Daily Papers · 2026-09-27
📄 论文日报 | 来源：Hugging Face Daily Papers
🗓 Episode：2026.09.25 | 世界模型物理推理可评测；大模型线性叠加双想法
1. 在世界模型中训练客体永久性
- 英文标题：Training Object Permanence in World Models
- arXiv：2609.28654v1
💡 一句话解读：视频生成模型（世界模型）是否像人一样理解“物体被挡住后依然存在”？研究者为此设计了150个认知科学小任务，用Blender生成大量随机变化的样本，构建了包含150万训练样本和300道考题的数据集。他们训练了一个160亿参数的世界模型PWM-WROP，在盲测中表现优异，在延续类模型中排名第一，总排名第三。
客体永久性（object permanence）和固体性是人类认知先验的标志。近期研究表明，视频生成模型——当前世界模型的一类典型代表——已开始展现出涌现的推理能力，使其成为构建类人物理智能的理想候选。视频模型中是否已涌现出客体永久性？若尚未涌现，我们能否用受核心认知启发的数据集来训练它们？我们提出了WROP（World Reasoning with Object Permanence），一个包含150个手工设计的、受认知科学启发的任务的数据基础设施，分为六个认知类别。我们构建了Blender生成器，在保持每个任务认知结构的同时，随机化速度、光照、相机角度及其他干扰参数，每个任务产生超过10,000个样本。我们发布了包含150万样本的训练语料库和一份300题的考试。在该考试上，我们评估了14个视频模型：3个参考到视频（reference-to-video）、7个编辑（edit）和4个延续（continuation）模型，其中包括我们的160亿参数世界模型PWM-WROP。在一项盲法成对Elo研究中，PWM-WROP在延续模型中排名第一，总体排名第三，仅次于两个参考到视频模型之间的统计平局。我们发布了数据、考试、模型答案、分数、权重以及PWM——我们在AWS Trainium2上的原生PyTorch训练栈。
📄 English Abstract
Object permanence and solidity are hallmarks of human cognitive priors. Recent studies show that video generation models, a paradigmatic class of current world models, have begun to show emerged reasoning abilities, making them ideal candidates for building human-like physical intelligence. Do video models have emerged object permanence in them? If not, could we train them with a core-cognition inspired dataset? We introduce WROP (World Reasoning with Object Permanence), a data infrastructure of 150 hand-designed cognitive science inspired tasks, divided into six cognitive categories. We build Blender generators that randomize speed, lighting, camera angle, and other nuisance parameters while preserving each task’s cognitive structure, yielding 10,000+ samples per task. We release a 1.5M-sample training corpus and a 300-question exam. On this exam we evaluate 14 video models: 3 reference-to-video, 7 edit, and 4 continuation, among which PWM-WROP, our 16B world model. In a blind pairwise Elo study, PWM-WROP ranks first among continuation models and third overall, behind only a statistical tie between two reference-to-video models. We release the data, exam, model answers, scores, weights, and PWM, our native-PyTorch training stack on AWS Trainium2.
2. 你的Transformer能同时容纳两个想法：LLM中线性叠加的证据
- 英文标题：Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs
- arXiv：2609.29845v1
💡 一句话解读：大语言模型通常被认为内部计算高度非线性，但本研究要解决的问题是：它能否像线性系统一样，把两个不同文本的输入叠加起来，输出也叠加？作者发现，当把两个文本的输入线性混合时，模型输出的下一个词概率分布也近似是两个单独输出的叠加，并称此为“叠加线性假设”。他们还发现这种线性是Transformer架构自带的，但预训练会削弱它，而轻量微调能恢复，最后用引导解码让一次计算同时生成两个连贯续写。
尽管大语言模型（LLM）依赖高度非线性的组件，本工作证明它们表现出基本的线性：当来自不同文本流的输入被线性组合时，模型输出是各个下一词元（next-token）分布的叠加。我们称之为“叠加线性假设”（Superposition Linearity Hypothesis）。我们提供证据表明，叠加是Transformer架构的内在属性，而非训练带来的涌现结果；事实上，我们观察到它随着预训练进行而趋于减弱。然而，我们证明通过轻量微调可以大幅恢复线性，显著降低预测的下一词元分布与各个下一词元分布平均值之间的差异。最后，我们引入一种引导解码（guided decoding）过程，解耦叠加的输出，从而在一次前向传播中同时生成两个连贯的续写。
📄 English Abstract
While Large Language Models (LLMs) rely on highly non-linear components, in this work we demonstrate that they exhibit fundamental linearity: when inputs from distinct text streams are linearly combined, the model outputs a superposition of the individual next-token distributions. We term this the \textit{Superposition Linearity Hypothesis}. We provide evidence that superposition is an intrinsic property of the Transformer architecture rather than an emergent consequence of training; in fact, we observe that it tends to diminish as pretraining progresses. However, we demonstrate that linearity can be substantially restored through lightweight fine-tuning, significantly reducing the divergence between the predicted next-token distribution and the average of the individual next-token distributions. Finally, we introduce a guided decoding procedure that disentangles superposed outputs, enabling the simultaneous generation of two coherent continuations from a single forward pass.
