OpenAI Releases LifeSciBench: Evaluating AI Systems' Capabilities in Real Scientific Research Scenarios

Deep Tide TechFlow reports that on June 18, OpenAI launched a new evaluation benchmark, LifeSciBench, aimed at assessing AI systems' capabilities in real scientific research scenarios. LifeSciBench is based on 750 expert-crafted tasks, covering 7 types of research workflows and 7 biological fields. The tasks were created by 173 researchers with PhD backgrounds and experience in the biotechnology or pharmaceutical industries. This benchmark emphasizes the evaluation of complex scientific research abilities, including evidence integration, experimental design, data analysis, scientific reasoning, and research communication, rather than focusing on single factual questions. Over 79% of the tasks involve multi-step reasoning, with an average of approximately 4 reasoning steps per task, and include 1,062 real research-related data attachments (such as papers, charts, sequence data, and structural files).