PaperlayerView exact revision on arXiv

Source-linked AI summary

Lingdan-V2: Large language models for clinically grounded reasoning in traditional Chinese medicine

Rui Hua, Yu Wei, Ziri Jin, Zhuo Sun, Kai Chang, Jianan Xia, Zixin Shu, Xiaodong Li, Dongmei Jia, Fei Dong, Runshun Zhang, Jian Yu, Hao Xu, Haibin Yu, Jiansheng Li, Baoyan Liu, Wenjia Wang, Xuezhong Zhou

arXiv:2610.14705v1cs.CLcs.AI

TL;DR

TCM prescription reasoning requires integrating clinical findings, syndrome differentiation, treatment principles, herbs, and dosages, yet existing general medical models and frameworks provide limited coverage. Lingdan-V2 combines TCM and clinical-data pretraining, reasoning supervision, and prescription-centered reinforcement learning. Its authors report prescription-related improvements across model scales, benchmark difficulties, and retrospective clinical evaluations.

  • Problem

    General medical models and evaluation frameworks provide limited coverage of Chinese-language TCM reasoning, where prescription decisions integrate clinical findings, syndrome differentiation, treatment principles, herbs, and dosage.

    Source: §1 · p. 2

  • Method

    Lingdan-V2 combines continued pretraining, reasoning-oriented supervised fine-tuning, and prescription-centered GRPO across 4B, 8B, and 14B models.

    Source: §1 · p. 2

  • Results

    Prescription-centered GRPO further improves prescription-related performance across model scales, benchmark difficulties, and retrospective clinical evaluations.

    Source: §4 · p. 12

  • Takeaways & Limitations

    The results support combining large-scale TCM clinical data, structured reasoning supervision, and prescription-oriented optimization to develop clinically capable TCM language models.

    Source: §5 · p. 14

    The study is limited to text-based models up to 14B parameters and cannot directly represent tongue images, pulse signals, medical imaging, or other multimodal evidence.

    Source: §4.6 · p. 14

Abstract

from arXiv · show

Large language models have advanced medical reasoning, but clinically grounded reasoning in traditional Chinese medicine (TCM) remains challenging because prescription decisions require integrating domain knowledge, syndrome differentiation, treatment principles, herbal prescription composition, and dosage determination. We develop a TCM language model family comprising models with 4 billion, 8 billion, and 14 billion parameters. The artificial intelligence contribution is a staged clinical alignment framework integrating large scale continued pretraining on TCM and clinical records, complementary reasoning supervision, and prescription-centered reinforcement learning (RL). We further formulate herbal prescription generation as a structured RL problem with prescription-level feedback, enabling direct optimization of clinical decision outputs beyond knowledge reproduction and generic instruction following. The engineering application is TCM clinical decision support, particularly herbal prescription recommendation. We evaluate the models on a public benchmark containing 25,620 expert-verified items across 13 TCM knowledge and clinical reasoning tasks and on a private retrospective dataset of 4,348 spleen and stomach disease prescription cases. The 14 billion parameter reinforcement aligned model achieves a prescription generation F1 of 37.0 on the public benchmark, compared with 32.3 for DeepSeek-R1 with 671 billion total parameters, and a herb-level F1 of 52.8 on the clinical dataset with improved dosage agreement. Reinforcement post-training consistently improves prescription recommendation across all scales, while physician-source analysis shows that expert heterogeneity affects reward stability. These findings demonstrate a scalable pathway from domain grounding and reasoning supervision to clinically oriented prescription decision support.

1. Introduction

TCM prescription decisions require integrating clinical findings, syndrome differentiation, treatment principles, herb composition, and dosage, yet general medical models and benchmarks provide limited coverage of this setting. Lingdan-V2 combines domain training, reasoning supervision, and prescription-centered reinforcement learning, and is evaluated across benchmark and retrospective clinical settings.

1. Introduction

  • Existing general medical models and evaluation frameworks provide limited coverage of Chinese-language, TCM-specific reasoning.Prior Chinese medical and TCM models and benchmarks address domain-specific language, knowledge, and clinical tasks, but structured prescription optimization remains insufficiently studied.Source: §1 · p. 2
  • TCM prescription recommendation requires integrating heterogeneous clinical evidence into syndrome differentiation, treatment principles, herb combinations, and dosage decisions.The paper frames prescription generation as a test of linking domain knowledge to coherent diagnostic and therapeutic decision making.Source: §1 · p. 2
  • Prescription decisions are not addressed by domain-data scaling alone: supervised learning imitates observed responses, while prescription-level feedback offers a verifiable alignment signal.Herb composition and dosage patterns can be normalized and compared with reference prescriptions.Source: §1 · p. 2
  • Lingdan-V2 pairs 4B, 8B, and 14B Qwen3 models with continued TCM pretraining, reasoning-oriented SFT, and prescription-centered GRPO.Continued pretraining uses 16.62 billion TCM-related tokens; SFT includes 240,000 TCM reasoning samples distilled from DeepSeek-R1.Source: §1 · p. 2, Figure 1 · p. 3
  • The study evaluates Lingdan-V2 on the LingLan benchmark and a private retrospective prescription dataset, analyzing performance across scales and identifying physician heterogeneity as a source of reward instability.The authors report broad TCM performance and consistent prescription gains.Source: §1 · p. 2

