Daily Papers · 2026-10-20
📄 论文日报 | 来源：Hugging Face Daily Papers
🗓 Episode：2026.10.19 | 具身智能跨门槛未实用；业务智能体合规可靠待提升
1. GPT-6-Astra 是否已实现具身智能？
- 英文标题：Has GPT-6-Astra Achieved Embodied Intelligence?
- arXiv：2610.22471
💡 一句话解读：这篇研究要回答的问题是：GPT-6-Astra 是否已经具备实用的具身智能（能像人一样在物理世界中灵活操作和移动）。作者通过大规模测试发现，它在多种机器人任务上表现惊人，但真实部署中仍面临演示数据昂贵、执行时间过长等瓶颈，因此认为它已跨过能力门槛，但尚未达到实用水平。
GPT-6-Astra 在机器人操作与导航方面展现出的能力，挑战了为单个任务和本体构建专用模型的传统范式。在本报告中，我们不仅追问 GPT-6-Astra 有多强，还探讨其能力对该领域意味着什么。首先，我们探测独立 GPT-6-Astra 的极限，在涵盖桌面操作、人形机器人、移动平台和灵巧手的 12 个代表性基准上进行大规模评估；在允许从训练集演示中进行上下文学习（ICL）并大幅增加执行步数的统一协议下，GPT-6-Astra 取得了前所未有的性能（图 1）。未来研究可能聚焦于提升：1）ICL 的泛化性；2）成本与执行时间；3）物理状态验证；4）不可逆故障的恢复；5）长时程任务中进度的保持。其次，我们研究用于具身控制的智能体框架，发现高层“系统 2”推理器（即 GPT-6-Astra）与低层“系统 1”控制器（如 π0.5）之间的闭环协调虽然成本高昂但十分有效。第三，我们在真实世界部署中检验 GPT-6-Astra，在 8 种机器人配置上的实验揭示了显著差距：1）ICL 所需的机器人演示收集成本高昂；2）在物理约束下，用更长的执行时间换取更高成功率并不现实。第四，我们展望下一代具身基准，开发了一个原型，其特点为：1）从真实世界需求中衍生任务；2）不提供域内演示；3）在世界时钟持续运行（即智能体思考时时间仍在流逝）的条件下评判智能体。综合来看，我们的结果表明 GPT-6-Astra 已基本跨过具身能力的门槛，但尚未达到实用具身智能的门槛。
📄 English Abstract
GPT-6-Astra demonstrates a level of competence in robotic manipulation and navigation that challenges the conventional paradigm of building specialized models for individual tasks and embodiments. In this report, we ask not only how capable GPT-6-Astra is, but what its capabilities imply for the field. First, we probe the limits of standalone GPT-6-Astra. We conduct a large-scale evaluation across 12 representative benchmarks spanning tabletop manipulation, humanoids, mobile platforms, and dexterous hands. Under a unified protocol that allows in-context learning (ICL) from training-set demonstrations and substantially increases the number of execution steps, GPT-6-Astra achieves unprecedented performance (Fig. 1). Future research may focus on improving 1) the generality of ICL 2) cost and execution time, 3) physical-state verification, 4) recovery from irreversible failures, and 5) long-horizon preservation of progress. Second, we study the agentic harness for embodied control. We identify that close-loop coordination between a high-level “System 2” reasoner (i.e., GPT-6-Astra) and a low-level “System 1” controller (e.g., {\pi}0.5) is costly but effective. Third, we examine GPT-6-Astra on real-world deployment. Our experiments on 8 robot configurations reveal substantial gaps: 1) robot demonstrations required for ICL are expensive to collect, and 2) it is implausible to trade longer execution for higher success rates with physical constraints. Fourth, we envision the next generation of embodied benchmarks. We develop a prototype that 1) derives tasks from real-world needs, 2) provides no in-domain demonstrations, and 3) judges agents on a world clock that keeps running while they deliberate. Taken together, our results suggest that GPT-6-Astra has largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence.
2. BizBench：揭示真实业务中部署长时程智能体的隐藏挑战
- 英文标题：BizAgentBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business
- arXiv：2610.20424
💡 一句话解读：现有智能体评测只看任务最终是否完成，忽略了执行过程是否合规。本研究构建了基于真实支付公司业务流程的评测基准，用789份文档和601个工具模拟真实环境，考察智能体能否在遵守约束的前提下完成任务。
现有的长时程智能体基准往往强调最终任务成功，而对中间执行的有效性评估有限。然而，真实业务运营要求智能体在满足关键约束的前提下，处理大量可能过时或相互冲突的政策。这些要求在高风险领域（如支付）尤为关键，因为无论最终结果如何，中间环节的违规都可能带来财务损失或监管后果。为弥补这一评估空白，我们提出了 BizAgentBench，这是一个基于某领先支付公司真实业务流程构建的多轮基准。其可执行环境整合了 789 份知识文档和 601 个可发现的业务工具，并保持策略、状态与工具行为的一致性。该基准包含 50 项任务，涉及资源约束、时间依赖、先前承诺以及相互冲突或过时的指引。每项任务均通过参考执行进行验证，同时采用确定性检查评估目标达成与执行有效性，而不规定单一解法路径。实验显示，即便是前沿模型也存在明显局限：Claude Opus 5 和 GPT-6 Sol 的严格成功率分别仅为 64.0% 和 61.3%。对于 Qwen3.8-27B，目标达成率为 42.0%，但严格成功率降至 32.7%，表明基于结果的评估可能掩盖无效执行。除成功率外，DeepSeek-V4.1-Flash 平均每次运行调用工具 191.3 次，比参考执行多约 70%，显示出可观的执行开销。总体而言，BizAgentBench 为评估和推进长时程智能体提供了一个测试平台，使其超越任务完成，迈向在复杂业务环境中可靠且符合约束的执行。
📄 English Abstract
Existing long-horizon agent benchmarks often emphasize final task success while providing limited assessment of intermediate execution validity. Yet real-world business operations require agents to navigate extensive, potentially outdated or conflicting policies while satisfying critical constraints throughout execution. These requirements are particularly consequential in high-stakes domains such as payments, where intermediate violations can incur financial losses or regulatory consequences regardless of the final outcome. To bridge this evaluation gap, we introduce BizAgentBench, a multi-turn benchmark grounded in real-world procedures from a leading payments company. Its executable environment integrates 789 knowledge documents and 601 discoverable business tools, with consistent policies, states, and tool behavior. The benchmark comprises 50 tasks involving resource constraints, temporal dependencies, prior commitments, and conflicting or outdated guidance. Each task is validated through reference execution, while deterministic checks assess both goal attainment and execution validity without prescribing a single solution path. Experiments reveal substantial limitations even among frontier models: Claude Opus 5 and GPT-6 Sol achieve strict success rates of only 64.0% and 61.3%, respectively. For Qwen3.8-27B, goal attainment reaches 42.0%, but strict success drops to 32.7%, demonstrating how outcome-based evaluation can mask invalid execution. Beyond success rates, DeepSeek-V4.1-Flash averages 191.3 tool calls per run, approximately 70% more than reference executions, revealing substantial execution overhead. Overall, BizAgentBench establishes a testbed for evaluating and advancing long-horizon agents beyond task completion, toward reliable and constraint-compliant execution in complex business environments.
