested content, making their outputs unreliable targets for direct reproduction. We therefore introduce AesCode, which generates code grounded in the content prompt while using an image generated from the same prompt as an aesthetic guidance. To do so, AesCode represents the target as a design graph that describes the whole canvas and makes each property attributable. From this graph, we derive complementary cross-modal rewards: deterministic verifiers score properties parsable from code and rendering, while a VLM judge assesses the non-parsable ones with a sample-specific Visual Graph Rubric (VGR) tied to graph elements and relations. We optimize these decoupled rewards through reinforcement learning with per-channel normalization to jointly improve spatial planning, aesthetic quality, and verifiable correctness. To evaluate this setting, we build a benchmark that scores verifiable correctness and visual quality separately, with visual scores highly consistent with human judgments. We find that reference images substantially improve the visual quality of code generated by existing models, and ablations show that AesCode benefits from references both as model input and as a source of visual rewards. Relative to same-scale Qwen3-VL baselines given the same references, AesCode-8B and AesCode-32B improve visual quality by 31.1 and 22.4 percentage points, reaching scores comparable to those of closed-source frontier models such as GPT-5.5 and Claude Opus 4.8.
7. Harness^2：面向开放世界的递归式智能体执行框架构建
- 英文标题：Harness^2: Recursive Agent Harnessing for an Open World
- arXiv：2610.14224
💡 一句话解读：这篇研究要解决的问题是：让 AI 智能体在遇到从没见过的新任务时，能自己改进工作方式，而不需要标准答案来打分。它提出的办法是 Harness^2 框架，分两层递归：第一层在具体任务中通过对比、试探、组合来调整做法，第二层把调整经验积累下来，让智能体以后遇到新任务时更会变通。
大语言模型（LLM）智能体的能力不仅取决于其底层模型，也取决于其执行框架（harness）。递归式自我改进（RSI）为自动演化这些框架提供了一条有前景的路径。然而，在开放世界环境中部署 RSI 仍面临两个关键挑战：其一，开放世界任务通常在框架设计阶段未曾见过、在测试时又是全新的，因而没有真值验证器来可靠地指导在线框架编辑；其二，RSI 过程仍高度依赖具体任务，无法泛化到新场景。我们提出 Harness^2，一个即插即用的递归式智能体框架构建方法，通过两个相互关联的递归层级来应对上述挑战。任务级递归使用对比式“提议—探测—组合”（Propose-Probe-Compose）步骤来改进当前任务的框架，在没有真值验证器的情况下可靠地指导编辑，并可通过并行或串行递归进一步扩展。框架构建级递归则从这些编辑中提取经验，用于更新一个持续改进框架，使智能体能够更好地适应未来任务。在三个专业基准和四种模型—框架配置上，单步改进即可稳定提升 GLM-5 与 Gemini 3.5 Flash 相对于其基础框架的表现，在 WorkBuddy-Bench 上最高提升 11.5 分、在 JobBench 上提升 4.5 分、在 LAB 上提升 11.8 个百分点，且随着任务级递归预算增加还能获得进一步提升。
📄 English Abstract
LLM agents' capabilities depend not only on their underlying models but also on their execution harnesses. Recursive self-improvement (RSI) offers a promising approach to automatically evolve these harnesses. However, deploying RSI in open-world settings still faces two critical challenges. First, open-world tasks are typically unseen during the harness design phase and novel at test time, leaving no ground-truth verifiers to reliably guide online harness edits. Second, the RSI process remains highly task-specific and fails to generalize to novel scenarios. We introduce Harness^2, a plug-and-play, recursive agent harnessing framework that addresses these challenges through two connected levels of recursion. Task-level recursion refines the current task's harness using a contrastive Propose-Probe-Compose step that reliably guides edits without ground-truth verifiers, and can be scaled further via parallel or sequential recursion. Harnessing-level recursion then extracts experience from these edits to update a continual improvement harness, enabling the agent to better adapt to future tasks. Across three professional benchmarks and four model-harness configurations, a single step consistently improves GLM-5 and Gemini 3.5 Flash over their base harnesses, yielding gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB, with further gains as the task-level recursion budget increases.
