Anthropic 提出 Agent4Sci：代码智能体在 20 项科学 ML 基准上超越已发表最优模型
Anthropic 让 Claude Opus 和 Codex 去刷 20 个科学 ML 基准，全部打败已发表最优结果，加三个技能后提升直接翻倍，做 AI 科研系统的必看。
Anthropic 发布 Agent4Sci 论文，让前沿编程智能体在 11 个领域的 20 项科学 ML 基准上挑战已发表最优模型，结果全部超越。Claude Opus 平均改进 10.4%，且没有任何一次运行被标记为学术不端。分析显示收益几乎全部来自标准 ML 工程技巧，智能体常误判新颖性、闲置算力并陷入小范围参数扫描。加入科研构思、实验管理、定时清单三个轻量技能后，Claude Opus 的提升从 10.4% 升至 20.1%，Codex 从 4.4% 升至 8.1%。模型本身不变，研究脚手架让收益大约翻倍。
Great paper from Anthropic on AI agents doing ML-for-science research.
(bookmark it)
If you build AI-scientist systems, this one is worth your time.
They propose Agent4Sci, which gives frontier coding agents 20 scientific ML benchmarks across 11 domains and asks them to beat the published best model.
They beat it on all 20. Claude Opus improves on the published best by 10.4% on average, and no run was flagged for misconduct.
The gains come almost entirely from standard ML-engineering tricks. The trajectories show agents misjudging what is novel, leaving compute unused, and settling into small local sweeps.
Adding three lightweight skills for research ideation, experiment management, and scheduled checklists raises Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1%, with more varied and more domain-relevant changes.
The models stay the same, and the research scaffolding roughly doubles the gain.
https://t.co/E9oOxkBjiA
Chat with Paper: https://t.co/9B0ITtV9bW