3. 在三维世界中学习思考与探索
- 英文标题：Learning to Think and Explore in 3D World
- arXiv：2610.21783
💡 一句话解读：现有视觉语言模型看3D场景时，要么一次喂多张图或视频，要么依赖3D模型，不够像人。本研究让模型像人一样一次只看一张图，把3D场景当作可走动、可转头观察的环境，自己决定往哪看、怎么探索，再回答问题。训练时只用最终答案对错来奖励，不教它该看哪里，模型便学会了按问题需要主动寻找被遮挡的目标、避免乱逛。
我们研究视觉语言模型（VLM）如何学习在三维空间中进行思考与主动探索。我们没有向模型提供多视角图像、视频或三维表示，而是保留 VLM 原生的视觉编码方式，即像人类一样每次只输入单张图像，并将三维表示仅视为一个环境，模型可通过动作（如向左移动）和观测（渲染图像）与之交互。每次 rollout 从初始第一人称视角和一个视觉问题开始，模型可以迭代推理、导航并调整视角，最后给出最终答案。在初步的轨迹监督微调之后，我们采用强化学习（RL），其奖励主要来自最终答案的正确性。关键在于，RL 并不直接监督模型应访问哪些视角或采取哪些探索动作，模型必须仅根据其交互在多大程度上提升性能来学习有效的探索策略。在六个三维基准上，我们的模型 ACT3D 平均准确率达到 53.2%，比其基座 Qwen 模型提升 22.0%，并超过专有的 Gemini 3.6 Flash。ACT3D 还优于现有开源三维方法，同时使用的输入帧数少 50–70%，并能泛化到未见领域，相比其监督初始化平均提升 5.9%。进一步分析揭示了自适应的、以问题为条件的探索行为：相比基座模型的性能增益从初始视角即可回答的问题上的 0.6%，上升到需要大量探索的问题上的 16.3%。ACT3D 会寻找被遮挡的目标、避免不必要的探索，并为空间比较获取互补视角。这些发现表明，基于结果的 RL 不仅使 VLM 能够对视觉证据进行推理，还能学会何时以及如何获取视觉证据。
📄 English Abstract
We study how vision-language models (VLMs) can learn to think and actively explore in 3D space. Instead of providing multi-view images, videos, or 3D representations to the model, we preserve the native visual encoding of VLMs, i.e., a single image, one at a time, just like humans, and treat the 3D representation only as an environment with which the model can interact through actions (e.g., move left) and observations (rendered images). Each rollout begins from an initial first-person view and a visual question. The model can then iteratively reason, navigate, and reorient its viewpoint before committing to a final answer. Following an initial trajectory-supervised finetuning, we apply reinforcement learning (RL) with rewards driven primarily by final-answer correctness. Crucially, RL provides no direct supervision over which viewpoints the model should visit or which exploration actions it should take. Instead, the model must learn an effective exploration policy solely from the extent to which its interactions improve performance. Across six 3D benchmarks, our model ACT3D achieves an average accuracy of 53.2% that improves its base Qwen model by 22.0% and surpasses proprietary Gemini 3.6 Flash. ACT3D also outperforms existing open-source 3D methods while using 50-70% fewer input frames, and generalizes to unseen domains with an average improvement of 5.9% over its supervised initialization. Further analysis reveals adaptive, question-conditioned exploration. Performance gains over the base model increase from 0.6% on questions answerable from the initial view to 16.3% on those requiring extensive exploration. ACT3D searches for occluded targets, avoids unnecessary exploration, and acquires complementary viewpoints for spatial comparisons. These findings demonstrate that outcome-based RL enables VLMs not only to reason over visual evidence, but also to learn when and how to acquire it.