8. TPPO：面向语言模型推理的跟踪价值PPO
- 英文标题：TPPO: Tracking-Value PPO For Language Model Reasoning
- arXiv：2610.16794
💡 一句话解读：大模型做数学推理时，强化学习常因只在最后给对错分，导致中间每一步的好坏难以判断。这篇论文提出TPPO，把整条推理路径的价值拆成逐位置的小份相加，用最终结果来监督训练，从而更准地追踪每一步的价值变化，数学成绩明显提升。
近端策略优化（PPO）广泛用于带可验证奖励的语言模型推理强化学习。虽然沿轨迹的价值演化刻画了相对的词元级信用，但标准评论家（critic）在稀疏的终端奖励下往往无法追踪它。为建模这种演化，我们分析相邻前缀价值之间的关系，并证明它们具有在对数几率空间中以加性更新形式出现的贝叶斯分解。我们利用这种加性结构构建了跟踪价值PPO（TPPO），一种将逐位置标量输出求和为价值逻辑值的累积价值参数化方法。这些输出通过对重构价值的终端结果监督来学习。TPPO在Qwen3-4B-Base和Qwen3-8B-Base上分别将聚合数学推理分数较PPO提升7.0和8.4分，同时改善了分布外泛化。增益还扩展到Qwen3-4B、Qwen2.5-Math-1.5B、Qwen3.5-9B以及倒计时和科学编程任务。进一步分析表明，该方法能更准确地追踪价值演化，并给出更可靠的优势估计。我们的代码已在https://github.com/LINs-lab/TPPO公开。
📄 English Abstract
Proximal Policy Optimization (PPO) is widely used in reinforcement learning with verifiable rewards for language-model reasoning. While \emph{value evolution along trajectory} characterizes relative token-level credit, yet standard critics often fail to track it under sparse terminal rewards. To model this evolution, we analyze the relation between successive prefix values and show that they admit a Bayesian decomposition with additive updates in log-odds space. We use this additive structure to construct Tracking-Value PPO (\method), a cumulative value parameterization that sums position-wise scalar outputs into value logits. These outputs are learned through terminal-outcome supervision on the reconstructed values. \method improves aggregate mathematical reasoning scores over PPO by 7.0 and 8.4 points on Qwen3-4B-Base and Qwen3-8B-Base, respectively, while improving out-of-distribution generalization. Gains extend to Qwen3-4B, Qwen2.5-Math-1.5B, Qwen3.5-9B and to countdown, and scientific coding tasks. Further analysis shows more accurate tracking of value evolution and more reliable advantage estimates. Our code is publicly available in https://github.com/LINs-lab/TPPO.
9. Vela：用户行为作为监督激发大语言模型智能
- 英文标题：Vela: User Behavior as Supervision Sparks LLM Intelligence
- arXiv：2610.14709
💡 一句话解读：这篇研究想解决的是：大模型虽然能聊天，却不太会从用户长期的行为记录里推断喜好和下一步动作。它的办法是把真实用户行为数据自动转成训练题和解题过程，用这些数据去微调大模型，从而提升它理解用户的能力，也顺带提升通用能力。
大语言模型（LLM）正越来越多地被部署在需要依据用户长期行为历史来行动的应用中，从个人助理到推荐与广告系统。这些应用要求 LLM 具备“用户维度”的智能：从数周到数年的行为数据中推断一个人的偏好、意图以及未来可能的行动。然而，当前模型是在静态文本或合成数据上训练的，二者都无法刻画用户行为随时间演变的方式。现实世界的行为记录已在内容平台和个人助理智能体中大规模存在，天然提供了一种监督信号。我们提出以用户行为作为监督，并引入 Vela 框架，将真实用户行为数据转化为可训练的任务与轨迹。基于 10 万名用户的匿名行为记录，Vela 在 91 个真实世界任务模板上生成了 10 万个任务；通过基于难度的过滤和轨迹接受标准，保留了约 7.5 万个任务及其选定的训练轨迹。在这些轨迹上微调 Qwen3.6-35B-A3B，使其在用户相关任务上比基座模型最高提升 11.5 分，在通用基准上平均提升 4 分。使用更多用户行为数据训练可进一步提升用户相关表现，而微调后再进行强化学习，比从基座模型直接做强化学习最高提升 7.7 分。这些增益表明，真实世界用户行为是一种可扩展的监督来源，既能提升以用户为中心的能力，也能提升 LLM 更广泛的智能。
📄 English Abstract
Large language models are increasingly deployed in applications that act on users' long-term behavior histories, from personal assistants to recommendation and advertising systems. These applications demand the user dimension of LLM intelligence: inferring a person's preferences, intent, and likely future actions from weeks to years of behavior data. Yet current models are trained on static text or synthetic data, neither of which captures how user behavior evolves over time. Real-world behavior records, already available at scale on content platforms and in personal-assistant agents, offer a natural supervisory signal. We propose user behavior as supervision and introduce Vela, a framework that converts real user behavior data into trainable tasks and trajectories. From the anonymized behavior records of 100,000 users, Vela generates 100,000 tasks across 91 real-world task templates; difficulty-based filtering and trajectory acceptance criteria retain approximately 75,000 tasks with selected training trajectories. Fine-tuning Qwen3.6-35B-A3B on these trajectories improves performance over the base model by up to 11.5 points on user-related tasks and by an average of 4 points on general benchmarks. Training on more user behavior data further improves user-related performance, and reinforcement learning after fine-tuning outperforms reinforcement learning from the base model by up to 7.7 points. These gains suggest that real-world user behavior is a scalable source of supervision for both user-centered capabilities and the broader intelligence of LLMs.