2. Materials and Methods

Lingdan-V2 is a 4B, 8B, and 14B model family trained through continued pretraining, supervised fine-tuning, and prescription-centered GRPO. The methods describe its data sources and processing, reward design, and benchmark and retrospective clinical evaluations.

2.1. Study Design and Model Overview

  • All three model scales follow the same CPT, SFT, and prescription-centered GRPO sequence, producing paired supervised and reinforcement-aligned checkpoints.The 14B models are used for primary evaluation; the 4B and 8B pairs characterize changes associated with GRPO.Source: §2.1 · p. 3, §2.4 · pp. 5–6

2.2.1. Continued Pretraining Corpus

  • The CPT corpus contains 16.62 billion tokens, while the consolidated SFT corpus contains 7.73 million samples and 2.08 billion tokens.SFT draws on instruction data, clinical reasoning, extraction, diagnosis completion, and prescription-generation tasks.Source: §2.2.1 · p. 3, §2.2.2 · p. 4

2.2.3. GRPO Training Data

  • The GRPO source pool contains 24,212 clinical prescription prompts, and training uses a 40,000-prompt stream sampled from physician-style and spleen-stomach sources.Because requested sampling sizes exceeded available source sizes, records were repeatedly sampled when constructing the stream.Source: §2.2.3 · p. 4

2.3.1. Clinical Data Standardization and De-identification

  • Clinical records are mapped across institution-specific schemas, normalized, de-identified, and organized into longitudinal sequences; shared quality control includes deduplication and screening.The clinical processing pipeline also standardizes and integrates diagnoses, medications, laboratory results, notes, and treatment records.Source: §2.3.1 · p. 5, §2.3 · p. 5, §2.3.2 · p. 5

2.5.1. Prescription-centered Composite Reward

  • GRPO optimizes prescription generation with group-relative advantages and a composite reward based on herb agreement, output format, and completion length.Herbs are normalized and matched one-to-one against references; the prescription reward is a normalized herb-level F1 score.Source: §2.5 · p. 6, §2.5.1 · p. 6

2.6.2. LingLan Benchmark Evaluation

  • LingLan contains 25,620 expert-verified items across 13 tasks, and LingLan-Hard is a 5,200-item high-difficulty subset.Models are compared under the same prompting, inference, and scoring protocols; six Lingdan checkpoints assess changes across scales.Source: §2.6.1 · pp. 6–7, §2.6.2 · p. 7

2.6.3. Private Clinical Prescription Evaluation

  • The private evaluation covers 4,348 spleen-stomach disease cases, assessing generated herb composition and dosages with herb-level and dose metrics.Matched herbs are aligned one-to-one for dosage evaluation; dose MAE is calculated over matched pairs.Source: §2.6.3 · p. 7

2.6.4. Inference and Scoring Protocol

  • Benchmark evaluation uses zero-shot task-specific prompts and standardized generation settings; the Overall Score averages 16 primary metrics, excluding dose MAE.The study also states that model checkpoints, training configurations, and evaluation resources will be released to support reproducibility.Source: §2.6.4 · p. 7, §2.7 · p. 7

3. Results

Lingdan-V2 was evaluated across LingLan benchmark tasks, model scales, and a private retrospective prescription set. Lingdan-14B-R1 led overall benchmark comparisons, while prescription-centered GRPO was associated with prescription-F1 gains across scales and stronger herb-level agreement in clinical cases.

