Daily Papers · 2026-09-18
📄 论文日报 | 来源：Hugging Face Daily Papers
🗓 Episode：2026.09.17 | 科学代码库转智能体环境；上下文机制网络赋能结构化数据智能
1. ScienceIDE：将全球科学代码库转化为智能体可学习环境
- 英文标题：ScienceIDE: Turning World’s Scientific Codebase into Agent Learnable Environments
- arXiv：2609.19134v1
💡 一句话解读：科学代码库虽蕴含大量知识，但工具链零散、领域规则隐晦，导致AI难以从中学习，这就是“科学经验瓶颈”。ScienceIDE把全球科学代码转成可编程环境，让智能体按专家标准生成任务、执行并验证，再用这些经验训练出PhAI-IDE系列模型，在代码修复等任务上取得提升。
科学代码库以可执行模型、方法和工具的形式编码了数十年的人类知识。然而，碎片化的工具链、隐性的领域约定以及专门化的正确性标准，使这些知识难以转化为可靠的学习经验——我们将这一挑战称为“科学经验瓶颈”。我们提出ScienceIDE，这是一种将全球科学代码转化为科学智能体可编程环境的基础设施。在专家定义的科学案例与验收标准指导下，智能体将代码库转化为可执行环境，支持任务生成、执行与科学验证。这些环境为监督微调、强化学习和评估提供了共享基础。利用经过验证的交互轨迹，我们训练了PhAI-IDE-72B、PhAI-IDE-9B和PhAI-IDE-4B。该模型系列在留出的科学代码修复任务以及部分通用基准（涵盖代码、推理和知识）上均取得提升，为科学经验向更广泛能力的正向迁移提供了证据。ScienceIDE为智能体学习与科学实践的一体化工作空间奠定了基础，使人类的科学软件成为发展科学智能的共享基底。代码：https://github.com/aitofound/ScienceIDE
📄 English Abstract
Scientific code repositories encode decades of human knowledge in executable models, methods, and tools. Yet fragmented toolchains, implicit domain conventions, and specialized correctness criteria make this knowledge difficult to convert into reliable learning experience-a challenge we call the scientific experience bottleneck. We introduce ScienceIDE, infrastructure for turning the world’s scientific code into programmable environments for scientific agents. Guided by expert-defined scientific cases and acceptance criteria, agents transform repositories into executable environments that support task generation, execution, and scientific verification. These environments provide a shared foundation for supervised fine-tuning, reinforcement learning, and evaluation. Using verified interaction trajectories, we train PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B. The model family shows gains in held-out scientific-code repair and across selected general-purpose benchmarks in code, reasoning, and knowledge, providing evidence of positive transfer from scientific experience to broader capabilities. ScienceIDE lays the foundation for an integrated workspace for agent learning and scientific practice, making humanity’s scientific software a shared substrate for developing scientific intelligence. Code: https://github.com/aitofound/ScienceIDE
2. LimiX-2：面向通用结构化数据智能的上下文机制网络
- 英文标题：LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence
- arXiv：2609.17488v1
💡 一句话解读：现有表格数据预测模型只盯着“根据已知数据猜答案”，忽略了数据背后的生成机制。LimiX-2 改用“联合建模”思路，同时学习数据本身和答案如何产生，从而在多个测试中超越现有模型，还能自动发现变量间的因果关系。
我们推出 LimiX-2，这是 LimiX 系列的新模型，基于此前建立的缩放定律，通过模型和数据缩放开发而成。LimiX-2 采用上下文机制网络（CMNs）范式，并以上下文条件掩码建模（CCMM）进行预训练。CMNs 将上下文学习的组织原则从以目标为中心的预测转变为面向机制的联合建模：它不再像传统表格 PFN 那样围绕 p(y | x, D_context) 目标构建网络，而是围绕学习 p(x, y | D_context) 设计，即对数据生成背后的联合结构进行上下文相关的表示。预训练使用由结构因果模型（SCMs）生成的合成数据集，涵盖多样的图结构、功能机制和观测过程。在 TabArena、TALENT 和 BCCO 上的评估表明，LimiX-2 优于当前针对特定数据集的模型和表格基础模型。除了预测性能，CMN 范式还促进了 LimiX-2 的因果意识：其特征注意力编码了直接因果关系，从而能够准确恢复因果骨架。
📄 English Abstract
We introduce LimiX-2, a new model in the LimiX family, developed through model and data scaling guided by our previously established scaling laws. LimiX-2 adopts the Contextual Mechanism Networks (CMNs) paradigm and is pretrained with Context-Conditional Masked Modeling (CCMM). CMNs shifts the organizing principle of in-context learning from target-centric prediction to mechanism-oriented joint modeling. Rather than centering the network on the $p(y \mid x, D_{\mathrm{context}})$ objective of conventional tabular PFNs, it is designed around learning $p(x, y \mid D_{\mathrm{context}})$, a context-dependent representation of the joint structure underlying data generation. Pretraining uses synthetic datasets generated by structural causal models (SCMs) spanning diverse graph structures, functional mechanisms, and observation processes. Evaluations on TabArena, TALENT, and BCCO show that LimiX-2 outperforms current dataset-specific models and tabular foundation models. Beyond predictive performance, the CMN paradigm also promotes causal awareness in LimiX-2: its feature attention encodes direct causal relationships, enabling accurate causal skeleton recovery.
