TCMAgent: A Multi-Agent Framework for General Traditional Chinese Medicine

Jinsheng Pan, Jiaqi Wei, Xiang Zhang, Chenyu You

Keywords: Large Language Models, Traditional Chinese Medicine, Multi-Agent Systems

Abstract:

A central challenge in artificial intelligence is designing systems that replicate expert cognition in domains where decisions require holistic data synthesis and deliberative reasoning. While large language models (LLMs) have achieved remarkable progress, their monolithic and sequential architectures impose cognitive bottlenecks that limit their ability to reason over multi-modal evidence or resolve competing hypotheses. Recent advances in multi-agent frameworks provide a new paradigm for overcoming these limitations by distributing reasoning across specialized agents and enabling structured deliberation.
We present \textsc{TCMAgent}, a novel multi-agent architecture that operationalizes a distributed and reflective reasoning workflow. Our framework introduces two key innovations: (\emph{i}) \emph{parallel evidence synthesis}, where agents process heterogeneous inputs concurrently to form a unified representation, and (\emph{ii}) a \emph{collaborative deliberation module}, inspired by clinical peer review, in which agents adversarially refine hypotheses to surface trade-offs and converge on robust decisions. This process is further enhanced by an experiential reflection mechanism that learns from historical reasoning traces, enabling continual self-improvement.
We validate \textsc{TCMAgent} on a multi-modal clinical benchmark in Traditional Chinese Medicine (TCM), a canonical domain where expert-level reasoning requires holistic integration of patient data and careful negotiation of conflicting principles. Experiments demonstrate that \textsc{TCMAgent} significantly outperforms strong LLM baselines in safety, coherence, and interpretability of treatment recommendations. These results provide the first empirical evidence that distributed, deliberative agentic architectures can overcome the cognitive bottlenecks of monolithic models, marking a step toward safer and more reliable AI in knowledge-intensive domains.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 2888