3. WanPE：面向现代文本到视频生成的电影级提示增强
- 英文标题：WanPE: Towards Cinematic Prompt Enhancement for Modern Text-to-Video Generation
- arXiv：2609.30221v1
💡 一句话解读：用户给视频生成模型写提示词时，往往说不清镜头怎么走、光线怎么变，导致成片效果差。本文提出 WanPE，一个用上百万真实视频训练出的“导演级”提示词增强模型，能把简单需求自动扩写成多镜头、带运镜和光影的电影级方案，并保证不偏离用户原意。
视频生成始于文本空间：先写出电影剧本，再转化为像素。当当代视频生成器扩展到 30 秒并能忠实遵循复杂条件时，文本提示在很大程度上主导了制作过程，规划动作、镜头轨迹、光照与声音如何在多镜头序列中展开。本文提出 WanPE，一个拥有 3970 亿参数、在 105 万条真实视频上训练的提示增强模型，旨在掌握导演级的电影化规划能力。WanPE 通过基于视频的反向构建来生成镜头级电影化方案，并采用语义一致性 GRPO（SC-GRPO）以在跨镜头和长时间跨度上忠实保留用户需求。为评估该能力，我们构建了 WanPEval——一个人工标注的测试平台，覆盖 5 至 30 秒的时长与不同意图粒度，并配有约 1.1 万次盲测成对评估。在驱动 Wan3.0 视频生成器时，WanPE-397B 相比原始用户提示在人类偏好上于 5-15 秒提升 10.66-18.84 分，在 30 秒场景中大幅提升 50.86 分。消融实验表明，反向构建明显优于正向改写，而 SC-GRPO 在不同模型规模下都能稳健保持语义保真度。最终，WanPE 在 5-15 秒区间领先所有受评估的商业方案，并在 30 秒区间与 Seedance 2.5 保持竞争力。
📄 English Abstract
Video generation begins in text space by authoring a cinematic screenplay, then materializes into pixels. As contemporary video generators scale to 30 seconds and faithfully follow complex conditions, the textual prompt largely directs the production, planning how actions, camera trajectories, lighting, and sound unfold across multi-shot sequences. In this paper, we present WanPE, a 397B-parameter prompt enhancement model trained on 1.05M real-world videos to master director-level cinematic planning. WanPE formulates shot-level cinematic plans via video-grounded reverse construction and employs Semantic-Consistency GRPO (SC-GRPO) to faithfully preserve user requirements across shots and over time. To benchmark this capability, we curate WanPEval, a human-annotated testbed covering durations from 5 to 30 seconds across varying intent granularities, supported by approximately 11K blind pairwise assessments. When powering Wan3.0’s video generator, WanPE-397B boosts human preference over raw user prompts by 10.66-18.84 points at 5-15 seconds and by a dramatic 50.86 points in the 30-second arena. Ablation studies show that reverse construction demonstrates clear superiority over forward rewriting, while SC-GRPO robustly preserves semantic fidelity across model scales. Ultimately, WanPE leads all evaluated commercial offerings at 5-15 seconds and remains competitive with Seedance 2.5 at 30 seconds.
4. OmniEcho：面向全模态具身智能体的音视频空间理解
- 英文标题：OmniEcho: Audio-Visual Spatial Understanding for Omni-Modal Embodied Agents
- arXiv：2609.23407v2
💡 一句话解读：这篇论文要解决的是：机器人（具身智能体）很难像人一样判断声音从哪个方向来，并把声音和看到的画面对应起来做推理。为此，作者先建了一个包含真实空间音频的评测基准 OmniEchoBench，又提出一个能同时处理声音和图像、并理解声音方位的模型 OmniEcho，让机器人能靠声音辅助导航。
人类可以毫不费力地定位声源方向，并将其与视觉线索结合进行推理，但这对具身智能体而言仍具挑战。尤其是，如何在具身场景中有效评估和建模空间音频理解仍不清楚。为填补这一空白，我们提出了 OmniEchoBench，一个面向空间音视频感知与音视觉语言导航（audio-vision-language navigation）的统一基准。OmniEchoBench 包含六个任务，覆盖 197 个真实世界空间音视频场景、2,972 组问答对，以及来自 30 个真实环境的 900 个一阶 Ambisonics（FOA）音频导航样本。为实现可扩展的训练监督，我们开发了一套可控的空间音频渲染流程，保持声源、视觉观测与智能体轨迹之间的几何一致性。在此基础上，我们提出 OmniEcho，一个具备空间感知能力的全模态模型，它引入 FOA 空间编码器以及预训练的语义音频通路。大量实验表明，OmniEcho 在空间音视频感知上达到当前最优性能；在声音引导导航任务中，其表现接近传统视觉语言导航的水平。这些结果说明空间音频可作为具身场景推理与导航的有价值信号，同时也凸显了细粒度空间定位与距离估计仍是重要的开放挑战。
📄 English Abstract
Humans can effortlessly localize the direction of a sound source and integrate it with visual cues for reasoning, yet this remains challenging for embodied agents. In particular, it is still unclear how to effectively evaluate and model spatial audio understanding in embodied settings. To address this gap, we introduce \textbf{OmniEchoBench}, a unified benchmark for spatial audio-visual perception and audio-vision-language navigation. OmniEchoBench comprises six tasks over 197 real-world spatial audio-visual scenes, 2,972 question-answer pairs, and 900 navigation samples with first-order ambisonics (FOA) audio collected from 30 real-world environments. To enable scalable training supervision, we develop a controllable rendering pipeline for spatial audio. It preserves geometric consistency among sound sources, visual observations, and agent trajectories. Building on this, we propose \textbf{OmniEcho}, a spatially aware omni-modal model. It introduces an FOA spatial encoder alongside a pretrained semantic audio pathway. Extensive experiments show that OmniEcho achieves state-of-the-art performance on spatial audio-visual perception. For our sound-guided navigation, OmniEcho reaches a performance level close to that of traditional vision-language navigation. These results demonstrate that spatial audio can serve as a valuable signal for embodied scene reasoning and navigation, while also highlighting fine-grained spatial localization and distance estimation as important open challenges. Our code and data will be available in https://github.com/PKU-VaLuE-Lab/OmniEcho/tree/main
