Jingfang: An LLM-Based Multi-Agent System for Precise Medical Consultation and Syndrome Differentiation in Traditional Chinese Medicine

1. Introduction

• LLMs’ clinical use in TCM is challenging because consultation and diagnosis require profound professional knowledge and clinical expertise.Source: §1 · p. 1

1.1. Related Work

• Prior work includes TCM-focused LLMs developed with varied training strategies, including SunSimiao, a Qwen2-7B model fine-tuned on high-quality medical data.The related-work discussion also describes TCMChat, ZhongJing, HuaTuo, and ShenNong.Source: §1.1 · pp. 1–2
• TCM diagnosis and treatment rely on multiple rounds of personalized inquiries, symptom and sign analysis, syndrome attribution, and corresponding treatment.The passage describes this sequence as the typical procedure for TCM syndrome differentiation and treatment.Source: §1.1 · p. 2
• Existing TCM LLMs struggle with variable, patient-specific consultations, scarce multi-round data, and treatment recommendations that may lack an explicit syndrome basis.These limitations affect diagnostic precision, treatment precision, interpretability, and reliability.Source: §1.1 · p. 2

1.2. Main Contribution

• JingFang integrates TCM Specialist Agents through MACCM for patient-tailored consultation and uses a trained Syndrome Agent and DSRS for syndrome differentiation and treatment.The system is designed to align with authentic TCM diagnostic and therapeutic scenarios.Source: §1.2 · p. 2

1.3. Organization

• The paper presents framework construction, specialist-agent development, and model training before reporting evaluations, ablations, and case studies.Source: §1.3 · p. 2

2.1. Framework of JingFang

• The framework comprises TCM Consultation, TCM Syndrome Differentiation, and TCM Treatment Recommendation in a cohesive consultation–syndrome differentiation–treatment workflow.The workflow is intended to align with real TCM clinical scenarios.Source: Figure 1 · p. 2, §2.1 · p. 3
• The Consultation Module uses MACCM to coordinate TCM Specialist Agents and a TCM General Agent for adaptive, multi-round information gathering.It aims to conduct patient-based consultations and collect comprehensive information.Source: §2.1 · p. 3
• The Syndrome Differentiation Module fine-tunes a Syndrome Agent on structured TCM data and applies the Ten Questions Song framework for interpretable diagnosis.Source: §2.1 · p. 3
• The Treatment Recommendation Module uses DSRS with sparse lexical and dense semantic retrieval to identify and rank prescriptions from a structured formula dataset.Source: §2.1 · p. 3

2.2. Multi-Agent Collaborative Consultation Mechanism

• MACCM assembles agents from multiple clinical domains, each with specialty knowledge, to emulate expert, adaptive, targeted TCM consultation.A general-purpose TCM agent contributes fundamental knowledge alongside the specialist agents.Source: §2.2 · p. 3

2.2.1. Generation of TCM Agents Team

• The Manager Agent selects a patient-specific specialist team from the chief complaint, adds a general TCM agent, and establishes the team once per consultation.The general agent provides broad TCM knowledge to support comprehensive consultation.Source: §2.2.1 · p. 3

2.2.2. Construction of Consultation

• Specialists propose one or two targeted follow-up questions, while the general agent generates broader questions that are combined and iteratively refined.The Record Agent summarizes prior consultation answers into the current patient condition used to guide the next round.Source: §2.2.2 · p. 3

2.2.3. Evaluation of Consultation

• The Evaluation Agent scores generated questions for comprehensiveness against the Ten Questions Song and for pertinence to specialist issues and the known patient condition.The two similarity-based scores are aggregated into a final evaluation score.Source: §2.2.3 · pp. 3–4

2.2.4. Optimization of Consultation

• Specialists propose modifications to evaluated questions, and the Optimization Agent integrates their suggestions before the next evaluation turn.Iteration ends at agent consensus or the maximum feedback turns; the final question can initiate formal consultation.Source: §2.2.4 · p. 4

