Daily Papers · 2026-09-26
📄 论文日报 | 来源：Hugging Face Daily Papers
🗓 Episode：2026.09.24 | 说话者双轨记忆；交互学习空间推理
1. SpeakerMem-R1：以说话者为中心的多方对话双轨记忆
- 英文标题：SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue
- arXiv：2609.26780
💡 一句话解读：多方聊天中，AI很难记住谁说了什么、谁和谁有关系、事情随时间怎么变。这篇论文提出双轨记忆法：一条记录原话并标注说话人，另一条整理出个人和群体的状态，查询时把两条信息按人、事、时间拼起来，还专门训练了一个小模型来减少记错。
多方场景下的长期对话记忆不仅需要从长期对话中检索相关内容，还必须区分谁说了什么、每句话涉及谁、个体如何相互感知、群体共享哪些信息以及状态如何随时间变化。近期对多方对话基准的研究表明，现有的通用大语言模型（LLM）记忆系统容易丢失人物与群体关系，或难以整合分散在成员、群体和时间中的线索。这些问题共同揭示了两个核心瓶颈：多方对话中的消息归属与关系理解，以及从交错历史中重建状态。为解决这两个问题，我们提出了 SpeakerMem-R1：其双轨记忆存储带说话人标记的逐字消息和派生状态，并组织为个人级和群体级视图，然后在查询时按实体、事件和时间融合两条轨道的证据。为了在结构化记忆构建过程中减少归属和更新错误并支持本地部署，我们使用 SpeakerLevenshtein 和以说话者为条件的 GRPO 训练了 Writer-R1。在 GroupMemBench、SocialMemBench 和 EverMemBench 上，SpeakerMem-R1 分别达到 47.9%、69.2% 和 61.9% 的二元准确率。在 EverMind-AI 公开报告的 EverMemBench 排行榜上，我们达到 62.33%，是近期最先进框架中报告的最佳结果。在全部 1,986 个 LoCoMo 问题上，我们达到 70.85%，我们将其用作两人长期对话的边界测试。在 305 个问题的受控评估中，强化学习将 SFT Writer 的平均准确率从 57.38% 提升至 68.20%。我们同时报告二元准确率和 token-F1，消融实验表明，在标准化评估接口下，逐字轨道和结构化轨道，以及个人级和群体级视图，是互补的。
📄 English Abstract
Long-term conversational memory in multi-party settings requires more than retrieving relevant content from long-term conversations: it must distinguish who said what, whom each statement concerns, how individuals perceive one another, what information is shared by the group, and how states change over time. Recent studies on multi-party dialogue benchmarks show that existing general-purpose LLM memory systems tend to lose person and group relations or struggle to integrate clues distributed across members, groups, and time. Together, these issues reveal two core bottlenecks: message attribution and relational understanding in multi-party dialogue, and state reconstruction from interleaved histories. To address both, we propose $\textbf{SpeakerMem-R1}$: its dual-track memory stores speaker-labeled verbatim messages and derived states organized into person-level and group-level views, then combines evidence from both tracks by entity, event, and time at query time. To reduce attribution and update errors during structured memory construction while enabling local deployment, we train Writer-R1 with SpeakerLevenshtein and speaker-conditioned GRPO. On GroupMemBench, SocialMemBench, and EverMemBench, SpeakerMem-R1 achieves binary accuracies of 47.9%, 69.2%, and 61.9%, respectively. On the publicly reported EverMemBench leaderboard from EverMind-AI, we achieves 62.33%, the best reported result among the latest state-of-the-art frameworks. It also achieves 70.85% on all 1,986 LoCoMo questions, which we use as a two-person long-term conversation boundary test. In a controlled evaluation of 305 questions, RL raises the SFT Writer's mean accuracy from 57.38% to 68.20%. We report both binary accuracy and token-F1, and ablations show that the verbatim and structured tracks, as well as person-level and group-level views, are complementary under the standardized evaluation interface.
