The authors assembled 20 benchmarks comprising 25 tasks across 11 scientific domains, including physical, life and environmental sciences.
Coding agents improve scientific models mostly through engineering, not insight
Across 20 scientific machine-learning benchmarks, agents exceeded published results, while workflow guidance roughly doubled mean relative gains for two agents.
Big Tech
Ming Zhong · Stacy Li · Nicholas Carlini · Matthew Jagielski
Anthropic Fellows Program · Anthropic
Research Digest··2 min read
Zhong and colleagues introduce Agent4Sci, a testbed for asking whether coding agents can improve scientific machine-learning models under fixed computing budgets.
§