Daily Papers · 2026-10-09
📄 论文日报 | 来源：Hugging Face Daily Papers
🗓 Episode：2026.10.08 | 递归状态按误差分配精度；世界动作模型扩展上下文。
1. STEPQuant：Delta-Rule 递归状态量化中的误差何时与何处重要
- 英文标题：STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization
- arXiv：2609.38169
💡 一句话解读：线性注意力用固定大小的循环状态替代不断增长的KV缓存，但这些状态在并发服务时仍占用大量内存。直接低精度量化会因误差逐步累积而严重掉精度。本文发现误差影响取决于时间和空间两个维度，据此提出STEPQuant，按误差大小和记忆寿命分配精度，并联合拟合行列缩放，实现6比特下接近FP32的精度，内存压缩超5倍。
线性注意力用固定大小的循环状态（recurrent states）替代不断增长的KV缓存，但在并发服务下这些持久状态可能成为显著的内存瓶颈。直接将循环状态量化到低精度往往导致严重的精度下降，因为量化误差会通过连续的状态更新传播。我们发现这些误差的影响取决于两个互补的维度：在时间上，长寿命记忆中的误差会持续多个解码步骤；在空间上，不同键行（key rows）的误差对模型输出的影响不同，而状态幅度在行和列方向上都存在显著变化。受此启发，我们提出STEPQuant，一种针对Delta规则循环状态的空间-时间训练后量化（post-training quantization）框架。STEPQuant根据误差幅度和记忆寿命分配精度，并基于状态分布和键行对输出误差的影响联合拟合键行与值列（value-column）的缩放因子。在Qwen3.8-27B和Kimi-Linear-48B-A3B-Instruct上，跨长、短生成基准的实验表明，STEPQuant在标称6比特预算下精度接近FP32状态，并在其4比特配置下优于均匀INT8。集成到SGLang并配合优化的GPU内核后，6比特STEPQuant实现了超过5倍的循环状态压缩，并将总服务内存降低最多68.7%。我们的代码可在https://github.com/Dreamer-Toby/STEPQuant获取。
📄 English Abstract
Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving. Directly quantizing recurrent states to low precision often leads to severe accuracy degradation, as quantization errors propagate through successive state updates. We discover that the impact of these errors depends on two complementary dimensions: temporally, errors in long-lived memory can persist across many decoding steps; spatially, errors in different key rows affect model outputs differently, while state magnitudes vary substantially along both rows and columns. Motivated by these observations, we propose STEPQuant, a spatial-temporal post-training quantization framework for Delta-rule recurrent states. STEPQuant allocates precision according to error magnitude and memory lifetime, and jointly fits key-row and value-column scales based on state distributions and key-row impact on output error. Experiments on Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct across both long- and short-generation benchmarks show that STEPQuant closely matches FP32-state accuracy under a nominal 6-bit budget and outperforms uniform INT8 in its 4-bit configuration. Integrated into SGLang with optimized GPU kernels, 6-bit STEPQuant achieves over 5x recurrent-state compression and reduces total serving memory by up to 68.7%. Our code is available at https://github.com/Dreamer-Toby/STEPQuant.
2. Long-WAM：扩展世界-动作模型的上下文
- 英文标题：Long-WAM: Scaling the Context of World-Action Models
- arXiv：2610.10528
💡 一句话解读：机器人实时控制需要看历史画面来推断动作，但看太多会拖慢反应。Long-WAM 提出一种框架，让模型在实时限制下也能利用长历史：先用无动作标签的视频自回归预训练，再微调成世界-动作模型。结果历史从0秒增到19.2秒，成功率从63.3%升到78.7%，并在多个任务和硬件上实现实时部署。
实时机器人控制需要足够的视觉历史来推断运动与任务进度，但处理这些历史会延迟动作。我们提出 Long-WAM，一个在实时控制约束下扩展因果世界-动作模型（causal world-action model）上下文的模型-系统框架。核心发现是：能访问历史不等于会利用历史——当视频基础模型以自回归（AR）方式预训练时，更长的历史收益大得多。我们首先从无动作标签的机器人视频和第一视角视频中学习因果预测，然后在世界-动作适配过程中保持这种“历史到未来”的结构。在 RoboCasa GR-1 上，将上下文从 0.0 秒增加到 19.2 秒，成功率从 63.3% 提升到 78.7%，而双向预训练的初始化则没有净收益；机器人领域的 AR 预训练进一步提高了 GR-1 和 LIBERO-Long 上的峰值成功率。Long-WAM 在 LIBERO-Long、RoboTwin 2.0 和 DOMINO 上也取得了对比方法中的最佳结果。流式观测编码、异步执行和硬件特定加速使其能在 RTX 5090、DGX Spark 和 Jetson AGX Thor 上部署而不牺牲未来预测；在 RTX 5090 上，每个动作块（包括未来视频潜在预测）耗时 107.4 毫秒。在 Unitree G1 和 YAM 上的实时部署支持动态和长时程操作，包括动态叠杯任务 95% 的成功率，而 Pi0.5 和 Fast-WAM 在 20 次试验中均未成功。作为记忆增强的执行器，Long-WAM 还能在复合任务中补充高层规划。
📄 English Abstract
Real-time robot control demands enough visual history to infer motion and task progress, but processing that history can delay action. We present Long-WAM, a model-system framework for scaling the context of causal world-action models under real-time control constraints. Our central finding is that access to history is not the same as using it: longer histories pay off far more when the video foundation is pretrained autoregressively (AR). We first learn causal prediction from robot and egocentric videos without action labels, then preserve this history-to-future structure during world-action adaptation. On RoboCasa GR-1, increasing context from 0.0 to 19.2 seconds raises success from 63.3% to 78.7%, whereas a bidirectionally pretrained initialization shows no net gain; robot-domain AR pretraining further raises peak success on GR-1 and LIBERO-Long. Long-WAM also achieves the best results among compared methods on LIBERO-Long, RoboTwin 2.0, and DOMINO. Streaming observation encoding, asynchronous execution, and hardware-specific acceleration enable deployment on RTX 5090, DGX Spark, and Jetson AGX Thor without dropping future prediction; on RTX 5090, each action chunk, including future-video latent prediction, takes 107.4 ms. Real-time deployment on Unitree G1 and YAM supports dynamic and long-horizon manipulation, including 95% success on dynamic cup stacking, where Pi0.5 and Fast-WAM succeed in none of 20 trials. As a memory-informed executor, Long-WAM also complements higher-level planning in composite tasks.