2. Spatial-Interactor：通过与可观测物理世界交互学习空间推理
- 英文标题：Spatial-Interactor: Learning Spatial Reasoning through Interaction with the Observable Physical World
- arXiv：2609.23038
💡 一句话解读：现有视觉语言模型在动态环境中难以理解物体移动和视角变化带来的空间状态变化。为此，研究者提出 Spatial-Interactor 框架，让模型通过与环境交互来学习空间推理，并设计三级课程和两阶段训练策略，显著提升了模型对局部状态变化和长轨迹整合的能力。
空间推理对于视觉语言模型（VLMs）理解和作用于物理世界至关重要。在动态环境中进行推理要求 VLMs 感知由物体运动和视角变化引起的局部状态转换，并在长轨迹上整合这些转换以维持更新的空间状态，然而现有 VLMs 在这两方面能力上仍然有限。当前的空间训练主要关注关于物体属性和空间关系的静态问题，对状态转换提供的直接监督有限；相比之下，交互轨迹自然连接了先前的观察、动作和后续观察，为局部状态转换提供了直接监督，而完整轨迹则揭示了连续转换之间的依赖关系。因此，我们提出了 Spatial-Interactor，一个通过交互训练 VLMs 建模物理世界状态转换的框架，将该学习过程组织为三级课程：L1 被动世界状态转换、L2 主动自身状态转换和 L3 长时程交互轨迹。相应地，我们从模拟和真实交互轨迹中构建了空间交互学习数据集（LSI-108K），其任务与每个级别的目标一致。我们的两阶段训练策略对 L1 和 L2 应用监督微调（SFT）以进行局部转换建模，然后使用在线策略蒸馏（OPD）进行特权自蒸馏：一个教师分支获得片段级转换描述，监督学生的在线策略思维链（CoT），帮助学生学会在 L3 长轨迹上整合连续转换。在多个 VLMs 和空间基准上的实验表明，在局部转换建模和长时程整合方面均取得了持续改进。
📄 English Abstract
Spatial reasoning is essential for vision-language models (VLMs) to understand and act in the physical world. Reasoning in dynamic environments requires VLMs to perceive local state transitions caused by object motion and viewpoint changes and integrate them over long trajectories to maintain an updated spatial state, yet existing VLMs remain limited in both capabilities. Current spatial training primarily focuses on static questions about object attributes and spatial relations, providing limited direct supervision for state transitions; in contrast, interaction trajectories naturally connect a preceding observation, an action, and a subsequent observation, offering direct supervision for local state transitions, while complete trajectories reveal dependencies among consecutive transitions. We therefore introduce Spatial-Interactor, a framework that trains VLMs to model physical-world state transitions through interaction, organizing this learning process into a three-level curriculum covering L1 passive world-state transitions, L2 active self-state transitions, and L3 long-horizon interaction trajectories. Accordingly, we construct the Learning from Spatial Interaction dataset (LSI-108K) from simulated and real interaction trajectories, with tasks aligned with the objective of each level. Our two-stage training strategy applies Supervised Fine-Tuning (SFT) to L1 and L2 for local transition modeling, and On-Policy Distillation (OPD) then uses privileged self-distillation: a teacher branch given segment-level transition descriptions supervises the student's on-policy CoT, helping the student learn to integrate consecutive transitions over L3 long trajectories. Experiments across multiple VLMs and spatial benchmarks show consistent gains in local transition modeling and long-horizon integration.
3. 快乐世界基准（HappyWorld-Bench）
- 英文标题：HappyWorld-Bench
- arXiv：2609.24308
💡 一句话解读：现有方法难以全面评估AI世界模型在生成质量、交互一致性和响应能力上的表现。为此，作者提出了HappyWorld基准，用于系统测试世界模型在探索、交互和修改时的综合能力。
评估世界模型需要同时考察其生成世界的质量，以及在探索、交互和修改过程中的一致性与响应能力。我们提出了HappyWorld基准。
📄 English Abstract
Evaluating world models requires assessing both the quality of the worlds they generate and their consistency and responsiveness under exploration, interaction, and modification. We introduce HappyWorld
4. 过往帧塑造未来：自回归视频生成中的记忆机制
- 英文标题：The Past Frames the Future: Memory for Autoregressive Video Generation
- arXiv：2609.28466
💡 一句话解读：这篇研究要解决的是：自回归视频生成时，模型只能记住有限的历史信息，导致生成久了会忘记角色身份、动态状态等关键内容。作者系统梳理了各种“记忆机制”，从记忆形式、功能、操作、学习和评估五个角度，为构建能长期记住历史的视频生成系统提供框架。
生成模型的进步提升了视频保真度，使长时程生成、交互式世界建模和演化视觉环境成为可能。自回归（AR）视频生成通过因果推演扩展视觉序列。然而，一个根本瓶颈随之出现：随着生成序列的扩展，实际模型必须在严格受限的上下文窗口、存储和计算限制下运行。因此，关键的历史信息（例如实体身份、动态状态和干预引发的因果变化）往往在其相关性减弱之前很久就已离开活跃上下文。克服这一限制并保持时间持续性构成了一个根本性的记忆问题。我们对AR视频生成中的记忆机制进行了系统而全面的综述。我们将记忆操作性地定义为跨外部AR步骤保持的持久历史信息，即使原始证据不再局部可访问，它也能影响未来的生成。基于这一统一框架，我们从五个互补视角组织文献：（I）形式，即历史的表征载体；（II）功能，即需要保留的特定语义和物理信息；（III）操作，即记忆的写入、读取、更新、管理和整合的生命周期；（IV）学习，即在闭环推演下对记忆行为的优化；（V）评估，即诊断真实记忆能力的范式。最后，我们综合了开放挑战，包括可组合且资源感知的记忆架构、可信的状态更新、自推演学习和标准化评估。通过连接表征、机制和学习范式，本文为开发可靠、以记忆为条件的视频生成系统建立了结构化基础。
📄 English Abstract
Advances in generative models have improved video fidelity, enabling long-horizon generation, interactive world modeling, and evolving visual environments. Autoregressive (AR) video generation extends visual sequences through causal rollouts. However, a fundamental bottleneck emerges: as the generated sequence expands, practical models must operate under strictly bounded context windows, storage, and computational limits. Consequently, critical historical information, e.g., entity identities, dynamic states, and intervention-induced causal changes, often leaves the active context long before its relevance diminishes. Overcoming this limitation and maintaining temporal persistence constitutes a fundamental memory problem. We present a systematic and comprehensive review of memory mechanisms in AR video generation. We formulate memory operationally as persistent historical information maintained across outer AR steps, capable of influencing future generation even after the originating evidence is no longer locally accessible. Building upon this unified framework, we organize the literature through five complementary perspectives: (I) Forms, the representational carriers of history; (II) Functions, the specific semantic and physical information requiring preservation; (III) Operations, the lifecycle of writing, reading, updating, managing, and integrating memory; (IV) Learning, the optimization of memory behaviors under closed-loop rollouts; and (V) Evaluation, the paradigms for diagnosing genuine memory capabilities. We conclude by synthesizing open challenges, including composable and resource-aware memory architectures, trustworthy state updating, self-rollout learning, and standardized evaluation. By bridging representations, mechanisms, and learning paradigms, this paper establishes a structured foundation for developing reliable, memory-conditioned video generation systems.
