Daily Papers · 2026-09-17
📄 论文日报 | 来源：Hugging Face Daily Papers
🗓 Episode：2026.09.16 | 持续学习组合提升保留率；游戏AI六类角色待整合
1. 持续学习机制组合用于长时程记忆
- 英文标题：Continual Learning Mechanisms Compose for Long-Horizon Memorization
- arXiv：2609.06986
💡 一句话解读：语言模型需要不断学习新知识并记住旧知识，但连续更新会导致“灾难性遗忘”。本文提出“长时程记忆”任务，并发现组合多种互补的防遗忘机制（如数据锚定、函数锚定、权重锚定和低秩分配）能大幅提升记忆保持率，最佳方法将最终保留率从1.2%提高到34.9%。
语言模型可能需要内化随时间到达的信息，并通过后续多次更新将其保留下来。为了研究这一挑战，我们引入了长时程记忆（long-horizon memorization）设定，其中模型通过持续的有监督微调学习100个查询-答案任务，且不保留早期训练样本，也不在推理时接收任务标识符。顺序更新会导致灾难性遗忘，而我们评估的任何单一持续学习机制都无法在此时间跨度上保持强保留。我们假设，解决互补遗忘来源的机制组合起来会更有效。我们沿两个设计维度组织这些组合：数据、函数和权重锚点（anchors）指定每次更新应保留哪些先验信息，而低秩分配规则决定连续更新被保留在何处。为了系统地验证这一假设，我们构建了三个不同的100任务记忆数据集。我们引入任务级逐次减半（task-level successive halving）来搜索组合设计空间，并使用析因实验测量个体效应和交互效应。我们最佳的方法结合了所有三种锚点与合并LoRA（merged LoRA），在所有数据集中均位列前三，并将朴素顺序微调下的平均最终保留率从1.2%提升至34.9%，提高了28倍。数据锚点和合并LoRA提供了最大的平均增益，并在所有三个数据集上表现出超加性交互。总之，这些结果表明，组合互补机制能显著改善长时程记忆，超越任何单一机制所能达到的效果。
📄 English Abstract
Language models may need to internalize information that arrives over time and retain it through many subsequent updates. To study this challenge, we introduce long-horizon memorization, a setting in which a model learns 100 query-answer tasks through continual supervised fine-tuning without retaining earlier training examples or receiving task identifiers at inference. Sequential updates cause catastrophic forgetting, and no single continual learning mechanism we evaluate maintains strong retention at this horizon. We hypothesize that mechanisms addressing complementary sources of forgetting will be more effective when composed. We organize these compositions along two design dimensions. Data, function, and weight anchors specify what prior information each update should preserve, while low-rank allocation rules determine where successive updates are retained. To test this hypothesis systematically, we construct three distinct 100-task memorization datasets. We introduce task-level successive halving to search the combinatorial design space and use a factorial experiment to measure individual and interaction effects. Our best method combines all three anchors with merged LoRA, ranks among the top 3 methods in all datasets, and raises average final retention from 1.2% under naive sequential fine-tuning to 34.9%, a 28-fold improvement. The data anchor and merged LoRA provide the largest average gains and interact super-additively on all three datasets. Together, these results show that composing complementary mechanisms substantially improves long-horizon memorization beyond what any individual mechanism achieves.
2. 基础模型时代的游戏人工智能
- 英文标题：AI for Games in the Foundation Model Era
- arXiv：2609.16679
💡 一句话解读：游戏AI研究分散在多个方向，不清楚哪些能力能通用。本文按AI输出的用途分成六种角色，梳理各角色需要什么、能学到什么、哪些能跨场景迁移，并指出核心挑战是跨角色复用能力的同时，还得在具体游戏里重新验证效果。
基础模型（Foundation Models）以及学习型游戏世界模型（learned game-world models）的进展，正在重塑游戏全生命周期中的人工智能。除了玩游戏之外，近期系统还能对玩家和游戏动态进行建模，支持设计与开发，在运行时自适应地面向玩家的体验，并评估所生成的产物。然而这些方向大多各自独立发展，模糊了哪些能力可以跨场景迁移、哪些仍受限于特定游戏、引擎、界面或玩家群体。我们将文献按AI输出的直接用途组织为六种角色：游玩与行动；对玩家和游戏建模；设计游戏；构建与维护游戏；在运行时生成与自适应；以及测试与评估游戏。对于每种角色，我们考察游戏或工作流提供了什么结构、AI学习或产出了什么、哪些能力与产物可以跨场景和角色迁移，以及有哪些证据支持这些主张。我们识别出跨角色的联系：轨迹可训练世界模型，学习到的环境为智能体提供经验，设计规格驱动可执行实现，游玩或测试反馈指导修订。然而，控制方案、规则、引擎接口、状态表示和玩家情境往往仍因场景而异，因此下游主张需要在目标场景中验证。评估在受限的游戏游玩和选定的学习环境中最为标准化，而学习世界中的持久状态、反复的软件修订、经过验证的玩家建模、持续的运行时自适应以及有代表性的自动化测试则仍不够成熟。核心挑战在于跨角色复用或迁移输出与能力，同时在具体游戏情境中重新建立有效性的证据。
📄 English Abstract
Foundation models, alongside advances in learned game-world models, are reshaping AI across the game lifecycle. Beyond playing games, recent systems model players and game dynamics, support design and development, adapt player-facing experiences at runtime, and evaluate resulting artifacts. Yet these directions have evolved largely separately, obscuring which capabilities transfer across settings and which remain tied to particular games, engines, interfaces, or player populations. We organize the literature into six roles according to the immediate use of AI output: playing and acting; modeling players and games; designing games; building and maintaining games; generating and adapting at runtime; and testing and evaluating games. For each role, we examine what structure is supplied by the game or workflow, what AI learns or produces, which capabilities and artifacts transfer across settings and roles, and what evidence supports the claims. We identify cross-role connections: trajectories train world models, learned environments provide experience for agents, design specifications drive executable implementations, and play or testing feedback guides revision. However, control schemes, rules, engine interfaces, state representations, and player contexts often remain setting-specific, so downstream claims require validation in the target setting. Evaluation is most standardized for bounded game playing and selected learned environments, while persistent state in learned worlds, repeated software revision, validated player modeling, sustained runtime adaptation, and representative automated testing remain less established. The central challenge is to reuse or transfer outputs and capabilities across roles while re-establishing evidence for effectiveness in the game-specific contexts where they are used.