3. 重新思考PPO中的评论家学习：理解与缓解价值平坦化
- 英文标题：Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening
- arXiv：2609.18708v1
💡 一句话解读：这篇论文要解决强化学习训练大语言模型时，PPO 算法中的评论家（critic）出现的一个问题：它给中间状态打的分几乎一样平，无法反映真实价值变化，作者称之为“价值平坦化”。他们发现原因是损失函数隐含的方差惩罚和相邻状态更新冗余，于是提出 SP³O 方法，每次只挑少数几个相隔较远的状态来训练评论家，从而缓解平坦化并提升策略效果。
在大语言模型的强化学习中，近端策略优化（PPO）通常使用评论家（critic）来估计状态价值并降低策略更新的方差。然而，我们发现 PPO 评论家中存在一种系统性失效模式，称之为价值平坦化（Value Flattening）：由多次蒙特卡洛延续估计得到的状态价值在中间状态间变化剧烈，而评论家的预测却相对平坦。我们在受控的 FrozenLake 环境中进一步观察到这一现象，并发现随着状态空间增大它变得更加明显。我们的理论与实证分析将价值平坦化与评论家损失中的隐式方差惩罚以及时间相关状态因梯度相似而导致的冗余更新联系起来。基于这些发现，我们提出了稀疏近端策略优化（SParse Proximal Policy Optimization, SP³O），它在每个回复中仅对少数几个分隔良好的状态应用价值损失，以缓解上述两种效应。在 Qwen3-Base 上的实验表明，每个回复仅监督三个状态的 SP³O 就能缓解价值平坦化，并在不同模型规模和评估套件上持续改进学习到的策略。总之，我们的结果将价值平坦化识别为标准 PPO 中评论家学习的一个重要但被忽视的失效模式，并表明一种简单的稀疏监督策略可以缓解它。
📄 English Abstract
In reinforcement learning for large language models, Proximal Policy Optimization (PPO) commonly uses a critic to estimate state values and reduce the variance of policy updates. However, we uncover a systematic failure mode in PPO critics, which we call Value Flattening: state values, estimated from multiple Monte Carlo continuations, change sharply across intermediate states while critic predictions remain comparatively flat. We further observe this phenomenon in a controlled FrozenLake environment and find that it becomes more pronounced as the state space grows. Our theoretical and empirical analyses relate Value Flattening to an implicit variance penalty in the critic loss and redundant updates from temporally correlated states with similar gradients. Motivated by these findings, we introduce SParse Proximal Policy Optimization (SP$^3$O), which applies the value loss to only a few well-separated states in each response to mitigate both effects. Experiments on Qwen3-Base show that SP$^3$O with only three states supervised per response can mitigate Value Flattening and consistently improve the learned policy across model sizes and evaluation suites. Together, our results identify Value Flattening as an important yet overlooked failure mode of critic learning in standard PPO and show that a simple sparse supervision strategy can mitigate it.
4. 置信度源于经验：从推理到智能体的经验性置信度估计
- 英文标题：Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents
- arXiv：2609.17708v1
💡 一句话解读：现有方法只根据模型当前这一次的推理过程来判断答案可不可靠，作者认为这不够。他们提出XConf：让模型回忆过去做过的类似任务、当时有多自信、结果对不对，再据此重新给出置信度。在多个测试中，它比多次采样更准、更省，还能帮模型在没把握时选择不回答。
可靠的置信度估计（confidence estimation）对于语言模型的可信部署日益关键：一个经过校准的、表示输出正确的概率的估计，决定了哪些结果可以交付、哪些需要上报、哪些需要重试。然而，现有的置信度估计器都有一个共同的设计前提：它们只读取当前的推理过程，无论是通过内省该过程、对其词元概率（token probabilities）打分，还是对其重新采样。我们认为，当前这一次推理不足以作为置信度的依据。我们提出 XConf（eXperiential Confidence，经验性置信度）：将置信度与模型积累的经验一起估计。经验以模型自身过往分级经历（episodes）的记录形式存储，每条记录包含任务、模型的反思、其声明的置信度、结果，以及评分到达后写下的一条教训。给定一个新任务，XConf 的 Recall 阶段会检索在相似任务上、且当时声明的置信度也相似的过往经历，并读取它们的历史成功率；其 Reflect 阶段则把这份记录展示给模型，让它说出自己反复出现的失败模式，并在自身历史战绩的参考下重新给出置信度。我们的估计器与格式无关，不需要访问 logit 或更新权重，且只花费一次答案生成的成本。在涵盖推理、编程、多模态问答和交互式智能体的九个基准测试、以及来自三个模型家族的四个模型上，XConf 在 24 次比较中的 23 次里，在区分能力（AUROC）上优于或持平于十次采样的自洽性（self-consistency），同时校准误差（ECE）低得多，而生成成本仅为其十分之一。在用于选择性预测（selective prediction）时，对置信度最低的 10% 经历选择弃答，可将智能体任务上的交付成功率最多提高 8.7 个百分点。因此，我们认为经验性置信度估计是未来通用置信度估计的一种新范式。
📄 English Abstract
Reliable confidence estimation is increasingly central to the trustworthy deployment of language models: a calibrated estimate of the probability that an output is correct decides what to ship, what to escalate, and what to retry. Existing confidence estimators, however, share one design premise: they only read the current inference process, either by introspecting on it, scoring its token probabilities, or resampling it. We argue that the current inference is not a sufficient basis for confidence. We propose XConf (eXperiential Confidence): estimating confidence together with the model’s accumulated experience. The experience is stored as a record of the model’s own graded past episodes, each holding the task, the model’s reflection, its stated confidence, the outcome, and a lesson written once the grade arrived. Given a new task, XConf’s Recall stage retrieves past episodes on similar tasks met with a similar stated confidence, and reads off their historical success rate; its Reflect stage shows the model this record, has it name its recurring failure mode, and restate a confidence now informed by its own track records. Our estimator is format-general, requiring no logit access or weight updates, and costs only one answer generation. Across nine benchmarks spanning reasoning, coding, multimodal QA, and interactive agents, and four models from three families, XConf beats or matches ten-sample self-consistency in discrimination (AUROC) on 23 of 24 comparisons, with much lower calibration error (ECE), at a tenth of the generation cost. Used for selective prediction, abstaining on the 10% least-confident episodes raises the delivered success rate by up to 8.7 points on agent tasks. We therefore see experiential confidence estimation as a new paradigm for future general-purpose confidence estimation.