5. 即时记忆：学习为LLM智能体整理任务自适应记忆
- 英文标题：Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents
- arXiv：2609.27334
💡 一句话解读：现有智能体记忆系统在任务完成后就急着把经验压缩成固定摘要，但此时还不知道未来会问什么，容易丢掉有用信息。本文改为先保留原始经历，等新任务来了再按需整理出针对当前问题的精简提示。这样还能直接用任务成败来训练整理器，在多个测试中成功率明显提升。
智能体记忆系统通过复用过去经验来提升未来表现，但多数现有设计在写入时（write time）就整理记忆：任务完成后，其轨迹被提炼为固定产物（如反思、工作流、技能或推理策略），之后按相似度检索。这迫使系统在未知未来查询前就决定什么值得记住，不可逆地丢弃信息，并产生一个与查询无关、却要服务多种下游任务的摘要。学习这种写入时整理器也很困难，因为某个存储决策的价值可能要到相关查询出现时才显现，甚至可能在很多任务之后，形成长时程的信用分配问题。我们转而保留原始轨迹，将整理推迟到读取时（read time），此时当前任务已知。给定检索到的轨迹和新任务，记忆整理器会合成一个紧凑、任务自适应的载荷，以贴合即时需求。由于该载荷在同一任务上被消费，整理器可直接根据即时任务成功来训练，避免延迟的效用信号，也无需人为分组相关任务。在 ALFWorld、WebShop 和 $\tau^2$-bench 上，我们的即时记忆（Just-in-Time Memory, JitMem）持续优于无记忆智能体以及启发式和学习的写入时记忆方法，分别比最强基线提升 16.2、16.3 和 3.9 个绝对成功率百分点。值得注意的是，即使未训练的整理器也已能与这些基线竞争或超越它们，表明任务自适应的读取时整理本身就是主要增益来源；训练整理器则进一步叠加提升。
📄 English Abstract
Agentic memory systems reuse past experience to improve future performance, yet most existing designs curate memory at write time: once a task is completed, its trajectory is distilled into a fixed artifact, such as a reflection, workflow, skill, or reasoning strategy, that is later retrieved by similarity. This forces the system to decide what is worth remembering before the future query is known, irreversibly discarding information and producing a query-independent summary that must serve many possible downstream tasks. Learning such a write-time curator is also difficult because the value of a storage decision may only become apparent when a relevant query arrives, potentially many tasks later, creating a long-horizon credit-assignment problem. We instead retain raw trajectories and defer curation until read time, when the current task is known. Given the retrieved traces and the new task, a memory curator synthesizes a compact, task-adaptive payload tailored to the immediate need. Because this payload is consumed on the same task, the curator can be trained directly from immediate task success, avoiding delayed utility signals and the need to artificially group related tasks. Across ALFWorld, WebShop, and $\tau^2$-bench, our Just-in-Time Memory (JitMem) consistently outperforms no-memory agents as well as heuristic and learned write-time memory methods, improving over the strongest baseline by 16.2, 16.3, and 3.9 absolute success-rate points, respectively. Notably, even an untrained curator is already competitive with or surpasses these baselines, showing that task-adaptive read-time curation itself is a major source of the gain; training the curator further compounds the improvement.
6. RewardVerse：面向视频奖励建模的评分量规引导策略优化
- 英文标题：RewardVerse: Rubric-Guided Policy Optimization for Video Reward Modeling
- arXiv：2609.22947
💡 一句话解读：视频生成模型需要奖励模型来打分，但现有模型直接给视频打一个分数，没有明确标准，导致分数在不同提示下忽高忽低，不可靠。本文提出RewardVerse，先让模型生成一套评分标准（量规），再按标准打分，并用两阶段训练算法优化，使分数更稳定、可解释。
强化学习（RL）对于优化视频生成模型至关重要，而一个稳健的奖励模型（RM）是其基石。然而，现有的视频奖励模型通常产生不稳定的标量分数，因为它们在没有明确评估标准的情况下，直接将复杂、主观的视频质量映射为单一分数。这导致标量漂移，即评分尺度在不同提示之间崩溃或偏移，使得奖励在RL中不可靠。受专业人工标注工程启发，我们用RewardVerse解决这个问题，这是一个基于量规（rubric）的视频奖励框架，在评估查询和评分器之间引入动态量规作为中间表示。RewardVerse不是无约束地直接评分，而是首先生成明确的评估标准，然后进行量规引导的评分，提供稳定的语义锚点以缓解标量漂移。为了高效优化这一协作流程，我们提出了量规引导策略优化（RGPO），一种两阶段训练算法。RGPO首先使用自演化的种子量规预热评分器，然后联合优化量规生成器以产生查询自适应的评估标准，同时持续将评分器与人类评分对齐。在16维EvalVerse基准和外部数据集上的大量实验表明，RewardVerse缓解了标量漂移，在逐点和成对评估上均达到最先进性能，并为视频生成中的RL提供了稳健且可解释的奖励信号。
📄 English Abstract
Reinforcement learning (RL) is vital for optimizing video generation models, with a robust reward model (RM) serving as the cornerstone. However, existing video reward models often produce unstable scalar scores because they directly map complex, subjective video quality into a single score without explicit evaluation criteria. This leads to scalar drift, where the scoring scale collapses or shifts across different prompts, making the reward unreliable for RL. Drawing inspiration from professional human annotation engineering, we address this problem with RewardVerse, a rubric-based video reward framework that introduces a dynamic rubric as an intermediate representation between the evaluation query and the scorer. Instead of unconstrained direct scoring, RewardVerse first generates explicit evaluation criteria and then performs rubric-guided scoring, providing a stable semantic anchor that mitigates scalar drift. To efficiently optimize this collaborative pipeline, we propose Rubric-Guided Policy Optimization (RGPO), a two-stage training algorithm. RGPO first warms up the scorer using self-evolving seed rubrics and then jointly optimizes the rubric generator to produce query-adaptive evaluation criteria while continuously aligning the scorer with human ratings. Extensive experiments on the 16-dimensional EvalVerse benchmark and external datasets demonstrate that RewardVerse mitigates scalar drift, achieves state-of-the-art performance on both pointwise and pairwise evaluation, and provides a robust and interpretable reward signal for RL in video generation.
