OpenAI Releases LifeSciBench With 750 Tasks, 19K Rubric Criteria

OpenAI released LifeSciBench with 750 expert-written tasks across seven biological domains and seven scientific workflows. Each task includes a prompt, artifacts, and a rubric with 25 criteria on average. Tasks require reasoning steps averaging four each and 53% need at least one artifact. Models pass only about one in three tasks. The benchmark includes 1,062 artifacts and 19,020 rubric criteria.