Stanford's Paper2Agent makes a paper answer as an AI agent

A Stanford team has built a tool that packs a research paper, along with its code and datasets, into an AI agent behaving like the paper's on-call author. It is called Paper2Agent, and its creators pitch the output as a kind of virtual corresponding author: a service that answers questions about a study and can rerun its methods on new data.

The test case Nature writes up is AlphaGenome, an AI model that predicts properties of DNA sequences. "Paper2Agent autonomously created an agent for the AlphaGenome paper in about 45 minutes, and the required computing power cost US$14," Nature reports. The resulting agent answered genetics questions with near-perfect accuracy and outscored Biomni, "a tool developed by academic researchers that draws from dozens of databases but scored much lower than the AlphaGenome agent."

Then it did something stranger. The team asked the agent to explain why a single change to a DNA "letter" is associated with "bad" cholesterol. It pointed to a different causal gene than the one the original AlphaGenome paper had pinpointed. According to James Zou, a computer scientist at Stanford University in California and a co-author, AlphaGenome's underlying data supported both hypotheses. Published conclusions, in other words, can be revisited by a paper's own agent without any new bench work.

Underneath, Paper2Agent leans on the Model Context Protocol: "The information is deposited onto a digital platform called an MCP server." Any large language model a scientist prefers can then be pointed at that server to hold the conversation. MarkTechPost's writeup adds that the AlphaGenome build produced its tool set without human intervention, and that in a separate demo three Paper2Agent-built agents were chained on a shared question.

The pitch, from Zou, is that the technology "can help us to reimagine what knowledge looks like in the future." Three of the researchers on our watchlist posted the Nature link within a day of it running.