AI 论文与研究

Rich Hickey 在 10 月 12 日发表的文章中指出，LLM 不是工具而是自动化。他认为工具需被手持、掌控并持续获得感官反馈，而 LLM 使用缺少控制、意图与持续引导，用户只是发出请求、由模型自行行动。他称用 LLM 读论文会错过学习写作、论证、思考等过程，并强调自动化具有与工具不同的内在属性。

Sakana AI 与 UC Berkeley 研究者提出多智能体自我监督方法 MASS，让单个 Qwen3.6-27B 模型身兼解题、评估、优化三角色，无需更强教师模型即可递归自我改进。

AWS 团队发布论文《Priming: Hybrid state space models from pre-trained Transformers》，提出 priming 方法，复用预训练 Transformer 权重构建混合架构，将预训练成本降低逾 99%，全流程消耗不足源模型 0.5% 的 token 预算。

为了量化模型处理陌生任务的能力，我们将后训练中训练过的任务的测试时计算扩展与其他任务进行了对比。 两个模型在分布外任务上的提升相似，但在相同的分布内得分下，Muse Wonder Full 在陌生任务上的表现优于 Muse Spark 1.3 趋势所预测的水平。我们将这一上移归因于更强的基座模型，它能更有效地将后训练带来的增益迁移到我们未曾预料到的任务上。

如果你只公布一个模型的基准测试成绩却不发布它（权重、产品、API 等"稍后"再说），那几乎可以肯定，在你把它发出来之前，会有人先发布一个你所在类别里更好的模型。

"我这可怜的人类大脑还得硬着头皮看这些"

StudentSim 是一个研究系统，用于在 AI 辅导工具接受学习者测试之前对其进行评估和改进。它模拟个体可能如何与辅导工具互动，从而提供反馈以提升辅导工具在 chess、写作和数学等学科上的表现：https://msft.it/6013aS3JF

Microsoft Research 研究发现，长时间运行的自主研究型智能体常陷入"idea basins"，即使增加算力也难以突破。

本周会在纽约的 @aiDotEngineer！🔥 - 构建自主协作伙伴 | 10/14 下午 2:55 - @GoogleDeepMind 展台 | 大部分时间 如果你在做智能体、评测，或者对 Gemini 有反馈，欢迎过来。

微软研究院提出 StudentSim，一种从公开去标识化学习记录训练个性化 AI 学生模拟器的研究系统，可在真实学习者测试前评估和改进 AI 辅导工具。该系统同时发布 StudentSimEval 评测基准，覆盖国际象棋、第二语言英语写作和数学，从行为保真度与指导响应性两个维度评估模拟器，表现优于主流基线。StudentSim 并非产品，不部署于课堂，也不用于替代教师。

Artificial Analysis 推出 Decision Index v1.0，用于衡量决策模型的表现。该指数同时考察决策模型的成本与时间效率，并指出决策模型存在过度自信问题。文章还对比了决策模型与 LLM 的差异，以及模型"知道何时不该回答"的能力。

Tessl 发文提出评估 AI 软件工厂开放性的关键问题是能否把组织上下文带进来、把生产中积累的知识带走。文章给出透明性、可复现性、互操作性和生态复用四个可验证维度，并建议从一个受限工作流入手，把反复纠正的问题沉淀为技能、架构决策和评测用例，再在其他智能体或 harness 上演练迁移。

NVIDIA 论文提出，相比逐步打分，只在首个让测试通过的决定性编辑处评估 base 模型更能预测其智能体后训练潜力。方法回放成功智能体轨迹并在每次编辑后重跑测试，检查 base 模型是否会写出该编辑、在失败编辑中选中它并写出可通过的实现。模型写出该编辑的概率排序与十个模型后训练后的 SWE-bench Verified 得分相关性达 0.964，且只需基准的成功轨迹和测试即可运行。

QwenGyre 解决了长程智能体 RL 的两大挑战。

千问发布 QwenGyre，为超长时程（XLong）智能体训练引入弹性强化学习，核心设计包括 Elastic Scheduler 与 Trajectory Processor。

Science 新研究显示，Google DeepMind 的 AlphaProof Nexus 独立解决了 353 个开放 Erdős 问题中的 9 个，其中 2 个悬置 56 年。

JetBrains 研究团队分析 4,782 个真实编程智能体会话后发现，功能开发与 bug 修复在真实请求中占比远低于一半，用户会在长会话中不断切换提问、规划、审查与编辑任务，而 SWE-Bench Pro 中这两类任务占比高达 93.8%。