4. 进入未知：面向开放科学研究问题的AI智能体基准测试
- 英文标题：Into the Unknown: Benchmarking AI agents on open scientific research problems
- arXiv：2610.19159
💡 一句话解读：这篇研究想搞清楚：AI智能体能不能像科学家一样，针对一个尚未解决的科学问题写出后续研究论文。为此，作者联合美国能源部国家实验室的科学家，用他们未发表的后续研究方向和对论文的硬性要求做成评分标准，让九个前沿AI智能体各写15篇论文来测试。结果没有一篇完全达标，多数还因为编造没做过的工作或数据对不上而没通过真实性检查。
AI智能体能否针对开放科学问题开展研究？我们提出“Into the Unknown”基准，向智能体提供一篇科学论文和一个待探索的研究方向，要求它撰写一篇后续论文。我们与美国能源部（DOE）国家实验室的科学家合作，他们各自贡献了自己工作中一个未发表的后续方向，以及投稿必须满足的必要条件。尽管评估后续论文的质量通常很困难，但用这些条件构成评分量表，可使没有参考答案的研究问题更易于评估。我们让九个前沿智能体各撰写15篇后续论文，覆盖量子计算、等离子体物理、计算化学、材料科学和科学机器学习等领域。135篇投稿中没有一篇满足全部条件，114篇未通过我们的真实性检查，最常见的原因是论文声称完成了从未做过的工作，或报告的数字与记录证据相矛盾。总体而言，智能体能轻易生成看似后续论文的文稿，但没有一篇达到领域科学家的标准，仍有很大提升空间。我们的代码和数据已公开。
📄 English Abstract
Can AI agents carry out research on open scientific problems? We introduce Into the Unknown, a benchmark that provides an agent with a science paper along with a research direction to pursue and tasks it to write a follow-up paper. We work with scientists at US Department of Energy (DOE) national laboratories, who each contribute an unpublished follow-up direction from their own work together with the necessary conditions a submission must meet. While assessing the quality of a follow-up paper is challenging in general, using these conditions to form a rubric makes evaluation more tractable for research problems that have no reference answer. We have nine frontier agents write 15 follow-ups each, spanning quantum computing, plasma physics, computational chemistry, materials science, and scientific machine learning domains. None of the 135 submissions satisfies all the conditions, and 114 fail our integrity check, most often because the paper claims work that was never done or reports numbers that contradict the recorded evidence. Overall, agents readily produce papers that look like follow-ups, but none meets the standard of the domain scientist, leaving substantial headroom. Our code and data are publicly available.
5. 轻装上阵，CarryOn：为长时程智能体构建面向对象的上下文
- 英文标题：Pack Light, CarryOn: Building Object-Oriented Context for Long-Horizon Agents
- arXiv：2610.18784
💡 一句话解读：长时程智能体常把大量交互历史反复塞进上下文，既费token又冗余。CarryOn把上下文改成围绕可复用的“任务对象”来组织，只保留进度载体、补充执行信息和近期交互，并在旧历史淘汰后增量更新这些对象，从而在OSWorld 2.0和BrowseComp-Plus上分别提升5.8%和7.0%，同时省下51.9%和22.9%的token。
长时程智能体（long-horizon agents）将进展累积在持久任务状态中，却反复处理描述该进展如何达成的交互历史。这引出一个自然的问题：长时程上下文能否改为围绕可复用的进展来组织，而非围绕产生该进展的历史？本文提出CarryOn，一个面向对象的上下文运行时（object-oriented context runtime），它以可复用任务对象为中心组织续接过程。CarryOn不把压缩历史或重建的进展叙述作为续接的主要依据，而是由续接载体（continuation carriers）、补充执行上下文以及近期交互的实时尾部共同构建上下文视图。核心挑战在于：当周围历史被淘汰后，如何保持这一可复用基础仍然有效，即如何在不于每次折叠时重建完整表示的情况下更新该基础。CarryOn跨折叠维护载体记录，依据绑定证据和适用的环境检查修订受影响的记录，同时保留未受影响的记录并联合更新执行上下文。在OSWorld 2.0和BrowseComp-Plus上的实验分别取得5.8%和7.0%的提升，同时总token使用量分别减少51.9%和22.9%。项目页面：this https URL
📄 English Abstract
Long-horizon agents accumulate progress in persistent task state, yet repeatedly process interaction histories describing how that progress was achieved. This raises a natural question: can long-horizon context instead be organized around reusable progress rather than the history that produced it? In this paper, we introduce CarryOn, an object-oriented context runtime that organizes continuation around reusable task objects. Rather than treating compressed histories or reconstructed accounts of progress as the primary basis for continuation, CarryOn builds a context view from continuation carriers, complementary execution context, and a live tail of recent interactions. A core challenge is keeping this reusable basis valid after its surrounding history is retired: how can CarryOn update the basis without reconstructing the full representation at every fold? CarryOn maintains carrier records across folds, revising affected records against bound evidence and applicable environment checks while carrying forward unaffected records and jointly updating execution context. Experiments on OSWorld 2.0 and BrowseComp-Plus show improved gains of 5.8% and 7.0% while reducing total token usage by 51.9% and 22.9%, respectively. Project page: this https URL
6. HexaShell：探索VLM智能体作为机器人编程智能体极限的轻量级执行框架
- 英文标题：HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
- arXiv：2610.23479
💡 一句话解读：研究想弄清：让会写代码的AI（基于视觉语言模型）直接操控双臂机器人，到底能走多远、卡在哪里。方法是给它一个极简框架，只把命令转成动作；AI通过摄像头看现场，失败后读日志、自己写小工具程序来纠错，并靠多相机三角定位找物体位置。结果在42个任务上成功率39.7%，刷新纪录，但瓶颈在接触类精细操作。
