Logo image
MediHive: A Decentralized Agent Collective for Medical Reasoning
Conference paper

MediHive: A Decentralized Agent Collective for Medical Reasoning

Xiaoyang Wang and Christopher C Yang
IEEE International Conference on Healthcare Informatics (ICHI)
12 Aug 2026
Featured in Collection :   Drexel's Newest Publications

Abstract

Single-agent large language models (LLMs) and centralized multi-agent systems (MAS) face scalability bottlenecks, single points of failure, and role confusion in medical reasoning. We propose MediHive, a decentralized multi-agent framework built around a passive shared memory pool. Agents self-assign specialist roles, debate conditionally when they disagree, and iteratively fuse peers’ analyses to reach consensus without a central coordinator. On MedQA and PubMedQA with a Llama-3.1-70B-Instruct backbone, MediHive reaches 84.3% and 78.4% accuracy, surpassing single-agent and centralized multi-agent baselines.

Metrics

1 Record Views

Details

Logo image