← All benchmarks

Agent4Sci mean relative improvement

Science · 2026-10-20

Agent4Sci asks frontier coding agents to hill-climb 20 scientific machine-learning benchmarks across 11 domains, starting from a human-verified reproduction of the published best model, for 12 hours on 8 H200 GPUs with held-out test scoring and misconduct auditing. The score is the mean over the 20 benchmarks of the signed relative improvement over the reproduced baseline, in percent. The authors present it as a discovery probe rather than a leaderboard, so it is kept outside the capability fit.

Top models (higher is better)