5. 智能体编辑世界模型：为LLM智能体重新思考世界建模
- 英文标题：Agent-Editing World Model: Rethinking World Modeling for LLM Agents
- arXiv：2609.28416v1
💡 一句话解读：这篇研究要解决LLM智能体在长任务中因错误假设和过时计划残留而决策跑偏的问题。它提出“智能体编辑世界模型”，不模拟工具反馈，而是判断哪些决策关键、哪些是噪声，并直接修改智能体的内部状态，让后续决策更可靠。
近期大语言模型（LLM）的进展使智能体能够在多样环境中处理长时程任务。为提升智能体表现，现有语言世界模型通常预测环境观测，但在真实反馈可用时，重建高熵且依赖执行的工具响应价值有限。同时，智能体面临任务状态污染，即无依据的假设和过时计划残留在历史中并扭曲后续决策。我们提出智能体编辑世界模型（AEWM），它建模推理与动作如何塑造未来任务进展，而非模拟工具响应。AEWM结合动作评判器（Action Judge）区分关键、探索和噪声决策，并通过状态修正（State Revision）从同一观测历史中编辑噪声推理-动作延续。EditAct将这些能力与真实执行集成，直接改变后续决策所依赖的状态，而不仅仅是提供批评。我们通过中期训练和监督微调在搜索、终端和软件工程领域训练AEWM。AEWM在动作评判基准上达到70.5%的宏F1，超过最强前沿基线10.6分。在六个基准和三个智能体骨干上，EditAct比最强基线平均提升3.2-6.7分。此外，对验证过的EditAct轨迹进行拒绝采样微调（称为AEWM-RFT），在三个领域比Self-RFT提升2.2-2.6分，且无需在线AEWM指导。
📄 English Abstract
Recent advances in large language models (LLMs) have enabled agents to tackle long-horizon tasks across diverse environments. To further improve agent performance, existing language world models typically predict environment observations, yet reconstructing high-entropy, execution-dependent tool responses offers limited value when real feedback is available. Meanwhile, agents suffer from \emph{task-state contamination}, where unsupported assumptions and outdated plans persist in history and distort subsequent decisions. We propose the \textbf{Agent-Editing World Model (AEWM)}, which models how reasoning and actions shape future task progress rather than simulating tool responses. AEWM combines \textbf{Action Judge} to distinguish \textsc{Critical}, \textsc{Exploratory}, and \textsc{Noisy} decisions with \textbf{State Revision} to edit noisy reasoning–action continuations from the same observed history. \textbf{EditAct} integrates these capabilities with real execution, directly changing the state underlying subsequent decisions rather than merely providing critiques. We train AEWM across Search, Terminal, and Software Engineering through mid-training and supervised fine-tuning. AEWM achieves 70.5% macro-F1 on our Action Judge benchmark, exceeding the strongest frontier baseline by 10.6 points. Across six benchmarks and three agent backbones, EditAct improves average scores by 3.2–6.7 points over the strongest baseline. Furthermore, rejection sampling fine-tuning on verified EditAct trajectories, termed \textbf{AEWM-RFT}, improves over Self-RFT by 2.2–2.6 points across three domains without online AEWM guidance.
6. 词性作为SAE潜空间中的涌现类别
- 英文标题：Parts-of-Speech as Emergent Categories in SAE Latent Space
- arXiv：2609.29362v1
💡 一句话解读：这篇研究想弄清：语言模型内部到底怎么表示词性（名词、动词等）这类语法信息。作者用稀疏自编码器（SAE）检查模型内部激活，发现词性信息确实能被还原出来，但并非一个特征对应一个词性，而是由一小群稀疏特征共同、分布式地表示，且不同词性差异很大。
稀疏自编码器（Sparse AutoEncoders, SAEs）为审视语言模型表征提供了一条有前景的途径，但其潜在特征究竟揭示了何种语言结构仍不清楚。我们以词性（part-of-speech, PoS）类别作为受控测试案例，研究形态句法信息是由单个潜在特征编码，还是由结构化的特征群组编码。我们发现，词性区分可从 SAE 激活中高度还原，但并不对应一对一的潜在特征／类别映射。这种可还原性不能归结为词汇记忆，且开放与封闭词性类别差异显著。各类别由紧凑的稀疏潜在特征群组支撑，不同词性标签之间存在明显差异。这些群组在留出数据上保持稳定，同时相关类别之间也表现出重叠。我们的结果表明，SAE 以分布式且依赖类别的方式定位形态句法信息，而非通过原子式的语法特征。
📄 English Abstract
Sparse AutoEncoders (SAEs) offer a promising way to inspect language model representations, but it is still unclear what kind of linguistic structure their latents expose. We use part-of-speech (PoS) categories as a controlled test case to study whether morpho-syntactic information is encoded by individual latents or by structured groups of features. We find that PoS distinctions are highly recoverable from SAE activations, but do not align with one-to-one latent / category mappings. This recoverability is not reducible to lexical memorisation, and Open and Closed PoS classes differ substantially. Categories are supported by compact groups of sparse latents, with substantial variation across tags. These groups remain stable on held-out data, while also showing overlap between related categories. Our results show that SAEs localise morpho-syntactic information in a distributed and category-dependent form rather than through atomic grammatical features.