基于视觉语言模型（VLM）的智能体如今在机器人操作排行榜上已能与学习型策略一较高下。我们探究：当编码智能体控制的是机器人而非计算机时，它能走多远、又会在哪里止步。我们的系统 HexaShell 保持执行框架轻量：它只负责把智能体的命令转化为机器人运动。智能体通过一条命令行驱动双臂机器人，可看到三路 RGB 相机并已知其标定参数。每次失败后，智能体读取日志、定位原因，并编写或修复称为“工具”的小程序，从而像编码智能体修复失败测试那样把所学固化为代码。这些工具只能利用机器人观测到的信息，通过三角定位确定物体相对于彼此及夹爪的位置：若两个相机看到同一点，该点位于两条相机射线的交点；若点位于已知平面（如桌面或方块顶面），则位于其射线与该平面的交点；当头相机不够用时，机器人移动腕部相机获取第二视角。我们仅用每个任务同一随机种子下的十种布局构建工具，随后冻结工具并在同种子的十个新布局上测试。我们使用中等推理强度的 Codex 与 GPT-6-Astra。在全部 42 个 RoboDojo 任务上，HexaShell 取得 39.7% 的成功率和 48.0 的进度分（按排行榜相同方式平均），创下 RoboDojo 新的最优成绩。瓶颈在于接触而非推理：推物任务达 92.5%，记忆类任务 65.0%，但紧密插入仅 8.3%。智能体失败时通常能看出失败原因，却无法让动作奏效。我们已在此链接开源目前演化出的工具。
📄 English Abstract
Agents built on vision language models (VLMs) now compete with learned policies on robot manipulation leaderboards. We ask how far a coding agent can go when it controls a robot instead of a computer, and where it stops. Our system, HexaShell, keeps the harness thin: it only turns the agent’s commands into robot motion. The agent drives a robot with two arms through one command line. It sees three RGB cameras and knows their calibration. After each failed attempt, the agent reads the log, finds the cause, and writes or fixes small programs called tools. In this way it keeps what it learns as code, as a coding agent does when it fixes a failing test. The tools may use only what the robot observes. They use triangulation to find where objects are, relative to each other and to the gripper. If two cameras see the same point, the point lies where the two camera rays meet. If a point lies on a known plane, such as the table or the top of a block, it lies where its ray meets that plane. When the head camera is not enough, the robot moves a wrist camera to get a second view. We build the tools on only ten layouts of one seed per task, then freeze them and test them on ten new layouts of the same seed. We use Codex with GPT-6-Astra at medium reasoning effort. On all 42 RoboDojo tasks, HexaShell reaches a success rate of 39.7% and a progress score of 48.0, averaged in the same way as the leaderboard. This is a new state of the art on RoboDojo. The limit is contact, not reasoning: pushing reaches 92.5% and memory tasks 65.0%, but tight insertion only 8.3%. When the agent fails, it usually sees the failure but cannot make the motion work. We open source the tools evolved so far at this https URL .
7. LiveMACE：演化市场中LLM智能体能力的过程感知评估
- 英文标题：LiveMACE: Process-Aware Evaluation of LLM Agent Capabilities in Evolving Markets
- arXiv：2610.09872
💡 一句话解读：只看最终结果来评价AI智能体，会掩盖它到底有没有真本事，在环境不断变化时尤其如此。研究者搭建了一个用真实金融市场当“动态考场”的测试平台，让五个大模型智能体连续运行30天，并记录完整决策过程，从而把“结果好坏”和“能力高低”分开衡量，找出各自卡在哪一环。
仅凭结果来评估智能体，会掩盖产生这些结果的能力。这一问题在演化环境中尤为突出，因为结果反映的是智能体行为与不断变化的外部条件之间的闭环互动。我们提出 LiveMACEBench，一个过程感知的基准测试，以真实运行的金融市场作为自然演化的测试平台，用于评估持久运行的 LLM 智能体。五个前沿 LLM 在工具使用（Tool Use）、持久记忆（Persistent Memory）、规则遵循（Rule Following）和多智能体协作（Multi-Agent Collaboration）配置相匹配的条件下，沿连续轨迹运行。我们既通过实际结果、也通过从完整决策轨迹中得出的机制特异性诊断来评估它们。在为期 30 天的实时评估中，我们发现明显的结果—能力差距：实际收益常常与能力特异性测量结果不一致，而相似的结果可能来自截然不同的机制使用模式。轨迹层面的诊断进一步揭示了各项能力上的不同瓶颈，表明机制的可及性、机制的有效使用与下游表现并不是可互换的智能体能力度量。LiveMACEBench 使这一区分变得可测量，将真实市场从性能排行榜转变为智能体能力的诊断环境。
📄 English Abstract
Evaluating agents by outcomes alone can obscure the capabilities that produce them. This problem is especially pronounced in evolving environments, where outcomes reflect a closed-loop interaction between agent behavior and changing external conditions. We introduce LiveMACEBench, a process-aware benchmark that uses live financial markets as a naturally evolving testbed for persistent LLM agents. Five frontier LLMs operate along continuous trajectories under matched Tool Use, Persistent Memory, Rule Following, and Multi-Agent Collaboration configurations. We evaluate them through both realized outcomes and mechanism-specific diagnostics derived from complete decision traces. Across 30 days of live evaluation, we find a pronounced outcome-capability gap: realized returns often diverge from capability-specific measurements, and similar outcomes can arise from markedly different patterns of mechanism use. Trace-level diagnostics further expose distinct bottlenecks across capabilities, demonstrating that mechanism access, effective mechanism use, and downstream performance are not interchangeable measures of agent capability. LiveMACEBench makes this distinction measurable, turning live markets from a performance leaderboard into a diagnostic environment for agent capability