7. PACT：从信用分配到评论家对齐
- 英文标题：PACT: From Credit Assignment to Critic Alignment
- arXiv：2609.26355
💡 一句话解读：这篇研究要解决强化学习训练大模型时，每个词对最终结果的贡献（信用）缺乏统一定义，导致各种训练方法关系不清的问题。作者先给出三条规则，证明它们能唯一确定这种贡献，并据此发现现有方法（如OPD、RLOO）其实都隐含地符合这个定义。他们还指出常用方法GAE中评论家的误差可能和真实贡献一样大，于是提出PACT：先更新演员再更新评论家，并用重要性采样纠正评论家训练，使其与最新策略对齐。在数学推理和代码修复任务上，PACT比GRPO、PPO等基线准确率更高。
强化学习已成为大语言模型（LLM）后训练的核心环节，但词元级信用（token-level credit）缺乏普遍接受的数学定义，使其与常用训练信号的关系不明确。我们提出三个正则性条件——完备性、前缀一致性和中性，并证明它们唯一地确定了词元级信用。这一刻画为解释现有算法中的现象提供了统一基础，并指导了改进的演员-评论家训练流程的开发。通过这一视角，在线策略蒸馏（OPD）中的理想教师充当隐式评论家，产生的期望策略梯度与词元级信用诱导的梯度成正比。响应级REINFORCE留一法（RLOO）信号尽管粒度更粗，但其期望策略梯度贡献与词元级信用相匹配。我们进一步在有界结果奖励下建立了近似信用稀疏性，并展示了广义优势估计（GAE）中中间评论家误差如何变得与底层信用相当。这些发现促使我们提出策略对齐评论家训练（PACT），它采用先演员后评论家的更新顺序，对评论家训练应用重要性采样校正，使评论家更好地与更新后的策略对齐。在智能体数学推理中，PACT在四个基准上平均准确率达到72.87%，分别比GRPO和PPO高出8.80和13.16个百分点。在SWE-bench Verified上，PACT通过率达到67.4%，分别比PPO、GRPO和SAO高出2.4、2.0和3.8个百分点。
📄 English Abstract
Reinforcement learning has become a central component of large language model (LLM) post-training, yet token-level credit lacks a generally accepted mathematical definition, leaving its relationship to commonly used training signals unclear. We formulate three regularity conditions, namely Completeness, Prefix Consistency, and Neutrality, and prove that they uniquely determine token-level credit. This characterization provides a unified basis for explaining phenomena across existing algorithms and guides the development of an improved actor-critic training procedure. Through this lens, an ideal teacher in On-Policy Distillation (OPD) acts as an implicit critic, yielding an expected policy gradient proportional to that induced by token-level credit. Response-level REINFORCE Leave-One-Out (RLOO) signals match the expected policy-gradient contribution of token-level credit despite their coarser granularity. We further establish approximate credit sparsity under bounded outcome rewards and show how intermediate critic errors in Generalized Advantage Estimation (GAE) can become comparable to the underlying credit. These motivate Policy Aligned Critic Training (PACT), which adopts an Actor-then-Critic update order to apply importance sampling correction to critic training and better align the critic with the updated policy. In agentic mathematical reasoning, PACT achieves 72.87% average accuracy across four benchmarks, outperforming GRPO and PPO by 8.80 and 13.16 percentage points, respectively. On SWE-bench Verified, PACT achieves a pass rate of 67.4%, outperforming PPO, GRPO, and SAO by 2.4, 2.0, and 3.8 percentage points, respectively.
8. 薛定谔的代码仓库：LLM 是学会了 SWE-bench，还是记住了它？
- 英文标题：Schrödinger’s Code Repository: Have LLMs Learned SWE-bench or Memorized It?
