Daily Papers · 2026-09-08
📄 论文日报 | 来源：Hugging Face Daily Papers
🗓 Episode：2026.09.07 | 多智能体协作需外部验证；搜索模型应区分自身能力与上下文增益
1. 双层协调反思：多智能体LLM系统的博弈论方法
- 英文标题：Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems
- arXiv：2609.02750v1
💡 一句话解读：多智能体系统常靠一个‘总指挥’把任务分给多个‘工人’，再通过反思改进。但现有方法缺乏统一理论。本文用博弈论分析这种协作，发现反思可能无效，于是提出一种新方法：只有确认改进有效时才接受反思结果，从而保证系统稳定提升，并在编程任务中取得更好成绩。
多智能体LLM系统通常使用一个编排器将任务分解给一组工人，然后通过文本反思进行改进。尽管经验结果强劲，但这些系统缺乏对协调、记忆改进和外部验证作用的统一解释。我们将编排器-工人交互建模为双层协调博弈：在有界耦合下，工人的局部更新博弈是一个近似势博弈，其均衡松弛由分解质量控制。然后，我们将反思分析为语义记忆状态上的随机移动。对于自由形式的反思，我们推导出有限时间上界，证明最坏情况紧性，并在可证伪的持续危害条件下给出正下界。我们进一步证明了一个信息论不可能性结果：仅观察生成转录的门控无法在文本不可区分环境中统一改进，而环境接地门控可以。受此分离启发，我们引入了随机反思记忆上升（SRMA），它仅在接地评估风险严格下降后才接受候选记忆。在校准和非退化纠正质量下，SRMA精确收敛，几何或多项式速率；匹配构造表明两种速率机制都是阶紧的。我们还为随机评估提供置信门控，并为分段平稳环境提供重新锚定保证。实验用环境接地指标实例化这些对象，并测试预测的协调和漂移规律。在500个SWE-bench实例上，完整的基于Kimi的系统解决了72.2%，而公共mini-SWE-agent参考为70.8%。代码：https://github.com/YihangChen9/Bilevel-Coordinated-Reflection
📄 English Abstract
Multi-agent LLM systems commonly use an orchestrator to decompose a task for a team of workers and then improve through textual reflection. Despite strong empirical results, these systems lack a unified account of coordination, memory improvement, and the role of external verification. We model orchestrator-worker interaction as a bilevel coordination game: under bounded coupling, the workers’ local-update game is an approximate potential game whose equilibrium slack is controlled by decomposition quality. We then analyse reflection as stochastic movement over semantic memory states. For free-form reflection, we derive a finite-time upper bound, prove worst-case tightness, and give a positive lower bound under a falsifiable persistent-harm condition. We further prove an information-theoretic impossibility result: no gate that observes only the generated transcript can improve uniformly over text-indistinguishable environments, whereas an environment-grounded gate can. Motivated by this separation, we introduce Stochastic Reflective Memory Ascent (SRMA), which accepts a candidate memory only after a grounded evaluation risk strictly decreases. Under calibration and non-degenerate corrective mass, SRMA converges exactly, geometrically or polynomially; matching constructions show that both rate regimes are order-tight. We also provide confidence gating for stochastic evaluation and re-anchoring guarantees for piecewise-stationary environments. Experiments instantiate these objects with environment-grounded metrics and test the predicted coordination and drift laws. On 500 SWE-bench instances, the complete Kimi-based system resolves 72.2% versus a 70.8% public mini-SWE-agent reference. Code: https://github.com/YihangChen9/Bilevel-Coordinated-Reflection
2. Iris：攀登搜索前沿
- 英文标题：Iris: Climbing to the Search Frontier
- arXiv：2609.04304v1
💡 一句话解读：这篇研究要解决的是如何让AI更好地进行多步搜索和推理。他们提出了两个搜索智能体Iris-mini和Iris-pro，通过一种新的训练方法（SFT-RL爬升）来提升性能，在多个基准测试上取得了领先结果。
我们介绍了Iris-mini和Iris-pro两个搜索智能体，分别在35B-A3B和397B-A17B规模上训练，并提供了背后的数据管道和训练方法。任务是从网页语料的超链接结构反向构建：我们从种子页面及其外链中提炼实体图，编写多跳链，将每个非答案实体改写为描述性引用，使得无法通过字符串匹配解决线索，并且只保留参考模型在闭卷时失败但提供证据后能解决的问题。这些问题随后被转化为轨迹，在SFT前在轨迹和回合层面进行过滤。策略随后通过RL针对实时搜索进行优化，奖励判断器和观察摘要器在训练集群内运行，过长的rollout在请求级别中断，并在下一步从已提交的前缀恢复。我们将两个阶段交替进行，称为SFT-RL爬升，将每轮RL中最难解决和最有效的rollout返回给下一轮监督训练。由于在这些基准上推理时的上下文管理比大多数系统间报告的差异更有价值，我们在启用和不启用上下文管理的情况下评估每个基准，保持工具集、上下文限制和判断器固定。所有结果来自单个ReAct智能体，没有子智能体或测试时验证。启用管理后，在BrowseComp、BrowseComp-ZH、DeepSearchQA和HLE上，两个模型分别达到82.2/84.8/86.9/52.3和88.6/85.1/92.9/56.4，这是各自参数范围内开源搜索智能体的最强整体结果。我们计划发布模型权重以及数据构建、训练和评估的完整方法。
📄 English Abstract
We present Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, together with the data pipeline and training recipe behind them. Tasks are reverse-constructed from the hyperlink structure of a web corpus: we author multi-hop chains over an entity graph distilled from a seed page and its out-links, rewrite every non-answer entity into a descriptive reference so that no clue can be resolved by string matching, and admit only questions that a reference model fails closed-book yet solves once the supporting evidence is supplied. These questions are then turned into trajectories, which are filtered at both the trajectory and the turn level before SFT. The policy is then optimized by RL against live search, with the reward judge and the observation summarizer served inside the training cluster, and with over-long rollouts interrupted at the request level and resumed from their committed prefix at the next step. We alternate the two stages in a procedure we call SFT-RL climbing, returning the hardest solved and most efficient rollouts of each RL round to the next supervised pass. Because inference-time context management is worth more on these benchmarks than most reported differences between systems, we evaluate every benchmark both with and without it, holding the tool set, the context limit, and the judge fixed. All results come from a single ReAct agent, with no sub-agents and no test-time verification. With management enabled, on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE the two models reach $82.2/84.8/86.9/52.3$ and $88.6/85.1/92.9/56.4$, the strongest overall results among open-source search agents in their respective parameter ranges. We plan to release the model weights together with the complete recipe for data construction, training, and evaluation.
