OpenAI releases LifeSciBench — do phd-level agents fail at real science?
openai introduced lifesicibench, a 750-task benchmark created by 173 phd scientists to test agentic capabilities in real-world workflows, signaling that current ai evaluations may be overestimating model utility in complex research environments.
openai introduced lifesicibench, a rigorous benchmark designed to assess whether ai systems can handle the complexity of real-world life science research, moving beyond simple fact-recall to applied workflows. the dataset comprises 750 expert-authored tasks spanning seven domains, created by 173 phd-level scientists with industry experience, featuring 1,062 artifacts such as pdfs, sequence files, and web references. 79% of tasks require multiple reasoning steps, and 53% demand interpretation of non-text artifacts, graded against rubrics requiring 90% expert agreement. this release highlights a shift in ai evaluation towards multi-step, agent-based reasoning capabilities necessary for drug discovery and biotech research.