Stanford's Paper2Agent converts papers into working AI agents

Stanford Medicine's Paper2Agent, published in Nature on Sept. 16, converts a scientific manuscript along with its data and code into an AI agent that answers questions about the paper and runs its methods on new datasets. A prototype turned 74 of 100 bioRxiv computational biology papers into working agents, with 593 of 599 generated tools passing automated validation.

Across 300 benchmark questions drawn from those 100 papers, the paper agents scored 91.2% accuracy, ahead of a Claude-plus-repository baseline at 80.3%. The team built an AlphaGenome agent in about 45 minutes for roughly $14 in compute and reported it beat the Biomni biomedical agent on genetics questions. Further demonstrations were built from Scanpy and TISSUE.

"An MCP lets AI essentially represent a paper PDF in a form that's easy for agents to access, almost like a filing system," senior author James Zou told the Stanford Report, describing the Model Context Protocol servers that the pipeline constructs. Lead author Jiacheng Miao worked with Joe R. Davis, Yaohui Zhang and Jonathan K. Pritchard on the framework.

The team also pointed two paper agents at each other. A genome-prediction agent applied its methods to a genome-wide association dataset for attention-deficit/hyperactivity disorder and flagged a variant near the gene MPHOSPH9 as linked to ADHD risk, a connection Zou said had not been reported before.

Zou described the ambition in broad terms: "This is an opportunity to fundamentally reimagine what knowledge looks like. Instead of having only passive artifacts, why don't we convert each static record into an active embodiment of knowledge?" He added that it is "still important to attribute the final discoveries and reference them back to original papers and original human authors."