GeneBench-Pro: Benchmarking Scientific Judgment in Computational Biology

The bottleneck in computational biology has shifted from generating data to making sense of it—analyzing messy, high-dimensional datasets requires a kind of scientific judgment that is hard to formalize.

Existing benchmarks mostly test recall or routine analysis steps, but real research depends on iterating, handling ambiguity, and knowing when a result is actually decision-ready.

GeneBench-Pro is designed to measure this elusive skill: 129 synthetic problems that force an agent to explore data, choose an appropriate analytical path, and close the inferential loop, all while avoiding common benchmark failure modes like multiple defensible answers or numerical insensitivity.