Organizations: School of Data Science, Fudan University, Shanghai, China · School of Life Sciences, Beijing University of Chinese Medicine, Beijing, China · The Affiliated People’s Hospital of Fujian University of Traditional Chinese Medicine, Fuzhou, Fujian, China · Institute of Science and Technology for Brain-Inspired Intelligence, Department of Neurology, Huashan Hospital, Fudan University, Shanghai, China · Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China · Key Laboratory of Computational Neuroscience and Brain-Inspired Intelligence, Fudan University, Ministry of Education, Shanghai, China
Large language models (LLMs) are advancing in multimodal understanding and reasoning, accelerating medical model development. Traditional Chinese medicine (TCM) uses evidence from the four examinations to determine disease location and nature, differentiate syndromes and guide treatment. This diagnostic process constitutes a complete chain of reasoning. We present Fu-TCM, a multimodal TCM LLM trained on data retained after cleaning and filtering an initial collection of approximately 7.12 million examples derived from classical texts, textbooks, medical images, clinical cases and publicly available datasets. The model was further optimized using Bianzheng-Grounded Policy Optimization (BGPO), which explicitly targets syndrome differentiation through rewards for response format, intermediate TCM reasoning fields and final syndrome predictions. Fu-TCM-27B led all six benchmarks spanning TCM reasoning, text and vision, achieving a macro-average accuracy of 81.01% and surpassing the strongest reference model with complete results across all six benchmarks by 7.13 percentage points. Its largest margin was 11.90 points on Fu-bianzheng. Case-level analyses further showed a positive association between intermediate-field accuracy and final syndrome-selection accuracy in most evaluated models. These findings suggest that TCM syndrome differentiation may be computationally learnable and that structured reasoning supervision can improve performance across TCM tasks.
Figures & tables
Figure 1: Fu-TCM framework. a, Processing of multisource traditional Chinese medicine texts, images and clinical cases to construct text-based and image–text question–answer pairs, case-reasoning data and test datasets. b, Two-stage training with TCM domain-specific learning and Bianzheng-Grounded Policy Optimization, using rewards for response format, syndrome prediction and intermediate bianzheng fields to optimize structured bianzheng reasoning. c, Performance comparison between Fu-TCM and leading models across text, vision and TCM reasoning benchmarks. d, Same-question parallel external validation comparing model-only responses, responses from participants answering independently and responses from participants assisted by Fu-TCM.
Figure 2: Benchmark performance of Fu-TCM and reference models. a, Macro-average accuracy across six benchmarks for the eleven models with complete results. b, Paired accuracy differences between Fu-TCM-27B and the strongest non-Fu-TCM model on each benchmark, with matched-item 95% bootstrap confidence intervals based on 20,000 resamples. c, Accuracy profiles across Fu-bianzheng, Fu-text, Ladder text, Shizhen exam, Fu-vision and Shizhen vision; radar axes denote models and radial distance denotes accuracy.
Figure 3: Final syndrome-choice accuracy and intermediate bianzheng field accuracy on the Fu-bianzheng benchmark. a, Final syndrome-choice accuracy of ten models evaluated on the same 311 test cases; error bars indicate Wilson 95% confidence intervals. b, Strict matching accuracy across 30 intermediate bianzheng fields on the same 311 cases; missing fields are scored as incorrect, and error bars indicate case-level bootstrap 95% confidence intervals based on 4,000 resamples. c, Model accuracy for the 15 most frequent syndromes; cell values indicate accuracy (%).
Figure 4: Intermediate bianzheng field accuracy and its association with final syndrome selection on the Fu-bianzheng benchmark. a, Exact-match accuracy across 30 intermediate bianzheng fields for ten models evaluated on the same 311 cases. Each cell shows accuracy (%). b, Association between mean intermediate-field accuracy and final syndrome-choice accuracy across ten models. Numbered points correspond to the model key. The dashed line shows the least-squares fit (Pearson r=0.89, P=0.0006; n=10 models). Within the eight reference models, the association was not statistically significant (Pearson r= − 0.04, P=0.93). c, Case-level logistic-regression fits for Fu-TCM-27B and Fu-TCM-9B. Lines show fitted probabilities and shaded regions denote Wald 95% confidence intervals. The odds ratios per 10-percentage-point increase in strict intermediate-field accuracy were 1.88 (95% confidence interval, 1.35 to 2.63) and 1.52 (1.08 to 2.14), respectively. d, Difference in mean intermediate-field accuracy between cases with correct and incorrect final answers (correct minus incorrect). Points show the mean difference and error bars indicate 95% bootstrap confidence intervals from 4,000 resamples. The intervals excluded zero for nine of the ten models.
Figure 5: Performance across textual and visual categories, external validation and training-objective ablation. a, Accuracy of eleven models across eight Fu-text categories and four TCM-Ladder categories. b, Accuracy of the same models across five Fu-vision categories based on the final item-level classification and eight Shizhen vision categories. Cells show accuracy (%). c, Individual intermediate-field and multiple-choice accuracy in the external evaluation. Intermediate-field accuracy was assessed on 20 syndrome-differentiation cases, and multiplechoice accuracy was assessed on 60 questions. Grey denotes independent answering, blue denotes model-assisted answering and orange denotes Fu-TCM-27B. Participants are ranked separately within each metric. d, Exact-match accuracy for non-zero intermediate bianzheng fields (reference values 1–3) and final syndrome-choice accuracy after domain learning, outcome-only group relative policy optimization (GRPO) and full Bianzheng-Grounded Policy Optimization (BGPO), evaluated on the same 311 cases. Missing or unparseable intermediate-field predictions were counted as incorrect. Error bars indicate case-level bootstrap 95% confidence intervals based on 4,000 resamples for non-zero field accuracy and Wilson 95% confidence intervals for final syndrome-choice accuracy.
Figure 6: Representative case comparisons of TCM diagnostic capabilities. a, Identification of the dominant syndrome from a clinical-text case. b, Recognition of subtle and conflicting tongue features. c, Recognition of compound pulse-waveform characteristics. d, Interpretation of complexion colour, lustre and their pathological meanings. Green and red mark the reasoning and predictions of Fu-TCM-27B and Qwen3.7-Max, respectively; ochre marks the benchmark reference answers.
Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Guangdong Provincial Laboratory of Traditional Chinese Medicine Hengqin, Guangdong, China
School of Computer Science & Technology, Beijing Jiaotong University, Beijing, 100044, China · Tianjin Tasly Digital Chinese Medicine Technology Co., Ltd., Tianjin, 300410, China · Tasly Biopharmaceuticals Co., Ltd., Tianjin, 300410, China +8