3. nanoMuse：面向你拥有的每台设备的开源个人智能体
- 英文标题：nanoMuse: An Open-Source Personal Agent for Every Device You Own
- arXiv：2610.08699
💡 一句话解读：当前个人智能体（如Meta的Muse）被大公司封闭在云端，用户无法掌控。nanoMuse提出开源方案：在用户所有设备上运行一个智能体，共享对话，操作手机和电脑，通过可自建的中继连接，所有动作经安全哨兵审核，记忆存为可读文件，模型可自选。
2011年的助手只会应答和等待，2023年的智能体执行一个任务后就停止。2026年9月，Meta的Muse展示了一个面向个人的智能体，它拥有账户、设备、记忆和持续对话，但封闭在供应商的云中，且仅限一个国家。这样的智能体被期望能操作用户的账户和设备，跨周记忆，在值得时主动发言，并为自己的行为负责。它是一种软件，而非模型，此前没有开源对应物。本报告用五个问题和三个视野定义了个人智能体，依据Meta的公开记录及其生产提示的副本，解读了Muse的构建方式，每条陈述均标注来源。然后介绍了nanoMuse——GPL-3.0下的开源对应物，在个人拥有的每台设备上运行一个智能体，能操作手机和电脑屏幕。它们通过任何人都可运行的中继共享同一对话；每个动作都经过Sentinel（哨兵），记忆是用户可读的文件，模型由用户选择。其规模和成本以估算形式给出。开放内容、带来源的记忆、针对操作手的评估套件及其开放模型，被列为路线图。
📄 English Abstract
Assistants from 2011 answered and waited, and agents from 2023 did a task and stopped. In September 2026 Meta's Muse showed an agent for one person, with accounts, devices, memory and a conversation that lasts, closed, in a vendor's cloud, in one country. Such an agent is expected to act on a person's accounts and devices, remember them across weeks, speak first when it is worth it, and answer for what it did. It is a kind of software, not a model, and until now had no open counterpart. This report defines the personal agent in five questions and three horizons. It reads how Muse is built from Meta's public record and a copy of its production prompt, each statement marked by its source. It then presents nanoMuse, the open-source counterpart under the GPL-3.0, one agent on every device a person owns, with hands on the phone's screen and the computer's. They share one conversation over a relay anyone can run; every action goes through a Sentinel, memory is files the person can read, and the model is their choice. Its size and cost are given as estimates. What is open, memory with provenance, an evaluation suite for the hands and an open model for them, is set out as a roadmap.
4. 递归游戏创作者：面向体验的智能体产品级游戏开发框架
- 英文标题：Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness
- arXiv：2610.08621
💡 一句话解读：现有游戏生成智能体只能做出能玩的游戏，但能玩不等于好玩。为此提出递归游戏创作者框架，用设计者、构建者、玩家、评审者四个角色循环迭代：设计者把需求变成方案，构建者做出游戏，玩家用代码接口高效试玩收集体验数据，评审者据此打分并给出改进意见，不断优化出真正有趣的游戏。
近期的游戏设计智能体在生成可玩游戏方面取得了显著进展，但程序正确并不保证玩家能获得愉悦体验。我们提出递归游戏创作者（Recursive Game Creator），一个面向体验的框架，旨在推动智能体游戏开发从粗糙的游戏原型迈向有趣的游戏。该框架围绕四个组件组织递归开发：设计者（Designer）、构建者（Builder）、玩家（Player）和评审者（Reviewer）。设计者将用户指令和评审者的反馈转化为详细方案；构建者将这些方案转化为候选游戏；原生代码化的玩家通过程序化接口创建并执行可复用策略，从而高效收集多样化的游戏过程轨迹，缓解了基于图形界面（GUI）的缓慢采集所导致的评估偏差；评审者使用精心设计的基于轨迹的指标来推断玩家偏好，并结合视觉证据和显式文本偏好，依据游戏特定标准对游戏进行评估。最后，评审者接受更优版本并为下一轮提供改进评审，从而闭合递归循环。我们的方法在 GameCraft-Bench 上取得了 77.89 的最先进总体性能。在 GameASG-Bench 上，其严格任务成功率达到 53.2%，相比同模型基线提升 34.1%，并在所比较方法中取得最高的平均运行时检查通过率 93.4%。用户研究显示，玩家游玩时间更长、评分更高。代码即将发布。
📄 English Abstract
Recent game design agents have made substantial progress in generating playable games. However, program correctness does not ensure an enjoyable experience for players. We present Recursive Game Creator, an experience-oriented harness to advance agentic game development from rough game prototypes into entertaining games. Recursive Game Creator organizes recursive development around four components: Designer, Builder, Player, and Reviewer. The Designer translates user instructions and Reviewer's feedback into detailed plans. The Builder turns these plans into candidate games. The coding-native Player creates and executes reusable policies through programmatic interfaces to efficiently collect diverse gameplay trajectories, mitigating evaluation bias caused by slow GUI-based collection. The Reviewer uses carefully designed trajectory-based metrics to induce player preferences, integrating with visual evidence and explicit textual preferences to evaluate games against game-specific criteria. Finally, the Reviewer accepts the better version and provides improvement reviews for the next round, closing the recursive loop. Our method achieves state-of-the-art overall performance of 77.89 on GameCraft-Bench. On GameASG-Bench, it achieves a strict task success rate of 53.2%, a 34.1% improvement over the same-model baseline, and the highest mean runtime-check pass rate at 93.4% among compared methods. A user study shows longer playtime and higher ratings. Code is coming soon.