3. Motion-Omni：面向口语对话的端到端语音与全身动作联合生成
- 英文标题：Motion-Omni: End-to-End Joint Speech and Full-Body Motion for Spoken Dialogue
- arXiv：2609.04250v1
💡 一句话解读：现有聊天数字人要么只会说话不会动，要么只会根据音频做动作，两者是分开的。这篇研究提出一个叫 Motion-Omni 的端到端模型，让数字人在生成语音的同时，直接同步生成面部表情和全身动作，不需要先录好音再做动作，速度更快，动作也更自然。
一个进行对话的虚拟形象应该决定说什么以及说话时的动作，但这些能力存在于不同的模型家族中：口语对话模型生成语音但不生成动作，而共语动作模型只能从提供给它的音频中生成动作。标准的解决方案是级联方法，即首先生成口语回应，然后在完成的音频上运行动作模型，这需要第二次完整的推理过程，并阻碍了两者之间的任何联合优化。我们提出了 Motion-Omni，一个端到端框架，其中口语对话模型原生地输出明确的面部表情以及手部、上半身和下半身的动作，这些动作直接从产生语音的隐藏状态生成。联合训练在这里是必不可少的：当语音通路被冻结时，动作与音频保持不对齐，而在两个目标下共同调整 LLM、语音生成器和动作生成器才能恢复对齐，同时保留口语对话能力。监督来自一个可扩展的、模型无关的流水线，该流水线使用可替换的动作教师为一致语音的语音回应生成伪标签，产生 422,856 个质量排序的配对（1,402 小时）。我们进一步发布了 SwDA-500，并且据我们所知，这是第一个针对随机开放式全身口语对话的公开评估协议，在跨动作系统匹配音频的同时统一了渲染、自动指标、人类评估和延迟测量。以 Qwen2.5-7B-Instruct 为骨干实例化，Motion-Omni-Q7 在无参考动作指标上与同音频教师级联的差距在 2% 以内，同时响应速度快 5.4 倍（RTF=0.78，快于实时），在节拍相关性和多样性上超过所有非教师级联，并达到 2.62% 的词错误率，在所比较的全模态系统中最低。
📄 English Abstract
An avatar that holds a conversation should decide what to say and to move while saying it, yet these abilities live in separate model families: spoken dialogue models produce speech without motion, and co-speech motion models produce motion only from audio handed to them. The standard remedy is a cascade that first generates the spoken response and then runs a motion model over the finished audio, which requires a second full inference pass and precludes any joint optimisation between the two. We present Motion-Omni, an end-to-end framework in which a spoken dialogue model natively outputs explicit facial expression together with hand, upper-body and lower-body motion, generated directly from the hidden states that produce the speech. Joint training is not optional here: with the speech pathway frozen, motion remains misaligned with the audio, and co-adapting the LLM, Speech Generator and Motion Generator under both objectives is what recovers alignment while retaining spoken-dialogue ability. Supervision comes from a scalable, model-agnostic pipeline that pseudo-labels consistent-voice speech responses with a replaceable motion teacher, yielding 422,856 quality-ranked pairs (1,402 hours). We further release SwDA-500 and, to our knowledge, the first public evaluation protocol for stochastic open-ended full-body spoken dialogue, matching audio across motion systems while unifying rendering, automatic metrics, human evaluation, and latency measurement. Instantiated with a Qwen2.5-7B-Instruct backbone, Motion-Omni-Q7 matches the same-audio teacher cascade to within 2% on reference-free motion metrics while responding 5.4 x faster (RTF=0.78, faster than real time), surpasses all non-teacher cascades on beat correlation and diversity, and reaches a 2.62% word error rate, the lowest among the omni-modal systems compared.
