Benchmark
ScienceAgentBench
Evaluate agents that turn scientific data and instructions into working analysis programs.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
102 tasks from 44 publications, covering bioinformatics, computational chemistry, geographical information science, and psychology/cognitive neuroscience.
Methodology
Agents produce self-contained Python programs; task-specific evaluators check execution and scientific outputs. The repository released a verified split and corrected artifacts on April 30, 2026.