Prescription Prediction on TCM-BEST4SDT (test) benchmark leaderboard · SOTA2 Research

Prescription Prediction on TCM-BEST4SDT (test)

31.11Herb-F1Dynamic residual H2, 3-seed mean

Evaluation Results

Method
Dynamic residual H2, 3-seed mean2026.07	31.11	21.07	3.8064
Dynamic residual H22026.07	30.9	20.96	3.8229
Learned fixed H22026.07	30.85	20.88	3.7933
Qwen2.5 RAG + candidate-herb reranker2026.07	30.78	20.91	3.7762
Static HGNN / hypergraph encoder2026.07	30.71	20.81	3.7716
Similar-case retrieval2026.07	30.59	21.19	3.5589
Label-wise attention classifier2026.07	30.36	20.75	3.7481
Qwen2.5 candidate-herb reranker2026.07	30.27	20.53	3.8124
BERT + syndrome + treatment multi-label classifier2026.07	30.18	20.64	3.7245
Qwen2.5 + candidate herb list2026.07	29.68	20.07	3.8611
R-GCN / knowledge graph baseline2026.07	29.56	20.18	3.6174
BERT + syndrome multi-label classifier2026.07	29.28	19.97	3.6684
BERT + treatment multi-label classifier2026.07	28.99	19.79	3.6522
Dense retrieval2026.07	28.97	19.84	3.5028
Qwen2.5 RAG generation2026.07	28.69	19.16	3.9188
BERT multi-label herb classifier2026.07	28.15	19.12	3.6031
BM25 retrieval2026.07	27.42	18.61	3.4317
Qwen2.5 few-shot generation2026.07	27.38	18.21	4.0025
Qwen2.5 direct generation2026.07	25.77	17.04	4.1268
Fixed H2 prior2026.07	24.22	15.51	3.2139
Syndrome + treatment herb frequency2026.07	23.48	14.86	3.1652
Treatment-specific herb frequency2026.07	22.19	13.98	3.0875
Syndrome-specific herb frequency2026.07	21.46	13.27	3.0418
Herb frequency2026.07	18.08	10.56	2.901