7. Qwen-Planner-Agent：面向真实世界移动规划智能体的闭环 AI-for-AI 框架
- 英文标题：Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents
- arXiv：2609.29892v1
💡 一句话解读：这篇论文要解决的是：如何让AI自己参与开发下一代AI系统，并应对手机等真实设备上复杂长任务规划难、开发成本高的问题。方法是搭建一个闭环框架，让AI自动造数据、训练模型、部署并收集反馈，再不断改进模型和运行环境。
大语言模型的快速进展正将AI从被动的内容生成扩展到工程与科学发现的主动工作流中。这一转变引出一个引人关注的问题：AI能否既是开发对象，又是构建下一代AI系统的积极参与者？我们在一个闭环的AI-for-AI框架内构建了Qwen-Planner-Agent，以探索可扩展开发与迭代改进。移动规划为该方案提供了严苛的测试：复杂的长时程任务挑战智能体的可靠性，而昂贵的真实设备交互限制了开发的可扩展性。该框架通过共享的动作-反馈-验证契约，将数据生产、模型训练和部署连接起来。(i) AI for Data构建了一个人工把关的智能体数据飞轮，其中专用智能体构造任务、收集交互轨迹、整理和平衡训练数据，并利用训练反馈指导后续数据生成。(ii) AI for Training将监督式规划冷启动与混合环境在线智能体强化学习相结合，其中我们引入能力感知奖励与优势工程（CARE），以在保持任务性能的同时降低推理和工具使用成本。(iii) AI通过执行证据驱动的循环推动模型与运行框架协同演化，该循环在运行时编排记忆、技能和工具，并将结构化的动作反馈和保留的失败轨迹反馈回协调的模型与框架适配中。Qwen-Planner-Agent在MobilePA-Bench上取得了所有评估模型和系统中的最佳综合性能，在工具使用、记忆、技能和子智能体协调方面均优于其基础模型。对我们模型的进一步评估显示，在非移动智能体基准上也取得改进，同时基本保持了通用能力。
📄 English Abstract
The rapid progression of large language models is extending AI from passive content generation into the active workflows of engineering and scientific discovery. This shift raises a compelling question: can AI be both the object of development and an active participant in building next-generation AI systems? We explore this question by building Qwen-Planner-Agent within a closed-loop AI-for-AI framework for scalable development and iterative improvement. Mobile planning offers a demanding test of this approach: complex, long-horizon tasks challenge agent reliability, while costly real-device interaction limits development scalability. The framework connects data production, model training, and deployment through a shared action-feedback-verification contract. (i) AI for Data builds a human-gated agentic data flywheel in which specialized agents construct tasks, collect interaction trajectories, curate and balance training data, and use training feedback to guide subsequent data generation. (ii) AI for Training combines a supervised planning cold start with hybrid-environment online agentic reinforcement learning, where we introduce Competence-Aware Reward-and-Advantage Engineering (CARE) to reduce reasoning and tool-use costs while preserving task performance. (iii) AI drives model–harness co-evolution through an execution-evidence-driven loop that orchestrates memory, skills, and tools at runtime and feeds structured action feedback and preserved failure traces back into coordinated model and harness adaptation. Qwen-Planner-Agent achieves the best overall performance among all evaluated models and systems on MobilePA-Bench, improving over its base model across tool use, memory, skills, and sub-agent coordination. Further evaluations of our model show improvements across non-mobile agentic benchmarks while largely preserving general capabilities.
8. IterSynth：通过角色解耦的迭代合成重新思考深度搜索智能体
- 英文标题：IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
- arXiv：2609.29444v1
💡 一句话解读：现有深度搜索智能体把规划、找证据、写答案全塞给一个模型，还不断堆积搜索历史，导致又乱又吵。IterSynth 把任务拆成“规划者”和“整合者”两个角色轮流干活，并用不断更新的摘要代替冗长历史，还配套了新的强化学习训练方法，让8B小模型在多个搜索测试上超过了同类最强模型。
深度搜索要求大语言模型智能体分解复杂查询、搜索证据并综合出有依据的答案，但现有的ReAct风格智能体存在两个局限：角色耦合，即一个策略必须同时处理规划、证据使用和综合；以及上下文累积，即不断增长的搜索历史引入噪声并掩盖有用信息。为解决这些问题，我们提出IterSynth，一种角色解耦且基于摘要的范式，它在负责识别信息需求的规划器（Planner）和负责将证据整合为不断演化的摘要状态的综合器（Synthesizer）之间交替。该设计将规划与综合分离，同时使用摘要作为搜索的持久状态，从而减少能力耦合和上下文噪声。为了有效训练IterSynth，我们进一步提出角色解耦策略优化（RDPO）用于强化学习，它将终端结果奖励与回合级评分标准评估相结合，并计算角色特定的优势以实现更精确的信用分配。在BrowseComp和Xbench-DS等五个长时程深度搜索基准上的实验表明，IterSynth-8B平均得分为50.7，超过此前最强的≤8B智能体4.2%。此外，IterSynth作为一种模型无关的提示范式，在前沿专有模型上相比ReAct及类似提示范式带来了显著的零样本提升。
📄 English Abstract
Deep search requires LLM agents to decompose complex queries, search for evidence, and synthesize grounded answers, yet existing ReAct-style agents suffer from two limitations: role coupling, where one policy must handle planning, evidence use, and synthesis; and context accumulation, where growing search histories introduce noise and obscure useful information. To address these issues, we propose IterSynth, a role-decoupled and summary-based paradigm that alternates between a Planner for identifying information needs and a Synthesizer for integrating evidence into an evolving summary state. This design separates planning from synthesis while using the summary as the persistent state of search, reducing both capability coupling and context noise. To train IterSynth effectively, we further introduce Role-Decoupled Policy Optimization (RDPO) for reinforcement learning, which combines terminal outcome rewards with turn-level rubric evaluations and computes role-specific advantages for more precise credit assignment. Experiments on five long-horizon deep-search benchmarks such as BrowseComp and Xbench-DS show that IterSynth-8B achieves an average score of 50.7, surpassing the strongest prior $\leq$8B agent by +4.2%. Moreover, IterSynth serves as a model-agnostic prompting paradigm, delivering substantial zero-shot gains over ReAct and similar prompting paradigms on frontier proprietary models.
