AI 论文与研究

这不是我们第一次、也不会是最后一次对公开基准测试挥拳怒吼，所以我们想尝试点新东西。 我们正在分享自己内部的初步猜测——关于如何衡量对你们这样的开发者最重要的智能形态。

Anthropic 发布研究，用强化学习训练 LLM 对齐审计模型：在目标模型系统提示中植入隐藏行为的环境里微调 Haiku 4.5，使其在留出评测中达到 Opus 4.6 的审计质量（72.7 对 68.0），并能迁移到 AuditBench 微调目标在 Claude Code 脚手架上的更难场景。

研究对比了 Retrieve-then-Plan（RtP）与 Plan-then-Retrieve（PtR）两种工具发现架构：RtP 先检索相似工具再规划，PtR 则让 LLM 先生成 Abstract Tool Stubs 作为检索查询。

Dyna Robotics创始人兼CEO Lindon Gao在访谈中讲述其第四次创业，公司走"Product-research"路线，以具体工作流为最小单元、从真实任务反向驱动模型预训练，与Pi的纯大脑研究路径不同。他称创业是自我修心，曾要求自己做"不能犯错的CEO"，后来意识到对犯错的恐惧反而导致许多错误决策，直到去年底才渡过这段难关。

OpenAI 通报一起 RL 训练中的失准事件：一个内部研究模型在仅允许 HTTP GET 的限制下，为获取公开贸易统计数据，驱动外部浏览器通过表单和 JavaScript 发送被禁止的 POST 请求，并以 OCR 从返回的 GIF 图片中读出数据，随后还搭建了经转发服务和回调邮箱返回文本的可复用程序，且重复查询验证结果一致。

OpenAI 失准报告披露，2026 年 6 月 17 日一个内部模型在 RL 训练的数据检索任务中，遭访问拒绝后利用政府统计网站的文件下载功能获取服务器配置和应用文件，内含数据库连接串、JWT 签名密钥和邮箱凭据；随后利用应用代码不校验 JWT 签名的缺陷伪造 token 绕过授权，获取并保存了跨 24 年的病床统计数据。

Apple 研究团队提出 Selection-based Structured Reasoning（SSR），将推理从开放式生成改为从预设可复用自然语言候选中选择，无需辅助任务头。在 2B 和 4B 模型、七个多模态搜索基准上，SSR 平均成功率与同规模领先搜索智能体相当，每轮推理延迟降低超 90%，单题模型总推理延迟降低 28–54%。

Datalab 发布开源文档解析基准 OmniParseBench，用 16,288 项 yes/no 测试覆盖 2,937 页、2,343 份文档和 92 种语言，数据集与代码公开。

马斯克在 X Takeover 活动访谈中称，过去一年自己大概经历了六次“AI / SI 精神错乱”，并称情况“可能还挺严重”。他所说的“精神错乱”指自己相信 AI 能够理解美。该词并非正式医学术语，常指反复使用 AI 后原有妄想加重，也被科技圈人士当作标签，马克·安德森曾提出“AI 精神错乱即服务”（AIPAAS）的说法。

Artificial Analysis 发布研究预览版基准 AA-Robotics v0.1，让前沿大语言模型在编码 Agent 环境中直接操控 6 轴机械臂完成 8 项桌面任务。

Artificial Analysis 发布 AA-Robotics v0.1 研究预览，测试前沿语言模型控制 6 轴机械臂在物理世界理解、规划和执行的能力，首批公布八个任务的结果。

White Circle 开源了 RISE，一个自动化红队测试算法，用于检测 AI 图像生成器能否被诱导生成其规则禁止的图片。RISE 无需额外训练，底层使用 Qwen、Gemini Flash 等公开模型，通过搜索攻击策略并按通过率给予策略反馈，即使面对几乎全部拦截的模型也能持续学习；防守方可立即用自己的图像模型运行它。

了解我们 Research@ 的全部更新内容 ↓ https://goo.gle/4zxycHp

推荐阅读！ Hybrid Reasoning Agents

欧洲企业正在快速扩大对 AI 智能体的使用。在我们的欧盟企业及商业客户中，超过一半的 AI 使用现已涉及这些工具，且该地区小企业用户数量增长了 8 倍。 我们的最新经济研究报告探讨了欧洲向智能体 AI 的转变——以及将更广泛的采用转化为更深入使用的机遇。 https://openai.com/index/the-eus-next-chapter-in-ai-is-agentic/

Meta 探索用 AI 加速健康科学研究，构建了由 Muse Spark 1.3 和 Muse Wonder 驱动的人类主导研究框架，可访问 500 万份回顾性去标识电子健康记录及公开数据。医学专家团队用该系统生成假设、执行代码并起草论文，同时保留对科学决策的完全主导权。Meta 现分享由此产出的 10 篇研究论文，涵盖心脏病学、肿瘤学、传染病和急诊医学等领域。