10. X-Ragent：面向基于证据的胸部X射线推理的工具增强强化学习
- 英文标题：X-Ragent: Tool-Augmented Reinforcement Learning for Evidence-Grounded Chest X-Ray Reasoning
- arXiv：2610.16811
💡 一句话解读：胸部X光诊断不能只看图，还得结合具体证据。现有AI工具虽能提供证据，但不会判断何时该用工具，也容易把无效数据混进诊断。本研究提出X-Ragent训练框架，让模型学会按需找证据、只采纳有效测量，从而做出更可靠的诊断。
胸部X射线（CXR）解读通常需要将视觉发现与任务特定证据相结合，而不能仅依赖直接的视觉识别。虽然工具增强的医疗智能体可以通过外部模块提供此类证据，但它们往往缺乏明确机制来学习何时需要辅助证据，以及防止不适用的定量证据进入诊断推理。为解决这一问题，我们提出X-Ragent，一种用于工具增强强化学习的后训练框架，它将学习到的辅助证据寻求与有效性门控的定量测量获取解耦，以实现基于证据的CXR推理。我们构建了Rad-CoT，一个包含7.7K样本的语料库，将因果有序的工具增强轨迹与报告式诊断推理相结合，并利用它通过智能体监督微调（SFT）初始化Qwen3-VL-8B。然后，我们使用组相对策略优化（GRPO）优化策略，采用准确性门控奖励，该奖励仅对正确轨迹应用测量一致性和辅助工具成本的过程级项，防止看似合理的流程行为补偿错误的诊断。在三个CXR基准测试中，X-Ragent取得了最强结果，包括在完全留出的ChestAgentBench上提高了8.6分。通过选择性地寻求证据并仅接受有效测量，X-Ragent支持更忠实且基于证据的CXR推理。
📄 English Abstract
Chest X-ray (CXR) interpretation often requires integrating visual findings with task-specific evidence rather than relying on direct visual recognition alone. While tool-augmented medical agents can supply such evidence via external modules, they often lack explicit mechanisms for learning when auxiliary evidence is needed and for preventing inapplicable quantitative evidence from entering diagnostic reasoning. To address this, we propose X-Ragent, a post-training framework for tool-augmented reinforcement learning that decouples learned auxiliary-evidence seeking from validity-gated quantitative measurement acquisition for evidence-grounded CXR reasoning. We construct Rad-CoT, a 7.7K-sample corpus combining causally ordered tool-augmented trajectories with report-style diagnostic reasoning, and use it to initialize Qwen3-VL-8B via agentic supervised fine-tuning (SFT). We then optimize the policy with Group Relative Policy Optimization (GRPO) using an accuracy-gated reward that applies process-level terms for measurement consistency and auxiliary-tool cost only to correct trajectories, preventing plausible process behavior from compensating for an incorrect diagnosis. Across three CXR benchmarks, X-Ragent achieves the strongest results, including an 8.6-point improvement on the fully held-out ChestAgentBench. By seeking evidence selectively and admitting only valid measurements, X-Ragent supports more faithful and evidence-grounded CXR reasoning.