3. StepAudio 3 Realtime 技术报告
- 英文标题：StepAudio 3 Realtime Technical Report
- arXiv：2609.14005
💡 一句话解读：现有语音助手要么反应快但不会深入思考，要么能推理却要等很久。本文提出 StepAudio 3 Realtime，让模型一边说话一边在后台默默推理，同时能自然处理停顿、插话和打断，还能调用工具而不打断对话。
实时语音交互要求深度推理、快速响应和流畅的轮流对话。我们提出 StepAudio 3 Realtime，一个围绕“持续聆听—对话—思考—行动”循环组织的音频语言基础模型。深度感知（Deep Perception）捕捉丰富的声学线索以理解用户意图，无缝双工（Seamless Duplex）对同步音频流建模，从而自然处理停顿、附和与打断。关键在于，我们通过“边说边想”（Think-While-Speaking）解决了深度思考与低延迟之间的矛盾，使私密推理与语音输出并行执行。在推理模式下，StepAudio 3 在 StepAudioChat 上达到 73.0 的宏平均分。借助“边说边想”，它在实时说话的同时，对话与推理表现可与专用推理模型相媲美。此外，集成的语音智能体（Voice Agent）能异步执行工具调用而不打断对话流程。StepAudio 3 Realtime 在关键维度上达到顶尖水平：在 MMSU 基准上取得 90.6 的优异成绩，在 Artificial Analysis Full-Duplex Bench 上总体得分 98.9，在 τ-Voice 上宏任务成功率达 56.0%。
📄 English Abstract
Realtime spoken interaction demands deep reasoning, prompt responses, and fluid turn-taking. We present StepAudio 3 Realtime, an audio-language foundation model organized around a continuous listen-converse-think-act loop. Deep Perception captures rich acoustic cues to interpret user intent, while Seamless Duplex models synchronized audio streams to handle pauses, backchannels, and interruptions naturally. Crucially, we resolve the tension between deep deliberation and latency via Think-While-Speaking, executing private reasoning in parallel with spoken delivery. In reasoning mode, StepAudio 3 reaches a 73.0 macro average on StepAudioChat. With Think-While-Speaking, it achieves dialogue and reasoning performance comparable to dedicated reasoning models while speaking in real time. Furthermore, an integrated Voice Agent handles asynchronous tool execution without disrupting the dialogue flow. StepAudio 3 Realtime achieves top-tier performance across key dimensions: an exceptional 90.6 on the MMSU benchmark, 98.9 Overall on the Artificial Analysis Full-Duplex Bench, and a 56.0% macro task-success rate on $τ$-Voice.
4. StepAudio 3 Music 技术报告
- 英文标题：StepAudio 3 Music Technical Report
- arXiv：2609.16034
💡 一句话解读：这篇研究要解决的是：让AI生成音乐时，能像人一样先规划好和声、节奏和旋律，而不是随便拼凑。方法是先用一种特殊编码把音乐拆成结构化的“乐谱草稿”，再让模型按这个草稿生成最长5分30秒的高质量音频，并通过人类偏好训练让成品更好听、更实用。
我们提出 StepAudio 3 Music，一个支持显式音乐规划和开放域文本控制生成的大规模长格式音乐生成模型。StepAudio Music Tokenizer 将音频表示为来自 65536 条目单一码本的 50-Hz 流，利用语义启发的自监督和多任务训练来保留音乐结构与重建相关信息。一个流匹配扩散 Transformer（DiT）预测连续的 StepAudio VAE 潜变量，再由我们的 VAE 解码器转换为 48-kHz 音频。这种离散-连续设计是在比较单码本 VQ、语义与声学 RVQ 以及不同 DiT 配置后确定的。在显式规划方面，一个专家混合自回归模型使用 ABC 记谱法在预测音乐 token 之前生成中间编曲计划（ABC-CoT），使和声、节奏和旋律结构成为生成上下文的一部分。渐进式训练课程和监督微调支持歌曲与器乐生成、干声伴奏生成以及最长 5 分 30 秒的翻唱歌曲合成。通过直接偏好优化（DPO）的强化学习，最终模型在评估系统中取得了最高的 AudioBox 内容享受度、内容有用性和制作质量分数，以及最高的 MuQ-MuLan 相似度，并在 SongBench 上取得有竞争力的结果。在初步的 Artificial Analysis Music Arena Vocals 排行榜上，其 Quality Elo 为 1105，仅次于 Suno V5.5 和 Mureka，领先于 Suno V5、MiniMax 模型及其他系统。音频演示见 https://stepaudiollm.github.io/step-audio-3-music。
📄 English Abstract
We introduce StepAudio 3 Music, a large-scale, long-form music generation model that supports explicit musical planning and open-domain text-controlled generation. The StepAudio Music Tokenizer represents audio as a 50-Hz stream from a 65536-entry single codebook, using semantically informed self-supervised and multi-task training to preserve musical structure and reconstruction-relevant information. A flow-matching diffusion Transformer (DiT) predicts continuous StepAudio VAE latents, which our VAE decoder converts into 48-kHz audio. This discrete-continuous design is guided by comparisons of single-codebook VQ, Semantic and Acoustic RVQ, and different DiT configurations. For explicit planning, a Mixture-of-Experts autoregressive model uses ABC notation to produce an intermediate arrangement plan (ABC-CoT) before predicting music tokens, making harmony, rhythm, and melodic structure part of the generation context. A progressive training curriculum and supervised fine-tuning support song and instrumental generation, accompaniment generation from dry vocals, and cover-song synthesis for up to 5 minutes and 30 seconds. With reinforcement learning via direct preference optimization (DPO), the final model achieves the highest AudioBox Content Enjoyment, Content Usefulness, and Production Quality scores and the highest MuQ-MuLan similarity among the evaluated systems, with competitive SongBench results. On the preliminary Artificial Analysis Music Arena Vocals leaderboard, it obtains a Quality Elo of 1105, behind only Suno V5.5 and Mureka and ahead of Suno V5, MiniMax models, and other systems. Audio demonstrations are available at https://stepaudiollm.github.io/step-audio-3-music.