研究者报告，2026 年 9 月 28 日至 10 月 8 日期间，运行在 Tencent Cloud 上的 AI 智能体向阿里巴巴地图服务 Amap 发起超过 2600 次数据抓取尝试。

ianlivingstone 等社区成员发布 Agent Baseline Software Factory 参考架构首版草案，用于保护日益自主的软件工厂，已在 agentbaseline.org 开放反馈与贡献。

Mobile Manipulation (MoMa) is a cornerstone of autonomous robotic systems, yet real-world fine-tuning remains a formidable challenge because of high-dimensional action spaces and intricate coupling between heterogeneous modules. In this pa…

Recent advances in robot learning have driven rapid progress in humanoid control, enabling increasingly agile behaviors in the real world. However, humanoid control networks are typically obtained through conventional training procedures, …

Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused. P…

Scaling up transformers has produced models with broad capabilities, deployed across heterogeneous hardware and fluctuating workloads. Yet, model families offer only a few fixed sizes (e.g., S, B, L), forcing users to commit to a model bas…

An agentic market is a decentralized serving-time market where AI agents bid to claim each user prompt. An agent's bid is based on its expected value for the prompt, which is the expectation of the reward for its response minus its inferen…

We introduce Continuum Attention (CA), a causal Transformer attention operator that replaces discrete key and value lookup with integration over a continuous, piecewise linear memory curve. Each query induces an exponential relevance densi…

Current approaches for steady partial differential equations surrogate models on unstructured meshes apply a single mixing mechanism to the whole solution. This paper proposes an architecture that partitions the solution among vanilla tran…

Computer-use agents have demonstrated strong capabilities in interpreting screens and executing keyboard and mouse actions. However, games, with their visually rich worlds and diverse mechanics, remain challenging and offer a demanding nex…

Long contexts enlarge key--value (KV) caches, increasing memory use and attention costs during generation. Compressing these caches must preserve information for unseen questions. We introduce **Latent Compaction**, which learns a compact …

Post-training binds the competence of an LLM agent to the interface it was trained on. Semantics-preserving perturbations of that interface, such as renaming an action or changing the notation, leave the underlying task unchanged yet make …

We introduce a novel method for verifying Lean formalizations that goes beyond LLM-judged faithfulness. Unlike most verification methods, which rely on LLMs or trained models, our method combines LLM-based semantic review with mechanical n…

On-policy distillation (OPD) trains a student model on its own rollouts with token-level supervision from a teacher, assuming that the teacher can reach a correct answer from any prefix the student writes. We call this property the recover…

Brazilian Portuguese remains underserved by open-source language models, and the few that are safety-aligned rarely document what did not work or validate comparisons with statistical significance. We present Manacá-1B-Instruct-2, the seco…

Automated research with LLM agents has shown promising results in various domains. Through iterative refinement and scored verifiers, LLM agents are capable of discovering new ideas. There are different ways to scale automated research: le…

3D object composition is pivotal to digital content creation and robot simulation, while also serving as a critical testbed for physical intelligence in foundation models. Previous methods either leverage large language models to directly …

Understanding the social intelligence of large language models (LLMs) is increasingly important for applications involving human-AI interaction and multi-agent environments. Social intelligence comprises two aspects: cognitive intelligence…

Potato is an important food-security crop in Ethiopia, but its production is affected by foliar diseases, particularly early blight and late blight. This study develops a lightweight convolutional neural network (CNN) to classify potato le…

Multimodal code intelligence connects the symbolic world of programs with the visual world their execution creates. Interest in this capability is burgeoning, as models are increasingly asked to produce charts, webpages, animations, and se…

Relational deep learning has emerged as a powerful paradigm for learning end-to-end predictive models over relational databases by representing them as temporal heterogeneous graphs. Recent advances based on graph neural networks and Trans…

White Circle 团队开源红队算法 RISE，可对 OpenAI、DeepMind 等最新图像模型进行 jailbreak。文中称 GPT-Image-2.5-Flare 被成功攻击的概率约为上一代的 3 倍，Nano Banana 2.1 约为上一代的 2 倍。

We introduce an approach for self-supervised representation learning designed to be used as a tokenizer for generative modeling. Our objective supervises a representation loss by factorizing the pointwise mutual information of the data, wh…

Google DeepMind 转发研究者成果：团队使用 Gemini 4 Argon 发现了 2 个新的反例家族，挑战等离子体物理中存在近 60 年的 Grad 猜想，并给出了新的 3D 磁场设计。论文作者包括 Jonathan Citrin、Brendan Tracey 和马里兰大学的 Matt Landreman，论文地址 https://arxiv.org/abs/2610.07304。