9. 面向泛化任务与运动规划问题的编码智能体
- 英文标题：Coding Agents for Generalized Task and Motion Planning Problems
- arXiv：2609.30233v1
💡 一句话解读：这篇论文要解决的问题是：机器人做复杂任务时，既要决定先做哪一步（比如先开门还是先拿杯子），又要考虑动作能不能做到（比如手臂够不够长、会不会撞到东西），这种规划很难，而且每换一个新场景就得重新费劲规划。作者想知道，能不能让会写代码的AI智能体自己写一个通用程序，以后遇到新场景直接拿来用。结果发现，这些智能体写出的程序比人类手工设计的规划器成功率更高，而且算得更省。
任务与运动规划（TAMP）问题即使在全可观测和以物体为中心的状态下仍然困难，因为离散决策与几何、运动学和动力学约束紧密耦合。泛化TAMP通过利用问题实例间的规律性来减少新实例上的规划开销，从而应对这一困难，但现有方法需要大量TAMP专用的工程工作。我们研究编码智能体（coding agents）能否通过合成可跨实例泛化的程序来自动化这一过程。给定任务描述和模拟器访问权限，每个智能体在固定的合成预算内自行决定如何与环境交互来开发程序；随后程序被冻结，并在未见过的实例上评估。我们在KinDER和PDDLStream的28个模拟环境中评估了Claude Code（Opus 5）和Codex（GPT-5.6 Sol与GPT-6 Astra），物体数量超出原始基准所评估的范围。在所有程序合成方法中，我们对980个生成程序各在100个留出实例上进行了评估，总计98,000个评估回合。总体而言，我们发现编码智能体在泛化TAMP上出奇地有效：在可提供规划器的16个环境中，三种智能体配置的平均成功率均超过手工设计的规划器、一次性生成以及基于LLM的泛化规划基线（56%至95%，而规划器为47%）。随着物体数量增长，智能体程序的成功率仍高于规划器，且平均每个实例的计算量少一个数量级。日志显示，智能体利用交互来校准物理模型、测试边界情况并改进策略。我们发布了所有代码，包括提供给智能体的完整提示词。这些发现表明，编码智能体是泛化TAMP的一个强基线。
📄 English Abstract
Task and motion planning (TAMP) problems remain difficult even with full observability and object-centric states because discrete decisions are tightly coupled to geometric, kinematic, and dynamic constraints. Generalized TAMP addresses this difficulty by exploiting regularities across problem instances to reduce planning effort on new instances. However, existing methods require substantial TAMP-specific engineering. We investigate whether coding agents can automate this process by synthesizing programs that generalize across instances. Given a task description and simulator access, each agent chooses how to interact with the environment while developing a program within a fixed synthesis budget. The program is then frozen and evaluated on unseen instances. We evaluate Claude Code (Opus 5) and Codex (GPT-5.6 Sol and GPT-6 Astra) on 28 simulated environments from KinDER and PDDLStream, with object counts beyond those evaluated in the original benchmark. Across all program synthesis methods, we evaluate 980 generated programs on 100 held-out instances each, 98,000 evaluation episodes in total. Overall, we find that coding agents are surprisingly effective at generalized TAMP: all three agent configurations outperform hand-engineered planners, one-shot generation, and an LLM-based generalized planning baseline in mean success (56% to 95% versus 47% for the planners, on the 16 environments where a planner is available). As object counts grow, the agents’ programs maintain higher success than the planner, using an order of magnitude less computation per instance on average. Logs show agents using interaction to calibrate physical models, test edge cases, and refine strategies. We release all code, including the full prompts given to the agents. These findings suggest that coding agents are a strong baseline for generalized TAMP.
10. 神经谱容量：仅凭网络规格测量与设计架构
- 英文标题：Neural Spectral Capacity: Measuring and Designing Architectures from Network Specification Alone
- arXiv：2609.23087v1
💡 一句话解读：现有衡量神经网络“大小”的指标只看参数量和计算量，无法区分结构不同的网络。本文提出“神经谱容量”，仅凭网络结构规格就能算出一个分数，并据此快速找到在资源限制下最优的架构。
现代Transformer的设计与压缩本质上都是在预算约束下分配容量。常用的标量指标#Params和#FLOPs能反映规模与计算量，却无法体现架构结构：两个参数量相同但深度-宽度、注意力头或前馈网络分配不同的架构会得到相同分数，但行为却不同。我们提出神经谱容量（NSC），一种基于每个权重矩阵奇异值谱的闭式标量。在标准随机初始化下，Marchenko-Pastur定律使得NSC仅凭架构规格即可计算，无需实例化模型、数据或梯度。其逐层可加结构支持NSC-DP，一种精确动态规划求解器，能在CPU上几秒内返回在资源约束下全局最大化NSC的架构——这是对现有免训练代理指标进行黑盒搜索无法提供的保证。实验上，NSC在七个Transformer和CNN家族中的排序表现优于#Params、#FLOPs及代表性免训练代理指标（在FlexiBERT上，对于参数量差异小于10%的配对，τ=0.505，而#Params降至0.082）；NSC-DP在2秒内于WikiText-103上发现了一个超越人工设计基线的Transformer-XL架构；并将LLaMA-7B修剪为在八个常识推理任务上表现最佳的5.7B模型，无需任何校准数据，比最强的免训练代理基线快约5900倍。
📄 English Abstract
Modern Transformer design and compression both reduce to allocating capacity under a budget. The standard scalars for these decisions, #Params and #FLOPs, capture size and compute but not architectural structure: two architectures with identical parameter budgets but different depth-width, head, or FFN allocations receive identical scores yet behave differently. We propose Neural Spectral Capacity (NSC), a closed-form scalar grounded in the singular-value spectrum of each weight matrix. Under standard random initialization, the Marchenko-Pastur law renders NSC computable from the architectural specification alone, with no model instantiation, data, or gradients. Its layer-wise additive structure admits NSC-DP, an exact dynamic-programming solver returning the architecture globally maximizing NSC under resource constraints in seconds on a CPU – a guarantee that black-box search over existing training-free proxies cannot provide. Empirically, NSC outperforms #Params, #FLOPs, and representative training-free proxies in ranking across seven Transformer and CNN families (on FlexiBERT, $τ= 0.505$ on pairs differing in #Params by less than 10%, where #Params collapses to 0.082); NSC-DP discovers a Transformer-XL architecture on WikiText-103 that beats the human-designed baseline in 2 seconds; and prunes LLaMA-7B to the best 5.7B model across eight commonsense reasoning tasks without any calibration data, about 5900x faster than the strongest training-free proxy baseline.