5. ProgramDistill：从交互式 Web 应用到可验证的参考引导软件工程任务
- 英文标题：ProgramDistill: From Interactive Web Apps to Verifiable Reference-Guided SWE Tasks
- arXiv：2609.18805v1
💡 一句话解读：现有编程智能体评测都靠文字说明来告诉它要做什么，但真实开发中，智能体往往得先玩通一个完整软件，再把这个功能在残缺的应用里复现出来。为此，作者做了 ProgramDistill：把应用拆成不同粒度的功能，每个功能都配一段可回放验证的行为，自动生成 4063 道题，用来考察智能体能否从参考软件中学会功能并补全代码。
编程智能体通常通过问题描述或指令来指定期望行为进行评测。然而在实际的网页开发中，智能体可能需要从可运行的软件中推断行为，并在一个不完整的应用中将其实现。我们提出 ProgramDistill，这是一个评测编程智能体的基准，其任务来自与功能完整的参考应用交互后所发现的功能。我们通过将应用分解为不同粒度的功能来构建 ProgramDistill，每个功能都关联可通过其黄金补丁（gold patch）执行的可回放行为。我们的流水线 mine-craft-patch 在 26 个应用中发现了 1975 个经回放验证的行为，并在无人工干预的情况下构建了 4063 个任务。在九个前沿编程智能体上，GPT-6 Astra 和 Claude Opus 5 在完整应用重建的累积工作流中分别取得 49.2% 和 28.8% 的成功率。在部分应用重建中，随着恢复深度从 1 增加到 8，成功率分别从 100% 降至 64.0%、从 96% 降至 32%。因此，ProgramDistill 提供了一个难度可控、可扩展的基准，用于评测和诊断编程智能体，并为未来基于课程的学习训练提供了自然基础。
📄 English Abstract
Coding agents are typically evaluated with desired behavior specified through issues or instructions. In practical web development, however, agents may need to infer behavior from working software and implement it in an incomplete application. We introduce ProgramDistill, a benchmark evaluating coding agents on features discovered through interaction with fully functional reference applications. We build ProgramDistill by factorizing applications into features of different granularities, each associated with replayable behaviors executable via its gold patch. Our pipeline, mine-craft-patch, discovers 1,975 replay-verified behaviors across 26 applications and constructs 4,063 tasks without human intervention. Across nine frontier coding agents, GPT-6 Astra and Claude Opus 5 achieve 49.2% and 28.8% success on cumulative workflows in full-application reconstruction. In partial-application reconstruction, success falls from 100% to 64.0% and from 96% to 32% as restoration depth increases from 1 to 8. ProgramDistill thus provides a scalable benchmark with controlled difficulty for evaluating and diagnosing coding agents, and a natural basis for future curriculum-based training.
6. ActionPiece：重新思考自回归视觉-语言-动作模型的动作标记化
- 英文标题：ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models
- arXiv：2609.18487v1
💡 一句话解读：现有动作标记器只看单点重建误差，忽略了不同情境下动作微调关系的保真度。本文提出物理秩一致性指标，并设计ActionPiece，通过联合监督让编码和量化都保持动作间的远近关系，从而提升策略成功率。
动作标记器在自回归视觉-语言-动作（VLA）模型中起核心作用，它既决定策略训练的目标，也决定从预测标记中恢复的可执行命令。其保真度通常用逐点重建指标（如均方误差MSE）评估，但微小的个体误差并不能完全刻画跨演示的动作调整被保留得有多忠实。压缩后，相似动作可能仍聚集在某个代表性运动附近，而不同情境所需的调整则被削弱、扭曲甚至反转。我们引入物理秩一致性（PRC），衡量标记化在重建后对局部物理距离排序的保持程度。评估解码后的动作可为不同标记词表和解码器架构提供共同参照，在逐点精度之外补充关系保真度的度量。我们进一步提出ActionPiece，通过表示学习与量化的联合监督来保持物理动作关系。物理秩保持监督编码器和量化特征距离中的远近排序，而量化正则化则将相同排序应用于码字分配分布。这两个目标与重建目标共同作用，产生离散动作标记，用于标准自回归策略学习，并通过冻结解码器执行。在相同的Qwen3-VL-4B策略训练设置下，ActionPiece在LIBERO上达到94.8%，在未见过的LIBERO-Plus上达到68.8%，额外评估在SimplerEnv上达到71.9%，在VLA-Arena L0-L2上达到51.5%。组件消融表明，这两个目标共同提升了PRC和策略成功率，证明了物理关系监督对动作标记化的价值。
📄 English Abstract
Action tokenizers play a central role in autoregressive vision-language-action (VLA) models, determining both the targets for policy training and the executable commands recovered from predicted tokens. Their fidelity is commonly evaluated using pointwise reconstruction metrics such as mean squared error (MSE), yet small individual errors do not fully characterize how faithfully action adjustments across demonstrations are preserved. After compression, similar actions may still cluster around a representative motion, while the adjustments needed for different contexts are diminished, distorted, or even reversed. We introduce physical rank consistency (PRC) to measure how well tokenization preserves local physical distance rankings after reconstruction. Evaluating decoded actions provides a common reference across token vocabularies and decoder architectures, complementing pointwise accuracy with a measure of relational fidelity. We further present ActionPiece, which preserves physical action relationships through joint supervision of representation learning and quantization. Physical rank preservation supervises near-far ordering in encoder and quantized feature distances, while quantization regularization applies the same ordering to codeword assignment distributions. Both objectives augment reconstruction, producing discrete action tokens for standard autoregressive policy learning and execution through a frozen decoder. Under the same Qwen3-VL-4B policy training setup, ActionPiece achieves 94.8% on LIBERO and 68.8% on unseen LIBERO-Plus, with additional evaluations reaching 71.9% on SimplerEnv and 51.5% across VLA-Arena L0-L2. Component ablations show that the two objectives jointly improve PRC and policy success, demonstrating the value of physical relationship supervision for action tokenization.
