Scholarly journals have served as the written record of scientific advancement since 1665, offering static text for people to read. A team of Stanford Medicine researchers, led by Jiacheng Miao and James Zou, aims to change that dynamic with a new tool called Paper2Agent, which converts research papers including text, figures and data into an interactive AI agent capable of discussing the paper and interacting with agents built from other papers. Paper2Agent is described in Nature.
Zou traced the idea to how knowledge has always been recorded. “For essentially all of human history, the way that we represent knowledge is in the form of these very passive artifacts,” he said. “In old times people carved knowledge into stones, and now we type knowledge into words on pages—but in some sense pages aren’t that much better.” He described the project as “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?”—something closer, he said, to a virtual author capable of explaining and extending its own findings.
Turning a paper into an agent starts with a group of AI “worker agents” that examine a published paper along with its code and data, then attempt to reproduce the original research from scratch inside a virtual environment. That reenactment lets the agents capture details a reader would otherwise have to dig out manually, from reagents to experimental setup and execution. The resulting knowledge is stored using a model context protocol, or MCP. “An MCP lets AI essentially represent a paper PDF in a form that’s easy for agents to access, almost like a filing system,” Zou said, with each section of the paper organized into its own folder. Humans still play a role in the process, since manuscripts don’t capture failed experiments or the judgment calls behind experimental setups; authors supply that missing context through conversational exchanges with the agent.
The team demonstrated what happens when paper agents talk to each other by converting two unrelated papers, one on predicting how genetic mutations affect the genome, the other a genome-wide association study of ADHD risk. Working together, the two agents flagged a molecular variant near a gene called MPHOSPH9 linked to increased ADHD risk, a connection Zou said had not previously been reported. “In the past, if there are two research groups that publish two different papers, those two research groups have to somehow find each other,” he said. The eventual goal, he added, is something closer to manuscript speed dating at scale, with millions of paper agents surfacing common ground on their own.
Zou emphasized that attribution remains essential to the system: “It’s still important to attribute the final discoveries and reference them back to original papers and original human authors.” He also said the conditions under which agents collaborate need close guidance to keep their work safe and ethical. The team has so far created more than 100 paper agents, with a longer-term goal of giving most manuscripts one. “Millions of papers are published every year,” Zou said. “There’s enormous potential here.”
Zou traced the idea to how knowledge has always been recorded. “For essentially all of human history, the way that we represent knowledge is in the form of these very passive artifacts,” he said. “In old times people carved knowledge into stones, and now we type knowledge into words on pages—but in some sense pages aren’t that much better.” He described the project as “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?”—something closer, he said, to a virtual author capable of explaining and extending its own findings.
Turning a paper into an agent starts with a group of AI “worker agents” that examine a published paper along with its code and data, then attempt to reproduce the original research from scratch inside a virtual environment. That reenactment lets the agents capture details a reader would otherwise have to dig out manually, from reagents to experimental setup and execution. The resulting knowledge is stored using a model context protocol, or MCP. “An MCP lets AI essentially represent a paper PDF in a form that’s easy for agents to access, almost like a filing system,” Zou said, with each section of the paper organized into its own folder. Humans still play a role in the process, since manuscripts don’t capture failed experiments or the judgment calls behind experimental setups; authors supply that missing context through conversational exchanges with the agent.
The team demonstrated what happens when paper agents talk to each other by converting two unrelated papers, one on predicting how genetic mutations affect the genome, the other a genome-wide association study of ADHD risk. Working together, the two agents flagged a molecular variant near a gene called MPHOSPH9 linked to increased ADHD risk, a connection Zou said had not previously been reported. “In the past, if there are two research groups that publish two different papers, those two research groups have to somehow find each other,” he said. The eventual goal, he added, is something closer to manuscript speed dating at scale, with millions of paper agents surfacing common ground on their own.
Zou emphasized that attribution remains essential to the system: “It’s still important to attribute the final discoveries and reference them back to original papers and original human authors.” He also said the conditions under which agents collaborate need close guidance to keep their work safe and ethical. The team has so far created more than 100 paper agents, with a longer-term goal of giving most manuscripts one. “Millions of papers are published every year,” Zou said. “There’s enormous potential here.”
Comment