Research watchResearchNew
Nature study turns research papers into AI agents that answer questions about their methods
A peer-reviewed Nature paper from Stanford describes Paper2Agent, which converts a paper and its code into an AI agent; it scored well on the authors' own tests, but new biological findings it suggested are unconfirmed.
Evidence: 3 sources · 1 independent report · 2 research papers · 3 open questions
What happened
- On 16 September, Nature published Paper2Agent, by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou.
- The system reads a paper and its code and builds a Model Context Protocol (MCP) server, a set of tools an AI model can call, then tests and refines it.
- In a batch run, it built agents for 74 of 100 computational biology papers from bioRxiv.
- A first version of the work was posted as a preprint on arXiv in 2025.
Why it matters
Much published research is hard to reuse because the code is difficult to run; if this approach holds up, it could make methods easier to apply, though it also raises new questions about checking AI-run analyses.
What is confirmed and what is claimed
Confirmed
Claimed, not independently verified
- The authors report that agents answered 300 questions drawn from the 100 papers with 91.2% accuracy, ahead of 80.3% for a baseline giving Claude direct access to the code repositories.
- The authors report that 593 of 599 generated tools passed automated validation.
- The authors say linked paper-agents pointed to a gene variant associated with ADHD risk that had not been reported before; this is a computational lead, not a confirmed biological finding.
Paper2Agent starts from a simple problem: a research paper describes a method, but using that method usually means installing someone else's code, finding the right data, and working out undocumented steps. The Stanford team's system automates that work by turning the paper and its code repository into tools an AI model can call, then testing those tools until they run.
What the study shows
The evaluation is the authors' own, published after peer review. In their tests, the resulting agents answered questions about the papers' methods more accurately than a general AI assistant given the same code. In one demonstration, the system built 22 validated tools from Google DeepMind's AlphaGenome paper in about 45 minutes.
Peer review checks that the methods and claims are reasonable; it does not confirm every downstream result. See our guide to preprints, peer review, and press releases.
Caveats
The most eye-catching results, new candidate genes for ADHD risk and psoriasis, came from connecting several paper-agents together. These are hypotheses generated by software, not experimental findings. Nature's news coverage also notes a case where the AlphaGenome agent nominated a different causal gene than the original paper did, with data that could support either answer. That shows both the promise and the risk: an agent can extend a method, but someone still has to judge whether its extension is right. Our guide to reading AI-for-science claims covers this distinction.
What to watch
- Independent groups applying the tool to their own papers.
- Lab follow-up on the gene associations.
- Whether journals or funders start asking for "agent-ready" code.
What we don’t know yet
- How well the approach works outside computational biology and on papers with messier code.
- Whether the new gene associations it nominated hold up in experiments.
- How users can tell when an agent's analysis diverges from what the original authors intended.
Evidence and sources
- Peer-reviewed study Reimagining research papers as interactive and reliable AI agents
- Independent reporting Turning scientific research papers into interactive AI agents
- Preprint (not peer reviewed) Paper2Agent: Reimagining Research Papers As Interactive and Reliable AI Agents