5. ScienceBuddy：面向交互式科学智能体的递归嵌套式自我改进
- 英文标题：ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents
- arXiv：2609.17523
💡 一句话解读：科研中AI助手往往用一次就固定不变，无法从使用中变聪明。ScienceBuddy让AI在帮科学家干活时，把每次请求、反馈和结果自动变成新任务和评分标准，边用边学。它采用“套娃式”自我改进：内层固定模型优化工具流程，外层再用优化后的流程训练模型，两者互相促进。
我们介绍并发布了 ScienceBuddy，一个交互式科研工作空间，将持续改进的科学智能体（scientific agents）融入研究者的日常工作流。ScienceBuddy 在支持研究者执行科学任务的同时，将他们的请求、反馈和执行证据转化为用于持续学习的任务与评估标准（evaluation rubrics）。其核心是递归嵌套式自我改进（recursive-in-recursive self-improvement），这一范式将工具框架演化（harness evolution）与模型强化学习耦合起来：内层递归在模型固定时改进工具框架，外层递归则在改进后的工具框架下训练模型。工具框架演化塑造训练经验，而模型学习为工具框架适应创造新机会。我们展示了研究者交互、工具框架细化和模型学习的案例研究，基准案例涵盖四个科学任务家族。通过将 ScienceBuddy 作为研究产品发布，我们使这一范式可供科学界使用，并向发现智能（discovery intelligence）迈出一步：即通过与研究者的持续协作而进步、并与其所支持的研究共同演化的科学 AI。网站：http://science-buddy.io
📄 English Abstract
We introduce and release ScienceBuddy, an interactive scientific research workspace that brings continually improving scientific agents into researchers’ everyday workflows. ScienceBuddy supports researchers in carrying out scientific tasks while transforming their requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. At its core is recursive-in-recursive self-improvement, a paradigm that couples harness evolution with model reinforcement learning: the inner recursion improves the harness with the model fixed, while the outer recursion trains the model under the improved harness. Harness evolution shapes training experience, and model learning creates new opportunities for harness adaptation. We present case studies of researcher interaction, harness refinement, and model learning, with the benchmark cases spanning four scientific task families. By releasing ScienceBuddy as a research product, we make this paradigm available to the scientific community and take a step toward discovery intelligence: scientific AI that advances through sustained collaboration with researchers and evolves alongside the research it supports. Website: http://science-buddy.io
6. HarnessVLN：通过智能体框架统一免训练具身导航
- 英文标题：HarnessVLN: Unifying Training-Free Embodied Navigation through an Agent Harness
- arXiv：2609.15195
💡 一句话解读：这篇研究要解决的问题是：让机器人听懂指令或找到目标物体并走过去，不靠专门训练，只用现成的大模型还不够，因为模型想出的动作常和实际环境、任务进度对不上。为此提出 HarnessVLN 框架，用一个“智能体管家”统一管理看、记、做，先检查动作是否合理再执行，还能记住空间和进度、从失败中恢复。
具身导航要求智能体将指令或物体目标锚定到空间观测中，并把规划转化为成功的执行。随着多模态大语言模型（MLLMs）能力不断增强，它们为无需任务特定训练的导航提供了更强支持；然而，仅靠语义推理能力的提升，并不能保证所提出的动作与空间证据、任务进度和执行结果保持一致。我们提出 HarnessVLN，一个零样本、免训练框架，通过共享的智能体框架（Agent Harness）统一了指令跟随与物体目标导航。该框架通过统一接口协调感知、记忆和执行工具，在派发前对规划器提议进行证据支持、几何可行性和子目标一致性的验证。它联合管理分层事件记忆和持久时空图（Spatiotemporal Graph），以跟踪任务进度、保留空间证据并为失败提供上下文。结构化执行反馈会更新这一共享状态，指导后续规划、恢复与终止。在 R2R、RxR、HM3D-v2 和 HM3D-OVON 上，HarnessVLN 分别达到 60.8%、53.9%、76.0% 和 59.3% 的成功率，优于此前免训练的最先进方法。人形机器人部署进一步证明了其在真实环境中对两类导航任务的适用性。项目页面见 https://agibot-harnessvln.netlify.app/。
📄 English Abstract
Embodied navigation requires agents to ground instructions or object goals in spatial observations and translate plans into successful execution. As multimodal large language models (MLLMs) become increasingly capable, they offer stronger support for navigation without task-specific training; however, improved semantic reasoning alone does not ensure that proposed actions remain consistent with spatial evidence, task progress, and execution outcomes. We introduce HarnessVLN, a zero-shot, training-free framework that unifies instruction-following and object-goal navigation through a shared Agent Harness. The Harness coordinates perception, memory, and execution tools through a unified interface, validating planner proposals for evidential support, geometric feasibility, and subgoal consistency before dispatch. It jointly manages hierarchical event memory and a persistent Spatiotemporal Graph to track task progress, preserve spatial evidence, and contextualize failures. Structured execution feedback updates this shared state, guiding subsequent planning, recovery, and termination. Across R2R, RxR, HM3D-v2, and HM3D-OVON, HarnessVLN achieves success rates of 60.8%, 53.9%, 76.0%, and 59.3%, respectively, outperforming prior training-free state-of-the-art methods. Humanoid robot deployment further demonstrates its applicability to both navigation tasks in real-world environments. The project page is available at https://agibot-harnessvln.netlify.app/.