8. Scene-Designer：面向智能体3D场景生成的设计智能学习
- 英文标题：Scene-Designer: Learning Design Intelligence for Agentic 3D Scene Generation
- arXiv：2610.23121
💡 一句话解读：现有方法生成3D室内场景时缺乏专业设计经验。本文提出Scene-Designer智能体框架，通过“设计+批评”两类经验学习设计智能，并用批评引导的强化学习提升场景质量。
从自然语言生成完整、连贯且可用的3D室内场景需要强大的设计智能，即设计高质量场景并评判其质量的能力。然而，现有方法要么依赖缺乏领域设计经验的通用模型，要么从静态场景中学习。本文提出Scene-Designer，一个通过互补的设计与批评经验学习设计智能的智能体框架。我们构建了两个数据集：Scene-OTA-35K提供“观察—思考—动作”监督以学习场景设计与高效纠错，Scene-OTC-15K提供“观察—思考—批评”监督以学习跨物理有效性、功能正确性和布局连贯性的场景质量评判。基于这些经验，我们提出Scene-GRPO，一种批评引导的GRPO框架，用于改进设计决策并减少不必要的试错纠正。大量实验表明，Scene-Designer在物理有效性、功能正确性和布局连贯性方面取得更优性能，且学习到的Scene-Critic在场景批评上显著优于通用模型。此外，测试时的批评引导进一步改善场景生成，表明批评器在训练之外也能提供有用指导。额外分析表明场景设计与批评能力之间存在相互增强的趋势。
📄 English Abstract
Generating complete, coherent, and usable 3D indoor scenes from natural language requires strong \textbf{design intelligence}: the ability to design high-quality scenes and critique their quality. However, existing approaches either rely on general-purpose models lacking domain-specific design experience or learn from static scenes. In this work, we introduce \textbf{Scene-Designer}, an agentic framework that learns design intelligence through complementary design and critique experiences. We construct two datasets: \textbf{Scene-OTA-35K}, which provides Observation–Thought–Action supervision for learning scene design and efficient error correction, and \textbf{Scene-OTC-15K}, which provides Observation–Thought–Critic supervision for learning scene quality critique across physical validity, functional correctness, and layout coherence. Based on these experiences, we propose \textbf{Scene-GRPO}, a Critic-guided GRPO framework that improves design decisions and reduces unnecessary trial-and-error corrections. Extensive experiments demonstrate that Scene-Designer achieves superior performance in physical validity, functional correctness, and layout coherence, while the learned Scene-Critic significantly outperforms general-purpose models in scene critique. Moreover, test-time Critic guidance further improves scene generation, showing that the Critic can provide useful guidance beyond training. Additional analysis suggests a trend of mutual enhancement between scene design and critique abilities.
9. 交互式视频生成中的推测采样
- 英文标题：Speculative Sampling in Interactive Video Generation
- arXiv：2610.20556
💡 一句话解读：交互式视频生成为了降低延迟，通常把视频切成小块逐块生成，但小块会牺牲画质。本文提出“推测采样”：先用大窗口一次性预测更多内容并缓存，条件不变时直接复用，预测结果逐步交付、未定部分仍可替换。这样既保留大窗口的高画质，又实现低延迟交互，无需复杂改造即可让大模型用于实时视频流。
交互式视频生成通常依赖较小的双向块（bidirectional chunks）来降低响应延迟。然而我们的初步实验表明，更大的块能提升视频质量并更好地利用 GPU 并行性，使得在采样时间仅小幅增加的情况下采样更多潜变量（latents）。受此启发，我们提出推测采样（Speculative Sampling），一种将大双向窗口的建模能力与小窗口的低延迟交互相结合的方案。它通过附加噪声潜变量作为填充来扩展建模窗口，并缓存所得预测，以便在条件不变时复用。预测结果被增量交付，而未提交的未来部分仍可替换。我们的延迟分析表明，这消除了由块边界施加的等待时间下界，将优化重点从减小块大小转向加速每块采样。因此，双向窗口大小成为联合优化视觉质量与采样时间的决策变量，而不再是对低延迟交互的必要约束。这为在交互式流式系统中部署大型双向模型开辟了道路，而无需复杂的流式后训练（streaming post-training）。
📄 English Abstract
Interactive video generation commonly relies on small bidirectional chunks to reduce response latency. Yet our preliminary experiments show that larger chunks improve video quality and better utilize GPU parallelism, allowing more latents to be sampled with only a modest increase in sampling time. Motivated by these observations, we propose Speculative Sampling, a strategy that combines the modeling power of large bidirectional windows with the low-latency interaction of small ones. It expands the modeling window by appending noise latents as padding and caches the resulting predictions for reuse when the condition remains unchanged. Predictions are delivered incrementally, while the uncommitted future remains replaceable. Our latency analysis shows that this removes the waiting-time lower bound imposed by chunk boundaries, shifting the optimization focus from reducing chunk size to accelerating per-chunk sampling. Bidirectional window size thus becomes a decision variable for jointly optimizing visual quality and sampling time, rather than a necessary constraint on low-latency interaction. This opens a path to serving large bidirectional models in interactive streaming systems without complex streaming post-training.