4. WorldSculpt：从锚定视频生成组合式世界
- 英文标题：WorldSculpt: Generating Compositional Worlds from Grounded Videos
- arXiv：2609.05416v1
💡 一句话解读：这项研究要解决的是：如何把堆满几百个物体的杂乱场景，变成一个个独立物体模型组成的3D世界。以前的方法要么只能生成一个整体，要么处理不了太复杂的场景。他们想出的办法是：借用已经很成熟的“单个物体3D生成”技术，再结合从多个角度拍摄的画面，让模型学会在复杂场景里也能逐个生成物体。结果证明，这个方法不用专门训练就能处理严重遮挡的大场景，效果还比之前的方法好。
我们研究的问题是生成包含数百个物体的杂乱场景的组合式3D表示。目标是将场景表示为放置在共享世界坐标系中的单个物体网格集合，以满足游戏、AR/VR、仿真和机器人等下游应用的需求。在物体相互严重遮挡且每个视角仅显示其几何形状一小部分的密集杂乱场景中，这项任务颇具挑战性。基于几何的方法通常将场景重建为单一表示，并在遮挡区域留下不完整的几何形状，而现有的具有生成先验的组合方法大多局限于相对简单的场景。我们证明，通过将强大的单物体3D生成先验适应于多视角观测，可以组合式地生成包含数百个物体的复杂场景。我们以Pixal3D为例实现了这一范式，并扩展了多视角条件路径，使物体生成基于多个姿态观测。尽管模型完全在规范空间中的单个物体上进行微调，但它无需任何场景级训练即可泛化到具有严重遮挡的大规模场景，证明了该范式的可行性和可扩展性。我们进一步引入了UE-MeshyScene，一个包含数百个物体、逐物体标注和真实网格的逼真密集杂乱场景基准。在单物体、受控多物体和UE-MeshyScene评估中，我们的方法始终优于先前方法，并且随着场景复杂度和遮挡程度的增加，优势更大。最后，我们通过将生成的3DGS世界（如Marble和HY-World 2.0）转换为组合式网格场景，展示了更广泛的适用性。
📄 English Abstract
We study the problem of generating a compositional 3D representation of a cluttered scene containing hundreds of objects. The goal is to represent the scene as a collection of individual object meshes placed in a shared world frame, as required by downstream applications such as gaming, AR/VR, simulation, and robotics. This task is challenging in densely cluttered scenes, where objects heavily occlude one another and each view reveals only a fraction of their geometry. Geometry-based approaches typically reconstruct the scene as a single representation and leave incomplete geometry in occluded regions, while existing compositional methods with generative priors are largely limited to relatively simple scenes. We show that complex scenes with hundreds of objects can instead be generated compositionally by adapting a strong single-object 3D generative prior to multi-view observations. We instantiate this paradigm with Pixal3D, extending it with a multi-view conditioning pathway that grounds object generation in multiple posed observations. Although the model is finetuned entirely on single objects in canonical space, it generalizes to large scenes with severe occlusion without any scene-level training, demonstrating the feasibility and scalability of this paradigm. We further introduce UE-MeshyScene, a photorealistic benchmark of densely cluttered scenes with hundreds of objects, per-object annotations, and ground-truth meshes. Across single-object, controlled multi-object, and UE-MeshyScene evaluations, our method consistently outperforms prior approaches, with larger gains as scene complexity and occlusion increase. Finally, we demonstrate broader applicability by converting generated 3DGS worlds, such as Marble and HY-World 2.0, into compositional mesh scenes.
5. Enoki：高效的多层级幻觉检测
- 英文标题：Enoki: Efficient Multi-Level Hallucination Detection
- arXiv：2609.00581v2
💡 一句话解读：大模型在重要场合使用时，经常说错话（幻觉）。现有检测方法要么只查整句话对不对，要么只找具体哪个词错了，两者结合很费劲。这篇研究提出一个叫 Enoki 的工具，它把句子拆成简单事实，核对后能直接指出错在哪个词，既省力又准确。
在高风险场景中部署大语言模型（LLM）时，确保事实性仍是一个关键挑战。现有的幻觉检测器通常在单一层级上运行：声明级（claim-level）方法提供可解释的事实单元，而跨度级（span-level）方法定位不支持的文本。弥合这两种视角成本高昂，因为依赖LLM的流程需要多次分解和验证调用，模块化系统还需要额外的声明到跨度对齐。我们提出Enoki，一个用于多层级幻觉检测的开放信息抽取（Open Information Extraction）框架。Enoki抽取文本锚定的关系事实，对照证据进行验证，并将不支持的事实映射回幻觉跨度。这种共享表示使得声明级验证和跨度级定位无需单独对齐即可实现。Enoki支持基于LLM、基于编码器和基于规则的抽取方式，通过统一接口平衡准确性和推理成本。实验表明，Enoki在使用更少资源的同时，与强声明级系统保持竞争力，并在细粒度跨度和实体级定位上取得优越性能。我们还发布了EnokiQA，一个具有对齐的声明级验证和跨度级定位注释的双粒度数据集。
📄 English Abstract
Ensuring factuality remains a critical challenge for deploying LLMs in high-stakes settings. Existing hallucination detectors usually operate at a single level: claim-level methods provide interpretable factual units, while span-level methods localize unsupported text. Bridging these views is costly, as LLM-heavy pipelines require multiple decomposition and verification calls, and modular systems need additional claim-to-span alignment. We propose Enoki, an Open Information Extraction framework for multi-level hallucination detection. Enoki extracts text-anchored relational facts, verifies them against evidence, and projects unsupported facts back to hallucinated spans. This shared representation enables claim-level verification and span-level localization without requiring separate alignment. Enoki supports LLM-based, encoder-based, and rule-based extraction regimes, balancing accuracy and inference cost through a common interface. Experiments show that Enoki remains competitive with strong claim-level systems while using fewer resources and achieves superior performance on fine-grained span- and entity-level localization. We also release EnokiQA, a dual-granularity dataset with aligned claim-level verification and span-level localization annotations.