- arXiv：2609.27891
💡 一句话解读：现有代码评测基准常因测试仓库被反复用于训练而存在数据泄露，导致高分可能只是模型记住了仓库特征而非真正会推理。为此，研究者提出SchrodingerRepo框架，在评测时动态改变仓库的命名、文件布局和代码写法，同时保持功能不变，以此检验模型是否真会解题。
仓库级编码基准已成为评估编码智能体的标准，但它们天生存在数据泄露问题，因为这些基准建立在被反复用于训练的热门开源仓库之上。因此，出色的表现可能反映的是对典型仓库线索的记忆，而非稳健的仓库推理能力。我们提出SchrodingerRepo（薛定谔的代码仓库），一个在动态实例化仓库表示下测试编码智能体的评估框架。SchrodingerRepo不重复使用测试仓库的静态表示，而是将测试仓库视为评估时的潜变量，仅在智能体进入评估环境时才动态实例化。实例化后的仓库保留原始可执行行为，同时通过四个变换层级——问题陈述重构、命名空间重映射、文件内布局重排和功能保持的代码重写——侵蚀命名约定、文件布局和实现模式等熟悉线索。我们在SWE-bench Verified和SWE-QA上评估了流行的大语言模型。结果表明，移除熟悉的仓库线索会持续降低智能体性能，并大幅增加各模型的交互成本。进一步分析显示，额外成本主要源于仓库探索和定位难度的增加。这些发现表明，当前编码智能体可能部分依赖记忆的仓库侧线索，凸显了在动态实例化仓库表示下进行评估的必要性。
📄 English Abstract
Repository-level coding benchmarks have become the standard for evaluating coding agents, yet they inherently suffer from data leakage because they are built upon popular open-source repositories repeatedly used for training. Consequently, strong performance may reflect memorization of canonical repository cues rather than robust repository reasoning. We propose SchrodingerRepo (Schrödinger’s Repository), an evaluation framework for testing coding agents under dynamically instantiated repository representations. Instead of repeatedly using a static representation of the test repository, SchrodingerRepo treats the test repository as an evaluation-time latent variable that is dynamically instantiated only when the agent enters the evaluation environment. The instantiated repository preserves the original executable behavior while eroding familiar cues such as naming conventions, file layouts, and implementation patterns through four transformation levels: problem statement reconstruction, namespace remapping, intra-file layout reordering, and functionality-preserving code rewriting. We evaluate popular LLMs on SWE-bench Verified and SWE-QA. Results show that removing familiar repository cues consistently degrades agent performance and substantially increases interaction costs across models. Further analysis reveals that the additional cost is primarily caused by increased difficulty in repository exploration and localization. These findings suggest that current coding agents may partially rely on memorized repository-side cues, highlighting the need for evaluation under dynamically instantiated repository representations.
9. GeoPair：用于免训练 Transformer 压缩的几何保持跨层因子分解
- 英文标题：GeoPair: Geometry-Preserving Cross-Layer Factorization for Training-Free Transformer Compression
- arXiv：2609.25963
💡 一句话解读：这篇研究要解决 Transformer 模型压缩时，现有方法要么单独处理每层、要么靠经验分组，忽略了各层激活的几何特性，导致压缩效果不佳。它提出一种免训练框架，通过优化跨层权重配对和共享字典分解，保留每层独特的校准几何，并结合结构化稀疏，实现高效压缩。
Transformer 架构存在跨层冗余，但训练后压缩流程通常孤立地优化各层，或依赖忽略层特定激活几何的启发式分组策略。我们提出一个原则性的免训练框架，顺序优化跨层权重配对和共享字典分解。与强制相邻层权重共享基或启发式合并激活统计不同，我们的方法识别结构兼容的投影，并学习一种共享表示，更好地保留每层独特的校准几何。结合结构化稀疏，这产生了高效的权重分解，且不牺牲功能保真度。在不同架构、规模和模态上，我们的方法取得了最先进的结果，始终优于独立的结构化权重分解和在启发式分组策略下运行的替代成对权重分解。通过用收敛的、优化驱动的流程取代启发式工程策略，我们为跨不同模态的可扩展 Transformer 压缩建立了有理论依据的基础。
📄 English Abstract
Transformer architectures exhibit cross-layer redundancies, yet post-training compression pipelines typically optimize layers in isolation or rely on heuristic grouping strategies that disregard layer-specific activation geometries. We introduce a principled, training-free framework that sequentially optimizes cross-layer weight pairings and shared-dictionary factorizations. Rather than forcing weights of adjacent layers to share a basis or heuristically merging activation statistics, our approach identifies structurally compatible projections and learns a shared representation that better preserves each layer's distinct calibration geometry. Coupled with structured sparsity, this yields highly efficient weight decompositions without sacrificing functional fidelity. Across diverse architectures, scales, and modalities, our method achieves state-of-the-art results, consistently outperforming independent structured weight decompositions and alternative pairwise weight factorizations, which operate under heuristic grouping strategies. By replacing heuristic engineering strategies with a convergent, optimization-driven pipeline, we establish a theoretically grounded foundation for scalable, transformer compression across different modalities.