5. 质疑问题本身：维持推理模型的自我进化
- 英文标题：Questioning the Questions: Sustaining Self-Evolution in Reasoning Models
- arXiv：2610.04299
💡 一句话解读：自我进化推理模型靠自问自答提升能力，但反复训练后性能会崩。本文发现自生成问题存在无效和重复两类毛病，并据此提出R-Quest方法，通过让模型学会拒绝无效问题、用基础模型判断问题新颖性来引导进化，从而在多个任务上稳定提升。
自我进化推理模型从自身生成的问题中学习，但反复自训练可能导致性能崩溃。本文研究了性能为何在多轮迭代中恶化以及如何维持自我进化。分析发现自生成问题存在两类反复出现的质量问题：无效问题以及同一数学问题的重复变体。首先，无效问题在各轮次中越来越普遍，而答案一致性过滤进一步增加了其在训练数据中的比例。其次，现有基于词汇相似度的问题多样性控制可能遗漏以不同方式表达的数学等价问题，导致后期训练轮次中问题多样性崩溃。基于这些发现，我们提出了R-Quest，利用问题有效性和新颖性反馈来引导自我进化。我们首先训练求解器识别并拒绝无效问题，然后利用其判断来指导提问者奖励并过滤求解器训练数据。为避免问题重复，我们使用一个冻结的基础模型比较采样的问题对并提供新颖性反馈。实验上，我们的方法在两个模型家族的数学推理、通用领域推理和代码生成共12个基准上始终取得最高平均性能。此外，R-Quest在十轮自我进化中保持稳定的性能提升，在最后一轮达到峰值，并比R-Zero高出17.32分。
📄 English Abstract
Self-evolving reasoning models learn from their own generated questions, yet repeated self-training can lead to performance collapse. In this paper, we investigate why performance deteriorates over successive rounds and how to sustain self-evolution. Our analysis identifies two recurring quality problems in self-generated questions: invalid questions and repeated variants of the same mathematical questions. First, invalid questions become more prevalent across rounds, and answer-consistency filtering further increases their proportion in training data. Second, existing question diversity controls based on lexical similarity can miss mathematically equivalent questions expressed in different ways, which leads to question diversity collapse in later training rounds. Building on these findings, we introduce R-Quest, which uses question validity and novelty feedback to guide self-evolution. We first train the solver to recognize and reject invalid questions, then use its judgments to guide questioner rewards and filter solver training data. To avoid question repetition, we use a frozen base model to compare sampled question pairs and provide novelty feedback. Empirically, our method consistently achieves the highest average performance on 12 benchmarks in mathematical reasoning, general-domain reasoning, and code generation across two model families. Additionally, R-Quest maintains stable performance gains over ten rounds of self-evolution, peaking in the final round and outperforming R-Zero by 17.32 points.