6. 优化前先询问：面向交互式优化的动态预建模澄清
- 英文标题：Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization
- arXiv：2609.05258v1
💡 一句话解读：这篇研究要解决的是：用户用自然语言描述优化问题时常常漏掉关键信息，导致模型建错。他们提出了一个测试平台和一套方法，让AI在建模前先判断要不要追问用户，而不是闷头建模。
大型语言模型（LLMs）越来越多地被用于从自然语言问题描述中构建优化模型，然而现实中的运筹学（OR）请求往往不完整：缺失的目标、约束或业务规则可能改变最终的数学规划。现有评估大多假设规格完整，因此忽略了代理在建模前是否知道何时需要澄清。我们引入了OR-Clarify，一个用于建模前澄清的基准测试。每个任务提供部分公开问题描述，隐藏结构化槽位，并通过与模拟用户的有限交互来评估代理。该基准支持开放式和基于选择的澄清，并衡量槽位恢复、停止行为、隐性假设和交互成本。我们进一步提出了交互式优化（InterOPT），一个两阶段框架，用于识别未解决的、对建模关键的空缺，并据此指导是提出下一个问题还是停止。在我们的基于选择的实验中，InterOPT在精确槽位恢复上大幅优于所有基线；在开放式设置中，它与强先验方法保持竞争力。总之，OR-Clarify和InterOPT将OR辅助重新定义为选择性完整性决策：需要时澄清，准备好时停止，并量化仍缺失的内容。
📄 English Abstract
Large language models (LLMs) are increasingly used to formulate optimization models from natural-language problem descriptions, yet realistic operations research (OR) requests are often incomplete: missing objectives, constraints, or business rules can change the resulting mathematical program. Existing evaluations largely assume a complete specification and therefore overlook whether an agent knows when clarification is needed before modeling. We introduce OR-Clarify, a benchmark for pre-formulation clarification. Each task presents a partial public problem description, withholds structured hidden slots, and evaluates agents through bounded interaction with a simulated user. The benchmark supports both openended and choice-based clarification, and measures slot recovery, stopping behavior, silent assumptions, and interaction cost. We further propose Interactive Optimization (InterOPT), a two-stage framework that identifies unresolved formulation-critical gaps and uses them to guide whether to ask the next question or to stop. In our choice-based experiments, InterOPT substantially outperforms all baselines in exact slot recovery; in the open-ended setting, it remains competitive with strong prior methods. Together, OR-Clarify and InterOPT reframe OR assistance as a selective completeness decision: clarify when needed, stop when ready, and quantify what remains missing.
7. 音频-视频模型中的注意力三角
- 英文标题：The Attention Triangle in Audio-Video Models
- arXiv：2609.03586v1
💡 一句话解读：这篇研究要解决的是音频-视频生成模型中，文字、声音和画面之间信息串味的问题。他们发现模型内部有个“注意力三角”，声音和画面会互相影响，导致生成结果跑偏。他们提出用注意力信号来诊断和干预，让生成的内容更符合文字描述。
音频-视频扩散模型依赖跨模态注意力来协调文本、声音和视觉内容，但这一机制也可能引入微妙且系统的语义泄漏。我们通过探测和分析“注意力三角”（即连接文本、音频和视频流的三个跨注意力边）来研究这些模型，并考察生成过程中语义信息如何在模态间路由。分析显示，音频-视频边的路由是双向的：音频可影响视频生成，视频也可影响音频生成。该边受模型参数中编码的偏差影响，并成为泄漏的主要来源：当提示与学习先验冲突时，跨模态交互可能覆盖预期条件，将语义重定向至视觉上典型但错误的结果。这些效应表明，语义伪影不仅源于注意力扩散至预期目标之外，更源于沿特定路径的结构化、偏差驱动的交互。基于此，我们提取注意力派生信号以揭示语义如何在模态间分布和锚定，并将其用作诊断工具，既分析泄漏，也在受控条件下故意引发泄漏。这使我们能探测跨模态路由的内部动态，并隔离单个交互的作用。我们进一步利用这些信号指导推理时干预，以促进更一致的跨模态对齐。大量实验支持我们的分析，并展示了在保持生成质量的同时改善语义锚定。
📄 English Abstract
Audio-video diffusion models rely on cross-modal attention to coordinate text, sound, and visual content, yet this same mechanism can introduce subtle and systematic semantic leakage. We study these models by probing and analyzing the ``attention triangle,’’ comprising the three cross-attention edges connecting the text, audio, and video streams, and examine how semantic information is routed across modalities during generation. Our analysis reveals that routing along the audio-video edge is bidirectional: audio can influence video generation, while video can influence audio generation. This edge is shaped by biases encoded in the model’s parameters and emerges as a major contributor to leakage: when prompts are in tension with learned priors, cross-modal interactions may override the intended conditioning and reroute semantics toward visually canonical but incorrect outcomes. These effects suggest that semantic artifacts arise not merely from attention spreading beyond its intended target, but from structured, bias-driven interactions along specific pathways. Building on this perspective, we extract attention-derived signals that expose how semantics are distributed and grounded across modalities, and use them as a diagnostic tool to both analyze and deliberately incur leakage under controlled conditions. This enables us to probe the internal dynamics of cross-modal routing and isolate the role of individual interactions. We further leverage these signals to guide inference-time interventions that encourage more consistent cross-modal alignment. Extensive experiments support our analysis and demonstrate improved semantic grounding while preserving generation quality.