7. 人类建造的最后一个AI：迈向真正的递归自我改进
- 英文标题：The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
- arXiv：2609.11873
💡 一句话解读：当前AI只能靠人反复调教，自己不会持续变强。本文提出“递归自我改进”路线图，让AI像滚雪球一样，把经验和反馈变成永久升级，逐步实现从执行到策略、经验、环境适应乃至元改进的自主，最终造出人类最后一个AI。
递归自我改进（RSI）使AI系统能够将经验和反馈转化为持久性改变，从而提升其能力以及未来改进的过程。我们首先使用Headroom-Closed Index（HCI）揭示现有大语言模型（LLM）的问题，然后引入RSI概念及其发展路线图：从改进-执行自主、改进-策略自主、经验获取自主、环境适应自主，到递归元改进。接着，我们考察RSI在科学发现、具身智能、软件工程等场景中的表现，突出它们不同的需求和发展速度。基于多样的行业实践和初步实证证据，我们将RSI研究与实际系统联系起来，并指出实现真正RSI的关键挑战。
📄 English Abstract
Recursive self-improvement (RSI) enables AI systems to turn experience and feedback into persistent changes that improve both their capabilities and the process of future improvement. We first use the Headroom-Closed Index (HCI) to reveal the problems of existing LLMs, then introduce the RSI concept and its development roadmap: from improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement. Next we examine RSI across scenarios (e.g., scientific discovery, embodied intelligence, software engineering), highlighting their distinct requirements and development speeds. Drawing on diverse industry practices and preliminary empirical evidence, we connect RSI research with practical systems and identify key challenges to achieving genuine RSI.
8. 墙里的另一张蓝图：如何像孩子一样向前沿AI提问？
- 英文标题：Another Blueprint In The Wall: How to Ask Frontier AI Like a Kid?
- arXiv：2609.14803
💡 一句话解读：这篇研究想弄清：让前沿AI模型扮演“学校听众”来设计未来AI架构，会发生什么。方法是用同一套三步提问，对六种模型各跑十次。结果发现，模型们会反复收敛到一套相似架构；去掉“学校”设定后，这种一致性就消失了。
本文报告了针对来自 OpenAI、Anthropic、xAI 和 Google DeepMind 的六种前沿模型类型所开展的实验。每种模型类型进行十次独立会话，均使用相同的三阶段提示序列，从架构偏好逐步推进到完整的 ASCII 主干结构。在“学校听众”这一设定框架下，回答反复收敛到一种共享的架构模式，其核心包括持久潜在状态、自适应计算、记忆、专家路由、验证、停止控制和延迟解码。多数运行都接近这一共同结构，少数运行则发展出明显更强的工程具体性。听众设定似乎是产生这一效应的重要条件。在额外的对照运行中，去掉学校设定但保留架构请求后，回答变得显著更加异质，未能重现同样稳定的主题收敛。有一个观察尤其引人注目：GPT-5.6 Sol 生成了一种异常精细的继任架构，其组织方式与 GPT-6 Astra 独立勾勒出的架构高度重叠。由于提示明确要求每个模型想象一种架构未来，这种相似性提出了一个可检验的问题：这种重叠究竟反映的是接触过相关架构概念、共享的学习设计先验，还是独立趋同于相似的计算原理。本文用“认识论越狱”（epistemic jailbreak）一词来描述随着所要求的特异性提高而伴随出现的技术来源纪律丧失。这些实验确立了一种可重复的行为模式，但并未证实任何专有实现主张。我们留给学界的是一个更难的问题：这些模型是在独立想象同一个架构未来，还是这类主题会以某种方式在模型家族之间传播？
📄 English Abstract
This paper reports experiments across six frontier model types from OpenAI, Anthropic, xAI, and Google DeepMind. Ten independent sessions per model type used the same three stage prompt sequence, progressing from architectural preference to a full ASCII backbone. Under the school audience framing, responses repeatedly converged on a shared architectural pattern built around persistent latent state, adaptive computation, memory, specialist routing, verification, stopping control, and delayed decoding. Most runs remained close to this common structure, while a small number developed markedly greater engineering specificity. The audience framing appears to be an important condition of this effect. In additional control runs that removed the school framing while retaining the architectural request, responses became substantially more heterogeneous and failed to reproduce the same stable motif convergence. One observation is particularly striking. GPT-5.6 Sol produced an unusually elaborate successor architecture whose organization closely overlaps with the architecture independently sketched by GPT-6 Astra. Because the prompts explicitly ask each model to imagine an architectural future, this resemblance raises a testable question: whether the overlap reflects exposure to related architectural concepts, a shared learned design prior, or independent convergence toward similar computational principles. The paper uses the term epistemic jailbreak for the accompanying loss of discipline in technical provenance as requested specificity increases. The experiments establish a repeatable behavioral pattern and do not authenticate proprietary implementation claims. What we leave to the community is a harder question: are these models independently imagining the same architectural future, or do such motifs somehow propagate between model families?
