MediHive: A Decentralized Agent Collective for Medical Reasoning

Researchers Xiaoyang Wang and Christopher C. Yang have introduced MediHive, a novel decentralized multi-agent framework designed to tackle complex medical reasoning tasks. Unlike traditional single-agent systems or centralized multi-agent architectures, MediHive employs a peer-to-peer network of LLM-based agents that autonomously self-assign specialized roles, conduct initial analyses, and engage in evidence-based debates. The system features a shared memory pool with iterative fusion mechanisms, allowing agents to detect divergences and locally integrate peer insights over multiple rounds to reach consensus. This decentralized approach specifically addresses the scalability bottlenecks, single points of failure, and role confusion common in centralized designs, particularly in resource-constrained environments.

Empirical results demonstrate MediHive's superior performance, achieving state-of-the-art accuracies of 84.3% on the MedQA dataset and 78.4% on PubMedQA. These scores significantly outperform both single large language models and centralized multi-agent system baselines. The framework's strength lies in its ability to handle interdisciplinary medical problems requiring robust management of uncertainty and conflicting evidence—areas where single-agent systems often falter. Accepted for presentation at IEEE ICHI 2026, this work represents a meaningful advancement toward scalable, fault-tolerant decentralized multi-agent systems for high-stakes healthcare applications, moving beyond theoretical promise to demonstrated performance gains in reasoning-intensive domains.