3.1. Benchmark Performance

  • Lingdan-14B-R1 scored 67.1 vs 64.6 for DeepSeek-R1 on full LingLan and 45.3 vs 38.9 on LingLan-Hard, leading by 2.5 and 6.4 points.It also exceeded its Qwen3-14B backbone, which scored 60.8 and 31.4, respectively.Source: §3.1 · p. 8, Figure 3 · p. 8, Table 3 · p. 10
  • On Chinese Patent Medicine Knowledge, Lingdan-14B-R1 exceeded DeepSeek-R1 in single-choice accuracy (78.1/38.3 vs 74.3/31.0) and multiple-choice F1 (87.7/76.6 vs 87.0/72.4).For diagnostic and therapeutic reasoning, syndrome-differentiation F1 was 12.1 full and 2.4 Hard; treatment-principle F1 was 15.5 and 2.1.Source: §3.1 · p. 9
  • Prescription-generation F1 reached 37.0 full and 24.9 Hard, exceeding the strongest external results of 32.7 and 20.1, respectively.Dosage cosine similarity was 35.4 and 23.7, above GPT-5's 31.4 and 19.7.Source: §3.1 · p. 9, Figure 4 · p. 9

3.2. GRPO-associated Changes Across Model Scales

  • Across all three model scales, prescription-recommendation F1 rose on both splits; at 14B it moved from 30.9 to 37.0 full and 17.4 to 24.9 Hard.Overall Score also increased at every scale and split, with larger absolute improvements on LingLan-Hard for both metrics.Source: §3.2 · p. 9, Figure 5 · p. 10

3.3. Herbal Prescription Recommendation

  • On the private spleen-stomach set, 14B-R1 reached herb-level F1 52.8, dose cosine similarity 50.6, and dose MAE 4.369, but dosage changes varied by scale.Herb-level F1 gains were primarily associated with higher recall, while precision remained similar.Source: §3.3 · p. 11, Table 5 · p. 11

3.4. Comparison with Existing TCM-oriented Language Models

  • Against domain baselines, 14B-R1 exceeded ShizhenGPT-32B by 7.3 and 9.0 Overall Score points and led the private-set comparison in herb F1, dose cosine similarity, and dose MAE.It scored highest or tied highest on 15 of 16 LingLan-Hard metrics; the exception was TLE accuracy, led by BianCang-14B.Source: Table 4 · p. 11, Table 5 · p. 11, §3.4 · p. 12

4. Discussion

Lingdan-V2 combines TCM clinical data, reasoning-oriented supervision, and prescription-centered reinforcement post-training. The discussion reports gains in prescription-related performance across model scales and evaluation settings, while identifying expert heterogeneity as a challenge for reward stability.

4.1. Domain Grounding and Reasoning Capability Formation

  • Lingdan-V2 uses 16.62 billion TCM-related continued-pretraining tokens, including approximately 10.4 billion clinical tokens, alongside TCM-specific and general reasoning supervision.The SFT stage used approximately 240K TCM reasoning samples and 110K general reasoning samples; full-parameter training entails higher computational costs than parameter-efficient alternatives.Source: §4.1 · p. 12

4.2. Prescription-centered GRPO and Task-specific Alignment

  • Prescription-centered GRPO improved prescription recommendation across 4B, 8B, and 14B models on LingLan and LingLan-Hard, with the largest 14B gain in prescription generation.Overall Score also rose on both evaluation splits, while clinical-record information extraction decreased slightly in the primary 14B lineage.Source: §4.2 · pp. 12–13

4.3. Physician-source Heterogeneity and Reward Stability

  • Single-physician training pools achieved the strongest GRPO reward improvement; optimization weakened as the pool expanded to 5, 10, and 20 physicians.The passage attributes the tension to diverse but clinically reasonable prescription targets making reference-based rewards less consistent; hyperparameter changes did not readily resolve the 10-physician pattern.Source: Figure 6 · p. 13, §4.3 · p. 13

4.4. Clinical Interpretation of Prescription Composition and Dosage

  • The clinical evaluation found closer agreement with reference dosage patterns in both relative structure and absolute amounts, complementing herb-level matching.Dose cosine similarity reflects relative dosage structure, while dose MAE measures absolute deviation.Source: §4.4 · p. 13

4.5. Evolution from Lingdan-V1 and Positioning among Existing TCM Language Models

  • Lingdan-V2 expands Lingdan-V1’s training-data volume to approximately 1.9% of the V2 total and combines domain grounding, reasoning supervision, and prescription-centered reinforcement training.The discussion positions this evolution as extending domain adaptation toward systematic optimization and evaluation of clinically oriented TCM reasoning and generation.Source: §4.5 · pp. 13–14
  • Lingdan-14B-R1 showed broader, more balanced performance across TCM capabilities than evaluated TCM-oriented models and maintained its advantage on the private spleen-stomach evaluation.Some existing models retained advantages on individual tasks.Source: §4.5 · p. 14

4.6. Limitations and Future Work

  • The study is limited to text-based models up to 14B, which cannot directly represent tongue images, pulse signals, medical imaging, or other multimodal clinical evidence.Future directions include heterogeneous clinical modalities, larger architectures, and world models representing disease progression, interventions, and outcomes.Source: §4.6 · p. 14

5. Conclusion

Lingdan-V2 combines large-scale TCM domain adaptation, reasoning-oriented supervision, and prescription-centered reinforcement post-training in a multi-scale model family. Lingdan-14B-R1 performs strongly across the LingLan benchmarks and clinical prescription recommendation, while physician-source heterogeneity affects reward stability.