10. PackLab：在机器人装箱中开发、训练与评估多模态大语言模型的综合框架
- 英文标题：PackLab: A Comprehensive Framework for Developing, Training, and Evaluating MLLMs in Robotic Bin Packing
- arXiv：2609.23784
💡 一句话解读：机器人装箱时，每放一个物体都会影响后面怎么放，现有方法要么靠人工规则，要么靠反复试错训练，遇到不同箱子或物体组合就不灵了。本研究搭建了一个叫 PackLab 的完整平台，专门训练和测试多模态大模型来做这种连续装箱决策，实验证明它比传统方法和通用大模型都表现更好。
机器人装箱需要长时程的序列决策，因为每个物体的放置都会影响后续装箱的可用空间。现有方法主要依赖手工设计的几何启发式规则来优化预设目标，或依赖在预设训练配置下通过试错学习得到的强化学习策略。尽管多模态大语言模型（MLLMs）在该任务上已有近期进展，但其在异构装箱配置下进行闭环序列决策的潜力仍未得到充分探索。为填补这一空白，我们提出了 PackLab，一个用于开发、训练和评估面向闭环机器人装箱的 MLLMs 的综合框架。PackLab-Suite 提供了一个基于物理的仿真平台，可规模化生成多样化的训练装箱轨迹并评估其物理结果。PackLab-VLM 是一个装箱专用 MLLM，能够理解不断变化的物体与容器状态，以闭环方式联合选择物体并预测放置位置。PackLab-Bench 提供了多个难度级别的标准化装箱场景，用于系统评估。大量实验表明，在平均意义上，PackLab-VLM 在不同物体集合和容器配置下均优于传统装箱启发式方法、传统强化学习方法以及通用 MLLMs，凸显了 MLLMs 在长时程机器人装箱中的潜力。代码、模型、数据集和基准测试已在 https://github.com/Correr-Zhou/PackLab 公开。
📄 English Abstract
Robotic bin packing requires long-horizon sequential decision-making, as each object placement affects the available space for subsequent packing. Existing methods primarily rely on hand-crafted geometric heuristics that optimize predefined objectives or reinforcement learning policies learned through trial and error over predefined training configurations. Despite recent advances in multimodal large language models (MLLMs) for this task, their potential for closed-loop sequential decisions across heterogeneous packing configurations remains underexplored. To address this gap, we introduce PackLab, a comprehensive framework for developing, training, and evaluating MLLMs for closed-loop robotic bin packing. PackLab-Suite provides a physics-based simulation platform for scalable generation of diverse training packing trajectories and evaluation of their physical outcomes. PackLab-VLM is a packing-specialized MLLM that understands the evolving object and container states to jointly select objects and predict placements in a closed-loop manner. PackLab-Bench provides standardized packing scenarios at multiple difficulty levels for systematic evaluation. Extensive experiments demonstrate that, on average, PackLab-VLM outperforms conventional packing heuristics, traditional reinforcement learning methods, and general-purpose MLLMs across object sets and container configurations, highlighting the potential of MLLMs for long-horizon robotic packing. The code, model, dataset, and benchmark are available at https://github.com/Correr-Zhou/PackLab .
11. MemBodied：面向视觉-语言-动作模型的循环联想记忆
- 英文标题：MemBodied: Recurrent Associative Memory for Vision-Language-Action Models
- arXiv：2609.28256
💡 一句话解读：机器人做家务时，常需要记住之前看到的东西（比如钥匙刚才放哪了），但现有模型只看当前画面，记不住过去。MemBodied 给模型加了一个固定大小的“小本子”，专门记录关键互动和初始场景，不用把历史画面全塞进去，就能让机器人记住该记的事，还省资源。
视觉-语言-动作模型（Vision-Language-Action models）为通用机器人控制提供了坚实基础，但绝大多数策略无法保留并利用当前观测之外的回合级信息。这一局限在依赖历史观测信息的历史相关操作任务中后果严重。将过去观测保留在上下文中虽有助于恢复此类信息，却会带来上下文不断膨胀和推理延迟的显著代价。为此，我们提出 MemBodied，一种固定大小的回合记忆，包含两个互补组件：记录跨策略调用交互的联想状态，以及保存初始场景紧凑表示作为参考的回合锚点。每次策略调用时，模型基于当前输入和记忆组件生成动作，而非直接使用过去观测。在五个需要记忆的 RMBench 任务上，MemBodied 的平均成功率是无状态策略的 7.81 倍、普通循环记忆的 2.98 倍，同时以少 10 倍的附加参数超越最强的记忆增强基线 1.3 倍。在完全可观测的 LIBERO-Long 套件上，它达到 90.6%，比无状态的 π0 策略提升 5.4%。这些发现支持 MemBodied 作为扩展策略上下文以处理历史相关操作任务的实用替代方案。
📄 English Abstract
Vision-Language-Action models provide a strong foundation for general-purpose robot control, yet a vast majority of policies do not preserve and leverage episode-level information beyond the current observation. This limitation is consequential in history-dependent manipulation tasks that depend on information available only in past observations. Retaining past observations in context can aid in recovering this information, but at the significant cost of ever-growing, bloated context and inference latency. We thus introduce MemBodied, a fixed-size episodic memory with two complementary components: an associative state that records interactions across policy calls and an episode anchor that preserves a compact representation of the initial scene as a reference. At each policy call, the model conditions action generation on the current input and the memory components, rather than directly using past observations. Across five evaluated RMBench tasks requiring memory, MemBodied achieves $7.81\times$ the mean success rate of a stateless policy and $2.98\times$ of vanilla recurrent memory, while outperforming the strongest memory-augmented baseline by $1.3\times$ with $10\times$ fewer added parameters. On the fully observable LIBERO-Long suite, it reached 90.6%, a 5.4% improvement over the stateless $\pi_0$ policy. These findings support MemBodied as a practical alternative to expanding the policy context for history-dependent manipulation.