11. AgentKernel：信任原生的智能体操作系统
- 英文标题：AgentKernel: The Trust-Native Agentic Operating System
- arXiv：2609.29647v1
💡 一句话解读：现在的 AI 智能体常把不可信的外部内容和高权限指令混在一起处理，还会记住中间信息、调用敏感工具，导致恶意内容能混进模型输入并触发危险操作。而现有的安全防护只是应用层插件，和智能体共享同一进程，容易被绕过。本文提出 AgentKernel，把身份、输入、记忆和执行四道防线做成操作系统级、无法绕过的强制服务，让安全贯穿智能体整个生命周期。
现代 AI 智能体经常跨越信任边界：它们摄入不可信内容，将其与特权指令结合，把中间信念持久化到长期记忆中，并调用特权工具。这形成了一个攻击面，恶意载荷可经由模型输入进入并引发有害的工具操作。然而，当前的治理体系仍停留在应用层中间件，与被监控的智能体共享同一进程信任边界。我们认为，智能体需要一个操作系统级基座，为身份、输入中介、记忆治理和执行控制提供强制性、不可绕过的服务。我们提出 AgentKernel，一个信任原生的智能体操作系统，其构建前提是安全必须成为一等设计约束。AgentKernel 将智能体生命周期包裹在强制执行的边界内，该边界由四大支柱组成：身份（Identity）、感知（Perception）、认知（Cognition）和执行（Execution）。每个支柱都将经典的操作系统安全原则适配到语义层面的失效问题，包括委托滥用、提示注入、记忆投毒和工具误用。AgentKernel 将结构性安全视为能力倍增器：内核管理的身份支持可信的跨组织协作；分级感知取代脆弱的单点过滤器；信息流受控的记忆在限制投毒的同时提升检索保真度；语义到内核的执行机制允许在不可绕过的边界之后授予更广泛的工具权限。我们将 AgentKernel 定位为编排框架、智能体运行时、治理平台和执行沙箱之下缺失的操作系统层，并通过系统性比较与安全分析表明，单一集成架构如何能在整个智能体生命周期内强制执行安全。
📄 English Abstract
Modern AI agents routinely cross trust boundaries: they ingest untrusted content, combine it with privileged instructions, persist intermediate beliefs in long-term memory, and invoke privileged tools. This creates an attack surface in which malicious payloads can enter through model inputs and cause harmful tool actions. Yet current governance stacks remain application-level middleware that share a process trust boundary with the agents they monitor. We argue that agents need an operating-system substrate providing mandatory, non-bypassable services for identity, input mediation, memory governance, and execution control. We introduce AgentKernel, a trust-native agent operating system built around the premise that security must be a first-class design constraint. AgentKernel wraps the agent lifecycle in a mandatory enforcement boundary organized into four pillars: Identity, Perception, Cognition, and Execution. Each pillar adapts classical OS security principles to failures at the semantic plane, including delegation abuse, prompt injection, memory poisoning, and tool misuse. AgentKernel treats structural security as a capability multiplier. Kernel-managed identity supports trustworthy cross-organization collaboration; graduated perception replaces brittle single-point filters; information-flow-controlled memory improves retrieval fidelity while limiting poisoning; and semantic-to-kernel enforcement permits broader tool privileges behind a non-bypassable boundary. We position AgentKernel as the missing OS layer beneath orchestration frameworks, agent runtimes, governance platforms, and execution sandboxes, and use systematic comparison and security analysis to show how a single integrated architecture can enforce security across the full agent lifecycle.
