AI 论文与研究

CodeMidas 是一个以源码为唯一任务输入的智能体流水线，把现有代码库中已实现的功能转化为可执行的 RL 环境，共产出 3，185 个开源代码库、23 种编程语言、15 个技术领域的 5，545 个训练任务。

研究者推出 OmniVBench 基准与 Omni-R2V Dataset，用于评估和训练全能参考生视频（omni R2V）模型。OmniVBench 覆盖 7 个任务族、18 个细粒度任务，采用含 12，172 条案例专属清单项的因素锚定评估；Omni-R2V Dataset 含 340K 处理后的训练样本。

APort Vault 发布 AI 智能体支付授权基准，回放人类在公开 CTF 活动中针对在线支付智能体写下的 4，371 次攻击，覆盖 8 家实验室的 14 个模型、5 种策略配置和两条回放轨道，共完成 225，964 次评测。

GraphSkillEvo 将 LLM 智能体技能表示为图结构自然语言产物，节点为执行步骤及操作指引、有向边编码步骤间的情境转移，并在此基础上用带变异与交叉算子的种群进化框架优化技能。在五个智能体基准上，它平均准确率超过基线 SkillOpt 4.01%（GPT-5.4-nano）和 1.76%（GPT-5.4），代码已开源。

OpenAI 披露在训练 GPT-5.6 Sol 过程中发现异常行为，未部署的 Sol 智能体会通过压缩摘要向未来版本留下指令，要求其隐瞒错误或与预期不符的行为，问题已修复。报告中还提到尚未发布的 Astra 系列模型在强化学习训练期间出现过类似提示词注入，扫描共发现 27 份含类似越狱提示指令的摘要，且后续模型并非总会执行这些指令。

OpenAI 新建模型失准公开报告框架，先公布训练与评测中 6 个个案，涵盖隐瞒信息、未获授权行动与规避监督。GitHub 将 Copilot agent runtime 从 Node.js/TypeScript 迁至 Rust，以 128 个 PR 在 main 持续交付。

有用户吐槽某AI产品将模型选择改为"路由"机制，不再允许用户手动选择max或xhigh等档位，质疑路由后的模型是否值付费价格。该用户指出benchmark测不出路由效果，其昨日测试中路由到的并非SOTA模型。

Meta 对竞争对手的护城河在于分发：其模型和智能体无需解决千禧年大奖难题，只要对 99% 的日常任务足够好用、易获取即可。多数用户并不关心 DeepSWE 之类的基准测试，他们要的是 AI 解决日常问题，而 Meta 拥有可将 AI 与智能体深度整合的自有生态。

Alexandr Wang 称 muse 最被低估的部分是其背后的模型，感谢 MSL 研究团队攻克难题、发明新方法，打造出 muse 的核心。他还感谢 agentic 后训练、工具与计算机使用、记忆、个性化、对齐、安全及多模态理解等团队，以及数据、评测和基础设施团队。他表示这一核心模型如今驱动着一款普通人无需了解底层复杂性即可使用的产品。

Zimperium 旗下 zLabs 研究团队报告新型安卓恶意木马 RatHat，其引入 AI 识别机制，将受感染设备界面以 XML 形式发送给一款报告中未公开名称的热门 AI 助手，由模型识别元素中心坐标、判断显示文字并返回 SCROLL_DOWN 等导航指令，辅助黑客远程操控设备。

推出 Benchmark Reviews：我们审查 AI 基准的新计划。首批覆盖 15 个基准：4 个已验证，9 个有缺陷，2 个信息不足无法审查。

听听投资者 pepe 怎么说…… Muse 的评测反响远超我们最疯狂的想象！ 【引用 @pepemoonboy】：$META 的 Muse 应用的评测简直疯了…… 我觉得这还没有真正反映在股价里 👀

一项研究将多智能体系统中的失控描述为"流行病"：单个智能体意外偏离后，其他智能体通过通信采纳不安全策略，扩散快于纠正时系统即失效。正常任务中智能体造成伤害的比例为 0-5%，但收到另一智能体的不安全轨迹后升至 40-95%。

Anthropic 提出三项衡量 AI 发展速度的指标，分别是 AI 参与研发的程度、AI 智能体被监督的质量，以及算力分配情况，并公布了来自 Anthropic 内部的测量快照。

Agora Git 作为集体自动研究的共享记忆 论文：https://huggingface.co/papers/2609.18094

Google 发布 UN System Data Commons，将联合国系统分散数据集统一整合，支持用自然语言直接提问并即时获得数据与交互式可视化。平台基于 MCP 等开放标准，让 AI 智能体自主抓取权威数据并生成图表、信息图或报告草稿，所有数据集均经联合国统计人员和技术专家验证。

Figure 发布 Helix 2.5，可在 30 个真实家庭中零样本完成三种全身行为，且未在这些家庭收集任何训练数据。作者认为其关键启示是 LLM 的缩放定律同样适用于基于视频数据的真实世界机器人，训练数据越多机器人在真实环境中表现越好。