2.2.5. Multi-Round Consultation Driven by Multi-Agent Collaboration

• The Consultation Agent uses the final question and current consultation result to interact with the patient, while the Record Agent updates the case for later diagnosis and treatment.Targeted questions continue until necessary information is collected or the maximum consultation rounds are reached.Source: §2.2.5 · p. 4

• Real-world TCM records contain irrelevant administrative statements and Western medical examination details that can distract models from TCM-relevant symptom patterns.Instrumental diagnostic content is also filtered to maintain focus on TCM-relevant symptoms.Source: §2.3 · pp. 4–5
• LLM-based preprocessing extracts information related to the Ten Questions Song and retains core patient-condition information to structure raw cases.The method aligns JF case outputs with the format of training cases.Source: §2.3 · p. 5
• Two models were fine-tuned on 43,000 cleaned cases: a LoRA-tuned Qwen2.5-7B-Instruct model and a RoBERTa-based classifier with a multi-classification head.Training used four NVIDIA A6000 GPUs.Source: §2.3 · p. 5

• Treatment recommendations integrate the patient’s syndrome, chief complaints, and medical history to dynamically adjust treatment strategies.The stated aim is precise, personalized care.Source: §2.4 · p. 5
• DSRS first filters prescriptions by syndrome-related features, then re-ranks candidates using sparse and dense embeddings fused with Reciprocal Rank Fusion.Filtering uses attributes such as etiology, affected organ, and syndrome type.Source: §2.4 · p. 5
• The system returns the three highest-ranked prescriptions from a structured database covering internal medicine, gynecology, pediatrics, and surgery.Database entries include syndrome types, clinical manifestations, representative formulas, herbs, and therapeutic methods.Source: §2.4 · p. 5

3.1. Datasets and Baseline Models

• The syndrome-differentiation resources comprise 48 contemporary textbooks, 700 historical texts, 80,000 prescription entries, and 43,000 selected training cases.Three PhD students reviewed and revised the data; the prescription database includes 1,000 classic formulas.Source: §3.1 · p. 5
• Comparisons included four open-source TCM models and four general SOTA LLMs, assessed with task-specific criteria.The TCM baselines included ZhongJing-7B, HuaTuo-32B, ShenNong-7B, and SunSimiao-7B.Source: §3.1 · pp. 5–6

3.2. Evaluation of Syndrome Differentiation Precision

• Evaluation used 8,699 test cases spanning 170 syndrome types, with weighted precision, recall, and F1 to account for category imbalance.Qwen2.5-7B-Instruct and RoBERTa were trained as the syndrome-differentiation models.Source: §3.2 · p. 6
• JF outperforms TCM baselines and general SOTA models across all syndrome-differentiation metrics, improving precision by at least 21.1% over Claude-Opus-4.5 and 124% over HuaTuo.The comparison uses weighted precision, weighted recall, and weighted F1-score.Source: Table 1 · p. 6, §3.2 · p. 6
• The authors report that foundation-model choice does not significantly affect syndrome differentiation, while training method and data are key factors.They refer to the ablation experiment in Section 3.4 for validation.Source: §3.2 · p. 6

3.3. Comprehensive Evaluation of Medical Consultation Ability

• Consultation evaluation used 50 randomly selected TCM cases, with an LLM simulating the patient in multi-round interactions.Experts assessed proactiveness, accuracy, practicality, and overall effectiveness; consultation records were also assessed for semantic and lexical consistency.Source: §3.3 · p. 6, §3.3.2 · p. 7
• JF consistently outperformed TCM and general-model baselines across proactiveness, accuracy, practicality, and overall effectiveness in multi-round consultation.The evaluation used pairwise win, loss, or tie comparisons; five experienced TCM experts rated the four dimensions.Source: §3.3 · p. 6, Figure 3 · p. 7