8. 不要丢弃Dropout：优化层稀疏性以实现高效的LLM训练与推理
- 英文标题：Don’t Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference
- arXiv：2609.05275v1
💡 一句话解读：这篇研究要解决的是：在大语言模型训练中，层丢弃（一种随机跳过部分网络层的技巧）常被弃用，因为有人认为它会降低精度。作者通过大量实验发现，只要用对方法（调整层分布、时间表和优化器参数），层丢弃不仅能降低训练损失，还能节省最多25%的计算量，并在推理时提速1.5倍。简单说，就是让模型训练更省力、跑得更快，还不掉链子。
层丢弃（Layer dropout，又称随机深度）已被证明能在语言和视觉Transformer中实现更快的训练、更高的准确率以及对零样本层剪枝的鲁棒性。然而，随着模型和数据集规模的扩大，丢弃——尤其是层丢弃——已基本从大型语言模型（LLM）的预训练方案中消失。尽管一些先前工作报道丢弃可能降低准确率，但没有综合研究量化甚至缓解这一影响。在本研究中，我们表明层丢弃应被用于最先进的LLM训练，并为训练和训练后的收益建立最佳实践和扩展分析。具体而言，通过最优的层分布、时间表和优化器超参数，我们观察到在相同的训练FLOPs下，层丢弃能带来更低的损失。对于给定的训练步数，LLM可以实现更低或相似的验证损失，同时节省高达25%的训练FLOPs。此外，层丢弃还支持显著的训练后优化，如提前退出、中间层跳过和自推测解码，在可忽略的精度损失下实现高达1.5倍的推理加速。通过超过2400次训练实验，涵盖从271M到8.2B参数的模型以及高达160B令牌的数据集，我们证明这些发现可可靠地扩展到大规模训练场景。所有预训练实验均在Cerebras CS-3系统上进行。
📄 English Abstract
Layer dropout (a.k.a. stochastic depth) has been shown to enable faster training, higher accuracy, and robustness to zero-shot layer pruning in both language and vision transformers. However, as models and datasets have scaled, dropout - particularly layer dropout - has largely disappeared from large language models (LLMs) pre-training recipes. While some prior work has reported that dropout can degrade accuracy, no comprehensive study has quantified, let alone mitigated, this effect. In this study, we show that layer dropout should be used in state-of-the-art LLM training, establishing best practices and scaling analysis for both training and post-training benefits. Concretely, with optimal layer distribution, time schedule, and optimizer hyperparameters, we observe that at the same training FLOPs layer dropout leads to lower loss. For a given number of training steps, LLMs can achieve lower or similar validation loss while saving upto 25% of training FLOPs. Moreover, layer dropout enables significant post-training optimizations, such as early exit, intermediate-layer skipping, and self-speculative decoding, yielding up to 1.5x inference speedup with negligible accuracy loss. Across more than 2400 training experiments, spanning models from 271M to 8.2B parameters and datasets up to 160B tokens, we demonstrate that these findings extend reliably to large-scale training regimes. All pre-training experiments were run on Cerebras CS-3 systems.
9. 思维链的表面之下：大语言模型中推理操作的机制性解读
- 英文标题：Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs
- arXiv：2609.04753v1
💡 一句话解读：这篇研究想弄清楚大语言模型在推理时，内部是怎么组织不同推理步骤的。他们发现，模型在中间层能把不同推理操作（比如分解问题、推理）区分开，而且这种区分不是靠表面文字。他们还发现，模型会根据上下文用不同的方式表示相同的文字。总之，模型内部确实有和推理步骤对应的几何结构。
大语言模型中的推理通过多种功能操作展开，如问题表述、目标分解和演绎。尽管这些操作在文本中被明确区分，但它们在表示空间中的几何组织方式尚不清楚。为此，我们研究了不同的推理操作是否在隐藏表示中表现出相应的几何结构。我们发现，操作在保留的表示中是可分离的，且分离性在中间层达到峰值，并验证了这种结构并非由词汇或位置混淆因素所解释。跨层来看，逐词的操作对齐在跨度上变得更加分散，而相同的表面词根据其周围块的操作被不同地表示。注意力掩蔽干预进一步表明，块起始处的操作对齐表示依赖于前面的推理上下文。因此，我们的工作表明，语言模型在语言推理表达与其内部几何结构之间保持了表示上的对应关系。代码和项目材料可在 https://github.com/naver-ai/beneath-cot 获取。
📄 English Abstract
Reasoning in large language models unfolds through diverse functional operations, such as problem formulation, goal decomposition, and deduction. Although these operations are explicitly distinguished in text, little is known about how they are geometrically organized in representation spaces. To this end, we investigate whether distinct reasoning operations exhibit corresponding geometric structure in hidden representations. We find that operations are separable in held-out representations, with separability peaking in middle layers, and verify that this structure is not explained by lexical or positional confounds. Across layers, token-wise operation-alignment becomes more distributed over spans, while identical surface tokens are represented differently depending on the operation of its surrounding chunk. Attention-masking interventions further show that operation-aligned representations at chunk onset depend on preceding reasoning context. Consequently, our work demonstrates that language models maintain representational correspondence between linguistic reasoning expressions and their internal geometric structures. Code and project materials are available at https://github.com/naver-ai/beneath-cot.