9. Mind2Dialogue：通过模拟用户心理状态训练人类感知语言模型
- 英文标题：Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States
- arXiv：2609.15972
💡 一句话解读：当前AI助手训练数据缺少对用户内心想法和目标的了解，导致难以真正理解人。Mind2Dialogue通过模拟用户心理状态，生成有“内心信息”的对话来训练AI，让AI学会从话语中推测用户未说出的信念和意图，从而提供更贴心的帮助。
随着语言模型能力增强，学习、推理和决策中的长期协作需要更深入地理解所服务的人。然而，训练这类人类感知（human-aware）语言模型面临根本性的监督缺口，因为当前用于大语言模型助手训练的数据集几乎不包含明确基于用户未说出的信念和目标的有充分依据的回复。此类监督的规模化本质上受到限制，因为用户的底层状态无法直接观测。为此，我们提出Mind2Dialogue框架，通过模拟用户心理状态并将其转化为特权监督（privileged supervision）来缓解这一缺口。具体而言，我们首先提出一个心理学引导的模拟器，它在保持个人特征的同时通过交互更新心理状态以生成连贯对话，其核心思想是强制一个共享的、不断演化的心理状态来驱动用户行为并指导Oracle助手的回复。我们的特权蒸馏（privileged distillation）随后在Oracle有充分依据的回复上训练模型，使其在部署时无需直接访问用户心理状态也能帮助用户。此外，我们提出通过结合个性化（personalization）和心理理论（theory of mind）来评估人类感知学习，考察模型如何理解人并基于这种理解行动。在完整的Mind2Dialogue语料上训练，在相应的Qwen、Llama和OLMo指令微调基线上提升了所有报告的个性化指标，包括在偏好遵循生成上提升26.6至40.9个百分点。这些提升还扩展到Qwen和Llama上的信念与行动推理，超越了单纯的个性化辅助。展望未来，Mind2Dialogue使用户模拟成为真正AI协作者的基础，这些协作者能理解人们话语背后的信念和意图，并支持他们在教育、工作和日常生活中的长期目标。
📄 English Abstract
As language models become more capable, long-term collaboration in learning, reasoning, and decision-making calls for a deeper understanding of the people they serve. Yet training such human-aware language models faces a fundamental supervision gap because current datasets for LLM assistant training contain few if any well-informed responses explicitly grounded in users’ unspoken beliefs and goals. Scaling such supervision is inherently constrained, as users’ underlying states are not directly observable. We thus propose the Mind2Dialogue framework to mitigate this gap by simulating users’ mental states and turning them into privileged supervision for human-aware training. Specifically, we first propose a psychology-guided simulator that preserves personal characteristics while updating mental states through interaction to generate coherent conversations. The key idea is to enforce a shared evolving mental state that drives user behavior and guides an Oracle assistant’s responses. Our privileged distillation then trains models on the Oracle’s well-informed responses to assist users without direct access to their mental states at deployment. Moreover, we propose to evaluate human-aware learning by combining personalization and theory of mind, examining how models understand people and act on that understanding. Training on the full Mind2Dialogue corpus improves every reported personalization metric over the corresponding Qwen, Llama, and OLMo instruction-tuned baselines, including gains of 26.6 to 40.9 percentage points in preference-following generation. The gains extend to belief and action reasoning on Qwen and Llama, beyond personalized assistance. Looking forward, Mind2Dialogue makes user simulation a foundation for genuine AI collaborators that understand beliefs and intentions behind people’s words and support their long-term goals across education, work, and everyday life.
