Social Chain of Thought Improves Medical Diagnosis with Multi-Agent LLMs

Given the critical nature of medical diagnostic reasoning and the increasing use of LLMs in healthcare, a new multi-agent architecture called Social Chain of Thought (SCoT) has been introduced. SCoT structures multi-agent interaction as a deliberative framework for collaborative LLM reasoning, drawing inspiration from the methodology of medical differential diagnosis. The research evaluates SCoT against single-agent baselines and other scaling methods, demonstrating that its recall advantage, particularly in the hardest diagnostic cases, cannot be replicated by monolithic inference alone. SCoT's success stems from multiple rounds of specialist conversation, which help recover ground-truth diagnoses and converge on higher-recall differentials. This approach provides greater transparency and robustness for complex medical scenarios where integrating diverse specialist reasoning is essential.