  • Lingdan-14B-R1 achieved strong, balanced performance on LingLan and LingLan-Hard, with clear improvements in prescription generation and retrospective clinical prescription recommendation.Source: §5 · p. 14
  • The study identifies physician-source heterogeneity as an important factor affecting reward stability in clinical reinforcement learning.Source: §5 · p. 14
  • The authors conclude that combining large-scale TCM clinical data, structured reasoning supervision, and prescription-oriented optimization supports development of clinically capable TCM language models.Source: §5 · p. 14
  • Future work will explore larger architectures, multimodal inputs, and world models representing longitudinal disease progression, interventions, and clinical outcomes.Source: §5 · p. 14

CRediT authorship contribution statement

The contribution statement assigns conceptualization and supervision across named authors, with data curation, validation, and other research roles distributed among the team.

Funding

The work reports support from national, Beijing, and central-university funding programs, and states that funders had no role in the study or publication decision.

  • The work was partially supported by national research and science foundations, the Natural Science Foundation of Beijing, and central-university funds.The statement lists grant numbers 2023YFC3502604, L232033, 82374302, 82274352, 82505376, U23B2062, and 2025JBZX064.Source: Funding · p. 15
  • Funders had no role in study design, data collection, analysis or interpretation, manuscript preparation, or the decision to submit.Source: Funding · p. 15

Declaration of competing interest

The authors declare no competing financial or nonfinancial interests.

A. Public Training Data Sources

Table A.1 lists the public datasets used to train Lingdan-V2.

A. Public Training Data Sources

  • Table A.1 presents the public datasets used for Lingdan-V2 training.Source: Table A.1 · p. 17

B.1. Group-relative Policy Optimization

The appendix describes group-relative policy optimization and a soft completion-length penalty.

B.1. Group-relative Policy Optimization

  • The group-relative advantage uses groups of eight sampled completions and their composite rewards, with within-group reward mean and standard deviation specified.Source: §B.1 · p. 18
  • The clipped GRPO objective applies no KL regularization and uses clipping parameters epsilon_low = 0.20 and epsilon_high = 0.25.Source: §B.1 · p. 18

B.2. Length Penalty

  • The length penalty stays zero through 2,944 completion tokens and becomes increasingly negative beyond that point, with Lmax = 3,072 and Lcache = 128.Source: §B.2 · p. 18

C. Complete LingLan Results for Smaller Models

The supplied captions describe LingLan task-level results and figures for information extraction and prescription recommendation, but do not provide score values.

C. Complete LingLan Results for Smaller Models

  • Table C.1 reports full-set and Hard-subset task scores for external models and the 4B, 8B, and 14B Lingdan-V2 R1 models, with a repeated Qwen3-32B scale-boundary anchor.Bold marks the best value within each table and split; lower MAE is better, and the final row reports Overall Score.Source: Table C.1 · p. 19

D. Information Extraction Performance

  • Figure D.1 compares full-benchmark and LingLan-Hard information-extraction F1, defined as the equal-weight mean of clinical-record and classical-text extraction F1.Source: Figure D.1 · p. 20

Prescription recommendation

  • The prescription-recommendation figure shows herb-level F1, dose cosine similarity, and dose MAE for Lingdan-R1, Lingdan-SFT, and Qwen3 models on the private spleen-stomach disease test set.Higher F1 and cosine similarity indicate better agreement; lower MAE indicates lower absolute dosage error.Source: Figure E.1 · p. 21

F. Reward Optimization under the 10-Physician Configuration

Four Lingdan-4B development runs under different hyperparameter settings showed fluctuating rewards and little sustained improvement; the authors link this behavior, alongside physician-source comparisons, to physician heterogeneity.

F. Reward Optimization under the 10-Physician Configuration

  • Across four runs, reward improvement was limited and not sustained, despite transient increases under some settings.The authors relate these trajectories, together with the physician-source comparison, to physician-source heterogeneity.Source: F. Reward Optimization under the… · p. 22
  • The runs varied hyperparameter settings to probe optimization behavior in the 10-physician configuration.Source: F. Reward Optimization under the… · p. 22
  • Figure F.1 shows raw rewards as faint curves and 100-step moving averages as solid curves.Source: Figure F.1 · p. 22

Related papers

Loading 2610.14705v1…