A research team developed Lingdan-V2, a family of large language models for clinical reasoning and herbal prescription recommendations in traditional Chinese medicine (TCM). The models have 4 billion, 8 billion and 14 billion parameters. The team tested them on a public benchmark with 25,620 expert-verified items across 13 TCM knowledge and clinical reasoning tasks, and on a private retrospective dataset of 4,348 prescription cases for spleen and stomach diseases.

Lingdan-V2 uses staged clinical alignment: continued pretraining on TCM and clinical records, complementary reasoning supervision, and prescription-centered reinforcement learning (RL), a training method that rewards desired outputs. The team framed herbal prescription generation as a structured RL problem, using feedback on complete prescriptions to optimize clinical decisions beyond reproducing knowledge or following general instructions.

On the public benchmark, the 14 billion parameter model aligned with RL scored 37.0 on prescription-generation F1, compared with 32.3 for DeepSeek-R1, which has 671 billion total parameters. On the clinical dataset, Lingdan-V2 achieved a herb-level F1 of 52.8 and improved dosage agreement. RL post-training improved prescription recommendations across all three model sizes. Analysis of physician sources found that differences among experts affected reward stability.

A research team developed Lingdan-V2, a family of large language models for clinical reasoning and herbal prescription recommendations in traditional Chinese medicine (TCM). The models have 4 billion, 8 billion and 14 billion parameters. The team tested them on a public benchmark with 25,620 expert-verified items across 13 TCM knowledge and clinical reasoning tasks, and on a private retrospective dataset of 4,348 prescription cases for spleen and stomach diseases.

Lingdan-V2 uses staged clinical alignment: continued pretraining on TCM and clinical records, complementary reasoning supervision, and prescription-centered reinforcement learning (RL), a training method that rewards desired outputs. The team framed herbal prescription generation as a structured RL problem, using feedback on complete prescriptions to optimize clinical decisions beyond reproducing knowledge or following general instructions.

On the public benchmark, the 14 billion parameter model aligned with RL scored 37.0 on prescription-generation F1, compared with 32.3 for DeepSeek-R1, which has 671 billion total parameters. On the clinical dataset, Lingdan-V2 achieved a herb-level F1 of 52.8 and improved dosage agreement. RL post-training improved prescription recommendations across all three model sizes. Analysis of physician sources found that differences among experts affected reward stability.