Autonomous LLM-Driven Research - from Data to Human-Verifiable Research Papers

Artificial intelligence (AI) promises to accelerate scientific discovery, but it remains unclear whether AI systems can perform fully autonomous research, and whether they can do so while adhering to key scientific values, such as transparency, traceability, and verifiability. The aim of this study was to develop and evaluate an AI-automation platform that performs transparent, traceable, and human-verifiable scientific research.

To mimic human scientific practices, we built "data-to-paper," an automation platform that guides interacting large language model (LLM) agents through a complete stepwise research process that starts with annotated data and results in comprehensive research papers, while programmatically backtracing information flow and allowing human oversight and interactions. The platform can run fully autonomously (in autopilot mode) or with human intervention (in copilot mode).

In autopilot mode, provided only with annotated data, data-to-paper raised hypotheses; designed research plans; wrote and debugged analysis codes; generated and interpreted results; and created complete, information-traceable research papers. Even though the research novelty of manuscripts created by data-to-paper was relatively limited, the process demonstrated the autonomous generation of de novo quantitative insights from data, such as unraveling associations between health indicators and clinical outcomes. For simple research goals and datasets, a fully autonomous cycle can create manuscripts that independently recapitulate the findings of peer-reviewed biomedical publications without major errors in about 80 to 90% of cases. Yet, as goal or data complexity increases, human copiloting becomes critical for ensuring accuracy and overall quality. By tracking information flow through the steps, the platform creates "data-chained" manuscripts, in which downstream results are programmatically linked to upstream code and data, thus setting a new standard for the verifiability of scientific outputs.

Our work demonstrates the potential for AI-driven acceleration of scientific discovery in data-driven biomedical research and beyond, while enhancing, rather than jeopardizing, traceability, transparency, and verifiability.