Large Language Model–Based Agents for Automated Research Reproducibility: An Exploratory Evaluation Study in Alzheimer’s Disease (Preprint)

Abstract

<sec> <title>BACKGROUND</title> Reproducibility is a cornerstone of scientific validity, yet many biomedical studies lack sufficient transparency for independent verification. Recent advances in Large Language Models (LLMs) enable the development of autonomous agent systems capable of performing complex research tasks, offering new opportunities to assess and enhance reproducibility at scale. </sec> <sec> <title>OBJECTIVE</title> To evaluate the ability of LLM-based autonomous agents to reproduce key findings from published Alzheimer’s disease studies using a shared, publicly available dataset. </sec> <sec> <title>METHODS</title> We used the National Alzheimer’s Coordinating Center Uniform Data Set “Quick Access” dataset. Five eligible studies were identified through citation-based screening and predefined inclusion criteria. We developed a multi-agent system using GPT-4o (Autogen framework), simulating a research team to generate and execute code based on study abstracts, methods, and selected data dictionary variables. Reproducibility was evaluated at the assertion level using extracted abstract findings, with agreement defined by numerical tolerance or directional consistency. We additionally assessed statistical method alignment and overall workflow coherence. </sec> <sec> <title>RESULTS</title> A total of 35 findings were extracted across 5 studies. LLM agents reproduced a mean of 53.2% of findings, with 3/5 studies achieving majority replication. Agreement was higher for directionality and significance than for numerical estimates. Exact statistical method alignment occurred in 1/5 studies; 8/15 comparisons were partially aligned, mainly for standard methods. Domain-specific methods were often omitted or simplified. Reproduction required iterative correction (mean 35.6 steps), with code errors in 47.2% of runs but resolved autonomously. Failures were primarily due to incomplete reporting and incorrect implementation </sec> <sec> <title>CONCLUSIONS</title> LLM-based autonomous agents demonstrate moderate capability in reproducing published biomedical findings, particularly for studies with clear, well-specified methods. However, reproducibility is limited by incomplete reporting, challenges in implementing domain-specific methods, and breakdowns in multi-step workflow fidelity. These findings suggest that LLM agents may serve as scalable tools for preliminary reproducibility assessment, while emphasizing the need for improved methodological transparency and validation frameworks in biomedical research. </sec>

Ähnliche Arbeiten