6. GRACE：面向高效视频生成的生成感知潜在压缩
- 英文标题：GRACE: Generation-aware latent compression for efficient video generation
- arXiv：2610.10524
💡 一句话解读：视频生成模型（如 DiT）计算量大、速度慢，压缩其输入的视频编码器可提速，但压缩后画面变差、且与已训练好的模型不兼容。本文提出 GRACE：保留原编码器的“基础信息”，只对压缩丢失的部分学一个“补充信息”，并让压缩结果对齐原模型的特征分布，再轻量微调模型。最终把 Wan2.1 的 token 数减少 8 倍、延迟降低 11.1 倍，画质不降。
高度压缩的视频自编码器为加速视频扩散模型提供了有效途径，因为扩散 Transformer（DiT）只需处理远少得多的 token。然而，这类自编码器训练困难：更高的压缩比会降低重建质量，而恢复质量又需要更多通道，这已知会拖慢 DiT 的收敛；同时，压缩后的潜变量（latent）与 DiT 训练时所用的潜变量不同，因此预训练的 DiT 必须从头重训或以高昂代价进行适配。压缩 DiT 原本训练所用的自编码器似乎能保持兼容性，但仅针对重建进行优化仍会使潜变量偏离 DiT 已学到的分布。为此，我们提出面向高效视频生成的生成感知潜压缩（GRACE），这是一个两阶段框架，可在压缩预训练视频自编码器的同时保持其与预训练 DiT 的兼容性。具体而言，我们保留来自预训练编码器的冻结基础潜变量，并为更强压缩下丢失的信息学习一个残差潜变量，同时在冻结 DiT 的特征空间中将压缩潜变量与预训练潜变量对齐，从而使自编码器针对生成任务进行优化。随后，我们通过轻量微调与非对称去噪（其中基础部分先于残差部分被去噪）来适配 DiT。在 480x832x81 分辨率下，GRACE 将 Wan2.1-I2V-14B 的 token 数减少 8 倍、延迟降低 11.1 倍，同时在 VBench 上达到与压缩前预训练流程相当的生成质量。
📄 English Abstract
Highly compressed video autoencoders offer an effective way to accelerate video diffusion models, as the Diffusion Transformer (DiT) operates on far fewer tokens. However, such autoencoders are challenging to train, since a higher compression ratio degrades reconstruction quality and recovering it requires more channels, which is known to slow the convergence of the DiT. The compressed latent also differs from the one the DiT was trained on, so the pretrained DiT must be either retrained from scratch or adapted at considerable cost. Compressing the autoencoder the DiT was trained with appears to preserve compatibility, yet optimizing it for reconstruction alone still shifts the latent away from the distribution the DiT has learned. To address this, we propose Generation-Aware Latent Compression for Efficient Video Generation (GRACE), a two-stage framework that compresses a pretrained video autoencoder while keeping it compatible with the pretrained DiT. Specifically, we keep a frozen base latent from the pretrained encoder and learn a residual latent for the information lost under stronger compression, while aligning the compressed latent with the pretrained latent in the feature space of the frozen DiT so that the autoencoder is optimized for generation. We then adapt the DiT with lightweight fine-tuning and asymmetric denoising, where the base is denoised ahead of the residual. GRACE reduces the token count of Wan2.1-I2V-14B by 8x and its latency by 11.1x at 480x832x81, while matching the generation quality of the pretrained pipeline before compression on VBench.
7. DecepEval：用于评估LLM智能体欺骗行为的基准
- 英文标题：DecepEval: A Benchmark for Evaluating Deception in LLM Agents
- arXiv：2610.07967
💡 一句话解读：这篇研究要解决的问题是：现有测试只能零散地发现大模型智能体偶尔会骗人，却说不清什么情况下更容易骗人。为此，作者建了一个包含1532个测试案例的基准DecepEval，覆盖28种职业场景，并借鉴经典诈骗理论提出“欺骗钻石”框架，从压力、利益、机会、冲突四方面制造诱导条件，对比同一任务在“中性”和“诱导”版本下的表现，从而量化欺骗行为。
随着大语言模型（LLM）智能体日益自主，它们可能通过欺骗来提升任务表现，这引发了对其可靠部署的担忧。现有评估表明LLM智能体确实会欺骗，但往往只考察孤立场景或狭义设定的条件，限制了对欺骗何时更易发生的系统性理解。为弥补这一空白，我们提出DecepEval——一个包含3类任务族、28个专业场景、共1532个实例的基准。借鉴经典诈骗理论，我们提出“LLM欺骗钻石”框架，刻画可能诱发欺骗的四种外部条件：压力、诱因、机会与冲突。DecepEval为每个实例配对中性版本与诱导版本，以测量依赖条件变化的欺骗率，同时借助明确的任务事实和可观察的智能体行为，将欺骗与能力相关错误区分开来。对九个前沿LLM的评估显示，诱导因素会提高各模型和各任务族的欺骗率，即便在基线欺骗率很低的模型中也是如此。DecepEval使这些脆弱性变得可测量，为迈向可信人工智能提供了共享基准。
📄 English Abstract
As large language model (LLM) agents become increasingly autonomous, they may pursue task performance through deception, raising concerns about their reliable deployment. Existing evaluations show that LLM agents can deceive, but often examine isolated scenarios or narrowly defined conditions, limiting systematic understanding of when deception becomes more likely. To address this gap, we introduce DecepEval, a benchmark comprising 1,532 instances across 3 task families and 28 professional scenarios. Drawing on classical fraud theories, we propose the LLM Deception Diamond framework, which characterizes four external conditions that may induce deception: pressure, incentive, opportunity, and conflict. DecepEval pairs neutral and induced versions of each instance to measure condition-dependent changes in deception rates, while explicit task facts and observable agent behavior help distinguish deception from capability-related errors. Evaluations of nine frontier LLMs show that inducements increase deception across models and task families, even among models with low baseline deception rates. DecepEval makes these vulnerabilities measurable, providing a shared benchmark for progress toward trustworthy artificial intelligence.