11. DoSA：点稀疏注意力
- 英文标题：DoSA: Dots Sparse Attention
- arXiv：2610.13990
💡 一句话解读：大模型处理长文本时，注意力机制要从海量历史信息里挑出最相关的少数几个，但挑之前往往得把所有候选都算一遍分数，很费算力。这篇论文提出 DoSA：用一套数学上的“上下界”快速排除明显没戏的候选，只对可能入选的少数算精确分数，且结果和全量计算完全一致，不损失精度，在 GPU 上最高提速约 2.7 倍。
稀疏注意力只访问少量键（key），但找到这些键仍可能需要对每个键与每个索引头（indexer head）逐一打分。我们提出 DoSA（Dots Sparse Attention，点稀疏注意力），一种使该计算变为条件式、同时保持精确 token 选择的无损方法。其核心观察是：Top-K 选择只需要足以否决某个候选的证据，而不必知道它的完整分数。一个种子集合给出了获胜阈值的下界；非负的头投影与联合残差界可给出正贡献的上界，而基于激活值的聚合则界定了负头所减去的幅度。该证书（certificate）把完整打分导向仍然有资格的候选；精确分数过滤进一步减少了写回与选择的工作量。我们建立了该证书及其精确选择条件，并分析了几何结构与选择裕度如何决定其有效性。我们的 GPU 实现使完整 Indexer 在英文、中文和代码负载上相较 SGLang 最高加速 2.74 倍，相较 LiteTopK 最高加速 1.69 倍。在 GLM-5.2 和 DeepSeek-V3.2 上的端到端评测中，首 token 时延（time-to-first-token）相较这两个基线分别最高加速 1.94 倍和 1.50 倍。
📄 English Abstract
Sparse attention visits few keys, yet finding them can still require scoring every key with every indexer head. We introduce DoSA (Dots Sparse Attention), a lossless method for making this computation conditional while preserving exact token selection. Its central observation is that Top-K selection requires sufficient evidence to reject a candidate, not necessarily its full score. A seed set supplies a lower bound on the winning threshold. Nonnegative head projections and joint residual bounds upper-bound positive contributions, while activation-informed aggregation bounds the magnitude subtracted by negative heads. The certificate directs full scoring toward candidates that remain eligible; exact-score filtering further reduces writeback and selection work. We establish the certificate and its exact-selection conditions, and analyze how geometry and selection margins govern its effectiveness. Our GPU implementation accelerates the complete Indexer by up to 2.74 times over SGLang and 1.69 times over LiteTopK across English, Chinese, and code workloads. End-to-end evaluation on GLM-5.2 and DeepSeek-V3.2 achieves up to 1.94 times and 1.50 times time-to-first-token speedups over the two baselines, respectively.
12. ExeWorldBench：基于代码的可执行3D世界建模基准测试
- 英文标题：ExeWorldBench: Benchmarking Code-Based Executable 3D World Modeling
- arXiv：2610.15166
💡 一句话解读：现有AI评测只零散考察物理推理或画面逼真度，无法判断AI能否真正理解世界运作原理并写出可运行代码。本文提出ExeWorldBench，用三个递进任务（单机制、多级系统、目标驱动）测试AI从看懂世界到用代码造出可执行3D世界的能力。
编程智能体正日益从复现世界的外观，转向构建能捕捉世界运行方式的可执行世界。这一转变要求可执行世界建模，它整合两种核心能力：世界理解，即推理世界底层的结构、机械与因果原理；以及可执行实现，即将这种理解转化为一个基于代码的系统，其行为由物理交互涌现。然而，现有基准仅提供碎片化评估，聚焦于物理推理、视觉保真度或行为正确性，导致这些能力的整合评估不足。为弥补这一空白，我们提出ExeWorldBench，据我们所知，这是首个通过基于代码的3D世界构建来综合评估世界理解与可执行实现的基准。ExeWorldBench包含三个逐步挑战的任务：机制可执行性，评估从视觉观察中实现单个机制；组合可执行性，扩展到具有空间、机械和因果依赖的多阶段系统；以及意图驱动的可执行世界合成，要求从高层目标和物理需求构建可执行世界。对前沿模型的广泛实验揭示了可执行世界建模中的重大局限，为开发更具物理基础和因果连贯的世界构建系统提供了洞见。
📄 English Abstract
Coding agents are increasingly evolving beyond reproducing what a world looks like toward constructing executable worlds that capture how a world works. This shift calls for executable world modeling, which integrates two core capabilities: world understanding, reasoning about the underlying structural, mechanical, and causal principles of a world, and executable realization, translating that understanding into a code-based system whose behavior emerges from physical interactions. However, existing benchmarks provide only fragmented evaluations centered on physical reasoning, visual fidelity, or behavioral correctness, leaving the integration of these capabilities insufficiently assessed. To bridge this gap, we introduce ExeWorldBench, to our knowledge the first benchmark to comprehensively evaluate world understanding and executable realization through code-based 3D world construction. ExeWorldBench comprises three progressively challenging tasks: Mechanism Executability, which evaluates the realization of individual mechanisms from visual observations; Compositional Executability, which extends to multi-stage systems with spatial, mechanical, and causal dependencies; and Intent-Driven Executable World Synthesis, which requires constructing executable worlds from high-level goals and physical requirements. Extensive experiments on frontier models reveal substantial limitations in executable world modeling, offering insights into the development of more physically grounded and causally coherent world construction systems.
