Let large language models judge each other: multi-agent peer-reviewed reasoning for medical question answering.

Abstract

To enhance the accuracy, interpretability, and robustness of large language models (LLMs) in medical question answering (MedQA). We designed a multi-agent peer-reviewed reasoning method in which multiple LLM agents independently generate chain-of-thought (CoT) reasoning with candidate answers, then act as peer reviewers to evaluate each other's reasoning for factual correctness and logical soundness. The highest-rated reasoning chain is selected to produce the final answer. Experiments were conducted with 5 state-of-the-art LLMs (Llama-3.1-8B, Qwen2.5-7B, Phi-4, DeepSeek-LLM-7B, and GPT-oss-20B) on 3 benchmark datasets: HeadQA, MedQA-USMLE, and PubMedQA. Performance was compared against single-model CoT reasoning and CoT-based majority voting. Peer-reviewed reasoning consistently outperformed both baselines. The best model combination achieved an average accuracy of 0.820 across datasets, exceeding the strongest single model (0.777) and majority voting ensembles (up to 0.789). The method also scaled effectively with more participating models, while peer assessments reliably distinguished high- from low-quality reasoning chains. The proposed multi-agent peer-reviewed reasoning method enables LLMs to act as both solvers and evaluators, yielding superior performance in MedQA. By emphasizing reasoning quality rather than answer agreement alone, this approach improves accuracy, interpretability, and robustness, offering a promising direction for trustworthy biomedical AI systems.