MEDMIRROR: Towards More Reliable Diagnosis in Traditional Chinese Medicine via Reflexive Interaction and Multi-Agent Collaboration

ACL ARR 2026 January Submission8546 Authors

Keywords: Traditional Chinese Medicine (TCM), Multi-Agent Systems, Syndrome Differentiation, Reflexive Interaction, Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), Multimodal Medical Diagnosis, Diagnostic Inquiry, Explainable AI, Clinical Reliability.

Abstract: Despite recent advancements in specialized Traditional Chinese Medicine (TCM) AI systems, they remain constrained by modal inflexibility, lack of reliability, and insufficient explainability. To address these issues, we propose MedMirror, a framework that focuses on two key components of medical consultation AI systems: diagnostic inquiry and evidence-based explanation. For the diagnostic inquiry phase, we propose User-Centric Reflexive Diagnostic Interaction. It leverages dual agents to perform dynamic, multi-turn inquiries, ensuring that comprehensive evidence is gathered to facilitate highly reliable diagnosis. For the evidence-based explanation phase, we propose Multi-Agent Collaborative Knowledge Synthesis \& Report Generation. This module synthesizes diagnostic data through multi-path parallel RAG, reflexive argumentation, and iterative drafting. It effectively transforms clinical findings into comprehensive, accessible, and evidence-backed reports for users. Experimental results demonstrate that MedMirror achieves superior performance in syndrome differentiation compared to existing baselines. Notably, its reflexive mechanism effectively mitigates information deficiency, while expert meta-evaluation confirms the system's effectiveness in producing high-quality and reliable diagnostic insights.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: multi-agent systems, clinical decision support, LLM agents, clinical dialogue systems, multi-modal dialogue systems, agent coordination and negotiation, retrieval-augmented generation

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings

Languages Studied: Chinese, English

Submission Number: 8546