Agent4Sci tests coding agents on 20 scientific ML benchmarks

What the research found

Agent4Sci paper tests frontier coding agents on 20 scientific ML benchmarks; Claude Opus gains 10.4% mean relative improvement, and targeted plug-in skills raise it to 20.1% with code and data public.

Claims and how well the evidence supports them

1. Agent4Sci benchmark spans 20 benchmarks across 11 domains 99%
2. Claude Opus reaches 10.4% mean relative improvement on published best models 99%
3. Plug-in skills raise Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1% 99%
4. Code and data are available on GitHub and Hugging Face 99%

Evidence

• huggingface.co https://huggingface.co/papers/2610.28763
• github.com https://github.com/maszhongming/agent4sci

Actionability rated low.