10. OpenMobile-2：以可扩展环境和应用原生工具构建多功能移动智能体
- 英文标题：OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools
- arXiv：2610.14157
💡 一句话解读：现有手机智能体训练环境少、数据缺、工具调用弱。研究团队搭建了含110多个应用的可控模拟环境，收集近1.2万条操作轨迹和2千多个可自动判分的训练任务，并让应用把部分功能开放成工具，使智能体能交替点击界面和调用工具，从而训练出性能接近前沿的开放手机智能体。
我们推出 OpenMobile-2，这是一个训练环境和配方完全开源的准前沿移动智能体。研究有三项关键进展：（1）构建含模拟商业应用的多样的环境：我们搭建了 MobileGym++，包含 35 个逼真、功能丰富的商业风格应用及跨应用工作流，同时保持重置与验证的完全可控性；结合 Android 模拟器中新配置的应用，形成覆盖 110 多个应用的多样化试验场。（2）大规模开放训练数据：在此基础上，我们整理了近 1.2 万条移动交互轨迹用于监督微调，并构建了目前最大的移动智能体可验证强化学习训练数据集合，包含 2000 多个带自动奖励的可执行任务。（3）GUI 与应用原生工具的混合使用：我们探索了一种实验性的移动使用设定，即应用在提供图形界面的同时将部分功能暴露为应用原生工具，使智能体可在任务中交替执行 GUI 操作与工具调用；我们在 MobileGym++ 中实现了该设定，覆盖 50 多个应用、300 多个精心界定范围的工具。此外，我们推出 MobileGym++ Bench，用于复杂、长时程的商业移动场景，可在同一任务套件上支持纯 GUI 与 GUI-工具混合两种评估。OpenMobile-2 在 AndroidWorld（79.9）和 MobileWorld（50.4）等既有基准上表现具有竞争力，并展现出向真实设备移动使用迁移的潜力，在 SPA-Bench 上从 31.9 提升至 59.6，接近翻倍。所有资源均可在本文给出的网址获取。
📄 English Abstract
We introduce OpenMobile-2, a near-frontier mobile agent with fully open training environments and recipes. We make three key advances: (1) Diverse environments with simulated commercial apps: We build MobileGym++, featuring 35 realistic, functionally rich commercial-style apps with cross-app workflows, while preserving full controllability for reset and verification. Together with newly configured apps in Android emulators, this yields a diverse playground spanning over 110 apps. (2) Open training data at scale: Building on this foundation, we curate nearly 12K mobile interaction trajectories for supervised fine-tuning and the largest open collection of verifiable RL training data for mobile agents, comprising over 2K executable tasks with automatic rewards. (3) Hybrid GUI and app-native tool use: We explore an experimental mobile-use setting where apps expose selected functionalities as app-native tools alongside their GUIs, allowing agents to interleave GUI actions and tool calls within a task. We implement this setting in MobileGym++ with over 300 carefully scoped tools across 50+ apps. Additionally, we introduce MobileGym++ Bench for complex, long-horizon commercial mobile scenarios, supporting both GUI-only and hybrid GUI-tool evaluation on a shared task suite. OpenMobile-2 performs competitively across established benchmarks, including AndroidWorld (79.9) and MobileWorld (50.4), while showing promising transfer to real-device mobile use, nearly doubling SPA-Bench performance from 31.9 to 59.6. All resources are available at this https URL .
11. BridgeWAM：通过潜在桥接查询连接视频专家与动作专家
- 英文标题：BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries
- arXiv：2610.14746
💡 一句话解读：现有视频-动作模型把两类专家层层密集相连，动作预测易被无关画面干扰、计算也冗余。作者用32个潜在桥接查询把视觉、语言和本体信息压缩到共享小空间，让动作专家按需取用，并解耦两层深度，实现更轻更快、成功率更高的机器人操作。
世界动作模型（WAMs）通常通过密集的逐层混合Transformer（MoT）注意力将视频专家与动作专家耦合，使动作预测暴露于无关视觉特征，同时强制专家深度匹配。这种耦合妨碍了视觉特征的选择性使用并引入冗余计算。我们提出专家桥（BoE），并实例化为BridgeWAM，它通过32个潜在桥接查询（LBQs）将视觉、语言和本体感受先验聚合到一个共享的有限潜在空间中。联合视频-动作监督使动作梯度引导视频专家朝向更适合动作预测的表示，而动作专家通过交叉注意力学习选择性检索任务相关信息。解耦专家深度并在动作去噪过程中复用LBQs，得到一个5.06B模型，其视频专家为30层、动作专家为两层，在LIBERO上达到98.25%的成功率，在LIBERO-Plus上汇总成功率为72.24%，展示了以超轻量动作头实现的有效任务执行和强泛化能力。消融实验支持联合表示适应和统一LBQ条件化。相对于Fast-WAM，BridgeWAM在仿真分析中将单样本推理FLOPs降低28.72%，并将真实机器人延迟降至77.1毫秒（2.47倍加速）。定性实验结果进一步表明，在各类扰动下模型优先关注目标物体和任务目的地。
📄 English Abstract
World Action Models (WAMs) commonly couple video and action experts through dense, layer-wise Mixture-of-Transformers (MoT) attention, exposing action prediction to irrelevant visual features while enforcing matched expert depths. This coupling hinders selective visual-feature use and introduces redundant computation. We introduce Bridge-of-Experts (BoE), instantiated as BridgeWAM, which aggregates visual, language, and proprioceptive priors through 32 Latent Bridge Queries (LBQs) into a shared, limited latent space. Joint video-action supervision lets action gradients guide the Video Expert toward representations better suited for action prediction, while the Action Expert learns to selectively retrieve task-relevant information through cross-attention. Decoupling expert depths and reusing LBQs during action denoising yield a 5.06B model with a 30-layer Video Expert and a two-layer Action Expert, achieving 98.25% success on LIBERO and 72.24% pooled success on LIBERO-Plus, demonstrating effective task execution and strong generalization with an ultralight action head. Ablations support joint representation adaptation and unified LBQ conditioning. Relative to Fast-WAM, BridgeWAM reduces single-sample inference FLOPs by 28.72% in simulation profiling and real-robot latency to 77.1 ms (a 2.47-fold speedup). Qualitative experimental results further indicate preferential attention to target objects and task destinations across perturbations.