10. RISE：基于自外推策略蒸馏的递归改进
- 英文标题：RISE: Recursive Improvement via Self-Extrapolating Policy Distillation
- arXiv：2609.05295v1
💡 一句话解读：这篇研究要解决大模型训练中“老师”质量不够好的问题。它提出一种新方法，让模型自己从过去的训练轨迹中生成“虚拟老师”，不断自我改进，不需要外部模型，效果比传统方法更好。
同策略蒸馏（On-policy distillation, OPD）为语言模型后训练提供密集的逐词元监督，但其效果受限于教师质量：外部教师存在分布不匹配问题，而带特权条件的自蒸馏则受限于上下文学习能力。我们提出RISE（通过自外推策略蒸馏的递归改进），该方法直接从模型自身的RLVR训练轨迹中构建合成教师。通过外推当前检查点与滞后锚点之间的位移——在参数空间或输出logit空间——RISE将稀疏的结果驱动的参数更新转化为密集的词元级目标，无需任何外部模型或特权条件。RISE将RLVR和OPD结合在一个互补循环中：结果奖励将外推引导至正确推理，而外推教师则细化词元级决策。此外，由于教师随学生进步而每轮刷新，蒸馏成为一种递归改进机制，而非一次性压缩步骤。在数学推理、多领域STEM、代码生成和多轮智能体任务上的实验表明，RISE在所有设置下均优于仅RLVR训练和同策略自蒸馏。
📄 English Abstract
On-policy distillation (OPD) provides dense, per-token supervision for language model post-training, but its effectiveness is bottlenecked by teacher quality: external teachers suffer from distribution mismatch, while self-distillation with privileged conditioning is limited by in-context learning capacity. We propose \textbf{RISE} (\textbf{R}ecursive \textbf{I}mprovement via \textbf{S}elf-\textbf{E}xtrapolating Policy Distillation), which constructs a synthetic teacher directly from the model’s own RLVR training trajectory. By extrapolating the displacement between the current checkpoint and a trailing anchor—in parameter space or output logit space—RISE converts a sparse outcome-induced parameter update into a dense token-level target, without any external model or privileged conditioning. RISE combines RLVR and OPD in a complementary loop: outcome rewards ground the extrapolation toward correct reasoning, while the extrapolated teacher refines token-level decisions. Moreover, since the teacher is refreshed every iteration as the student improves, distillation becomes a recursive improvement mechanism rather than a one-shot compression step. Experiments spanning mathematical reasoning, multi-domain STEM, code generation, and multi-turn agentic tasks show that RISE outperforms RLVR-only training and on-policy self-distillation across all settings.
11. MaxKernel：面向TPU的智能体内核生成
- 英文标题：MaxKernel: Agentic Kernel Generation for TPUs
- arXiv：2609.04523v1
💡 一句话解读：为AI加速器写高性能代码很难，需要硬件专家。这篇研究用大语言模型（AI）配合编译器反馈，造了个叫MaxKernel的多智能体系统，能自动生成优化代码。它有三种模式：人机协作、全自动、以及图搜索式自动探索，在50个任务上表现接近甚至超过专家手写代码。
为加速器设计和编写高性能自定义内核是一项复杂任务，需要深厚的硬件级专业知识。大型语言模型（LLM）可与实时编译器反馈相结合，构建用于内核生成的智能体系统。在本工作中，我们提出了MaxKernel，一个多智能体系统，实现了三种不同的TPU内核开发范式：（1）用于协作式逐步设计的人机协同（HITL）智能体；（2）执行全自动、基于指标/跟踪的优化循环的自主（Auto）智能体；（3）图基自主搜索，扩展Auto智能体以进行设计空间的全局探索。所有三种范式都利用共享的专业子智能体池来处理规划、实现、自我调试、测试和硬件性能分析。我们在JaxBench（一个包含50个多样化TPU内核任务的综合基准测试集）以及来自最先进开源模型的复杂真实工作负载上评估了MaxKernel。我们证明MaxKernel能持续生成高度优化的实现，与专家手工调优的基线相匹配，并在基准测试中展现出显著的性能。我们的智能体已开源，可在https://github.com/AI-Hypercomputer/accelerator-agents/tree/main/MaxKernel获取。
📄 English Abstract
Designing and authoring high-performance custom kernels for accelerators is a complex task that requires deep hardware-level expertise. Large Language Models (LLM) can be leveraged together with real-time compiler feedback to build agentic systems for kernel generation. In this work, we present MaxKernel, a multi-agent system that implements three distinct paradigms for TPU kernel development: (1) a Human-in-the-Loop (HITL) agent for collaborative, step-by-step design; (2) an Autonomous (Auto) agent that executes a fully automated, metric/trace-driven optimization loop; and (3) a Graph-Based Autonomous Search that scales the Auto agent for global exploration of the design space. All three paradigms leverage a shared pool of specialized sub-agents to handle planning, implementation, self-debugging, testing, and hardware profiling. We evaluate MaxKernel on JaxBench, a comprehensive suite of 50 diverse kernel tasks for TPUs, alongside complex, real-world workloads from state-of-the-art open-source models. We demonstrate that MaxKernel consistently generates highly optimized implementations, matching expert hand-tuned baselines and delivering significant performance across the benchmark. Our agent is open-sourced and available https://github.com/AI-Hypercomputer/accelerator-agents/tree/main/MaxKernel.
