Multi-stage inference problems in genomics and quantitative biology, scored on the 10-problem public set.
Ten simulated analysis problems from the GeneBench-Pro public release, spanning statistical and population genetics, single-cell eQTL mapping, functional and 3D genomics, and clinical and cancer genomics. Each gives the agent a short prompt and messy staged data generated from a known simulation, so the correct answer is fully determined. The agent has to clean the data, choose and check a model, and report a final estimate. A run passes only if every reported value matches the hidden answer within tolerance, and scores are pass rates averaged over 10 runs per problem.