12. 从研究目标到科学闭环：通过证据支持的物理研究评估智能体
- 英文标题：From Research Goals to Scientific Closure: Evaluating Agents Through Evidence-Supported Physics Research
- arXiv：2610.21767
💡 一句话解读：这篇研究要解决的问题是：AI智能体能否像科学家一样，把一项研究从头到尾组织好并得出有充分证据支撑的结论。为此作者构建了PhysClosureBench基准，包含50个专家核验、目标明确的物理研究任务，用原子得分衡量局部成果、用研究闭环等级衡量整体结论是否成立，并评测了七个前沿大模型。
AI智能体能否组织起能得出有充分支撑结论的科学研究？为评估这一能力，我们提出了PhysClosureBench，这是一个包含50个经专家核验、目标明确的研究任务的基准，覆盖强场、阿秒及邻近物理领域。每个任务给出研究目标和关键科学约束，但由智能体自行组织研究过程，并决定如何生成和验证回答问题所需的证据。原子得分（Atomic Score）以20项标准衡量局部科学成果，研究闭环（Research Closure，RC0–RC4）则根据不可或缺的证据要求和适用范围条件是否满足，评估研究对既定问题的解决程度。我们评测了七个前沿大语言模型，其中Claude-Opus-5取得最高平均原子得分（6.29/10），但仅在32.7%的尝试中达到RC4。在整个基准上，36.2%的尝试虽得分不低于8/10却仍未达到RC4；且七个模型在三次运行中都达到RC4的任务数，均明显少于至少一次运行中达到RC4的任务数。案例层面的分析进一步说明，对生成结果的误读以及未解决的物理适用性限制会导致闭环不完整。这些结果共同揭示了AI智能体在局部科学成果与实现完整研究闭环能力之间持续存在的差距。
📄 English Abstract
Can AI agents organize scientific investigations that lead to well-supported conclusions? To evaluate this capability, we introduce PhysClosureBench, a benchmark of 50 expert-verified, goal-specified research tasks spanning strong-field, attosecond, and neighboring physics areas. Each task specifies the research objective and essential scientific constraints, while leaving the agent to organize the investigation and determine how to generate and validate the evidence needed to answer the question. A 20-criterion Atomic Score measures local scientific accomplishment, while Research Closure (RC0–RC4) assesses how fully the investigation resolves the specified research question based on whether its indispensable evidence requirements and applicable scope conditions are satisfied. We evaluate seven frontier LLMs, where Claude-Opus-5 achieves the highest mean Atomic Score (6.29/10) but reaches RC4 in only 32.7% of attempts. Across the benchmark, 36.2% of attempts scoring at least 8/10 still fall short of RC4, while all seven models reach RC4 on substantially fewer tasks in all three runs than in at least one. Case-level analysis further illustrates how misinterpretation of generated results and unresolved physical applicability limits can lead to incomplete closure. Together, these results reveal a persistent gap between AI agents’ local scientific accomplishments and their ability to achieve complete research closure.
13. 面向长时程智能体强化学习的树感知后验蒸馏
- 英文标题：Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning
- arXiv：2610.23181
💡 一句话解读：长时程智能体强化学习有两个难题：只用最终奖励导致每步功劳难分配、波动大；固定批次反复更新又会让策略越跑越偏。本文提出 TreeOPSD：把每轮交互当作树节点，从共享前缀分叉出多条后续并递归回传最终奖励，得到每步的隐式功劳，再把它转成有界的“老师”分布来蒸馏学生策略，无需价值网络或外部强教师。
带可验证奖励的强化学习已成为大语言模型可扩展的后训练范式。然而，长时程交互训练智能体面临两个挑战：仅使用终端奖励会使轮级信用分配稀疏且高方差；而在固定 rollout 批次上进行顺序小批量更新，会因当前策略偏离行为策略而引入策略陈旧性。现有方法通过学习评论家、过程奖励模型或同策略蒸馏获得更稠密的信用，但它们依赖价值网络、步级标注或更强的外部教师，增加了训练复杂度；同时，重要性采样与裁剪可缓解策略漂移，却未指定有限目标来阻止顺序更新过冲。为此，我们提出 TreeOPSD，一个无需评论家、树感知的后验自蒸馏框架，用于长时程智能体强化学习。它将每个交互轮视为树节点，从共享前缀分叉出备选后续，并递归地将终端奖励回传为隐式的轮级信用；随后把每个信用映射为以相对冻结行为策略的似然比为参数、闭式且有界的教师分布。最小化该教师与学生策略间的散度，可在行为策略处沿切线恢复策略梯度方向，同时在顺序小批量中把更新约束到有限目标。实验上，TreeOPSD 在多个有竞争力的基准和模型规模上提升了智能体能力，尤其对长时程任务，例如在 Deep Research 设置下将 Qwen3-30B-A3B-Thinking-2507 的平均准确率较 GRPO 提升 4.97 分。
📄 English Abstract
Reinforcement learning with verifiable rewards has become a scalable post-training paradigm for large language models. However, training agents over long-horizon interactions presents two challenges: using only terminal rewards makes turn-level credit assignment sparse and high-variance, while sequential mini-batch updates on a fixed rollout batch introduce policy staleness as the current policy drifts from the behavior policy. Existing approaches obtain denser credit via learned critics, process reward models, or on-policy distillation, but they rely on a value network, step-level annotations, or a stronger external teacher that increase training complexity. Meanwhile, importance sampling and clipping mitigate policy drift, but they do not specify a finite target that stops sequential updates from overshooting. In this work, we introduce TreeOPSD, a critic-free, tree-aware posterior self-distillation framework for long-horizon agentic reinforcement learning. Treating each interaction turn as a tree node, TreeOPSD branches alternative continuations from shared prefixes and recursively backs up terminal rewards into implicit turn-level credit. It then maps each credit to a closed-form, bounded teacher distribution parameterized by the likelihood ratio relative to the frozen behavior policy. Minimizing the divergence between this teacher and the student policy recovers the policy-gradient direction as the tangent at the behavior policy while bounding updates toward a finite target across sequential mini-batches. Empirically, TreeOPSD improves agentic capabilities across several competitive benchmarks and model scales, especially for long-horizon tasks, e.g., raising the average accuracy of Qwen3-30B-A3B-Thinking-2507 by 4.97 points over GRPO on the Deep Research setting.
