← All benchmarks

EurekaBench

Science · 2026-09-30

26 partially solved research problems across neuroscience, geophysics, plasma physics, astrophysics, computer science and chemistry, in which an AI agent runs up to four hours of experiments in a simulator environment and submits a mechanism with an executable implementation and experiment log. Per problem, the conditioned score pools predictive-accuracy points on held-out data and scientific-insight points from expert-written yes/no questions, scoring zero unless the mechanism passes every scientific-constraint test; two agentic judges score each submission, their scores are averaged, and the score is the mean over problems.

Top models (higher is better)