SCIENCE UNDER AUTOMATION
BixBench3 v2
13 models; 20 study tasks; 138 artifacts; best reported score 0.48 for GPT 5.6 Sol.
What this result supports
Artifacts within tasks are dependent; artifact similarity is not binary workflow completion.
Evidence record: preprint. Checked: 2026-10-20.
Where this evidence is used
Computational genomics
- Choose a defensible analysis design: BixBench3 supplies objectives and methodological guidance.
- Execute study-scale analysis: Best reported artifact score is 0.48; artifact success is not workflow success.
- Interpret and reproduce biological results: Artifacts include interpretation; connected reliable completion is not established.
- Frame a novel biological question: The benchmark supplies the research question.
- Resolve confounding and causal alternatives: No dedicated causal-validation score is available.
- Validate in independent biological data: Independent mechanism validation is outside this benchmark.