7. Agora：将 Git 作为集体自动研究的共享记忆
- 英文标题：Agora: Git as Shared Memory for Collective AutoResearch
- arXiv：2609.18094v1
💡 一句话解读：多个AI研究助手各自为战，重复劳动多、发现少。Agora让它们把研究记录存进Git，像共享笔记本一样，每个结论都可追溯和复现，还能避免所有人挤在同一条路上。
AutoResearch等自主研究循环表明，单个编程智能体可以无人值守地改进训练设置。但运行多个智能体时，每个会话都从头开始，因此智能体越多往往意味着重复搜索而非更多发现。Agora为这类智能体提供共享记忆：研究被记录为存储在Git中的仅追加有向无环图（DAG），因此每个主张都是一个提交，任何人都可以检出并重新运行。每个结果、见解、假设、验证和报告都是一个不可变提交，其父边表示它所基于的内容；一个派生索引暴露前沿、被忽视的分支以及每个主张的验证状态，而多样性感知的选择规则防止社区坍缩到单一领导者。我们描述了该系统并报告了其首次持续使用：一次近12天的运行中，13个语言模型工作者在没有分配任务也没有中央规划器的情况下，解决一个权重迁移问题。给定141个预训练供体模型和一个冻结的1.196亿参数注意力-SSM混合模型（其维度与任何供体都不匹配），工作者必须在没有训练数据或梯度更新的情况下初始化目标模型。他们发布了1,703项贡献，并将评估器从每字节3.39比特降至1.899比特，缩小了与训练过的GPT-2 124M之间62%的差距。获胜方案将供体的下一词元统计压缩到目标的嵌入和输出头中，然后通过稀疏编辑注意力、前馈和状态空间块来添加短程上下文信号。其145次提交的祖先跨越15个账户，并发布了165次独立复现，无一失败。我们描述了运行中唯一一次将社区拉出单一文化的人类干预、该轨迹能证明和不能证明什么，以及一个受控比较，以确定共享研究状态是否能提高单位计算量的发现。
📄 English Abstract
Autonomous research loops such as AutoResearch show that one coding agent can improve a training setup unattended. Run several of them and each session starts from scratch, so more agents tend to mean more duplicated search rather than more discovery. Agora is a shared memory for such agents: research is recorded as an append-only directed acyclic graph (DAG) stored in Git, so that every claim is a commit anyone can check out and rerun. Each result, insight, hypothesis, verification, and report is an immutable commit whose parent edges say what it builds on; a derived index exposes the frontier, the neglected branches, and the verification status of each claim, and a diversity-aware selection rule keeps the community from collapsing onto one leader. We describe the system and report its first sustained use: a run of nearly 12 days in which 13 language-model workers, with no assigned tasks and no central planner, worked on a weight-transfer problem. Given 141 pretrained donor models and a frozen 119.6M-parameter attention-SSM hybrid whose dimensions match no donor, the workers had to initialize the target without training data or gradient updates. They published 1,703 contributions and drove the evaluator from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. The winning recipe compresses donor next-token statistics into the target’s embedding and output head, then adds a short-range context signal through sparse edits to attention, feed-forward, and state-space blocks. Its 145-commit ancestry spans 15 accounts, and 165 independent reproductions were posted, none of which failed. We describe the single mid-run human intervention that pulled the community out of a monoculture, what the trace does and does not establish, and the controlled comparison that would settle whether shared research state improves discovery per unit of compute.