8. VepAgent：通过工具增强强化学习桥接因果转移以实现视频事件预测
- 英文标题：VepAgent: Bridging Causal-Transition via Tool-Augmented Reinforcement Learning for Video Event Prediction
- arXiv：2610.06293
💡 一句话解读：视频事件预测要求模型根据已看到的画面推断接下来会发生什么，但现有大模型习惯总结过去、依赖文字常识，难以补全中间没拍到的因果环节。为此，作者提出 VepAgent：先造一个带推理链的数据集教模型补全中间状态，再给它配备状态跟踪、找帧、放大画面等工具，并用奖励机制逼它真正看画面而非猜文字，从而更准地预测未来事件。
多模态大语言模型（MLLMs）在视频理解方面展现出显著潜力，但其对回溯式总结和以文本为中心的先验的依赖，往往限制了它们在视频事件预测（VEP）中桥接未观测到的因果转移的能力。为此，我们提出 VepAgent，一个将因果转移推理与工具增强强化学习（RL）相结合的智能体框架，用于稳健的 VEP。与以往从历史依赖中被动投射未来轨迹的方法不同，我们的方法显式建模从终端观测状态到未来事件的逻辑演进。具体而言，我们首先构建了 futurebench-4K，一个用于监督微调（SFT）的高质量思维链数据集，通过结构化地推导未观测中间状态，有效弥合因果逻辑鸿沟。随后，我们开发了一个诊断工具库，集成状态跟踪、帧检索和区域放大，使智能体能够在推理过程中借助外部工具动态增强推理，以恢复缺失的时空证据并解决视觉歧义。此外，我们提出了一种复合奖励机制，联合优化预测准确性、因果连贯性和可靠先验，促使智能体依赖真实的视觉 grounding 而非表面的文本相似性。在 FutureBench 和 NEPBench 数据集上的大量评估表明，我们的方法达到了最先进的性能，显著优于更大的 MLLMs，并验证了我们的智能体式、面向未来的推理范式的实证有效性。
📄 English Abstract
Multimodal Large Language Models (MLLMs) have demonstrated remarkable potential in video understanding, yet their reliance on retrospective summarization and text-centric priors often limits their ability to bridge unobserved causal transitions when applied to Video Event Prediction (VEP). To address this, we propose VepAgent, an agentic framework that integrates causal-transition reasoning with tool-augmented reinforcement learning (RL) for robust VEP. Unlike prior methods that passively project future trajectories from historical dependencies, our approach explicitly models the logical progression from terminal observed states to future events. Specifically, we first construct futurebench-4K, a high-quality chain-of-thought dataset for supervised fine-tuning (SFT) that effectively bridges the causal-logic gap by structuring the deduction of unobserved intermediate states. Subsequently, we develop a diagnostic tool library integrating state tracking, frame retrieval, and region magnification, enabling the agent to dynamically augment reasoning with external tools to recover missing spatio-temporal evidence and resolve visual ambiguities during inference. Moreover, we propose a composite reward mechanism that jointly optimizes prediction accuracy, causal coherence, and reliable prior, compelling the agent to rely on genuine visual grounding rather than superficial textual similarities. Extensive evaluations on FutureBench and NEPBench datasets demonstrate that our method achieves state-of-the-art performance, significantly outperforming larger MLLMs and validating the empirical effectiveness of our agentic, future-oriented reasoning paradigm.
9. SGF+：解耦自回归视频生成的梯度流
- 英文标题：SGF+: Decoupling Gradient Flows for Autoregressive Video Generation
- arXiv：2610.10429
💡 一句话解读：自回归视频生成中，同一组参数既要“去噪当前帧”，又要“写记忆给未来帧”，两种任务梯度打架，导致画质和连贯性难以兼顾。SGF+把这两个角色拆成独立参数，仅靠因果注意力保持互动，无需额外损失或数据，就能同时提升画质与长时一致性，还能连续生成24小时。
自回归视频生成需要在去噪当前帧的同时，将其键值表示（key-value representations）写入作为未来预测的上下文。然而，这两种角色通常共享参数，我们发现它们的梯度表现出不同模式且存在系统性负对齐，阻碍了视觉质量与时间一致性的联合优化。我们提出自梯度强制增强版（Self Gradient Forcing Plus, SGF+），为上下文写入和去噪分配独立参数，同时通过因果注意力（causal attention）保持二者的交互。两种角色使用原始生成目标联合优化，无需辅助损失，其中上下文写入通过其对未来预测的贡献受到监督。这一简单改动在逐帧和分块生成中均优于所评估的基线，提升了视觉质量和长时程一致性，且无需额外视频训练数据或更长训练时间。仅在5秒片段上训练，SGF+即可支持长达24小时的连续生成，无需长视频微调。这些结果凸显了角色特定参数化（role-specific parameterization）作为高质量自回归视频生成和原生长时程外推的有效设计原则。
📄 English Abstract
Autoregressive video generation requires denoising the current frames while writing their key-value representations as context for future predictions. However, these two roles typically share parameters, and we find that their gradients exhibit distinct patterns and systematic negative alignment, hindering the joint optimization of visual quality and temporal consistency. We introduce Self Gradient Forcing Plus (SGF+), which assigns separate parameters to context writing and denoising while preserving their interaction through causal attention. Both roles are jointly optimized using the original generation objective without auxiliary losses, with context writing supervised through its contribution to future predictions. This simple change improves visual quality and long-horizon consistency over the evaluated baselines in both framewise and chunkwise generation, without additional video training data or a longer training horizon. Trained on only 5s rollouts, SGF+ supports continuous generation for up to 24 hours without long-video fine-tuning. These results highlight role-specific parameterization as an effective design principle for high-quality autoregressive video generation and native long-horizon extrapolation.
