Related Experiment Video
Updated: Jan 11, 2026
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
1000
Reinforcement learning for LLM-based explainable TCM prescription recommendation with implicit preferences from small
Xinyu Wang1,2,3, Xiaohe Sun4, Lei Yang1
1School of Artificial Intelligence and Information Technology, Nanjing University of Chinese Medicine, Nanjing, 210023, China.
Chinese Medicine
|November 19, 2025
Summary
This study introduces a novel two-stage framework for Traditional Chinese Medicine (TCM) prescription recommendations, enhancing accuracy and interpretability. The approach leverages knowledge distillation and reinforcement learning for improved clinical decision support.
Area of Science:
- Artificial Intelligence
- Computational Medicine
- Traditional Chinese Medicine
Background:
- Traditional Chinese Medicine (TCM) prescription recommendations require enhanced interpretability and accuracy.
- Existing intelligent decision-support systems often lack transparency in their recommendations.
Purpose of the Study:
- To develop a two-stage training framework for improving TCM prescription recommendation interpretability and accuracy.
- To integrate knowledge distillation and implicit preference-driven reinforcement learning into a compact model.
Main Methods:
- Utilized GPT-4o to parse TCM clinical records for distillation samples.
- Employed Low-Rank Adaptation (LoRA) for fine-tuning the Qwen2.5-7B model to generate explainable outputs.
- Trained a lightweight BART model and used Direct Preference Optimization (DPO) for reinforcement tuning.
Main Results:
- Achieved P@30 of 35.62% and F1@30 of 37.36%, outperforming baselines.
- Knowledge distillation improved generalization and explainability.
- Reinforcement learning further enhanced F1@30 by 2.01%.
Conclusions:
- The proposed approach enhances the quality and transparency of TCM prescription recommendations.
- Offers a strategy for building trustworthy and clinically applicable intelligent TCM decision-support systems.