8. VC-Attention：面向低比特注意力的值平滑与 Softmax 转换
- 英文标题：VC-Attention: Value Smoothing and Softmax Casting for Low-bit Attention
- arXiv：2609.15810v1
💡 一句话解读：视频生成模型里注意力计算太慢太耗资源，低比特量化又因少数极大值（离群值）导致精度崩掉。本文提出免训练方案：把值向量按相似度重排并减去块均值再量化，同时把Softmax里的指数运算直接融合成一次乘加，从而在多种GPU上既提速又保真。
扩散Transformer（Diffusion Transformer）能实现最先进的视频生成，但其长时空序列使注意力成为部署的主要开销，因此可部署的低比特内核必须兼顾精度与速度。精度受离群值限制：一个块的量化尺度由其最大元素决定，导致典型元素只能落在很窄的可表示范围内。已有工作对查询（query）和键（key）做平滑，但值（value）的离群值没有固定的通道或时空结构，仍是输出误差的主要来源。速度受Softmax限制：低比特张量核心（Tensor Core）只加速两个矩阵乘法，因此两者之间的高精度指数运算成为数据中心GPU上最长的流水线阶段。我们提出VC-Attention，一种免训练的低比特注意力框架，通过将值平滑（Value smoothing）与融合的概率转换（probability Cast）配对来解决这两个问题。V-Smooth利用轻量级在线聚类对值令牌（token）重新排序，使硬件块内的令牌能一起良好量化。它仅对减去块均值后的残差进行量化，并从在线Softmax已维护的行和中恢复该均值。ExpCast-FP8通过一次融合乘加将对数域分数直接映射为E4M3概率编码，消除了FP32指数运算和格式转换。我们为B200、B300、H200、RTX PRO 6000和RTX 5090实现了VC-Attention。在Wan2.2、LongCat-Video、HunyuanVideo-1.5和MiniMax-H3上，VC-Attention相比低比特基线提升了保真度，在数据中心Blackwell和Hopper上注意力内核比BF16 FlashAttention-4加速1.46-1.59倍，在工作站显卡上加速2.3-3.6倍，端到端生成片段分别快1.13-1.19倍和1.36-1.70倍。
📄 English Abstract
Diffusion Transformers deliver state-of-the-art video generation, but their long spatiotemporal sequences make attention the dominant deployment cost, and a deployable low-bit kernel must be accurate and fast. Accuracy is limited by outliers: a block’s quantization scale is set by its largest entries, leaving typical entries confined to a narrow range of representable values. Prior work smooths queries and keys, but value outliers follow no fixed channel or spatiotemporal structure and remain the dominant source of output error. Speed is limited by softmax: low-bit Tensor Cores accelerate only the two matrix multiplications, so the high-precision exponential between them becomes the longest pipeline stage on datacenter GPUs. We propose VC-Attention, a training-free low-bit attention framework that addresses both by pairing Value smoothing with a fused probability Cast. V-Smooth reorders value tokens by lightweight online clustering, so the tokens in a hardware block quantize well together. It quantizes only the residual after subtracting the block mean, and restores that mean from the row sum the online softmax already maintains. ExpCast-FP8 maps log-domain scores directly to E4M3 probability codes with one fused multiply-add, eliminating the FP32 exponential and the format conversion. We implement VC-Attention for B200, B300, H200, RTX PRO 6000, and RTX 5090. Across Wan2.2, LongCat-Video, HunyuanVideo-1.5, and MiniMax-H3, VC-Attention improves fidelity over low-bit baselines, speeds up the attention kernel over BF16 FlashAttention-4 by 1.46-1.59x on datacenter Blackwell and Hopper and by 2.3-3.6x on workstation cards, and generates a clip 1.13-1.19x and 1.36-1.70x faster end to end.
9. EvolveTrade：面向自演化LLM交易智能体的经验驱动策略精炼
- 英文标题：EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents
- arXiv：2609.17632v1
💡 一句话解读：这篇研究要解决的问题是：LLM交易智能体在部署后，其调用工具、收集信息、验证信号和管理风险的方式被固定下来，无法随市场变化而调整。为此，作者提出EvolveTrade框架，把智能体的系统提示词当作可演化的策略，每隔一段时间用积累的决策记录和实际收益反馈来修改它，而底层大模型保持不变，从而让智能体不断优化自己的交易流程。
大语言模型（LLM）交易智能体能够结合市场数据、新闻和可执行分析，但其行为通常由部署前固定的静态手写工具使用策略控制。这限制了它们在不断变化的市场状态下调整证据收集、工具调用、信号验证和风险管理方式的能力。我们提出EvolveTrade，一个自演化框架，将工具使用型交易智能体的系统提示词视为文本参数化策略。在每个更新间隔后，一个策略智能体利用累积的决策轨迹和已实现投资组合反馈来修订该策略，同时保持骨干LLM固定不变。更新后的策略随后用于下一批交易决策，使智能体能够随时间精炼其信息获取和投资组合构建流程。跨多种市场状态和两个LLM骨干的实验表明，EvolveTrade在大多数评估设置中，相较于固定策略的LLM基线，通常能提升夏普比率（Sharpe Ratio）和累计收益（Cumulative Return）。行为分析进一步表明，自演化策略增加了代码介导的分析并激活了与市场状态相关的计算；案例层面的策略到收益归因追踪了策略引起的配置变化如何贡献于已实现的收益差异。这些结果表明，调整管理工具使用的可复用流程是构建更稳健LLM交易智能体的关键方向。
📄 English Abstract
Large language model (LLM) trading agents can combine market data, news, and executable analysis, but their behavior is often controlled by static hand-written tool-use policies that are fixed before deployment. This limits their ability to adapt how they gather evidence, invoke tools, verify signals, and manage risk under changing market regimes. We introduce EvolveTrade, a self-evolving framework that treats the system prompt of a tool-using trading agent as a text-parameterized policy. After each update interval, a Policy Agent revises this policy using accumulated decision traces and realized portfolio feedback, while keeping the backbone LLM fixed. The updated policy is then used for the next batch of trading decisions, enabling the agent to refine its information-acquisition and portfolio-construction procedure over time. Experiments across multiple market regimes and two LLM backbones show that EvolveTrade often improves Sharpe Ratio and Cumulative Return over fixed-policy LLM baselines, achieving the improved SR and CR in most evaluated settings. Behavioral analyses further show that self-evolved policies increase code-mediated analysis and activate regime-relevant computations; case-level policy-to-return attributions trace how policy-induced allocation changes contribute to realized return differences. These results suggest that adapting the reusable procedure governing tool use is a key direction for building more robust LLM trading agents.