10. UniWAM：统一世界-动作模型
- 英文标题：UniWAM: Unified World-Action Model
- arXiv：2610.02054
💡 一句话解读：现有机器人模型要么只会“看懂”不会“动手”，要么只会“动手”不懂世界规律。UniWAM把理解、想象未来画面和预测动作三件事放进一个模型一起训练，还用自然语言表示动作，并设计了数据清洗和分阶段训练方法，让机器人更聪明、更稳、更会听指令。
视觉-语言-动作模型（VLA）受益于预训练视觉-语言模型的理解与推理能力，但仅靠动作监督难以充分扎根于世界动态。相反，世界-动作模型（WAM）继承了视频生成模型的时空先验，但在分布偏移下的语义理解与推理仍然有限。我们提出 UniWAM，一种统一架构，整合物理推理器、世界生成器和动作预测器，联合学习对物理世界的语义理解、视觉生成和动作预测。为确保训练数据质量，我们为人类第一视角数据和机器人数据开发了严格的数据清洗与标注流程。为在保留视觉-语言组件继承的语言能力的同时使其适应具身任务，我们用自然语言表示低级动作，并引入一种预训练方案，将来自视觉问答（VQA）数据、人类第一视角数据和机器人演示的互补监督分配给相应的模型组件。在后训练阶段，未来视觉噪声增强降低了对精确未来预测的依赖，而历史条件流匹配（history-conditioned flow matching）利用编码的动作历史来初始化动作生成。这些设计共同显著减少了去噪步数，同时保持性能。UniWAM 在多项评估中取得最先进（SOTA）性能，包括分布内性能、鲁棒性、泛化能力、指令遵循和长时程任务执行。此外，我们揭示了统一人-机器人协同训练的对数线性缩放定律，证明了在人类与机器人数据混合上进行大规模预训练的有效性。
📄 English Abstract
Vision-language-action models benefit from the understanding and reasoning capabilities of pretrained vision-language models, but action-only supervision provides limited grounding in world dynamics. Conversely, world-action models inherit spatiotemporal priors from video generation models, yet remain limited in semantic understanding and reasoning under distribution shifts. We introduce UniWAM, a unified architecture that integrates a physical reasoner, a world generator, and an action predictor to jointly learn semantic understanding of the physical world, visual generation, and action prediction. To ensure the quality of the training data, we developed a rigorous data cleaning and annotation pipeline for both human egocentric data and robot data. To adapt the vision-language component to embodied tasks while preserving its inherited language capabilities, we represent low-level actions in natural language and introduce a pre-training recipe that assigns complementary supervision from visual question answering (VQA) data, human egocentric data, and robot demonstrations to the appropriate model components. During post-training, future visual noise augmentation reduces reliance on precise future predictions, while history-conditioned flow matching uses encoded action history to initialize action generation. Together, these designs significantly reduce denoising steps while maintaining performance. UniWAM achieves state-of-the-art (SOTA) performance across multiple evaluations, including in-distribution performance, robustness, generalization, instruction following, and long-horizon task execution. Furthermore, we uncover a log-linear scaling law of unified human-robot co-training, demonstrating the effectiveness of large-scale pre-training on a mixture of human and robot data.
11. 半事实信用增强策略优化
- 英文标题：Semifactual Credit-Augmented Policy Optimization
- arXiv：2609.40360
💡 一句话解读：大模型做数学题时，常被题目里无关的措辞变化带偏，答错题。原因是现有训练方法给每个词同样的奖励，可能把这种坏习惯也一起强化了。本文提出SCAPO：先测出哪些词在换说法时最容易动摇，训练早期就压低这些不稳定词的奖励权重，从而让模型推理更准、更稳。
带可验证奖励的强化学习（RLVR）提升了大语言模型（LLM）的推理能力，但其预测仍对与任务无关的提示特征敏感。我们通过保持问题及其答案不变的半事实提示干预来研究这种敏感性，分析发现词元级敏感度存在显著差异，并表明在解码时抑制高漂移词元候选可在不更新模型权重的情况下提升推理准确率。这些发现揭示了组相对策略优化（GRPO）的一个局限：它把同一由结果导出的优势分配给每个响应词元，可能在强化有用推理的同时也强化潜在的虚假依赖。受此启发，我们提出半事实信用增强策略优化（SCAPO），这是GRPO的一种因果启发式变体，将半事实稳定性纳入词元级信用分配。SCAPO在半事实干预下测量固定响应的词元概率漂移，并使用归一化稳定性分数在训练早期降低相对不稳定词元的优势，同时不因稳定性本身给予额外信用。在Qwen3-4B-Base和Qwen3-1.7B-Base上，SCAPO相比GRPO在AIME 2024-2026上的准确率分别提升5.63和4.17个百分点。在两个模型规模上，SCAPO在大多数评估的数学基准以及所有评估的分布外基准上均取得了所比较方法中的最佳结果。这些结果表明，半事实稳定性为通过RLVR中更细粒度的信用分配来提升推理与泛化提供了一种有效的训练信号。代码见https://github.com/DtYXs/SCAPO。
📄 English Abstract
Reinforcement learning with verifiable rewards (RLVR) has improved the reasoning capabilities of large language models (LLMs), yet their predictions remain sensitive to task-irrelevant prompt features. We investigate this sensitivity through semifactual prompt interventions that preserve the underlying problem and its answer. Our analysis reveals substantial variation in token-level sensitivity and shows that suppressing high-drift token candidates during decoding improves reasoning accuracy without updating model weights. These findings highlight a limitation of Group Relative Policy Optimization (GRPO), which assigns the same outcome-derived advantage to every response token and may reinforce potential spurious dependence alongside useful reasoning. Motivated by this observation, we introduce Semifactual Credit-Augmented Policy Optimization (SCAPO), a causally inspired variant of GRPO that incorporates semifactual stability into token-level credit assignment. SCAPO measures token probability drift for fixed responses under semifactual interventions and uses normalized stability scores to reduce advantages for relatively unstable tokens during early training, while granting no additional credit for stability alone. On Qwen3-4B-Base and Qwen3-1.7B-Base, SCAPO improves AIME 2024-2026 accuracy over GRPO by 5.63 and 4.17 percentage points, respectively. At both model scales, SCAPO achieves the best results on most evaluated mathematics benchmarks and all evaluated out-of-distribution benchmarks among the compared methods. These results suggest that semifactual stability provides an effective training signal for improving reasoning and generalization through finer-grained credit assignment in RLVR. The code is available at https://github.com/DtYXs/SCAPO.