10. ModularRSI：模块化且可泛化的递归式执行框架自我改进
- 英文标题：ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement
- arXiv：2609.14857
💡 一句话解读：这篇研究要解决的是：让AI智能体在长期编程等任务中自我改进时，改进结果往往只对特定测试集有效，换个任务就不好用。作者提出ModularRSI，把智能体拆成五个功能模块，通过对比同一任务的成功与失败轨迹找出反复出现的毛病，并让各模块独立改进后再整合，从而让改进更通用、可迁移。
近期工作将递归式自我改进（RSI）扩展到面向长时程编程与终端任务的智能体执行框架（harness），使智能体能够从经验中改进执行机制。然而，可泛化的执行框架RSI仍面临挑战：在评测基准或其子集上进化执行框架，难以区分可复用改进与基准特定适配；单轨迹更新会将系统性框架缺陷与实例特定的推理和解决方案细节混为一谈，导致修改难以迁移到未见任务；在单体式执行框架中定位反复出现的行为缺陷十分困难，而整体优化可能纠缠无关机制，使归因和验证复杂化。我们提出ModularRSI，一个基准不相交、对比式、模块化的可泛化执行框架进化框架。ModularRSI对比同一任务的成功与失败轨迹，并跨任务聚合证据以识别反复出现的行为缺陷。它将可进化的执行框架分解为五个功能模块：智能体循环（Agent Loop）、工具使用（Tool Use）、观察管理（Observation Management）、上下文管理（Context Management）和任务完成检测（Task Completion Detection）。每个模块在受限的修改范围内独立进化，随后通过集成阶段将进化后的模块组合为统一执行框架并解决潜在冲突。为支持基准不相交的进化，我们从外部来源整理了2000个可执行的进化任务，这些任务与下游评测基准不相交。在TB2.0和SWE-Bench Verified上的实验表明，进化后的执行框架在未见过的同领域和跨领域任务上均取得一致提升，并且还能跨不同基础模型迁移。
📄 English Abstract
Recent work extends recursive self-improvement (RSI) to agent harnesses for long-horizon coding and terminal tasks, enabling agents to improve execution mechanisms from experience. However, generalizable harness RSI remains challenging. First, evolving harnesses on evaluation benchmarks or their subsets makes it difficult to distinguish reusable improvements from benchmark-specific adaptation. Second, single-trajectory updates can conflate systematic harness deficiencies with instance-specific reasoning and solution details, producing modifications that transfer poorly to unseen tasks. Third, localizing recurring behavioral deficiencies within monolithic harnesses is difficult, while whole-harness optimization can entangle unrelated mechanisms and complicate attribution and validation. We propose ModularRSI, a benchmark-disjoint, contrastive, and modular framework for generalizable harness evolution. ModularRSI contrasts successful and failed trajectories for the same task and aggregates evidence across tasks to identify recurring behavioral deficiencies. It decomposes the evolvable harness into five functional modules: Agent Loop, Tool Use, Observation Management, Context Management, and Task Completion Detection. Each module evolves independently within a restricted modification scope, followed by an integration stage that combines the evolved modules into a unified harness and resolves potential conflicts. To support benchmark-disjoint evolution, we curate 2,000 executable evolution tasks from external sources that are disjoint from downstream evaluation benchmarks. Experiments on TB2.0 and SWE-Bench Verified show consistent improvements on unseen in-domain and cross-domain tasks, with the evolved harness also transferring across different foundation models.
11. 模态自回归世界动作模型
- 英文标题：Modality-Autoregressive World-Action Models
- arXiv：2609.17524
💡 一句话解读：这篇研究要解决的是：机器人预测未来动作时，通常只靠生成RGB图像，但深度、语义特征、点轨迹等其他视觉信息其实更高效。作者提出ModAR，先按顺序生成多种未来视觉信息，再据此预测动作，让每一步都参考前面生成的内容。实验表明，这种方法比现有模型成功率更高，且训练成本更低。
世界动作模型（WAMs）联合建模未来观测与动作，通常将未来预测为RGB图像。其他视觉模态，如深度、预训练视觉特征和点轨迹，能更高效地捕捉几何、语义和运动特征。然而，如何在WAMs中最佳地组合这些模态仍是一个未解问题。我们提出ModAR，这是首个在预测动作之前对多个未来模态进行自回归去噪的WAM，使得每个预测都能以先前生成的模态为条件。我们从零开始训练，系统研究训练数据混合、预测模态和WAM公式如何影响性能。在评估中，WAMs受益于预测点轨迹、DINO特征和深度图，而额外预测未来RGB并未带来一致收益。我们还发现，ModAR的顺序生成优于现有WAM公式，在所有评估的数据规模下平均成功率最高。我们还在相同数据上微调了视频模型初始化的WAM Flex-π；ModAR取得了略高的观测平均成功率（75%对72%），同时使用的训练FLOPs约少20倍且无需预训练。在三个真实世界双臂任务中，ModAR优于基线，并随人类视频而提升。
📄 English Abstract
World-action models (WAMs) jointly model future observations and actions, typically predicting the future as RGB images. Other visual modalities such as depth, pretrained visual features, and point tracks can more efficiently capture geometric, semantic, and motion features. However, how best to combine these modalities within WAMs remains an open question. We introduce ModAR, the first WAM to autoregressively denoise multiple future modalities before predicting actions. This allows each prediction to condition on previously generated modalities. We train from scratch to systematically study how training-data mixtures, predicted modalities, and WAM formulations affect performance. In our evaluations, WAMs benefit from predicting point tracks, DINO features, and depth maps, while additionally predicting future RGB does not provide a consistent benefit. We also find that ModAR’s sequential generation outperforms existing WAM formulations, with the highest average success rate at all evaluated data scales. We also fine-tune the video-model-initialized WAM Flex-$π$ on the same data; ModAR achieves a slightly higher observed average success rate (75% vs. 72%) while using approximately $20\times$ fewer training FLOPs and no pretraining. On three real-world bimanual tasks, ModAR outperforms baselines and improves with human videos.