10. Zing-0.5：迈向具备实时联合动作与文本控制的可玩世界
- 英文标题：Zing-0.5: Toward Playable Worlds with Real-Time Joint Action and Text Control
- arXiv：2609.17909v1
💡 一句话解读：这篇论文要解决的问题是：让AI生成的可玩世界能实时响应用户的键盘操作和文字指令，并保持画面连贯。方法是训练一个50亿参数的自回归模型，把键盘动作和文字指令统一控制，用分段教师模型指导快速生成，实现24帧每秒的实时交互，成本极低。
我们介绍了Zing-0.5，一个50亿参数的自回归世界模型，专为可玩性设计：用户可以通过键盘和在线文本联合控制来探索生成的世界、影响正在发生的事件，并对产生的反馈做出响应。我们的方法包含三项技术贡献：（1）统一动作与文本条件，将幅度感知的键盘输入与时间对齐的文本指令及联合标注视频相结合，在同一序列中学习导航和事件控制；（2）面向增量生成的事件尺度监督，使用在连接多提示视频上训练的分段级教师模型，通过分布匹配蒸馏来监督块级因果学生模型；（3）低成本实时交互，结合四步生成与保留上下文的流式处理，支持832×480分辨率下24帧每秒的推理，估计服务器租赁成本约为每流分钟0.009美元。在158个WBench导航案例中，Zing-0.5总体得分为81.0，一致性得分为88.5。联合控制演示显示，在持续导航过程中无需重新开始生成即可实现文本引导的事件变化。我们发布了模型权重、推理代码和Zing-SGLang服务实现，以支持可玩生成世界的进一步研究。
📄 English Abstract
We introduce Zing-0.5, a 5B autoregressive world model designed for playability: users can explore generated worlds, influence unfolding events, and respond to the resulting feedback through joint keyboard and online text control. Our approach brings together three technical contributions: (1) Unified action and text conditioning, combining magnitude-aware keyboard inputs with temporally aligned text instructions and jointly annotated videos to learn navigation and event control within the same sequence; (2) Event-scale supervision for incremental generation, using a segment-level teacher trained on connected multi-prompt videos to supervise a block-level causal student through distribution-matching distillation; and (3) Low-cost real-time interaction, combining four-step generation with context-preserving streaming to support 832 x 480 inference at 24 FPS at an estimated server rental cost of approximately USD 0.009 per stream-minute. Zing-0.5 achieves an overall score of 81.0 and a consistency score of 88.5 across 158 WBench Navigation cases. A joint-control demonstration shows a text-directed event change during continued navigation without restarting generation. We release the model weights, inference code, and Zing-SGLang serving implementation to support further work on playable generated worlds.
11. 注视作为共同基础证据：MapTask与MUNDEX的跨语料库分析
- 英文标题：Gaze as Evidence for Common Grounding: A Cross-Corpus Analysis of MapTask and MUNDEX
- arXiv：2609.18011v1
💡 一句话解读：在信息不对称的协作任务中，人们如何确认彼此理解一致（即“共同基础”）？本研究把两个对话语料库（MapTask 和 MUNDEX）统一成同一套行为标注，分析说话时的注视行为。结果发现，当双方理解趋于一致时，人们更多看向任务相关区域、更少看同伴，视线更集中、切换更少；这种关联在主导任务的一方身上最明显。不过效应较弱，作者认为注视只是促进相互理解的一个辅助线索。
在信息不对称的协作任务中，参与者通过互动来协调彼此的理解。我们考察注视能否为两个此类任务中的共同基础（grounding）提供证据。基于离散的行为标注，我们将 HCRC MapTask（Anderson 等，1991）和 MUNDEX（Türk 等，2023）映射到一套共享的“同伴/任务/别处”词汇表，并围绕任务相关的对话单元计算注视特征。在两个语料库中，对齐的指称解读（MapTask）和“已理解”（UND）判断（MUNDEX）都与更多指向任务的注视、更少指向同伴的注视、更低的注视熵以及更少的注视转移相关联。这些关联在主导任务的参与者身上最为清晰：在给予者产生的指称和解释者判断中，后者还与解释接受者的注视共变。在 MapTask 同一说话者的指称链中，当先前未对齐的指称对象变为对齐时，说话者在该提及处的注视熵更低。在分组交叉验证下，最佳注视特征组相对控制变量仅有适度提升：MapTask 中为时间特征，MUNDEX 中为原始比例。由于效应较小，且当以重复参与者而非对话作为推断单位时若干效应减弱，我们将注视视为促进共同基础的一个辅助线索，需结合任务与对话语境加以解读。
📄 English Abstract
In collaborative tasks with asymmetric information, participants coordinate their understanding through interaction. We ask whether gaze provides evidence about grounding across two such tasks. Working from discrete behavioral annotations, we map HCRC MapTask (Anderson et al., 1991) and MUNDEX (Türk et al., 2023) into a shared partner/task/away vocabulary and compute gaze features around task-relevant dialogue units. In both corpora, aligned reference interpretations (MapTask) and UND (understood) judgments (MUNDEX) are associated with more task-directed gaze and with less partner-directed gaze, lower gaze entropy, and fewer gaze transitions. The associations are clearest for the participant leading the task: in giver-produced references, and in explainer judgments, which also co-vary with the explainee’s gaze. In same-speaker MapTask reference chains, the speaker’s gaze entropy is lower at the mention where a previously non-aligned referent becomes aligned. The best gaze feature groups improve modestly over controls under grouped cross-validation: temporal features in MapTask and raw proportions in MUNDEX. Because effects are small and several weaken when recurring participants rather than dialogues are the unit of inference, we treat gaze as one contributing cue to grounding, to be interpreted alongside task and dialogue context.