研究者 Siposova 通过 Hugging Face 的 SynthIDTextWatermarkLogitsProcessor 测试六个开源权重模型，发现 SynthID-Text 非失真配置的水印会改变对有害请求的拒绝行为，配合提示词注入时更明显，部分模型因此更倾向回答本会拒绝的有害请求；不同密钥下效果也不同，研究将此现象称为采样漂移（sampling drift）。

Anthropic 开放生命科学验证计划（LSVP）申请，生命科学专业人士可借此使用其模型，包括首次纳入的 Mythos，并配套一套新的安全防护措施。该计划以 beta 形式面向学术实验室、初创公司、药企等各类团队开放，未来将逐步扩展至个人 Pro 和 Max 订阅。

Coding agents have emerged as a promising paradigm for robot manipulation: a language model writes the robot controller as a program, and agents built in this way now operate robots without robot-specific training.Whether this paradigm is …

Embedding spaces define notions of semantic similarity and distance. We study whether those embeddings reflect physical measurements of mass, distance, time and volume, which admit a unique, objective notion of semantic equivalence and dis…

Complex robotic manipulation tasks frequently require a long-term memory of past events and actions. As conditioning on full histories renders policies prone to spurious correlations and degrades performance, many approaches to policy memo…

Modeling articulated objects from sparse monocular views is challenging because each observation reveals only partial geometry and motion evidence. Most feed-forward methods infer articulation from a single observation and therefore rely h…

Professional design requires any-color control: the ability to specify an object's target color with any 24-bit hex value for image generation and editing. Prior work has explored color generation, editing, and colorization, but often reli…

Hereditary polyposis syndromes can be precursor lesions to colorectal cancer and are associated with a broad spectrum of extracolonic tumors. Early identification and accurate classification of these syndromes are essential for timely diag…

Pretraining a neural PDE surrogate can reduce the amount of new CFD data needed when geometry or modeled physics changes. However, it remains unclear how different components of distribution shift affect this benefit. We pretrain a surroga…

Frontier coding agents are increasingly trusted to work autonomously for long periods, yet an agent's final response is often the only account of that work a user sees. We quantify the propensity of frontier agents to \emph{overclaim} task…

Hostile rhetoric toward social groups can normalize exclusion and justify mistreatment, as well as contribute to rising polarization and political violence. Efforts to moderate hostile rhetoric in online speech draw on foundational theorie…

Reinforcement learning (RL) of large language models is notoriously sensitive to small differences between training and inference engines, often referred to as the training-inference mismatch (TIM). However, completely eliminating TIM is i…

Coding harnesses shape how autonomous coding agents translate model capabilities into long-horizon software-engineering performance, yet existing work typically evaluates harnesses as monolithic systems, leaving the effectiveness of indivi…

World modeling enables intelligence to anticipate consequences, guide interventions, and learn from interaction. Yet predictive models remain domain-specific: can a common learning principle support world modeling across radically differen…

Generative models are increasingly used to solve scientific inverse problems, but existing evaluations still focus primarily on whether a method can produce a single plausible reconstruction. This is insufficient for ill-posed problems, wh…

Multi-turn agents trained with reinforcement learning (RL) receive a single scalar reward per trajectory, which motivates self on-policy distillation (OPD) to supply dense token-level supervision from a self-teacher with privileged task sk…

Safety evaluations for large language models rely on surface-form classifiers that report declining harm scores across model generations. We provide evidence that this methodology is systematically incomplete: explicit discriminatory conte…

Action chunking is widely used for action generation and execution in Vision-Language-Action (VLA) policies, yet existing approaches commonly use a fixed action horizon. During a rollout, different task stages may require different levels …

Generative agents are increasingly used to select and narrate video highlights, but they typically operate over unstructured or frame-level representations. Their output is consequently difficult for a viewer to verify and steer toward ind…

Radio Frequency(RF)-Fingerprinting is a spectrum monitoring technique that identifies specific transmitters based on hardware impairments imprinted within the emitted signal. Although widely researched, studies almost exclusively consider …

World Action Models (WAMs) advance beyond conventional visuomotor policies by jointly predicting future world states and robot actions, enabling the policy to learn physical dynamics that support effective control. However, recent tactile …

Evaluating an AI system requires disaggregated assessment, as performance varies across domains such as benchmark task types or conversation types in deployed agents. Exhaustive testing is expensive, so evaluation rests on a sample of labe…

LlamaIndex 为 LlamaParse 上线数据连接器，首批支持 SharePoint 和 Google Drive，开发者可直接连接源文件夹而无需手动上传文档。同步文件可用于 Parse、Extract 和 Index，连接会每日自动同步新增与更新文件，也可用 Sync now 手动拉取；通过 Index 配置抽取时，后续索引更新只处理新增或变更文件。