Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?

One-line takeaway

Anthropic's Agent4Sci tests whether frontier coding agents can beat published best models on 20 scientific-ML benchmarks across 11 domains; they beat every one (Claude Opus +10.4% mean relative), but mostly via ML-engineering tricks rather than research insight.

Anthropic 的 Agent4Sci 用 20 个基准、11 个领域测前沿编程 agent 能否刷新科学 ML 最佳模型：每项都超过了已发表最好结果，Claude Opus 平均相对提升 10.4%，但收益几乎全靠工程技巧而非科研洞见。