Capability benchmark · Biology · Medicine
LifeSciBench evaluates agents on expert life-science research workflows
An expert-authored benchmark of realistic life-science research tasks spanning evidence, analysis, design, validation, translation, and communication workflows.
Summary
OpenAI introduced LifeSciBench, a 750-task evaluation for AI systems handling realistic life-science research requests with supporting figures, documents, sequences, structures, and other artifacts. The benchmark reports expert validation of task relevance and measures partial but still limited frontier-model performance.
AI role
Models interpret research artifacts, reason through uncertainty, make domain-specific judgments, and provide expert-facing answers scored against detailed rubrics.
Narrative role
This adds a high-signal measure of whether models can assist in applied biology and drug-discovery work that requires evidence handling, practical judgment, and uncertainty-aware communication rather than clean-answer biology questions.
Caveat
The benchmark and initial results are reported by the developer rather than peer review; rubric design, model-access conditions, and task selection may limit direct comparison with other scientific-agent evaluations.