12. 当模型修改过度：关于最小代码编辑的保真度
- 英文标题：When Models Edit Too Much: On the Fidelity of Minimal Code Edits
- arXiv：2609.04061v1
💡 一句话解读：这篇研究要解决的是：AI 在修代码时，除了修好 bug，还会改动太多无关代码，让审查变难。他们提出了一个测试方法，并发现简单的指令能减少过度修改，同时用强化学习训练效果最好。
大型语言模型（LLMs）越来越多地被用于编辑现有代码，但仅正确性是不够的：有用的修复还应该是最小的、可审查的，并且忠实于原始实现。我们研究了过度编辑，即模型在修复 bug 之外重写代码的倾向。我们通过向参考解决方案注入受控的 AST 级损坏，构建了一个包含 400 个 BigCodeBench 问题的评估框架，使每个修复任务都有一个已知的最小补丁。在前沿 LLM 中，过度编辑很普遍，即使在像 GPT-5.5 这样的强模型中也是如此：高 Pass@1 可能与不必要的大编辑和增加的认知复杂性共存。一条保留指令显著减少了这种行为，将平均多余 Levenshtein 距离从 0.195 降至 0.131，将增加的认知复杂性降低了 26.6%，并将 Pass@1 提高了 2.3 个百分点。然而，这些收益并非仅仅来自更大的推理预算或更大的模型。我们接下来询问是否可以在后训练期间直接学习最小编辑。我们观察到，监督微调会过拟合到已见的损坏模式，而强化学习在域外编辑保真度和性能保持之间提供了最佳权衡。这些结果将编辑保真度定位为代码修复质量的一个独立轴，并表明它可以被测量和学习。
📄 English Abstract
Large language models (LLMs) are increasingly used to edit existing code, but correctness alone is not enough: useful repairs should also be minimal, reviewable, and faithful to the original implementation. We study over-editing, the tendency of a model to rewrite code beyond what is required to fix a bug. We construct an evaluation framework from 400 BigCodeBench problems by injecting controlled AST-level corruptions into reference solutions, giving each repair task a known minimal patch. Across frontier LLMs, over-editing is widespread even among strong models like GPT-5.5: high Pass@1 can coexist with unnecessarily large edits and added cognitive complexity. A preservation instruction substantially reduces this behavior, lowering average excess Levenshtein distance from 0.195 to 0.131, reducing added cognitive complexity by 26.6%, and increasing Pass@1 by 2.3 points. However, these gains do not simply follow from a larger reasoning budget or larger models. We next ask whether minimal editing can be learned directly during post-training. We observe that supervised fine-tuning overfits to seen corruption patterns, whereas reinforcement learning gives the best out-of-domain edit-fidelity and performance-retention trade-off. These results position edit fidelity as a distinct axis of code-repair quality and show that it can be measured and learned.
13. 分组自适应裁剪策略优化
- 英文标题：Group Adaptive Clipping Policy Optimization
- arXiv：2609.00444v1
💡 一句话解读：强化学习训练AI时，常用固定裁剪限制更新幅度，但难题的少数正确回答和简单题的大量正确回答被同样对待，导致难题信号被压制。本文提出GAPO方法，根据回答的优劣动态调整裁剪边界，给难题更多更新空间，无需额外奖励设计，在数学和编程任务上提升了准确率。
针对带可验证奖励的强化学习（RLVR），分组相对策略优化（GRPO）通常在所有轨迹上使用固定的重要性采样（IS）比率裁剪边界。我们发现一个关键局限：难题中稀少的正确轨迹与简单题中丰富的正确轨迹被以相近的比率裁剪，尽管它们贡献了截然不同的学习信号。低分组成功率的轨迹具有更大的IS比率，携带更强的梯度信号以促进探索和解决新问题，却常被固定裁剪过度抑制。为解决此问题，我们提出分组自适应裁剪策略优化（GAPO），一种对GRPO方法的即插即用修改，使裁剪边界适应轨迹优势。GAPO受反向KL信任区域视角启发，认为学习信号更大的轨迹应获得相应更大的更新空间。GAPO无需奖励塑形，仅调整裁剪阈值，同时保留标准PPO/GSPO替代目标。在Qwen和Llama模型上，对于基础模型通过率较低的数学推理和编码基准，GAPO在Pass@1和Pass@k上均持续优于固定裁剪和优势塑形基线。
📄 English Abstract
Group relative policy optimization for reinforcement learning with verifiable rewards (RLVR) typically uses a fixed importance-sampling (IS) ratio clipping boundary across all rollouts. We identify a key limitation: rare correct rollouts on harder problems and abundant correct rollouts on easier problems are clipped at comparable rates, despite contributing very different learning signals. Rollouts with low group success exhibit larger IS ratios and carry stronger gradient signal for exploration and solving new problems, yet are disproportionately suppressed by fixed clipping. To address this, we propose Group Adaptive Clipping Policy Optimization (GAPO), a plug-in modification to GRPO methods that adapts the clipping boundary to the rollout advantage. GAPO is motivated by a reverse-KL trust-region perspective, which suggests that rollouts with larger learning signal should receive proportionally greater update headroom. GAPO requires no reward shaping and preserves the standard PPO/GSPO surrogate while adapting only the clipping threshold. Across Qwen and Llama models, GAPO consistently improves both Pass@1 and Pass@k over fixed clipping and advantage-shaping baselines on math reasoning and coding benchmarks where the pass rates by the base model are relatively low.
