BixBench3: AI agents on research-study-scale computational biology

Zane Koch, Asmamaw T. Wassie and collaborators | arXiv preprint | not peer reviewed | submitted 26 August 2026

Problem: agent benchmarks rarely test end-to-end biological studies on raw data. Contribution: 20 tasks and 138 graded artefacts; the authors report best score 0.48 across 13 frontier models, with degradation on larger data and longer workflows. Significance: exposes orchestration and data-scale limits. Limitations: preprint, author-reported, model/version sensitive and not independently reproduced.