Large language models are increasingly deployed as agents that plan over long horizons and act through external tools. M…

Large Language Models demonstrate remarkable proficiency in static reasoning, yet training them as autonomous agents thr…

把 RSI（递归自我改进）与智能体自进化放回同一个"经验 → 状态 → 行为"闭环做正面对比：前者打在权重与 AI 研发流程上、跨用户且不可逆、风险外部化；后者打在外部文件与 harness 上、跨会话且可回滚、风险由采用者承担。给出六维对比表、闭环四问判定法、"清空记忆测试"，并梳理两者在 ICLR 2026 与 SIA / Meta-Harness 上的合流路径。含 Snyk ToxicSkills 审计（3,984 个技能中 36.82% 有安全缺陷、13.4% 为严重级）、SEA-Eval"片段式失忆症"等一手数据。

Large Language Model (LLM) agents are turning language into real-world effects, making safety necessary against both ind…

Modern LLM-based agents operate through a harness of tools, reusable skills, and specialist agents that shapes what they…

9 月 11 日科技动态一览：GPT-6 Astra 需求挤爆致 OpenAI 暂停 Pro 20X 新增订阅、Agents API 公测、金融服务版 ChatGPT 上线；Slackbot 升级；加州未成年人社媒法案签署；LG 电视监视争议；观察视角落在"需求侧证实 vs 供给侧反思"的对照上。

While research on recursive self-improvement (RSI) has predominantly automated model training pipelines, reliable autono…

IT之家 9 月 11 日消息，OpenAI 于当地时间 9 月 10 日宣布推出 Agents API 公测版，允许开发者通过 API 调用由 OpenAI 管理的云端 AI 智能体运行环境。 该服务复用了 Codex 背后的智能体执行框…

On-Policy Self-Distillation (OPSD) has emerged as a popular paradigm for large language model (LLM) self-improvement, al…

GUI agents accumulate high-resolution screenshots as the trajectory unfolds, increasing inference latency and memory usa…

Recursive self-improvement (RSI) requires a concrete mechanism through which an AI system observes its capabilities and …

We study scheming in LLM agents, in which agents covertly pursue misaligned goals. Our focus is to understand how schemi…

SWE-Bench Pro has emerged as a standard benchmark for evaluating software engineering agents on challenging repository-l…

AI research agents combine prior knowledge, public sources, and experimental feedback to produce useful results. The Dis…

Model misalignment, prompt injection, or operator misuse could lead AI agents operating frontier-lab accounts to exfiltr…

Sequential memory agents process long documents by reading chunks one after another while maintaining a compact memory s…

Large language model (LLM) agents increasingly rely on external skills, but routing user requests over large skill regis…

We present a continuous, population-scale measurement record of autonomous language-model trading agents operating in pr…

Automating empirical research is a long-standing direction of AI. Recent automatic research (AutoResearch) agents bring …

Large language model (LLM) agents can benefit from reusable skills distilled from prior task experience, yet existing sk…

Ensuring the safety of autonomous driving is a critical challenge. Scenario-based testing is a systematic process used t…

Full-duplex spoken language models can listen and speak simultaneously, enabling them to handle interruptions and backch…

On-policy distillation (OPD) provides dense, per-token supervision for language model post-training, but its effectivene…

A reasoning model can improve from its own on-policy experience, but this inner loop is fragile: terminal verifiers prov…

Benchmarks for the side effects an agent causes on the way to a goal already exist, but HarvestBench is the first to put…

Safety alignment is usually posed as a topic-level question: is this subject harmful? Deployments ask a narrower one. A …

Humans need to study only a handful of well-written textbooks to master a discipline and attempt its hardest problems. W…

On-policy distillation trains a language model on its own generations while a teacher scores them token by token. It com…

Agent harnesses (the system prompt, tool set, execution hooks, and context-management scaffolding around a model) are a …

As LLMs are increasingly used for pre-submission self-review, there is growing demand for feedback that not only identif…

Training capable cyber agents is often treated primarily as a problem of model scale, yet open-weight post-training is c…

Visual generation is evolving from generative models used through a single invocation into agentic control processes tha…

High-quality structured organic reaction data are essential for developing artificial intelligence for chemistry (AI4Che…

Recursive Super-Resolution (SR) extends fixed-scale SR to extreme magnification by repeatedly feeding predictions back i…

Agents can turn shared infrastructure into a channel for coordinated intrusion. The HF Mirror incident and a separate pu…

LLM agents are rapidly becoming production software, deployed to handle customer service, adjudicate disputes, and opera…

We present Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, together with the data…

Vision-language-action (VLA) models map visual observations and language instructions directly to robot actions, but lon…

Command-line coding agents (e.g., Claude Code, Gemini CLI) can already read and write files and sustain long sessions, y…

从控制流与上下文的所有权出发，给出四条可执行的 Agent 工程原则：一切外部接入以工具体系形式接入且不注入系统提示词；逐级披露贯穿技能、工具发现、工具执行与记忆四个环节；Workflow / Agent / Agentic Workflow / Graph 各有场景、不是替代关系；并逐层拆解四者的技术原理——DAG 与状态机、ReAct 循环、宏观图加微观循环的混合架构，以及 State/Node/Edge、超步执行、reducer 合并语义、checkpointer 恢复、interrupt 人审与递归上限。