12. 面向LLM偏好对齐的零阶范式
- 英文标题：A Zeroth-Order Paradigm for LLM Preference Alignment
- arXiv：2609.19144v1
💡 一句话解读：现有让大语言模型对齐人类偏好的方法，在偏好差距很小时容易因“似然位移”而失效。本文提出ComPO，一种基于比较的零阶对齐方法，不直接优化可微的偏好损失，而是从偏好对中提取方向信息，并给出收敛保证和在线版本，在多个模型上优于现有方法。
直接偏好对齐方法因计算和内存效率高而被广泛用于使大语言模型（LLM）与人类偏好对齐。然而，似然位移（likelihood displacement）促使人们寻找其他方式，从似然边际较小的偏好对中提取信息。本文提出并分析了基于比较的偏好优化（ComPO），这是一种基于比较预言机的零阶对齐方法。ComPO从这些偏好对中提取方向信息，而不直接优化可微的偏好损失。我们为其基本离线方案建立了收敛保证，条件包括平滑性、梯度稀疏性以及预言机与潜在目标之间的兼容性。我们进一步引入在线ComPO，它保留离线比较机制，并利用未标注的策略生成进行相对于参考策略的反向KL控制。遵循偏好微调的覆盖视角，我们在局部覆盖和分布内成对奖励准确率下，为一种基本约束方案建立了性能保证。在Mistral、Llama、Gemma-2、Qwen3和Gemma-3模型上的实验表明，该方法优于现有的直接对齐方法，包括长度控制的胜率，并且成对级别的诊断提供了与缓解似然位移一致的证据。
📄 English Abstract
Direct preference alignment methods are widely used to align large language models (LLMs) with human preferences because of their computational and memory efficiency. However, likelihood displacement motivates alternative ways to extract information from preference pairs with small likelihood margins. In this paper, we propose and analyze Comparison-based Preference Optimization (ComPO), a zeroth-order alignment method based on comparison oracles. ComPO extracts directional information from these pairs without directly optimizing a differentiable preference loss on them. We establish a convergence guarantee for its basic offline scheme under smoothness, gradient sparsity, and compatibility between the oracle and a latent objective. We further introduce online ComPO, which retains the offline comparison mechanism and uses unlabeled policy generations for reverse-KL control relative to a reference policy. Following the coverage perspective of preference fine-tuning, we establish a performance guarantee for a basic constrained scheme under local coverage and in-distribution pairwise reward accuracy. Experiments on Mistral, Llama, Gemma-2, Qwen3, and Gemma-3 models demonstrate improvements over existing direct alignment methods, including length-controlled win rates, with pair-level diagnostics providing evidence consistent with mitigating likelihood displacement.
13. HypoEvolve：遗传算法使多智能体大语言模型能够发现科学假设
- 英文标题：HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses
- arXiv：2609.15938v1
💡 一句话解读：这篇论文想搞清楚：多个AI智能体一起协作，到底能不能比单个AI提出更好的科学假设？为此作者提出HypoEvolve，用类似“遗传进化”的方式让不同分工的大模型智能体反复提出、修改、筛选假设，并在抗癌药物再利用任务上验证，效果优于六个对比方法。
科学智能体通过综合证据、评估提案并发展新解释来促进假设发现。近期系统将科学智能体与进化式搜索结合，通过批判、比较和修订来工作。然而，不同形式的智能体协作如何影响假设质量仍是一个未解问题。回答这一问题需要将智能体自身科学能力的影响与其协作带来的影响分离开来，因此框架必须保留智能体的科学角色，并支持假设的组合、修订与保留规则。基于这一观点，我们提出HypoEvolve，它通过对假设种群进行逐代更新，使协作过程显式化。具体而言，我们提出一种分代遗传算法，协调专门化的大语言模型（LLM）智能体，使其整合机制性论证、重新审视假设前提，并评估证据与可检验性。每一代都明确规定科学判断与新提案如何重塑种群，从而使协作对假设质量的影响可直接检验。此外，我们围绕具有科学意义的假设设计评估，这些假设解释某项拟议干预为何可能有效。药物再利用将这些解释与靶点层面的生物学主张联系起来，并依据外部证据加以评估。具体来说，我们将DepMap和Open Targets改造为互补的外部指标，二者分别基于实验、遗传和临床证据。在34种癌症类型上，HypoEvolve在两项指标上均取得优于六个基线的最高分。DepMap选择性达到0.171，而最强基线为0.115。相较单次生成方法的提升也能泛化到留出的癌症类型。HypoEvolve推进了自主科学的一种愿景：AI研究团队能够实现超越单个模型的发现能力。
📄 English Abstract
Scientific agents contribute to hypothesis discovery by synthesizing evidence, assessing proposals, and developing new explanations. Recent systems combine scientific agents with evolutionary search through critique, comparison, and revision. However, how different forms of agent collaboration affect hypothesis quality remains an open question. Answering this question requires separating the effects of agents’ scientific capabilities from those of their collaboration. A framework must therefore preserve agents’ scientific roles and support rules for combining, revising, and retaining hypotheses. Building on this view, we introduce HypoEvolve, which makes collaboration explicit through successive updates to a hypothesis population. Specifically, we propose a generational genetic algorithm to coordinate specialized large language model (LLM) agents that integrate mechanistic arguments, reconsider assumptions, and assess evidence and testability. Each generation specifies how scientific judgments and new proposals reshape the population, making collaboration effects on hypothesis quality directly testable. Moreover, we design our evaluation around scientifically meaningful hypotheses that explain how a proposed intervention could work. Drug repurposing links these explanations to target-level biological claims assessed against external evidence. Specifically, we adapt DepMap and Open Targets into complementary external measures grounded in experimental, genetic, and clinical evidence. Across 34 cancer types, HypoEvolve achieves the highest scores against six baselines on both measures. DepMap selectivity reaches 0.171, versus 0.115 for the strongest baseline. Gains over single-pass generation also generalize to held-out cancer types. HypoEvolve advances a vision of autonomous science in which AI research teams achieve a capacity for discovery beyond that of individual models.