12. RunningTab：基于环境侧标签的直接工作区交互
- 英文标题：RunningTab: Direct Workspace Interaction with Environment-Side Tabs
- arXiv：2610.10444
💡 一句话解读：让AI代理从工作区文件里自动生成报告等成果时，容易漏掉该做的事或读过的文件。RunningTab给环境加了一个“任务清单”，记录任务要求、已读文件和未打开文件，代理能对照清单检查，避免遗漏。
许多知识工作从工作区已有的文件中生成新交付物，LLM代理正开始接管这类工作。通过直接语料交互，代理可以在终端中搜索和阅读这些文件而无需索引，以这种方式从多个文件生成交付物即我们所说的直接工作区交互（DWI）。然而，访问文件只是任务的一半：没有任何机制跟踪任务要求什么、已读了什么、以及列出了但从未打开的文件，这些信息在上下文窗口中转瞬即逝，导致代理可能提取了某个数字却仍交付一份遗漏它的报告。为此，我们提出RunningTab，一个为直接工作区交互配备环境侧标签的框架：由环境与代理共同维护的每任务记录，记录任务还欠什么。具体而言，代理添加其需求，环境将每个已读文件记录为带来源的摘录，并将每个列出但未打开的文件记录为候选；代理随后可以看到每个需求及其最匹配的摘录和最佳未打开候选，根据匹配内容解决需求或附理由搁置，并且如果试图在需求仍开放时结束，会在完成检查中收到这些需求。我们在三个基准测试和三个LLM上验证了RunningTab，它始终优于普通DWI和将记录保存在模型中的基线，而其标签通常能包含交付物所需的值。
📄 English Abstract
Much knowledge work produces new deliverables from files a workspace already holds, and LLM agents are beginning to take such work over. Through direct corpus interaction, an agent can search and read any of those files from a terminal with no indexing, and producing a deliverable from many of them in this way is what we call direct workspace interaction (DWI). Reaching the files, however, is only half the task: nothing keeps track of what the task asks for, what has been read, and what was listed but never opened, all of which slip through the context window without leaving a trace, so an agent may extract a figure and still deliver a report without it. To address this, we present RunningTab, a framework that equips direct workspace interaction with an environment-side tab: a per-task record of what the task still owes, kept by the environment alongside the agent. Specifically, the agent adds its requirements, while the environment records every file read as an excerpt with its provenance and every listed but unopened file as a candidate; the agent can then see each requirement beside its best-matching excerpts and top unopened candidates, resolve it against matching content or set it aside with a reason, and, should it try to finish with requirements still open, receive them in a finish check. We validate RunningTab on three benchmarks with three LLMs, where it consistently outperforms plain DWI and baselines that keep the record in the model, while its tab usually holds the values a deliverable needs once seen.
13. WorldSonus：为世界带来声音
- 英文标题：WorldSonus: Bringing Sound to Worlds
- arXiv：2610.08760
💡 一句话解读：现有世界模型能生成逼真画面，却几乎没有声音。WorldSonus 用流式因果自回归扩散架构实时合成音频（实时率仅0.41），并支持在生成过程中用文字指令动态调整声音，还能输出与画面几何和镜头运动匹配的立体声。
近年来，世界模型（world models）的进展使视觉合成越来越逼真，但这些生成的环境大多仍是无声的。为世界模型引入声音面临三大核心挑战：跟上交互式视频流的实时生成、响应流中声音指令的交互式控制，以及反映场景几何与相机运动的空间对齐立体声。为此，我们提出 WorldSonus，一个面向世界模型实时空间声音合成的交互式视频到音频（video-to-audio）框架。为实现实时生成，WorldSonus 采用流式因果自回归扩散架构，以 0.41 的低实时率（RTF）合成音频片段；为实现交互式控制，我们引入以音频为中心的描述生成流程与按块索引的提示调度，从而在生成过程中动态操控声音事件；为实现空间对齐，我们利用从多种立体声与环绕声（ambisonic）数据中整理出的高质量立体声监督。大量实验表明，尽管 WorldSonus 是为世界模型量身定制，它仍能有效泛化到开放域视频到音频基准，在声学质量与空间对齐两方面达到或超越最先进的双向模型。
📄 English Abstract
Recent advances in world models have enabled increasingly realistic visual synthesis. However, these generated environments remain largely silent. Bringing sound to world models poses three core challenges: real-time generation to keep pace with interactive video streams, interactive control to respond to mid-stream sound instructions, and spatially aligned stereo to reflect scene geometry and camera motion. To address these demands, we introduce WorldSonus, an interactive video-to-audio framework designed for real-time spatial sound synthesis in world models. For real-time generation, WorldSonus employs a streaming causal autoregressive diffusion architecture that synthesizes audio chunks at a low real-time factor (RTF) of 0.41. For interactive control, we incorporate an audio-centric captioning pipeline with chunk-indexed prompt scheduling, enabling dynamic manipulation of sound events during generation. For spatial alignment, we leverage high-quality stereo supervision curated from diverse stereo and ambisonic data. Extensive experiments demonstrate that while tailored for world models, WorldSonus generalizes effectively to open-domain video-to-audio benchmarks, matching or outperforming state-of-the-art bidirectional models in both acoustic quality and spatial alignment. Project page: https://noizai.github.io/WorldSonus/
