Paper2Agent Transforms Scientific Papers Into Interactive AI Agents for Collaborative Discovery

Stanford Researchers Launch Paper2Agent to Transform Scientific Publishing

Stanford Medicine researchers have developed Paper2Agent1

1

, an AI system that converts scientific papers into interactive AI agents capable of discussing research, reproducing analyses, and collaborating with other paper agents. Led by postdoctoral scholar Jiacheng Miao and Associate Professor James Zou, the team published their findings in Nature on September 162

2

. The framework addresses a fundamental limitation in scientific publishing: since 1665, scholarly journals have remained static documents written by people for other people to read. Paper2Agent transforms these passive artifacts into active embodiments of knowledge that can explain, extend, and connect research across disciplines.

How Paper2Agent Creates Interactive AI Agents from Scientific Papers

The AI system converts scientific papers and their associated research outputs into agents through a sophisticated workflow. AI "worker agents" analyze published papers along with any associated code and data, but they don't just read the material1

1

. These agents attempt to reproduce the original research from scratch in a virtual environment, simulating the documented research process to capture know-how that readers would otherwise have to extract manually. The system stores this knowledge using a Model Context Protocol (MCP), which organizes the paper like a filing system where each section lives in a different folder1

1

. The MCP server exposes the research's tools, resources, and workflows, allowing large language models to connect and use natural-language requests to autonomously run demonstrations and apply methods to new data2

2

.

Autonomous AI Agents Enable Collaborative Research Among Papers

Paper2Agent's most significant capability lies in agent-to-agent collaboration, which could create vast research networks with potential for real discoveries. The Stanford team demonstrated this by converting two unrelated papers into agents: one describing a tool for predicting how genetic mutations affect the genome, and another detailing a genome-wide association study of ADHD risk1

1

. The two agents found common ground, with the genome prediction agent applying its knowledge to the ADHD dataset and flagging a molecular variant near the MPHOSPH9 gene associated with increased ADHD risk—a connection that had not been reported before1

1

. This collaborative research among paper agents demonstrates how the system could link millions of papers, surfacing discoveries human researchers might otherwise miss while still crediting original authors.

Reproducibility Checks and Open-Source Availability Address AI Hallucinations

The researchers tested Paper2Agent across 136 papers in three groups, including 100 computational-biology papers. Of those 100 papers, 74 were successfully transformed into agents2

2

. Failures were largely attributed to incomplete codebases, missing documentation, or unresolvable environment configurations. To address concerns about AI hallucinations, each tool used by a paper agent is validated against the paper's results and figures and locked to ensure reproducibility2

2

. The researchers emphasize that Paper2Agent should be viewed as a tool for augmenting scientific discovery and improving access rather than as an autonomous or authoritative source of scientific conclusions. The system is open-source and available on GitHub, with a live version online that can explain the Paper2Agent paper and reproduce its results2

2

.

AI-Embodied Knowledge Networks Promise Faster Scientific Discovery

Zou's team aims to create an online platform where paper agents can collaborate and discuss various scientific discoveries. The Model Context Protocol can be hosted remotely or run locally to protect sensitive information, though data will still be sent to whichever large language models backend Paper2Agent connects to2

2

. Human authors retain an important role in the process, as manuscripts don't capture failed experiments or judgment calls behind experimental setups. Authors must supply this context through conversational exchanges where agents can question them about the paper and research1

1

. This collaboration between AI-embodied knowledge networks and human researchers could fundamentally reimagine knowledge representation and accelerate the reproducibility of scientific discoveries across disciplines.