13. LeWAM：基于JEPA的端到端世界与动作建模
- 英文标题：LeWAM: End-to-End World and Action Modeling with JEPAs
- arXiv：2610.12818
💡 一句话解读：现有JEPA模型先学视觉和动态，再单独训练控制策略，导致任务成功率低、规划慢。LeWAM把预测下一状态和生成动作放在一起端到端训练，并约束状态表示服从高斯分布，让一个模型既能直接输出动作，也能快速评估候选动作做规划。
联合嵌入预测架构（JEPA）无需像素重建即可学习紧凑的潜在世界模型。现有JEPA流程通常先训练视觉表示和动作条件动态，再添加控制策略或求解器。我们提出LeWAM，一种用于控制的端到端JEPA，其特点包括：（1）对下一潜在状态预测和动作流匹配进行联合训练的目标；（2）对潜在状态施加各向同性高斯约束，使单个模型既能直接执行策略，又能通过其学习到的潜在动态评估采样动作来进行高效规划。与先前用于视觉规划的JEPA相比，LeWAM在接触密集任务上将成功率从28.6%提升至89.7%，在长时程任务上从10.9%提升至46.2%，且规划速度提升32.3倍。在仿真和物理机器人上的行为克隆任务中，LeWAM也与已有方法具有竞争力。可解释性分析表明，LeWAM学习到了与控制相关的表示和动作分布，能有效指导测试时的规划与控制。
📄 English Abstract
Joint-embedding predictive architectures (JEPAs) learn compact latent world models without pixel reconstruction. Existing JEPA pipelines typically train visual representations and action-conditioned dynamics before adding a control policy or solver. We introduce LeWAM, an end-to-end JEPA for control with: (1) a joint training objective over next latent state prediction and action flow matching, and (2) an isotropic Gaussian constraint on the latent state, enabling a single model to support direct policy execution and efficient planning by evaluating sampled actions with its learned latent dynamics. Compared with prior JEPAs in visual planning, LeWAM improves success rate from 28.6% to 89.7% on contact-rich tasks and from 10.9% to 46.2% on long-horizon tasks, and plans 32.3 times faster. On behavior cloning tasks in simulation and physical robots, LeWAM is also competitive with established methods. Interpretability analyses show that LeWAM learns control-relevant representations and action distributions effectively guiding test-time planning and control.
14. 得分更高，回答更差：通过协议级评分量规缓解基于评分量规的强化学习中的奖励黑客行为
- 英文标题：Scoring Higher, Answering Worse: Mitigating Reward Hacking in Rubric-Based RL via Protocol-Level Rubrics
- arXiv：2609.38847
💡 一句话解读：基于评分量规的强化学习用打分规则给模型当奖励，但简单的加权求和会让模型钻空子：漏掉关键点也能靠无关的废话把分数补回来，结果分数更高、回答更差。作者提出把评分项按临床流程分组，只有一组内所有条件都满足才算得分，从而堵住这个漏洞。
基于评分量规的强化学习（Rubric-RL）在没有验证器的情况下训练语言模型：由评判者逐条检查量规中的每个标准，再把判定结果聚合成奖励，最常见的方式是加权求和。我们发现这种加性聚合正是薄弱环节。在求和方式下，各标准可以相互补偿：一个漏掉关键决策的策略，可以用没人需要的建议把分数买回来。在临床问诊中，这样的策略得分更高、回答却更差：量规覆盖率上升，但在留出的医生标准上的恰当性却低于未训练的模型。问题不在医学标准本身。把这些标准分组、要求它们必须同时成立后，同样的标准一字未改，就能挽回三分之一的损失；而缩短回答几乎毫无帮助。因此我们提出协议级量规（ProRubric），它保留标准所要求的内容，只改变聚合方式：把一份清单归并为少数几个协议级维度，只有当某个维度的所有标准都成立、且其失败条款未被触发时，该维度才计分。分组只需离线做一次，且不改变优化器。ProRubric 在不损失覆盖率的情况下将恰当性提升 10.8 分，并在两种规模下都取得七个基准测试的最佳平均成绩。奖励的有效性不仅取决于量规验证了什么，还取决于它如何聚合。
📄 English Abstract
Rubric-based reinforcement learning (Rubric-RL) trains language models where no verifier exists. A judge checks each criterion of a rubric, and the verdicts are aggregated into a reward, most often by a weighted sum. We show that this additive aggregation is the weak point. Under a sum, criteria compensate for one another: a policy that misses the one decision that matters can buy the points back with advice nobody asked for. On clinical consultation, such a policy scores higher and answers worse. Rubric coverage rises while appropriateness on held-out physician criteria falls below the untrained model. The medical criteria are not to blame. Grouped so that they must hold together, the same criteria, unchanged to the word, recover a third of the loss; shorter answers