12. WhatWorkedBench：基准测试AI智能体的实验理解能力
- 英文标题：WhatWorkedBench: Benchmarking Experimental Understanding in AI Agents
- arXiv：2609.27490
💡 一句话解读：AI做实验时，常搞不清改动哪个部分会带来什么效果。为此研究者建了一个测试基准，让AI看代码、做少量实验后，预测所有配置组合的得分。结果发现，用高斯过程等统计方法能明显提高预测准确度。
AI研究智能体需要可靠地了解其实验如何改变结果。我们提出WhatWorkedBench来测量实验理解能力，即在有限预算的实验后，对组件变更所产生结果的预测准确性。智能体检查代码、选择测量方式，并提交一个响应曲面——一张预测所有组件设置组合得分的表格。穷举式CPU执行提供了在固定其他组件时改变每个组件的参考效应。这些效应涵盖了来自30个数据源、8种工作流类型的36个任务中的变更组合，共1248条配置记录。核心评估结合了全部八个家族的4206条数值控制记录，以及原始六个家族的108个智能体回合。在八次新测量中，配对效应岭回归在22个数据源中的15个上选出了最优配置，并在三个数据源上将每个效应误差限制在得分范围的10%以内。将高斯过程（GP）拟合到相同的智能体观测上，将效应恢复率（相对于真实效应大小的准确性）从原始Flash队列的0.632提高到0.698，并在另一个队列中从0.621提高到0.720。在六个已完成的节拍检测和图提交上，相同观测的GP将家族宏观恢复率从0.303提高到0.455。在六个具有六个二元选项、20次新测量的工作流上，编码代码等价性（即行为相同的配置）将GP恢复率从0.248提高到0.462。WhatWorkedBench支持关于实验智能体、自适应实验设计、数值推断以及程序结构利用的研究。
📄 English Abstract
AI research agents need reliable knowledge of how their experiments change outcomes. We introduce WhatWorkedBench to measure experimental understanding, the accuracy of predictions about component changes after budgeted experimentation. Agents inspect code, select measurements, and submit a response surface, a table predicting scores for every configuration of component settings. Exhaustive CPU execution supplies reference effects for changing each component while holding the others fixed. These effects capture combinations of changes across 36 tasks from 30 data sources and 8 workflow types, with 1248 configuration records. Core evaluation combines 4,206 numerical-control records across all eight families and 108 agent episodes across the original six. At eight new measurements, pair-effect ridge selects an optimum on 15 of 22 sources and limits every effect error to 10% of score range on three. Fitting a Gaussian process (GP) to the same agent observations raises effect recovery, accuracy relative to true effect magnitude, from 0.632 to 0.698 in the original Flash cohort and from 0.621 to 0.720 in an additional cohort. On six completed beat-detection and graph submissions, the same-observation GP raises family-macro recovery from 0.303 to 0.455. On six workflows with six binary options at 20 new measurements, encoding code equivalences, configurations with identical behavior, raises GP recovery from 0.248 to 0.462. WhatWorkedBench supports research on experimental agents, adaptive experimental design, numerical inference, and use of program structure.
13. 混元-A13B 技术报告
- 英文标题：Hunyuan-A13B Technical Report
- arXiv：2609.27284
💡 一句话解读：大模型通常又大又贵，跑起来很费算力。这篇报告提出一个开源模型：总参数800亿，但每次只激活130亿，像只叫醒需要的专家，省算力又保能力；还让模型按问题难易切换快慢思考。
我们提出 Hunyuan-A13B，一个基于混合专家（Mixture-of-Experts）架构的开源大语言模型。它总参数量为800亿，但推理时仅激活130亿参数，从而在模型能力、计算效率与部署成本之间取得平衡。该模型在严格筛选的20万亿token语料上预训练，并加强了STEM数据整理，提升了事实可靠性与推理能力；高质量的监督微调与大规模强化学习进一步增强了整体性能。Hunyuan-A13B 还引入了双模式思维链（Chain-of-Thought）框架，使推理深度可随任务复杂度调整：对常规查询采用快速思考，对复杂多步问题采用慢速思考。评估显示，该模型在数学、科学、编程、通用语言理解及智能体任务上表现具有竞争力，常接近更大模型。其高推理吞吐量使其适合对延迟敏感的应用。我们发布 Hunyuan-A13B，以支持开放研究与实际的大语言模型部署。
📄 English Abstract
We present Hunyuan-A13B, an open-source large language model based on a Mixture-of-Experts architecture. It contains 80 billion total parameters but activates only 13 billion during inference, balancing model capability, computational efficiency, and deployment cost. The model is pretrained on a rigorously filtered 20T-token corpus with enhanced STEM data curation, improving factual reliability and reasoning ability. High-quality supervised fine-tuning and large-scale reinforcement learning further enhance its overall performance. Hunyuan-A13B also introduces a dual-mode Chain-of-Thought framework that adapts reasoning depth to task complexity: fast thinking for routine queries and slow thinking for complex, multi-step problems. Evaluations show competitive performance across mathematics, science, programming, general language understanding, and agent tasks, often approaching that of much larger models. Its high inference throughput makes it suitable for latency-sensitive applications. We release Hunyuan-A13B to support open research and practical LLM deployment.