UC San Diego 新论文提出把心理理论（ToM）写进训练奖励，让 7B 智能体学会推断伙伴的信念与目标，部署时无需额外模型即可独立运行。该方法训练一个推断伙伴信念、意图和想法的模型，在 RL 阶段为候选回复打分，最终智能体单独运行。论文题为《Mental-Models for Multi-Agent Systems》，编号 arXiv:2610.12453。

AWS 与汇丰的论文提出 Double Ratchet，让评分器与智能体技能同步进化，仅用 10 个标注样本即可训练出接近真实标注的效果。该评分器由针对单一失败类型的规则化小检查逐步构建，版本选择只依据与 10 个标注样本的一致性，而非提高智能体得分。

In-silico optimization of nucleic acids relies on predictive oracles that imperfectly capture biological function, so sequences optimized against one model may exploit model-specific artifacts and fail to generalize to independent models. …

As large language models are increasingly adopted as multi-party communication facilitators, there is a growing demand for agents capable of balancing diverse group dynamics across contexts. However, existing LLM facilitators are often fin…

Network decision-making tasks, including adaptive bitrate streaming, load balancing, and congestion control, require policies to process heterogeneous states and actions while balancing near- and long-term consequences. Existing solutions …

Text-to-image models have made remarkable progress in visual fidelity and instruction following, yet complex compositional generation remains challenging. Existing agentic approaches address this limitation through iterative generation and…

蚂蚁 inclusionAI 在 GitHub 发布 BiDex-100，一个覆盖 106 个仿真任务的双臂灵巧操作基准，用于评测视觉-语言-动作（VLA）策略。该基准包含 8,080 条成功验证演示、6.25M 帧、58 维统一动作空间，并提供匹配的 ID/OOD 评测协议与 LeRobot v2.1/v3.0 数据集。

Participant-disjoint cross-validation prevents direct participant overlap, but it may still produce optimistic estimates when acquisition site and diagnosis are structurally dependent. We examined Parkinson's disease classification using 1…

Time series (TS) anomaly detection (AD) lacks a universal detector, with the best choice determined by the series rather than by the algorithm. Existing methods therefore either select one detector for each series or average a pool of them…

Embedded platforms increasingly host safety-related streaming functions next to less critical ones. Whether a safety architecture pattern, such as a doer/checker pair, protects a function depends on the cores that run its components. Their…

Spatial transcriptomics jointly captures molecular state, tissue morphology, and spatial organization, yet spatial-domain methods are typically compared only at late training checkpoints, obscuring differences in convergence efficiency. We…

Language models trained only on text from before a given year are useful for studying hindsight and historical language, but the open collections of pre-modern Chinese carry no publication dates and unreliable dynasty labels. We present Lü…

Freely swiveling casters are widely adopted in industrial AMRs/SDVs for their mechanical simplicity and load capacity. However, caster misalignment during low-speed or high-curvature maneuvers induces strong contact nonlinearities, widenin…

A model may select the correct answer yet fail to construct it through a sequence of local decisions. We introduce Choice2Seq, a diagnostic benchmark that measures correctness retention across complete-answer selection and sequential const…

Existing cumulative regret guarantees for Bayesian optimization (BO) do not necessarily characterize the quality of the algorithm's final output. This gap is critical under noisy evaluations, where the goal is to return a single high-quali…

Flagship large language models (LLMs) offer rich general knowledge and strong reasoning capabilities, but their high inference cost and latency limit their use for large-scale data labeling. We study zero-shot text classification with sema…

Real-world studies (RWS) analyze routinely collected healthcare data, such as electronic health records and insurance claims, to estimate treatment effects, monitor safety, and identify risk factors outside randomized trials. Automating RW…

Recent multimodal large language models (MLLMs) achieve strong results on video question answering, yet often fail to recover the temporal structure that makes a video coherent. They may merge neighboring events, miss single-event progress…

This manuscript presents DeepThink, a knowledge-addressable framework for recurrent language modeling. A Knowledge Bank stores typed Knowledge Units, a Knowledge Activator selects units from the current execution state, Knowledge Assembly …

We propose the task of Interactive 4D Volumetric Liquid Forecasting: predicting the spatiotemporal evolution of a future dense 3D liquid velocity field conditioned on how a solid object moves through it. This task provides an upstream phys…

Large language models (LLMs) have demonstrated remarkable capabilities in solving complex tasks in coding, mathematical logic, and iterative reasoning, especially by coordinating effectively within multi-agent systems (MAS). However, langu…

A fruitful approach towards understanding generalization in machine learning involves characterizing how models separate generalizable structure from idiosyncratic or even corrupted features of training data. Fisher information has emerged…

Interactive-agent evaluations commonly rely on standardized LLM-based user simulators that exhibit broadly cooperative behavior, potentially overlooking failures that arise under diverse but realistic user interactions. This can lead to ov…