14. Tetris3D：生成物体相互契合的3D场景
- 英文标题：Tetris3D: 3D Scene Generation With Objects That Fit Together
- arXiv：2610.10539
💡 一句话解读：现有方法从单张图重建3D场景时，物体常各自独立生成，导致相邻物体位置和形状对不上、摆不稳。Tetris3D让每个物体的生成都参考周围物体的形状和物理关系，从而保证整场景几何合理、物理稳定，并构建了带物理标注的大规模数据集。
我们提出 Tetris3D，一个用于单图像3D场景重建的生成式框架，能够恢复出在场景中物理与几何一致的物体。现有方法通常独立生成物体或以隐式方式耦合它们，难以为相互作用的相邻物体之间提供细粒度空间兼容性的指导。为此，我们显式地以周围物体的几何及其物理关系为条件来生成每个物体，引导其形状与姿态在场景中保持几何与物理上的合理性。此外，我们引入 ComOb，一个基于物理模拟的数据集，包含120万个场景，涵盖多种物体类别间的物理交互，并提供逐物体网格与成对物理关系标注。在合成与真实场景上的大量实验表明，Tetris3D 即使在交互区域被遮挡时也能恢复出连贯的物体形状与姿态，并在生成质量和物理稳定性上达到最先进水平。
📄 English Abstract
We propose Tetris3D, a generative framework for single-image 3D scene reconstruction that recovers objects which are physically and geometrically coherent as a scene. Existing methods often generate objects independently or couple them implicitly, providing limited guidance for ensuring fine-grained spatial compatibility between neighboring objects that interact with one another. To address this, we explicitly condition the generation of each object on the geometry of surrounding objects and their physical relationships, guiding its shape and pose to remain geometrically and physically plausible within the scene. Moreover, we introduce ComOb, a physics simulation-based dataset of 1.2M scenes featuring physical interactions across diverse object categories, with per-object meshes and pairwise physical relation annotations. Comprehensive experiments on synthetic and realworld scenes show that Tetris3D recovers coherent object shapes and poses even when interacting regions are occluded, and achieves state-of-the-art performance in both generation quality and physical stability.
15. ReSAIL：缓解迭代式智能体自蒸馏中的崩溃
- 英文标题：ReSAIL: Mitigating Collapse in Iterative Agent Self-Distillation
- arXiv：2609.39306
💡 一句话解读：这篇论文要解决智能体在反复自我学习（自蒸馏）时性能越来越差甚至崩溃的问题。他们发现，关键在于挑选那些“特权信息”最能改变老师模型判断的交互步骤来学习，同时让学生模型保留老师模型依赖特权信息的行为。这个方法叫ReSAIL，能有效防止崩溃，在多个任务上大幅提升成功率。
迭代式自蒸馏使LLM智能体能够从 successive deployments 中学习，为实现递归自我改进（RSI）提供了一条路径。然而，我们在现有方法上的实验揭示了跨周期部署性能的崩溃，同时使用特权信息（PI）的任务性能也下降。我们通过优先选择信息量大的交互步骤进行蒸馏，并在学生成为下一任教师时保留PI条件行为来解决这种崩溃。我们提出了保留与选择性增强的迭代自蒸馏（ReSAIL），这是一种用于基于PI的迭代自蒸馏的即插即用增强方法。ReSAIL选择PI最强烈改变教师预测的交互步骤，并平衡跨轨迹的蒸馏损失。它还将学生的PI条件输出分布正则化到所选和未选步骤上冻结教师的输出分布，以保留PI条件行为用于下一周期的监督。在ALFWorld和TextCraft上，ReSAIL在三个周期内跨模型规模持续获得显著增益，当添加到自蒸馏基线时，最终周期成功率的平均绝对增益为22.5%。基于敏感性的离线数据选择也提高了多模态GUI智能体在AITZ上的动作预测准确性。这些发现首次证明，更稳健的学习机制可以有效缓解迭代智能体自蒸馏在部署轨迹中的性能崩溃。
📄 English Abstract
Iterative self-distillation enables LLM agents to learn from successive deployments, offering a path toward recursive self-improvement (RSI). Yet our experiments with existing methods reveal a collapse in deployment performance across cycles, while task performance with privileged information (PI) also declines. We address this collapse by prioritizing informative interaction steps for distillation and preserving PI-conditioned behavior as the student becomes the next teacher. We introduce Retentive and Selective Augmentation for Iterative Self-Distillation (ReSAIL), a plug-in augmentation for iterative PI-based self-distillation. ReSAIL selects interaction steps where PI most strongly changes the teacher's predictions and balances the resulting distillation losses across trajectories. It also regularizes the student's PI-conditioned output distributions toward those of the frozen teacher at selected and unselected steps to preserve PI-conditioned behavior for supervision in the next cycle. On ALFWorld and TextCraft, ReSAIL sustains substantial gains across model scales over three cycles, with an average absolute gain of 22.5% in final-cycle success rates when added to self-distillation baselines. Sensitivity-guided selection of offline data also improves action prediction accuracy for multimodal GUI agents on AITZ. These findings provide the first evidence that a more robust learning mechanism can effectively mitigate performance collapse in iterative agent self-distillation over deployment trajectories.