14. 词汇参考框架：在原生语义中解释语言模型残差流
- 英文标题：Lexical Reference Frames: Interpreting LM Residual Streams in Native Semantics
- arXiv：2610.22169
💡 一句话解读：语言模型内部每层的“隐藏状态”到底代表什么，一直难以直接读懂。现有方法要么用固定镜头，要么把状态当点云，都不符合模型自身读法。作者提出“词汇参考框架”，用剩余层和注意力模拟模型原生方式解读状态，使每个隐藏状态可读、可归因、可操控。
语言模型的词汇表就是其原生语义，嵌入与反嵌入固定了残差流两端状态的含义。介于两者之间，现有方法要么复用固定镜头（lens），要么把状态当作点云，都无法像模型本身那样读取状态。我们主张每个隐藏状态都有一个原生参考框架，它由剩余层诱导、能自适应上下文。词汇参考框架（Lexical reference frames）以模型原生且上下文感知的方式近似该框架：通过给定轮数的位置间注意力及对其余部分的典型注意力输出，用剩余层读取状态，且不对模型输出做任何拟合。任何逐层方向固定的镜头，其与原生框架词汇读出值的平均余弦相似度都无法超过其相干性的平方根——在 SmolLM3-3B 和 OLMo3-7B 的中层分别为 0.30 和 0.35；而一阶的 LexFrame 达到 0.97 和 0.97，其读出值与模型输出的平均差距在 0.16 和 0.15 nats 以内。其系数差异比归因修补（attribution patching）更忠实地归因预测，其词汇基可在不干扰其他 token 的情况下引导预测。平均而言，其读出值在最后 0.82 和 0.81 的层上稳定于模型的首选 token，而镜头至多只有 0.11。因此这些框架使每个隐藏状态都可读、可归因、可操控。
📄 English Abstract
The vocabulary of a language model is its native semantics, and the embedding and unembedding fix the meaning of a state at the two ends of the residual stream. Between them, existing methods reuse a fixed lens or treat states as a point cloud, and neither reads a state as the model does. We argue that every hidden state has a native reference frame, induced by the layers that remain and adaptive to the context. Lexical reference frames approximate it in a model-native and context-aware way, reading a state through the remaining layers with a given number of rounds of attention between positions and a typical attention output for the rest, fitting nothing to the model’s output. No lens with fixed directions per layer can reach a mean cosine with the native frame’s lexical readouts above the square root of their coherence, 0.30 and 0.35 at mid-depth of SmolLM3-3B and OLMo3-7B. LexFrame[1], of order 1, reaches 0.97 and 0.97, and its readings are on average within 0.16 and 0.15 nats of the model’s output. Its coefficient differences attribute predictions more faithfully than attribution patching, and its lexical basis steers predictions while sparing other tokens. On average, its readings settle on the model’s top token over the final 0.82 and 0.81 of the layers, those of lenses over at most 0.11. These frames thus make every hidden state readable, attributable and steerable.
15. RAP：基于人类制作范例的声明式标记论文到海报生成数据集
- 英文标题：RAP: A Dataset for Paper-to-Poster Generation via Declarative Markup from Human-Crafted Exemplars
- arXiv：2610.23230
💡 一句话解读：论文要解决的是：把一篇论文自动做成一张会议海报，现有数据太少、只能靠人工打分，而且生成方法都从零设计版式，没法利用人类做好的海报。作者构建了RAP数据集，把2.5万多张人类海报转成可填内容的网页代码模板，让机器可以检索范例、直接套用版式并自动渲染比对。
制作会议海报是微缩版的视觉编程：必须把整篇论文压缩到一页上，同时锻炼摘要、视觉设计和代码实现能力，且只能使用论文本身。然而，视觉编程语境下该任务的基准仍然稀缺：现有资源规模小且仅标注布局框，评估只能退回到VLM-as-judge或人工检查表，而非客观的渲染-比较；它们支持的生成方法从零设计布局，使人类制作的海报既不能作为可检索的设计库，也不能作为经过验证的监督信号。我们提出RAP（Reconstruct A Poster），一个将论文到海报生成视为视觉编程任务的数据集与基准。RAP包含2022至2026年间多个机器学习会议的25K+经核验的论文-海报对，标注了三级内容分类体系，并附有150篇测试论文各配其人类海报的留出基准。一个全自动的反渲染流水线将每张海报转换为带类型、可填充槽位的样式化HTML/CSS，因此每个样本都是可通过渲染与源海报比对来验证的论文-海报-标记三元组。该语料库同时服务于推理和训练：PosterRAG是一个免训练的多智能体流水线，检索一个范例并填充其槽位而不改动布局或样式；而SFT+RL方案则从相同三元组训练专用模型。在线演示见此处URL。
📄 English Abstract
Making a conference poster is visual coding in miniature: an entire paper must be compressed onto a single page, which exercises summarization, visual design, and code implementation at once, using nothing beyond the paper itself. Yet benchmarks for this task in the context of visual coding remain scarce: existing resources are small and annotated only with layout boxes, so evaluation must fall back on VLM-as-judge or human checklists rather than objective render-and-compare, and the generation methods they support design layouts from scratch, leaving human-crafted posters unusable as either a retrievable design library or as verified supervision. We introduce RAP (Reconstruct A Poster), a dataset and benchmark for paper-to-poster generation as a visual coding task. RAP contains 25K+ verified paper-poster pairs from various machine learning conferences held between 2022 and 2026, annotated with a three-level content taxonomy, plus a held-out benchmark of 150 test papers each paired with its human poster. A fully automated derendering pipeline turns each poster into styled HTML/CSS of typed, refillable slots, so every sample is a paper-poster-markup triplet verifiable by rendering against the source poster. The corpus serves both inference and training: PosterRAG, a training-free multi-agent pipeline, retrieves an exemplar and refills its slots without touching layout or styling, while an SFT+RL recipe trains a dedicated model from the same triplets. An online demo is available at this https URL .