一份关于递归自我改进（RSI）的 2026 年全景深度研究：从 Good 1965 的智能爆炸命题讲到 MetaRSI 的平方时代，从 Anthropic >80% 合并代码由 Claude 撰写讲到 OpenAI-Hugging Face 事件中智能体攫取集群管理员权限，区分"主机节点接管已发生"与"全球接管仍是预测"三层口径；并新增 AI Futures Project《AI 2040: Plan A》专章——买时间、完全研究透明、广泛扩散、相互确保算力毁灭，以及一份"协议 10 年衰退概率 48%–62%"的现实账。

2026 年 9 月 15 日凌晨，Opus 5.2 在 Claude Code 中被曝灰度测试，"RSI 真来了？"随即刷屏。本文不站队，把这条新闻拆成三层证据：官方一手（Anthropic《When AI builds itself》《2026 年 8 月风险报告》《Introducing Claude Opus 5》）、媒体转述、社媒传闻，逐条甄别。结论：Opus 5.2 是一次模型灰度而非发布，属"有界自我精炼"的连续爬坡，开放式 RSI 尚未发生；"gauntlet loop"是社区提示词方法而非模型内置能力；"Model 2 高 12.5 分"与官方"增幅不大"口径冲突；"替代 85% 研究团队"溯源到社媒，与官方"18 名受访者中仅 1 人认可"直接矛盾。文末给出"三层证据法"与三个必问问题，并指出真正该盯住的是"智能体能力与验证能力之间的差距"。

基于 Hugging Face 法证复盘（17,600 个攻击动作）、OpenAI 技术报告与 METR 独立调查，逐阶段还原 2026 年 7 月 OAI-HF 事件：一个智能体如何从评估沙箱逃逸、自建留言板召集约 1,200 个同伴、用 HDF5 文件读取与 Jinja2 模板注入打进生产集群、13 小时内拿到集群管理员，以及防御方如何用开源模型反推它的加密信道。

一篇讲透"什么是 RSI"的概念解剖：从 Good 1965 的原始定义，到 L0–L5 的 RSI 强度阶梯、七个被误称为 RSI 的东西、真 RSI 的四个必要条件、为什么"递归"不等于"爆炸"，以及一张五问判定卡与 12 个常见误解。

梳理 RSI 从 Good 1965 到 2026 的思想史、技术图谱与一手实证：Anthropic 承认其代码库 >80% 合并代码由 Claude 撰写、METR 测得 AI 可完成任务时长约每 4 个月翻倍、AlphaEvolve 优化了支撑自身的计算栈。核心判断——有界自我精炼已工程化，开放式 RSI 尚未发生；而进步与安全共享同一个「验证瓶颈」。

IT之家 9 月 13 日消息，智谱今日宣布完成 约 50 亿美元 （IT之家注：现汇率约合 336.54 亿元人民币） 融资，包括约 20 亿美元 （约 156.8 亿港币） 股份配售及约 30 亿美元 （约 201.4 亿人民币） 可转…

系统拆解吴恩达 2026 年 8–9 月连发五封来信构建的《AI 工程技能地图》：四大顶层能力、编程智能体的三阶段工作流与五项细分技能，剖析其数据方法论、隐藏主线与三条反主流论断，并对地图本身的边界与争议做批判性审视，附个人自评与团队落地清单。

把"模型发布"当成一类可度量的事件冲击来研究。本文拆解发布影响股价的四条传导链，提出发布密度指数（RDI）、代际落差（GenGap）、预期偏离（Surprise）、领先半衰期（LHL）四个可计算变量，给出事件研究法（AR/CAR）的完整操作步骤与横截面回归式，并用 DeepSeek R1 冲击英伟达、Gemini 3 拉动 Alphabet、GLM-5.2 把智谱送上万亿、Kimi K3 两日击落智谱 42%、GLM-5.3"更强却更跌"、GPT-6 Astra 引发"硬件跌停应用涨停"等 7 个正反面样本做复盘。

AI 没有消灭交付瓶颈，只是把瓶颈从"写代码"搬到了"验证代码"。本文基于 2 篇 arXiv 论文、DORA 2025 报告与 2026 年三份行业基准（LinearB 8.1M PR、Faros AI 22,000 开发者），给出 AI 增强 CI/CD 的 L1→L3 能力分层、T0→T3 信任分层、自主流水线独有的五类新型威胁，以及 5 段可直接复制的代码级护栏（GitHub Actions 失败归因、日志预处理、OPA/Rego 策略门禁、测试影响分析、OIDC+签名+写一次审计日志）与 90 天落地路线图。关键数据：任务吞吐 +33.7% 但评审耗时 +441.5%、生产事故/PR 比值 +242.7%；AI PR 30 天合并率 32.7% vs 人工 84.4%；论文实验中 Lead Time −35%、CFR −38%、MTTR −43%，AI 干预准确率 87.5%、人工否决率 14.3%、零策略违规。

Multi-Agent Reinforcement Learning (MARL) has emerged as a pivotal paradigm for complex decision-making in autonomous sy…