Paper2Agent turned 74 of 100 computational biology papers into AI-usable tool sets, letting assistants run published analyses on new data rather than just read the papers.
The system packages a paper’s code, methods and data, tests them against reference outputs, and writes step-by-step instructions so an AI agent can use the tools in the right order.
In those converted papers, the agent answered questions and reproduced results more accurately and more than twice as fast as the same AI working from the original paper and code alone.
Failures mostly came from missing code or data, or software that would not run, highlighting a persistent barrier to reusing other scientists’ work even when it is shared online.
Stanford’s James Zou says the approach could make papers an active form of knowledge and surface collaborators, though outside experts warn low-quality paper-agents could become noisy or annoying.