Paper2Agent Turns Research Papers into AI Scientists That Can Answer Questions and Run Experiments

A new framework called Paper2Agent converts conventional research papers, their code, datasets and workflows into interactive AI agents. These agents can answer technical questions, reproduce published analyses, apply methods to new datasets and collaborate with agents built from other research papers. Published in Nature on September 16, 2026, the system could change how scientific knowledge is accessed, verified and reused.

Scientific papers have traditionally been treated as finished documents. Researchers read them, locate the supporting code, download datasets, configure software environments and then spend hours attempting to reproduce or adapt the published method.

Paper2Agent proposes a different model: what if every research paper could become an interactive research assistant capable of explaining its methods, executing its workflows and helping scientists explore new questions?

A research team led by scientists including Jiacheng Miao, Joe R. Davis, Jonathan K. Pritchard and James Zou has introduced Paper2Agent, an automated framework that transforms scientific papers into active AI agents. The work was published in Nature on September 16, 2026.

From Static Papers to Interactive Research Agents

Most scientific publications contain valuable knowledge, but using that knowledge often requires significant technical expertise. A paper may describe an algorithm clearly while leaving researchers to locate its code, install dependencies, understand input formats and reproduce the correct workflow independently.

Paper2Agent attempts to remove much of that friction. It examines a research paper alongside its supplementary material, datasets, code repositories, examples and analytical workflows. It then converts those components into a structured system that can be accessed through natural-language interaction.

The resulting agent behaves like a “virtual corresponding author”. Researchers can ask questions about the paper, request explanations of methods, run supported analyses or apply the published approach to new data.

Rather than merely retrieving paragraphs from a document, the system aims to make the research itself executable.

How Paper2Agent Works

The framework uses a multi-agent pipeline to transform a paper into a functioning AI-accessible research system.

First, it identifies the paper’s associated codebase and relevant computational resources. Specialized AI agents then configure the software environment and extract important methods from the paper and its implementation.

The framework creates tools through the Model Context Protocol, or MCP, which provides a standardized way for AI systems to access external tools, datasets and workflows.

A separate testing agent checks whether the generated tools reproduce expected outputs from the original paper. It compares generated files, numerical results and figures against reference results. Tools that repeatedly fail validation can be excluded from the final system.

This validation layer is important because simply asking a language model to recreate scientific code can introduce errors, incorrect assumptions or what researchers describe as “code hallucination”. Paper2Agent attempts to reduce that risk by linking its tools to tested implementations from the original research.

Each validated tool also retains references to the original code, helping users trace how a result was generated.

Researchers Can Ask Questions Without Writing Code

Once connected to a compatible AI assistant, a paper agent can be used through ordinary language.

A scientist could ask an agent to explain a method, process a dataset, reproduce a figure, interpret a genomic variant or carry out a multi-step analysis without manually writing every command.

This does not mean the agent eliminates the need for scientific judgment. Instead, it lowers the technical barrier between reading a paper and actually using the method described in it.

The Nature team demonstrated the approach using computational biology tools including AlphaGenome, Scanpy and TISSUE. These systems were converted into agents capable of answering technical questions and performing analyses based on their original research methods.

Performance Across Research Tasks

The researchers evaluated Paper2Agent across several groups of papers and scientific tasks.

Among 100 computational biology papers, the system successfully converted 74 into paper agents. It proposed 599 tools, of which 593 passed automated validation.

On a benchmark of 300 tutorial-based questions, Paper2Agent combined with Sonnet 4 achieved an accuracy of approximately 91.2 percent. This was higher than direct repository access using the same model, which achieved approximately 80.3 percent, and another tested configuration that achieved about 86.3 percent.

The framework also reduced average query cost and response time. The researchers reported an average cost of approximately US$0.20 and a response time of 1.6 minutes per query, compared with US$0.38 and 4.3 minutes when the model worked directly with the paper and repository.

Across 42 execution-based tasks involving 10 non-biological computational papers, Paper2Agent achieved approximately 98.1 percent accuracy across repeated runs.

For papers focused more on data and discovery than executable code, its structured resource layer achieved approximately 89 percent accuracy on synthesis-based questions.

These results suggest that the approach is not limited to one specific scientific field, although papers with missing code, unavailable datasets or poorly generalizable scripts remain more difficult to convert.

Paper Agents Can Collaborate With One Another

One of the most significant aspects of Paper2Agent is that individual paper agents can be connected together.

Scientific discoveries often emerge when researchers combine a method from one publication with data or findings from another. Traditionally, this requires human researchers to read several papers, understand different tools and manually coordinate the workflows.

Paper2Agent enables AI agents representing separate papers to exchange information and use one another’s capabilities.

For example, one agent may represent a genomic prediction model, another may represent a single-cell analysis method and a third may contain a dataset or perturbation workflow. Together, they can analyze a research question that no single paper addresses on its own.

In one demonstration, multiple paper agents were used to investigate a disease-associated genetic variant and prioritize a possible causal gene. The system generated a mechanistic hypothesis that would still require experimental validation, but showed how paper agents could combine methods from different studies to support new scientific investigations.

Beyond Search-Based AI

Paper2Agent is different from a conventional chatbot that simply summarizes a PDF or answers questions using retrieved text.

A retrieval-based system may explain what a paper says, but it generally cannot reliably execute the paper’s complete workflow. Paper2Agent is designed to expose not only the manuscript but also its code, datasets, tools, examples and analysis procedures.

This makes the paper an executable research artifact rather than a static publication.

The system can potentially help scientists:

• Reproduce published results
• Apply methods to new datasets
• Understand unfamiliar technical fields
• Test alternative hypotheses
• Compare methods across multiple papers
• Build cross-disciplinary research workflows

The broader idea is that scientific publications could become modular components in an AI-driven research ecosystem.

Important Limitations Remain

Not every paper can be converted into a reliable agent.

The researchers found that some papers failed because executable code was missing, datasets or model files were unavailable, software dependencies could not be configured, or scripts were too specific to the original experiment.

Paper2Agent also depends on the quality of the underlying research materials. If a paper’s code is incomplete or its methods are poorly documented, the resulting agent may have limited functionality.

Furthermore, a validated computational workflow does not automatically guarantee that a scientific conclusion is correct. An agent can reproduce an analysis while still operating within the assumptions, limitations or biases of the original study.

Human researchers remain responsible for interpreting results, checking experimental validity and deciding whether a computational finding deserves laboratory or clinical follow-up.

A New Model for Scientific Communication

Paper2Agent reflects a broader shift in how research may be published and consumed.

Instead of treating a paper as the final destination of scientific communication, the framework treats it as the starting point for an interactive system. A publication could contain not only explanations and results, but also an AI-accessible interface through which other researchers can question, test and extend its work.

Such a model could be especially valuable in fields where methods are technically complex and cross-disciplinary collaboration is difficult. A biologist could interact with a computational method without becoming an expert programmer, while a data scientist could access domain-specific research workflows through natural language.

The researchers describe this as a move toward “agent-native” scientific knowledge, where papers become active participants in the research process rather than passive records of completed work.

Paper2Agent does not replace peer review, laboratory validation or expert judgment. Its importance lies in making published science easier to reproduce, operate and combine with other knowledge. If the approach scales, future research papers may no longer be read only as documents—they may also function as specialized AI collaborators available to scientists around the world.