12. Rufus-Air：开放的大语言模型后训练配方
- 英文标题：Rufus-Air: An Open LLM Post-Training Recipe
- arXiv：2609.29421v1
💡 一句话解读：这篇研究要解决的问题是：如何公开、可复现地训练一个强大的大语言模型，让其他研究者也能照着做。它提出了一套名为 Rufus-Air 的八阶段训练配方，从基础对话到推理、编程、搜索等能力逐步提升，并公开了数据、奖励设计和训练细节，最终效果超过官方版本。
Rufus-Air 是一个在 GLM-4.5-Air-Base（106B-A12B）上开放且可复现的后训练配方，由八个阶段串行组成：SFT、推理强化学习、编程强化学习、指令遵循强化学习、通用智能体、编程智能体、搜索智能体和 RLHF。我们记录了复现该配方所需的数据、奖励设计、基础设施、阶段顺序及各阶段结果。各阶段从基础能力推进到高级能力，并从硬性可验证奖励过渡到较软的基于评判模型的信号。训练基于开源组件和公开数据，其中大部分按原样使用，无需新的人工标注或内部蒸馏教师模型。我们的主要发现是：(i) 多样且高质量的 SFT 奠定了坚实的能力基础；(ii) 难度过滤使强化学习提示保持在有效的学习范围内；(iii) 奖励可靠性为阶段排序提供了实用原则；(iv) 基础设施和工程选择是配方的一部分，而不仅仅是实现细节。Rufus-Air 优于官方 GLM-4.5-Air 后训练版本，并与同等规模的开源模型相比具有竞争力。
📄 English Abstract
Rufus-Air is an open and reproducible post-training recipe on GLM-4.5-Air-Base (106B-A12B), organized as a serial pipeline of eight stages: SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF. We document the data, reward design, infrastructure, stage order, and stagewise results needed to reproduce the recipe. Stages progress from basic to advanced capabilities and from hard, verifiable rewards to softer judge-based signals. Training builds on open-source components and public data, much of it used as released, without new human annotation or an in-house distillation teacher. Our main findings are that (i) diverse, high-quality SFT establishes a strong capability floor; (ii) difficulty filtering keeps RL prompts within a productive learning range; (iii) reward reliability provides a practical principle for ordering stages; and (iv) infrastructure and engineering choices are part of the recipe, not just an implementation detail. Rufus-Air improves over the official GLM-4.5-Air post-trained release and is competitive with similarly sized open models.
13. PUBG Ally：作为AI队友的对话式具身智能体
- 英文标题：PUBG Ally: A Conversational Embodied Agent as an AI Teammate
- arXiv：2609.29837v1
💡 一句话解读：这篇研究要解决的是：如何让AI在《绝地求生》里像真人队友一样，既能听懂玩家说话、又能实时操作角色一起打游戏。方法是把语言模型和游戏控制结合起来，让AI负责思考和说话，再指挥一个快速控制层完成移动、战斗等动作，并用近3.9万局真实玩家对局数据来训练和评估。
我们介绍了 PUBG Ally，一个用于《PUBG：绝地求生》的具身智能体（embodied agent），它能够推理、自主行动，并作为支持语音的队友与玩家并肩作战。构建这样的队友需要结合两项困难能力：它必须在严格的延迟约束下感知并响应不断变化的游戏世界，同时与玩家自然互动，并使其语音与动作保持同步。因此，Ally 将智能体工具使用（agentic tool use）与实时游戏控制相结合。一个语言模型智能体通过受控接口检查游戏信息、理解玩家语音、维护上下文、决定说什么，并发出高层动作选择，以驱动一个更快的控制层进行移动、战斗和恢复。由于玩家和 Ally 的语音与动作会持续相互影响并改变比赛进程，训练需要来自实际对局的数据。因此，我们收集了近 3.9 万场真实玩家与 Ally 同场游玩的对局数据，记录游戏过程、玩家语音、智能体决策、工具使用、动作和玩家反馈，并用这些记录进行迭代训练。为评估队友质量，我们使用玩家反馈和偏好比较来识别离线评估与玩家偏好之间的差距，并迭代改进评估标准。将 Ally 部署到在线服务中，还需要低延迟的设备端执行以及面向玩家交流的安全保障，我们通过模型压缩、上下文压缩、针对性安全训练、运行时护栏和记忆脱敏来解决这些问题。在在线服务期间，我们调查了 141 个国家的玩家。在游戏记录确认与 Ally 同场游玩的受访者中，当被问及是否会推荐 Ally 时，正面回应比负面回应高出 25.1 个百分点，玩家们不仅将 Ally 描述为工具，还将其描述为队友或伙伴。
📄 English Abstract
We introduce PUBG Ally, an embodied agent for PUBG: BATTLEGROUNDS that can reason, act autonomously, and play alongside players as a voice-enabled teammate. Building such a teammate requires combining two difficult capabilities: it must perceive and respond to a constantly changing game world under strict latency constraints while interacting naturally with players, keeping its speech synchronized with its actions. Ally therefore combines agentic tool use with real-time game control. A language-model agent uses a controlled interface to inspect game information, interpret player speech, maintain context, decide what to say, and issue high-level action choices that steer a faster control layer for movement, combat, and recovery. Because the player’s and Ally’s speech and actions continually shape each other and the course of the match, training requires data from actual gameplay. We therefore collect data across nearly 39k sessions in which real players play alongside Ally, recording gameplay, player speech, agent decisions, tool use, actions, and player feedback, and use these records for iterative training. To evaluate teammate quality, we use player feedback and preference comparisons to identify gaps between offline evaluations and player preferences, and iteratively refine the evaluation criteria. Deploying Ally in live service further requires low-latency on-device execution and safeguards for player-facing communication, which we address through model compression, context compaction, targeted safety training, runtime guardrails, and memory redaction. During the live service, we surveyed players in 141 countries. Among respondents whose play with Ally was confirmed in game records, positive responses exceeded negative responses by 25.1 percentage points when asked whether they would recommend Ally, with players describing Ally not only as a tool but also as a teammate or companion.
