Stanford’s Paper2Agent Turns Published Papers Into an Attribution Problem

A team of Stanford researchers has published a framework in Nature that turns scientific papers into AI agents capable of answering questions about their own findings, reproducing their own analyses, and applying their own methods to new data. In one demonstration, two of those paper agents, built from studies that had nothing to do with each other on paper, collaborated and surfaced a gene variant linked to ADHD risk that neither original paper had reported. The team’s own senior author says credit for a finding like that should still trace back to the human authors involved, though it is not obvious which authors that is when two separate papers’ agents make the connection together.

The framework, called Paper2Agent, was published in Nature on September 16, 2026, led by postdoctoral scholar Jiacheng Miao and senior author James Zou, an associate professor of biomedical data science at Stanford Medicine. The Register was first to cover the publication, and Zou answered follow-up questions from the outlet by email.

A paper that can answer its own questions

Zou’s framing for why this matters starts with a complaint about how science has always stored what it learns. “For essentially all of human history, the way that we represent knowledge is in the form of these very passive artifacts,” he told Stanford Medicine’s news office. “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?”

Paper2Agent’s mechanism for doing that is the Model Context Protocol (MCP), the same open standard that has recently shown up in products as different as Google’s Home smart-device control and enterprise tool-calling agents. Paper2Agent analyzes a paper alongside its supplementary materials, code repository, and datasets, then uses a coordinated set of specialist agents to construct an MCP server that exposes the paper’s own tools, workflows, and results. “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.

Once that server exists, any MCP-compatible chat agent, Claude Code among them, can connect to it and query the paper in natural language: ask what a figure means, rerun the analysis behind it, or point the paper’s own methods at a new dataset. Critically, the system does not just trust whatever code it generates. Each candidate tool is run and checked against the paper’s own reported results before it gets included in the final server: according to a technical breakdown of the framework’s validation process, a tool only passes if its “expected files appear and numbers match within 3%” of what the original paper reported. Tools that keep failing that check are dropped rather than shipped.

The evaluation got a lot bigger between the preprint and the journal

Paper2Agent first appeared as a preprint in September 2025, and at that stage the evidence for it was three hand-picked case studies: an agent built from an AlphaGenome genomic-variant paper (22 validated tools, 100 percent accuracy on 15 original and 15 novel test queries), one built from a TISSUE spatial-transcriptomics paper (6 tools), and one built from a ScanPy single-cell-analysis paper (7 tools, a focused subset of the full package). That is a reasonable proof of concept. It is not a scale test.

By the time Nature accepted the paper roughly a year later, the evaluation had grown considerably. According to The Register, the team ran Paper2Agent across 136 papers in three groups, including 100 computational-biology preprints pulled from bioRxiv. Of those 100, 74 were successfully converted into working agents, with the researchers attributing most failures to incomplete codebases, missing documentation, or environment configurations the system could not resolve on its own.

The tool-level numbers, independently confirmed via a separate technical write-up of the paper, are more granular: 593 of 599 proposed tools passed that same-answer-within-3-percent validation check. Tested against a 300-question benchmark built from the bioRxiv corpus, Paper2Agent answered correctly 91.2 percent of the time, ahead of a baseline that just handed a general-purpose coding agent the paper’s raw repository and asked it to figure things out (80.3 percent using Claude Sonnet 4, 86.3 percent using Sonnet 4.6). On a separate set of 10 non-biology papers, including ones behind the tools TabPFN, SAM 2, and SAELens, the system hit 98.1 percent accuracy across 42 execution tasks, and it rejected 100 percent of a batch of deliberately out-of-scope questions rather than fabricating an answer.

Two papers that had never met, working together

The most consequential part of the paper is not the benchmark score. It is what happened when the researchers converted two unrelated studies, one about predicting genetic mutations and one about ADHD risk factors, into agents and let them talk to each other. “In the past, if there are two research groups that publish two different papers, those two research groups have to somehow find each other,” Zou said. Here, the agents found the connection on their own: a molecular variant near the gene MPHOSPH9 that both agents’ combined analysis tied to elevated ADHD risk, a link Zou says had not been previously documented in the literature.

That is a genuinely new result, not a reproduction of an old one, and it happened because two static PDFs became queryable, executable systems that could be pointed at each other’s data. Stanford’s own release counts more than 100 paper agents built to date, and Zou is characteristically unbothered by the scale of what is still ahead. “Millions of papers are published every year,” he told TechXplore. “There’s enormous potential here.”

The attribution problem nobody has solved

Enormous potential is also where the harder question starts. If an AI agent built from Paper A and an AI agent built from Paper B combine their outputs and surface a finding that neither human author claimed, whose discovery is it? Zou’s own answer points toward the humans: “It’s still important to attribute the final discoveries and reference them back to original papers and original human authors,” he said. What he does not say is which original paper, when the finding did not come from either one on its own, and science’s existing machinery, author lists, citation counts, peer review, was not built to answer that either.

The published paper hedges in the same direction on a related point: reliability. Its authors are explicit that Paper2Agent is not a stand-in for scientific judgment. Per the paper’s own text, cited by The Register, the team wants Paper2Agent “viewed as a tool for augmenting scientific discovery and improving access, reproducibility and reuse of papers, rather than as an autonomous or authoritative source of scientific conclusions.” Stanford’s release adds a related, more mundane limitation: a manuscript cannot automatically capture the failed experiments or judgment calls a human researcher made along the way, so a paper agent is only ever as complete as what made it into print, and letting agents collaborate toward new discoveries is something the team says still needs close monitoring for safety and research-ethics compliance, not something they consider solved and shippable at full autonomy.

What is already available

None of that caution has slowed the release. Paper2Agent is open source on GitHub under an MIT license, where it has drawn roughly 2,900 stars and 416 forks. It installs as a skill inside Claude Code or Codex, copying a skills/paper2agent folder into the host agent’s skill directory and invoking it with a /paper2agent command pointed at a paper’s GitHub repository. Pre-built, already-validated MCP servers for the three original case-study papers, AlphaGenome, ScanPy, and TISSUE, are hosted on Hugging Face Spaces for anyone to connect to without running the conversion themselves. There is also a public agent at paper2agent.ai that can discuss the Paper2Agent paper itself, reproduce its own results, and, per Zou, ingest other papers a visitor wants to try it on.

Scale versus oversight

The gap between where Paper2Agent is and where its creators want it to go is the real story here. A year of scaling took the evaluation from three case studies to 136 papers and a genuinely new scientific finding. Getting from 100-odd paper agents to the “millions of papers” Zou envisions is a much bigger jump, and it runs straight into two questions this launch does not settle: how a credit system built around single human authors is supposed to keep up with a tool designed to combine their work, and how much autonomy to give agents that are, by design, capable of finding things their human authors never looked for, the second of which the team itself says still needs close human oversight.