Introducing LifeSciBench

Agentic AI systems are increasingly adept at scientific tasks, but their utility in life science research hinges on their ability to manage complex, real-world scenarios. Traditional benchmarks often fall short, focusing on narrow domains or isolated skills, thereby failing to capture the full spectrum of research-level work where scientists interpret incomplete evidence, reconcile conflicts, and make difficult decisions under uncertainty.

LifeSciBench addresses this by offering 750 expert-authored tasks, spanning seven workflows and biological domains. These tasks are crafted by practicing life scientists with Ph.D.-level training and direct experience in drug discovery. The benchmark measures how well AI systems support realistic research rather than just answering biology questions, mirroring the complexity of actual scientific work through tasks that require multiple reasoning and decision-making steps.

Each task is structured as a request to a knowledgeable collaborator, including a scientific prompt, relevant context, and requiring a free-response answer. Expert-written rubrics, with an average of 25 criteria per task, evaluate not only scientific correctness but also the detail, justification, caveats, and formatting expected by scientists. This granular assessment reflects how scientific work is evaluated in practice, often prioritizing the validity of the process and usefulness for research decisions over just the final answer.