14. 可验证隐藏动力学游戏：从已求解机制生成智能体强化学习环境
- 英文标题：Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms
- arXiv：2609.27321
💡 一句话解读：现有AI训练环境常先搭场景再定评分规则，导致动态与评估脱节。该研究反过来先解一个数学题，再让AI把解题过程包装成可交互工具，从而自动生成大量带标准答案的训练环境，让智能体在交互中学习，效果大幅提升。
语言模型智能体日益面临具有演化状态、相互依赖决策和延迟结果的长时程任务。扩展其训练需要多样化的智能体环境、可靠的结果信号以及较低的扩展成本。现有生成流程通常先构建环境，再定义其结果规则或标注轨迹，导致动力学与评估事后对齐。VHD-Play 反转了这一依赖关系：先采样并求解一个数学模型，再由基于语料库的设定器将其决策过程呈现为有状态工具。可执行的动力学和轨迹评分参考均继承自同一个已求解模型。该流程以每个环境几美分的成本生成了 3,300 个多样化的智能体环境。在三个家族上训练 Qwen3.6-35B-A3B，使其在五家族诊断中的平均智能体得分从 0.204 提升至 0.815。增益还出现在所有三个训练家族的留出实例和八个未见机制家族上，并超越生成基底扩展到通用函数调用、旅行规划和 365 天电子商务的外部基准。在 E-Commerce Bench 上，训练后的检查点每次运行均未破产，且超过 Qwen3.7-Max。我们比较了书面问题与揭示或隐藏其参数的有状态版本。比较表明，大部分可学习差距在于有状态交互，而非底层问题求解。一个冻结的 35B 设定器实现了更大的环境，且规模匹配的训练在机制规模和时程增长时仍保持增益，表明存在一个可演化训练基底的潜力。
📄 English Abstract
Language-model agents increasingly face long-horizon tasks with evolving state, interdependent decisions, and delayed outcomes. Scaling their training requires diverse agentic environments, dependable outcome signals, and low extension cost. Existing generation pipelines commonly construct an environment before defining its outcome rule or annotating its trajectories, leaving dynamics and evaluation to be aligned post hoc. VHD-Play reverses this dependency by sampling and solving a mathematical model before a corpus-grounded setter renders its decision process as stateful tools. The executable dynamics and trajectory-scoring reference are inherited from the same solved model. The pipeline produces 3,300 diverse agentic environments at a cost of a few cents each. Training Qwen3.6-35B-A3B on three families raises its mean agentic score from 0.204 to 0.815 in a five-family diagnostic. Gains also appear on held-out instances from all three training families and eight unseen mechanism families, then extend beyond the generated substrate to external benchmarks for general function calling, travel planning, and 365-day e-commerce. On E-Commerce Bench, the trained checkpoint completes every run without bankruptcy and exceeds Qwen3.7-Max. We compare written-out problems with stateful versions that reveal or hide their parameters. The comparison shows that most of the learnable gap lies in stateful interaction rather than underlying problem solving. A frozen 35B setter realizes larger environments, and scale-matched training retains gains as mechanism size and horizon grow, indicating the potential for an evolving training substrate.
15. 所有模态都是平等的，但视频更平等：弥合联合视频生成中的交叉注意力差距
- 英文标题：All modalities are equal, but video is more equal: Closing the Cross-Attention Gap in Joint Video Generation
- arXiv：2609.27901
💡 一句话解读：视频生成时，其他模态（如动作、音频）与视频的相互关注不对称：它们能影响视频，但视频对它们的约束弱。研究提出RecCAR，用强方向对齐弱方向，提升生成质量。
视频是物理事件的丰富表示，捕捉外观、几何、运动和时间演化。其他模态（如3D人体运动或音频）编码同一事件的较窄方面。我们发现联合多模态扩散Transformer在跨模态对应中表现出相应的不对称性：伴随模态与视频形成强对应，但它们约束视频的反向对应仍然弱得多。我们将两个方向表示为视频标记上的可比较对应分布，并将它们的不一致定义为反向对应差距。我们引入RecCAR（反向跨模态注意力正则化），一种KL正则化器，使用已建立的视频到模态对应作为固定参考，将较弱的模态到视频对应朝其对齐。在联合视频-运动和视频-音频生成中，RecCAR将人体解剖学得分从0.69提高到0.75，并将音视频不同步从0.804降低到0.752，同时改善整体生成。
📄 English Abstract
Video is a rich representation of a physical event, capturing appearance, geometry, motion, and temporal evolution. Other modalities, such as 3D body motion or audio, encode narrower aspects of the same event. We find that joint multimodal diffusion transformers exhibit a corresponding asymmetry in cross-modal correspondence: companion modalities develop strong correspondences to video, but the reciprocal correspondences through which they constrain video remain substantially weaker. We express both directions as comparable correspondence distributions over video tokens and define their disagreement as the reciprocal correspondence gap. We introduce RecCAR, standing for Reciprocal Cross-modal Attention Regularization, a KL regularizer that uses the well-established video-to-modality correspondence as a fixed reference and aligns the weaker modality-to-video correspondence toward it. Across joint video-motion and video-audio generation, RecCAR improves the Human Anatomy score from 0.69 to 0.75 and reduces audio-video desynchronization from 0.804 to 0.752, while improving overall generation