Abstract
Medical Decision-Making (MDM) is a multi-faceted process that requiresclinicians to assess complex multi-modal patient data patient, oftencollaboratively. Large Language Models (LLMs) promise to streamline thisprocess by synthesizing vast medical knowledge and multi-modal health data.However, single-agent are often ill-suited for nuanced medical contextsrequiring adaptable, collaborative problem-solving. Our MDAgents addresses thisneed by dynamically assigning collaboration structures to LLMs based on taskcomplexity, mimicking real-world clinical collaboration and decision-making.This framework improves diagnostic accuracy and supports adaptive responses incomplex, real-world medical scenarios, making it a valuable tool for cliniciansin various healthcare settings, and at the same time, being more efficient interms of computing cost than static multi-agent decision making methods.