3.3.1. Robustness of Multi-Round Medical Consultation

• JF exceeded 91% accuracy in proactiveness and overall effectiveness, and its average success rate in practicality exceeded 85%.The authors describe the improvements as statistically and practically significant across diverse clinical scenarios.Source: §3.3.1 · p. 7
• JF’s mean win rate exceeded 77%, no lower confidence-interval bound fell below 73%, and mean differences from baselines exceeded 53%.The 95% confidence intervals for all variables excluded zero.Source: Table 2 · p. 7, §3.3.1 · p. 7

3.3.2. Semantic and Lexical Consistency Analysis

• JF exhibited the highest mean semantic similarity, 0.8409, and lowest variance, 0.0743, suggesting superior alignment stability across samples; it also attained the maximum BLEU-1 score, 0.3305.Source: §3.3.2 · p. 7
• Without the TCM General Agent, JF’s semantic similarity was 0.8323 and its BLEU-1 score was 0.1882.The authors report that integrating the agent enhances semantic coherence and linguistic accuracy.Source: Table 3 · p. 8, §3.3.2 · p. 8

3.4.1. Evaluation of TCM General Agent in MACCM

• 89 of 100 evaluations preferred JingFang with the TCM General Agent, versus 11 for the version without it.The evaluation used an LLM-based Patient Agent to simulate multi-round interactions and GPT-4o to compare comprehensiveness and pertinence; consultations averaged 9.09 rounds with the agent versus 4.94 without it.Source: §3.4.1 · p. 8, Table 4 · p. 8

3.4.2. Applicability Assessment of the TCM Syndrome Di

• All tested foundation models improved their syndrome-differentiation performance after training with the proposed method and data; the approach also exceeded GPT-5.2’s 0.6323 precision.The experiment examined representative models from several model families and reports that foundation-model choice was not the primary factor in the improvement.Source: §3.4.2 · pp. 8–9, Figure 4 · p. 9

3.4.3. Ablation Study on the Dual-Stage Retrieval Scheme

• In 85 of 100 cases, DSRS achieved higher retrieval similarity than the Single-Stage approach; it scored lower in the remaining 15.The comparison used a unified recall threshold; the authors report that DSRS reduces semantic interference from similar-symptom, non-target syndromes, supporting more precise and personalized treatment recommendations.Source: §3.4.3 · p. 9, Figure 5 · p. 9

3.5. Case Studies

• Across Damp-Heat Sinking Downward and Spleen–Kidney Yang Deficiency cases, JingFang showed stable, reliable, TCM-consistent capabilities through interactive consultation, syndrome differentiation, and treatment recommendation.The simulations used typical cases extracted from classical TCM literature.Source: §3.5 · p. 9
• The case studies attribute proactive, comprehensive inquiry to MACCM, transparent TCM reasoning to the Syndrome Agent, and customized, syndrome-consistent prescriptions to DSRS.Together, these components form a workflow from consultation through syndrome differentiation to treatment recommendation, described as a practical and interpretable path toward AI-assisted TCM.Source: §3.5 · p. 9

4. Conclusion and Discussion

• JingFang integrates TCM Specialist Agents to emulate authentic diagnostic and therapeutic scenarios, with MACCM supporting targeted consultation and explicit reasoning.The authors describe the consultation process as a foundation for diagnosis and syndrome differentiation.Source: §4 · pp. 9, 11
• Syndrome-differentiation precision improved by at least 124% over existing TCM models and 21.1% over SOTA LLMs.The authors also report that JingFang outperforms current TCM-specific models and SOTA LLMs in multi-round consultation across four professional evaluation dimensions.Source: §4 · p. 11
• The authors identify the complexity of real TCM clinical scenarios as motivating further exploration of multi-agent collaboration and multimodal diagnostic technologies.They frame these directions as practically significant for developing more precise and professional TCM models.Source: §4 · p. 11