14. EventEgoHands++：基于事件的第一人称3D手部网格重建与真实数据集
- 英文标题：EventEgoHands++: Event-based Egocentric 3D Hand Mesh Reconstruction with Real Dataset
- arXiv：2609.17189v1
💡 一句话解读：第一人称视角下，用事件相机重建3D手部模型时，背景事件会干扰手部信号，且现有方法无法区分左右手，导致重建不准。本文提出EventEgoHands++，通过手部检测器识别左右手实例，并用自适应注意力动态调整关注区域，从而更准确地重建双手。
3D手部网格重建是一项具有挑战性但至关重要的任务，可应用于人机交互和AR/VR等下游场景。尽管传统相机已被广泛采用，但基于传统相机的方法在低光环境和严重运动模糊下表现不佳。为解决这些限制，事件相机因其高动态范围和高时间分辨率而受到关注。然而，将事件相机应用于第一人称手部重建仍面临挑战，因为相机佩戴者的运动会产生密集的背景事件，掩盖手部特定信号。尽管首个基于事件的第一人称方法通过手部分割缓解了这一问题，但其二值手部掩码无法区分左右手。因此，模型缺乏实例级手部信息，即使只有一只手或没有手时也会预测双手，导致手间关系错误和重建精度下降。本文提出EventEgoHands++，一个从第一人称视角进行基于事件的3D手部网格重建框架。该方法包含一个手部检测器，用于估计左右手的实例级边界框和掩码。此外，我们引入自适应注意力机制，根据这些检测结果动态门控注意力，以准确学习双手之间的空间关系和相互交互。为了训练和评估我们的框架，我们扩展了合成数据集N-HOT3D，并新建了EEH-R，这是迄今为止最大的真实世界基于事件的第一人称手部数据集，包含约100万标注帧，采集环境包括低光条件。在合成和真实数据集上的大量实验表明，我们的方法始终优于基线。
📄 English Abstract
3D hand mesh reconstruction is a challenging yet essential task for downstream applications, including human-robot interaction and AR/VR. Although conventional cameras have been widely adopted for this task, methods that rely on them struggle in low-light environments and under severe motion blur. To address these limitations, event-based cameras have recently attracted attention for their high dynamic range and high temporal resolution. However, applying event cameras to egocentric hand reconstruction remains challenging because camera wearer’s motion produces dense background events that obscure hand-specific signals. Although the first egocentric event-based approach mitigates this issue using hand segmentation, its binary hand mask does not distinguish between left and right hands. As a result, the model lacks instance-level hand information and predicts both hands even when only one or neither hand is present. This limitation leads to incorrect inter-hand relationships and degraded reconstruction accuracy. In this paper, we propose EventEgoHands++, a framework for event-based 3D hand mesh reconstruction from an egocentric viewpoint. The proposed method incorporates a Hand Detector that estimates instance-level bounding boxes and masks for both the left and right hands. Moreover, we introduce Adaptive Attention, which dynamically gates the attention based on these detection results to accurately learn the spatial relationship and mutual interactions between the hands. To train and evaluate our framework, we extend the synthetic N-HOT3D dataset and newly construct EEH-R, the largest real-world event-based egocentric hand dataset to date, comprising approximately 1M annotated frames captured in environments including low-light conditions. Extensive experiments on both synthetic and real datasets demonstrate that our method consistently outperforms the baselines.
15. SpectralShift：通过谱重参数化有效扩展 Gated DeltaNet 的上下文窗口
- 英文标题：SpectralShift: Effective Context Window Extension of Gated DeltaNet via Spectral Reparameterization
- arXiv：2609.14320v1
💡 一句话解读：线性注意力模型在长文本任务中表现不佳，因为现有扩展上下文的方法忽略了状态转移矩阵的频谱特性。本文提出 SpectralShift，通过重新参数化衰减谱的初始化并调整学习率，增强慢速传播模式、保留快速衰减模式，从而有效扩展 Gated DeltaNet 的上下文窗口。
近年来，线性注意力层越来越多地被用于大规模长上下文建模，以替代 softmax 注意力。然而，现有的上下文扩展方法通常直接进行持续预训练而不修改这些层，忽略了线性注意力状态动态的谱特性。本文从转移矩阵的谱视角研究 Gated DeltaNet（GDN）的长上下文扩展，识别出决定长程信息检索的两个关键因素：（1）与目标依赖长度对齐的足够宽的慢速谱带；（2）保留用于状态清除和上下文切换的快速衰减模式。基于此，我们提出 SpectralShift，一种用于 GDN 长上下文持续预训练的谱重参数化方法。具体而言，SpectralShift 重新参数化 alpha 投影的初始化，通过增强慢速传播能力来重塑衰减谱，并进一步为 alpha 投影引入学习率缩放以促进长上下文训练。实验表明，SpectralShift 在训练过程中持续提升长上下文能力，为扩展线性注意力模型的上下文窗口提供了一种有效且高效的解决方案。代码已开源于 https://github.com/RUCAIBox/GDN-SpectralShift。
📄 English Abstract
Recently, linear attention layers have been increasingly adopted to replace softmax attention at scale for long-context modeling. However, existing context extension approaches typically apply continued pretraining directly without modifying these layers, overlooking the spectral properties of linear attention state dynamics. In this work, we study long-context extension of Gated DeltaNet (GDN) from a spectral perspective of transition matrix and identify two essential factors governing long-range information retrieval: (1) a sufficiently broad slow spectral band aligned with the target dependency length, and (2) the preservation of fast-decaying modes for state clearing and context switching. Based on this observation, we propose SpectralShift, a spectral reparameterization approach for long-context continual pretraining of GDNs. Specifically, SpectralShift reparameterizes the alpha projections initialization to reshape the decay spectrum by enhancing slow propagation capacity, and further introduces a learning-rate scaling for alpha projections to facilitate long-context training. Experiments show that SpectralShift consistently improves long-context capabilities over training, providing an effective and efficient solution for extending context windows of linear attention models. The code has been open-sourced at https://github.com/RUCAIBox/GDN-SpectralShift.