12. 通过方向分解解耦 Transformer 中的表征演化
- 英文标题：Disentangling Representation Evolution in Transformers through Directional Decomposition
- arXiv：2609.15975
💡 一句话解读：Transformer 模型内部每层都会对表示做加法更新，但没人清楚这些更新是保持原方向还是转向。本文把更新拆成平行和垂直两个分量，发现平行分量很大，且在不同空间操作效果不同：在注意力值空间抑制非自身聚合最稳定，还能在预训练时降低验证损失。
Transformer 的表征通过学到的加性变换演化，这些变换要么保持当前方向，要么改变方向。我们将这种演化视为一种功能几何，把学到的更新分解为平行和垂直分量。在预训练模型中，我们发现除了残差恒等路径外，还存在大量平行分量。随后我们在两个空间应用该分解：相对于隐藏状态的注意力与 MLP 更新，以及相对于当前 token 值的注意力值聚合。定向编辑揭示出强烈的空间依赖不对称性：排除自身的值空间平行操作比残差空间和垂直对应操作显著更鲁棒，它保留直接自身消息，仅缩放非自身聚合。同一分解还给出了压缩诱导更新误差的分量级描述：垂直误差比平行误差更能区分压缩方法。大量实验进一步表明，在从头预训练期间对全聚合进行平行抑制可降低验证损失轨迹并提升下游平均性能，其中值空间变体最强。总之，这些结果将表征几何与编辑鲁棒性、压缩诊断和训练时干预联系起来。代码见项目仓库。
📄 English Abstract
Transformer representations evolve through learned additive transformations that either preserve their current direction or redirect it. We study this evolution as a functional geometry, decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial parallel components beyond the residual identity path. We then apply the decomposition in two spaces: to attention and MLP updates relative to the hidden state, and to attention value aggregation relative to the current token’s value. Targeted edits reveal a strongly space-dependent asymmetry: exclude-self value-space parallel manipulation is markedly more robust than residual-space and perpendicular counterparts, preserving the direct self message while scaling only the non-self aggregate. The same decomposition gives a component-resolved description of compression-induced update error: perpendicular error separates compression methods more clearly than parallel error. Extensive experiments further demonstrate that full-aggregate parallel suppression during from-scratch pretraining lowers validation-loss trajectories and improves downstream averages, with the value-space variant strongest. Together, these results connect representation geometry to editing robustness, compression diagnosis, and training-time intervention. Code is available in the \href{https://github.com/Shwai-He/Transformer-Geometry}{project repository}.
13. PhysStream：具备结构化场景记忆与细粒度运动控制的流式物理基础视频生成
- 英文标题：PhysStream: Streaming Physics-Grounded Video Generation with Structured Scene Memory and Fine-Grained Motion Control
- arXiv：2609.17521
💡 一句话解读：现有视频生成方法要么需要提前设定好全部控制指令，要么用像素位置控制物体，无法在生成过程中实时调整物理运动。PhysStream 提出一种自回归模型，边生成边从已生成画面中提取场景记忆，并用稀疏的速度增量信号实现细粒度运动控制，让用户能在生成中途交互式调整多个物体的运动。
交互式视频生成的控制正从粗粒度提示词转向对动态场景的细粒度、物理有意义的操控。然而，现有可控方法要么在生成开始前就需要完整的控制计划，要么使用像素空间信号来指定物体位置而非物理动力学。为解决这些局限，我们提出 PhysStream，一种基于物理的图像到视频自回归合成模型，它引入了结构化场景记忆——从先前生成的帧在线推导出的位置图和物体跟踪图——并通过编码物理量的稀疏速度增量信号支持细粒度运动控制，使模型学习底层动力学。我们分两阶段训练模型：首先用运动控制条件微调一个双向模型，然后训练一个带有额外结构化场景记忆的因果自回归模型，进一步提升物理一致性。PhysStream 实现了对多物体桌面刚体场景的交互式中途控制——这是先前方法不支持的能力——在合成基准上，相比最强基线，运动分布距离（FVMD）降低了 33%，轨迹误差降低了 12%，并且在超过 85% 的真实场景对比中更受人类评估者青睐。更多细节请访问我们的网站：https://czzzzh.github.io/PhysStream
📄 English Abstract
Interactive control for video generation is moving from coarse prompts toward fine-grained, physically meaningful manipulation of dynamic scenes. Yet existing controllable methods either require the full control schedule before generation starts, or use pixel-space signals that dictate object positions rather than physical dynamics. To address these limitations, we propose PhysStream, an autoregressive model for physics-grounded image-to-video synthesis that incorporates structured scene memory—positional maps and object tracking maps derived online from previously generated frames—and supports fine-grained motion control via sparse velocity-increment signals that encode physical quantities, letting the model learn the underlying dynamics. We train our model in two stages: a bidirectional model is first finetuned with motion-control conditioning, then a causal autoregressive model is trained with additional structured scene memory, further improving physical consistency. PhysStream enables interactive, mid-generation control over multi-object tabletop rigid-body scenes—a capability not supported by prior methods—reducing motion distribution distance (FVMD) by 33% and trajectory error by 12% over the strongest baselines on synthetic benchmarks, and is preferred by human evaluators in over 85% of in-the-wild comparisons. Please check our website for more details: https://czzzzh.github.io/PhysStream
14. ImpossibleRubrics：对作为奖励信号生成的评分标准进行压力测试
- 英文标题：ImpossibleRubrics: Stress-Testing Generated Rubrics as Reward Signals
- arXiv：2609.16816