14. 世界动作智能体：利用视觉语言模型通过世界动作预演实现机器人操作
- 英文标题：World Action Agent: Harnessing VLMs for Robot Manipulation via World Action Rehearsal
- arXiv：2609.29964v1
💡 一句话解读：现有视觉语言模型虽知识广博，却只能间接帮机器人（如预测限制或写程序），或只给机器人看画面而非可动手操作的世界。本文提出“世界动作智能体”，让模型在可视化的动作工作区里直接驾驶机器人：自动选接触视角、预演并修改每个动作、在视野内修正偏差，还能从视频和教学中学习技能。
通用视觉语言模型（VLM）为机器人操作带来了广泛知识和空间推理能力，但现有系统要么间接使用它们来预测约束或编写程序，要么只让它们观察场景，而非进入一个可行动的世界。我们提出世界动作智能体（WAA），一种多智能体框架，让 VLM 借助基本工具驾驶机器人，并在视觉动作工作区内做出每一个决策。该工作区具有三个特性：接触视图根据场景几何自动选取，呈现当前交互周围的场景；动作预演将每个动作转化为可编辑的提案，智能体可独自或通过想象智能体根据规划反馈进行预览和修改，然后再执行；视图内修正则闭合了观察、预演与底层执行之间的回路，让智能体在观察到残差偏移的视图中将其消除。通过同一工作区，WAA 以两种方式获取具身程序性知识：它在基于证据的审查下从专家视频和人类教学中演化出多模态技能，并通过技能智能体进行查询；其交互轨迹则用于训练更小的 VLM 来驾驶同一框架。在 LIBERO-Pro 上，仅从 LIBERO-90 演化出技能的 WAA 达到了 75.6% 的平均成功率，为当前最优，超越了端到端 VLA、代码即策略智能体以及使用相同骨干网络的视觉框架基线；这些技能无需进一步学习即可在 robosuite 上保持有效。在框架轨迹上微调 Qwen3.5-9B，可将其域外成功率从 1.7% 提升至 43.3%。
📄 English Abstract
General-purpose vision-language models (VLMs) bring broad knowledge and spatial reasoning to robot manipulation, yet existing systems either use them indirectly, to predict constraints or write programs, or give them a view of the scene rather than a world in which to act. We present World Action Agent (WAA), a multi-agent harness through which VLMs pilot robots with basic tools, making every decision within a visual action workspace. The workspace has three properties. Contact views, selected automatically from the scene geometry, present the scene around the current interaction. Action rehearsal turns each action into an editable proposal that the agent, alone or through an Imagination Agent, previews and revises against planning feedback before execution. In-view correction closes the loop between observation, rehearsal, and low-level execution, letting the agent remove residual offsets in the view where it observes them. Through the same workspace, WAA acquires embodied procedural knowledge in two ways: it evolves multimodal skills from expert videos and human teaching under evidence-based review and consults them through a Skill Agent, and its interaction traces train smaller VLMs to pilot the same harness. On LIBERO-Pro, WAA with skills evolved only from LIBERO-90 reaches a state-of-the-art 75.6% average success, outperforming end-to-end VLAs, code-as-policy agents, and a visual-harness baseline with the same backbone; the same skills remain effective on robosuite without further learning. Fine-tuning Qwen3.5-9B on harness traces raises its out-of-domain success from 1.7% to 43.3%.
15. ExplorationBench：测量 AI 系统在可验证异星世界中的探索能力
- 英文标题：ExplorationBench: Measuring AI Systems’ Exploration in Verifiable Alien Worlds
- arXiv：2609.30199v1
💡 一句话解读：科学发现难在未知领域，AI 需要自己探索。但评估这种能力很难：新假设对不对？是探索来的还是背出来的？本文提出 ExplorationBench，用规则可执行、与常识冲突的“异星世界”来测试，让 AI 无法靠记忆，只能真探索。
科学发现始于已知问题的边界之外。在那里，AI 系统必须进行探索：提出假设、设计实验并基于结果迭代。然而，评估这种能力很困难：（1）如何验证一个真正新的假设是否成立；（2）如何判断系统是通过探索发现了它，还是仅仅从预训练数据中回忆了相关知识。为此，我们提出了 ExplorationBench，它将评估科学探索这一棘手问题转化为一个具体且可处理的框架，该框架建立在可验证的异星世界（Alien Worlds）之上：其规则可执行，因此每个答案都能被精确检查；同时它们与熟悉的知识相冲突，因此仅靠回忆无法解决任务。该基准包含两个沙盒：AlienCode（31 个发现目标，70 个任务）和 AlienLogic（24 个发现目标，70 个任务）。每个沙盒提供一份有缺陷的手册、针对任务的环境反馈以及专用的工具调用模式。系统利用这些资源探索沙盒，然后解决留出的任务。我们评估了 10 个 AI 系统，发现最强的系统能够获取并应用不熟悉的规则，而性能在不同轨迹间差异很大，持续探索可能停滞或逆转先前的进展。ExplorationBench 朝着能够通过探索在未知环境中获取并应用真正新知识的 AI 系统迈出了一步。
📄 English Abstract
Scientific discovery begins where known problems end. There, AI systems must engage in exploration: framing hypotheses, designing experiments, and iterating on the results. However, evaluating this ability is difficult: (1) how to verify whether a genuinely new hypothesis holds, and (2) how to determine whether a system has discovered it through exploration or merely recalled related knowledge from pre-training data. To this end, we introduce ExplorationBench, which turns the wicked problem of evaluating scientific exploration into a concrete and tractable framework built on verifiable Alien Worlds: their rules are executable, so every answer can be checked exactly, and they conflict with familiar knowledge, so recall alone cannot solve the tasks. The benchmark contains two sandboxes, AlienCode (31 discovery targets, 70 tasks) and AlienLogic (24 discovery targets, 70 tasks). Each sandbox provides a flawed manual, task-specific environmental feedback, and a dedicated tool-call schema. Systems use these resources to explore the sandbox, then solve held-out tasks. We evaluate 10 AI systems and find that the strongest systems can acquire and apply unfamiliar rules, while performance varies substantially across trajectories and continued exploration can stall or reverse earlier gains. ExplorationBench represents a step towards AI systems that can acquire and apply genuinely new knowledge through exploration in unknown environments.