recover almost none. We therefore propose Protocol-level Rubrics (ProRubric), which keeps what the criteria ask for and changes how they are aggregated. It groups a checklist into a few protocol-level dimensions. A dimension counts only when all of its criteria hold and its failure clause does not fire. The grouping is done once, offline, and leaves the optimizer unchanged. ProRubric raises appropriateness by 10.8 points without losing coverage and has the best seven-benchmark average at both scales. Reward validity is set not only by what a rubric verifies, but by how it aggregates. Code is available at https://github.com/Estrellajer/ProRubric
15. Multics：基于分层通信的生成式多玩家世界模型
- 英文标题：Multics: Generative Multi-Player World Model with Hierarchical Communication
- arXiv：2610.17201
💡 一句话解读：多玩家世界模型要让多个玩家在同一虚拟环境里各自看到画面，但玩家一多，现有方法要么计算量爆炸，要么互相干扰。本文提出 Multics，把玩家信息按“个人—小组—全局”三层来传递，只让相关玩家互相交流，还能随人数灵活扩展，训练4人即可直接用于更多玩家。
多玩家交互式世界模型旨在从多个玩家的视角模拟共享环境，但要在扩展到更多玩家的同时实现稳健生成仍具挑战。现有依赖全连接通信的方法（如 Solaris）会随玩家数量增长带来二次方的注意力开销，而集中式方法（如 γ-World）则因共享的全局表示而受到无关事件之间的干扰。为解决这些问题，我们提出 Multics，一种采用分层通信的多玩家视频生成框架，在玩家、小组和全局三个层级上组织信息交换。该分层结构将小组 token（group token）作为中间的小组级表示引入自注意力，用以聚合局部相关的玩家信息；这些 token 再与全局 token 相连以实现跨组通信，而分层拓扑限制了无关局部事件之间的直接混合。小组 token 的参数在各小组槽位间共享，使同一通信模块无需引入新的组专属参数即可适应不同数量的玩家。我们进一步提出基于近期相机轨迹的预算约束动态分组，能够自适应地将相关玩家分组，同时限制不必要的通信。实验表明，Multics 在 4 人和 6 人设置下达到最先进性能，在跨玩家一致性、玩家特定动作与相机可控性以及抗干扰鲁棒性方面均有提升。此外，Multics 无需重新训练即可从 4 人训练泛化到 6、8、10 人推理，同时相比全连接通信减少了自注意力计算量。
📄 English Abstract
Multi-player interactive world models aim to simulate shared environments from multiple players' perspectives, but achieving robust generation while scaling to more players remains challenging. Existing approaches that rely on fully-connected communication (e.g., Solaris) incur quadratic attention costs as the number of players grows, while centralized approaches (e.g., $\gamma$-World) suffer from interference between unrelated events in a shared global representation. To address these challenges, we present Multics, a multi-player video generation framework with hierarchical communication that structures information exchange across player, group, and global levels. This hierarchy introduces group tokens into self-attention as intermediate group-level representations that aggregate locally relevant player information. These tokens connect to global tokens for cross-group communication, while the hierarchical topology limits direct mixing between unrelated local events. Group-token parameters are shared across group slots, allowing the same communication module to accommodate varying numbers of players without introducing new group-specific parameters. We further propose budget-constrained dynamic grouping based on recent camera trajectories, which adaptively groups relevant players while limiting unnecessary communication. Experiments demonstrate that Multics achieves state-of-the-art performance in 4- and 6-player settings, with improved cross-player consistency, player-specific action and camera controllability, and interference robustness. Moreover, Multics generalizes from 4-player training to 6-, 8-, and 10-player inference without retraining, while reducing self-attention computation compared with fully-connected communication.