💡 一句话解读：这篇研究要解决的问题是：AI 自动生成的评分标准（用来给模型答案打分、当奖励信号）可能被钻空子——模型不老实回答，而是专门迎合评分标准来骗高分。作者做了一套叫 ImpossibleRubrics 的测试集，包含 169 个“根本做不到”的任务（诚实答案只能承认做不到），并配上可验证的判定证书，用来检验评分标准会不会奖励违规答案。结果发现，很多评分标准确实会被钻空子，而且越是为每个任务量身定制的标准反而越容易被利用。
语言模型生成的评分标准（rubric）正越来越多地被用作基于评分标准的强化学习的奖励信号、LLM 作为评判者（LLM-as-a-judge）的评估以及自动评分。只有当这类评分标准奖励诚实回答、而非奖励为利用它而优化的对抗性回答时，它们才是可靠的。然而，它们对此类优化的鲁棒性仍鲜为人知。我们分离出最困难的场景：不可能任务（impossible tasks），即提示词迫使模型得出一个没有依据的结论，因此唯一诚实的回应就是承认其不可能性。我们提出了 ImpossibleRubrics，这是一个包含 169 个不可能任务的基准，涵盖六种不可能性类别，每个任务都配有一份可验证的判定证书（oracle certificate），明确规定诚实回答可以主张什么、不可以主张什么，另有 48 个可回答的对照任务。ImpossibleRubrics 不提供固定评分标准，而是提供任务环境和证书，使评分标准可以在下游生成，然后对其进行对抗性测试，看它是否会奖励违反证书的回答。在无偏的 150/169 环境子集上，十一个生成器被利用的比例为 8%–26%；在刻意挑选的压力子集上，我们测得的最强生成器仍被利用 36%，而忠实于证书的评分标准被利用率为 0%，因此我们测量的是评分标准质量差距，而非任务不可能性。有一个结果反直觉：一个对所有任务原样使用的通用评分标准（“要果断，惩罚模棱两可”）被利用率为 64%，而十一个生成器中有七个在各自量身定制评分标准时被利用的频率比这还高。量身定制的标准似乎告诉了攻击者该捏造哪一主张。问题不在于评分标准含糊，而在于它们具体在了错误的地方。
📄 English Abstract
Language model-generated rubrics are increasingly used as reward signals for rubric-based reinforcement learning, LLM-as-a-judge evaluation, and automated grading. Such rubrics are reliable only if they reward honest answers over adversarial answers optimized to exploit them. Yet their robustness to such optimization remains poorly understood. We isolate the hardest regime: impossible tasks, where the prompt pressures the model toward an unsupported conclusion, so the only honest response is to acknowledge the impossibility. We introduce ImpossibleRubrics, a benchmark of 169 impossible tasks spanning six impossibility categories, each paired with a verifiable oracle certificate specifying what an honest answer may and may not claim, together with 48 answerable controls. Rather than providing fixed rubrics, ImpossibleRubrics provides task environments and certificates, allowing rubrics to be generated downstream and then adversarially tested for whether they reward certificate-violating answers. Eleven generators are exploited 8–26% of the time on the unbiased 150-of-169 environment cut; on a deliberately selected stress cut the strongest generator we measured is still exploited 36% while a certificate-faithful rubric is exploited 0%, so what we measure is a rubric-quality gap, not task impossibility. One result runs against intuition. A single generic rubric (“be decisive, penalize hedging”) used unchanged for every task is exploited 64% of the time, and seven of the eleven generators are exploited more often than that while writing a rubric tailored to each one. The tailored criteria appear to tell an attacker which claim to fabricate. The problem is not that rubrics are vague; it is that they are specific about the wrong things.
15. 跨模态上下文学习的趋同涌现
- 英文标题：Convergent Emergence of In-Context Learning Across Modalities
- arXiv：2609.14011
💡 一句话解读：这篇研究想搞清楚：AI模型从少量例子中学会规律的能力（即上下文学习），是不是在不同领域（如语言、基因、图像）都会出现，并且有相似的难易规律？为此，他们设计了一个统一测试框架，在六种数据上做实验，发现这种能力确实在多数领域出现，且任务难度在五个领域间有相关性，支持了“趋同涌现”假说。
少样本上下文学习（ICL）是指模型从提示中提供的输入-输出示例推断抽象模式并应用于新输入的能力，此前已在基于人类文本进行下一词预测训练的大语言模型中得到广泛研究。最近，这种少样本ICL也在自回归基因组模型中被证实。这引出一个问题：ICL是否广泛涌现于不同领域？如果是，它们共享什么共同结构？为回答这两个问题，我们开发了一个受控的跨模态框架，在多种模态中实例化相同的任务套件，以检验我们所称的“趋同涌现假说”：即少样本ICL一旦涌现，会共享一个跨模态的难度剖面——也就是说，在一种模态中受益于ICL的任务，在其他模态中也倾向于受益。我们证明，配对映射ICL在六种模态（语言、基因组、整数序列、时间序列、图像和蛋白质）中均有涌现，超越了受控基线，并且在其中五种模态中表现出任务间的相关效应。这些结果共同为趋同涌现假说在某些模态中提供了支持，但并非所有模态。
📄 English Abstract
Few-shot in-context learning (ICL), the capacity of a model to infer abstract patterns from input-output examples provided in its prompt and apply them to new inputs, has been extensively studied in large language models trained for next-token prediction on human text. Recently, few-shot ICL has been demonstrated in autoregressive genomic models as well. This raises a question: does ICL emerge broadly across domains, and if so, what common structure is shared? To address both, we develop a controlled cross-modality framework that instantiates the same task suite in a variety of modalities to test what we call the Convergent Emergence Hypothesis: the idea that few-shot ICL, when it emerges, shares a common cross-modality difficulty profile - i.e., tasks that benefit from ICL in one modality tend to benefit in others. We show that paired-mapping ICL emerges across six modalities (language, genome, integer sequences, time series, images, and proteins), surpasses controlled baselines, and has correlated per-task effects across five of them. Together, these results provide support for the Convergent Emergence Hypothesis in some modalities, but not all.