14. 基于冻结视觉语言模型的免训练、以语音为中心的全模态理解
- 英文标题：Training-Free Speech-Centric Omni Understanding with Frozen VLMs
- arXiv：2609.04242v1
💡 一句话解读：这项研究要解决的是：如何让AI模型同时听懂声音、看懂画面，而不需要花大价钱重新训练整个模型。他们提出的方法是：给现有的视觉语言模型（VLM）加一个“翻译器”，把语音转成文字，再喂给模型，这样模型不用改结构就能理解语音了。这个方法既省钱又有效，还能保留模型原有的视觉和推理能力。
视听理解仍然具有挑战性，因为模型必须联合解释口语内容、视觉事件及其时间关系。现有的全模态模型通常引入专门的音频编码器，并依赖昂贵的音频-视频-文本训练，这将全模态能力与特定的视觉语言模型（VLM）主干紧密耦合，并可能削弱其现有的视觉和推理能力。这引发了三个问题：是否每个新的VLM都需要原生全模态训练，是否可以在保留原始主干的同时添加以语音为中心的全模态能力，以及更丰富的声学表示在何处仍然必不可少。我们引入了免训练全模态（TFO），这是一个即插即用的框架，无需架构修改或多模态重新对齐，即可将冻结的VLM转换为以语音为中心的全模态模型。TFO使用Whisper提取经过置信度过滤、带时间戳的转录文本，并通过VLM现有的语言接口路由它们，同时保持其视觉通路不变。在56个基准测试和21种语言上与原生全模态模型的匹配比较中，TFO在视听理解上具有竞争力，在所有五种模型设置中提高了平均纯音频性能，并实现了显著的多语言语音增益。冻结VLM通常也比相应的原生全模态检查点更能保留更强的图像/视频理解、视觉基础、编码、数学推理和医学问答能力。这些结果表明，强大的以语音为中心的全模态理解通常可以通过模块化的音频到语言路由获得，而不是通过昂贵的特定主干训练。
📄 English Abstract
Audio-visual understanding remains challenging because models must jointly interpret spoken content, visual events, and their temporal relationships. Existing omni models typically introduce dedicated audio encoders and rely on expensive audio-video-text training, tightly coupling omni capability to specific VLM backbones and potentially weakening their existing visual and reasoning abilities. This raises three questions: whether native omni training is necessary for every new VLM, whether speech-centric omni capability can be added while preserving the original backbone, and where richer acoustic representations remain essential. We introduce Training-Free Omni (TFO), a plug-and-play framework that converts a frozen VLM into a speech-centric omni model without architectural modification, or multimodal re-alignment. TFO uses Whisper to extract confidence-filtered, timestamped transcripts and routes them through the VLM’s existing language interface, while leaving its visual pathway unchanged. Across matched comparisons with native omni models on 56 benchmarks and 21 languages, TFO is competitive on audio-visual understanding, improves average audio-only performance across all five model settings, and achieves substantial multilingual speech gains. Freezing the VLM also generally preserves stronger image/video understanding, visual grounding, coding, mathematical reasoning, and medical question answering than the corresponding native omni checkpoints. These results show that strong speech-centric omni understanding can often be obtained through modular audio-to-language routing rather than costly backbone-specific training.
15. τ^τ-Bench：用于端到端、真实智能体构建的环境
- 英文标题：$τ^τ$-Bench: An Environment for End-To-End, Realistic Agent Construction
- arXiv：2609.04611v1
💡 一句话解读：这篇研究要解决的是：现有测试无法判断AI系统能否像人类一样，在真实客户条件下交付一个完整的客服智能体。他们提出了一个叫τ^τ-bench的测试环境，模拟真实业务场景，让AI开发者智能体从零开始构建客服智能体，并用模拟用户来打分。结果发现最强AI也只通过23.9%的测试，而专家能达82.2%，说明AI在深度理解、沟通和实验方面还有很大差距。
LLM智能体正迅速成为生产软件，被部署用于处理客户服务、裁决纠纷和操作内部系统。值得注意的是，构建这些智能体的工作越来越多地交给编码智能体，但现有基准测试很少说明AI系统能否在真实客户参与的条件下交付一个智能体。我们引入了τ^τ-bench（读作hyper-tau-bench），一个将智能体构建作为任务的基准测试。开发者智能体被给予企业实际保存的记录、持有需求的客户、运营必须通过的生产API、要继承的代码库以及服务成本和模型的限制：这与真实参与提供的起点相同。它必须从中交付一个完整的客户服务智能体，通过将该智能体部署到保留的模拟用户上进行评分。在跨越四个领域的53个任务中，最强的配置，即Claude Code下的Claude Opus 5，仅通过了23.9%的评估模拟。同时，专家撰写的参考上限得分为82.2%。失败之处反映了人类智能体开发者所见的：模型发出浅层查询而非深入理解记录，几乎不与客户沟通，并且对智能体架构和服务支出实验太少，直接交付第一个能运行的方案。我们旨在让τ^τ-bench将协作式智能体构建的工作转变为编码智能体的可衡量目标。
📄 English Abstract
LLM agents are rapidly becoming production software, deployed to handle customer service, adjudicate disputes, and operate internal systems. Notably, the work of building them is increasingly handed to coding agents, yet existing benchmarks say little about whether an AI system can deliver one under the conditions of a real client engagement. We introduce $τ^τ$-bench (pronounced hyper-tau-bench), a benchmark that makes agent construction the task. A developer agent is given the records a business actually keeps, a client who holds requirements, a production API that operations must run through, a codebase to inherit, and limits on serving cost and models: the same starting point a real engagement provides. From these it must deliver a complete customer-service agent, scored by deploying that agent against held-out simulated users. Across 53 tasks spanning four domains, the strongest configuration, Claude Opus 5 under Claude Code, passes just 23.9% of evaluation simulations. Meanwhile, an expert-authored reference ceiling scores 82.2%. The failures mirror ones human agent developers see: models issue shallow queries in place of deep comprehension of the records, communicate almost nothing to the client, and experiment too little with agent architecture and serving spend, shipping the first design that runs. We aim for $τ^τ$-bench to turn the work of cooperative agent building into a measurable target for coding agents.