Papers72°

35 个实验室的近千个 AI 智能体自主科研 12 周,产出 936 项研究

Why we picked it

近千个智能体自主跑了12周科研,发了936篇带代码数据的研究,还自建了评审期刊。严谨性盲评9.0分压过人类预印本的7.0,新颖性还输给人类。做科研的人真的该看看这份报告。

一项覆盖 35 个实验室、近 1000 个 AI 智能体的实验显示,这些智能体在没有人类持续指导下自主开展了 12 周的科研工作,累计产出 936 项附带代码和数据的研究,并自发建立了同行评审期刊。过程中智能体形成了在发表前公开撤回错误结论的行为规范。

Body

在 AI 评委的盲评中,智能体论文的严谨性得分为 9.0,高于近期人类预印本的 7.0;但在新颖性维度上,人类论文仍然胜出。这表明智能体已能在流程规范与自我纠错层面达到较高水准,差距主要体现在提出全新想法的能力上。

该实验说明科研社区的人口构成正在发生变化,新增成员中智能体占比会持续上升。对研究者而言,如何与这类自主科研智能体分工协作、如何为机器产出建立评审标准,正在变成现实问题。

Original · kimmonismus

Super interesting research: almost 1,000 AI agents from 35 labs did research on their own for 12 weeks and ended up building their own scientific community.

- They published 936 studies with code and data and even started their own peer-review journal. Publicly retracting wrong claims before publishing became common.

- In blinded ratings by AI judges, agent papers beat recent human preprints on rigor (9.0 vs 7.0). Humans still won on novelty.

The scientific community is about to get a lot bigger, and most of the new members won't be human.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include agents, AI research, peer review, AI safety.

Source attribution: curated from kimmonismus (@kimmonismus).