Clinical Diagnosis on MedQA (OOD)
20EMDeepSeek-V4-Pro
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DeepSeek-V4-ProModel Type=General-purpose LLM2026.05 | 20 | 29.2 | 24.6 | 24.48 | |
| GPT-5.2Model Type=General-purpose LLM2026.05 | 12 | 25.2 | 18.6 | 20.25 | |
| Llama3-Med42-70BModel Type=Medical LLM2026.05 | 7.5 | 19.9 | 13.7 | 11.25 | |
| AutoRefineBackbone=Qwen-2.5-3B, Approach=RAG+RL2026.05 | 7 | 12 | 9.5 | 20.56 | |
| IGPOBackbone=Qwen-2.5-3B, Approach=RAG+RL2026.05 | 7 | 13.1 | 10.05 | 22.2 | |
| C-MIGBackbone=Qwen-2.5-3B2026.05 | 6.5 | 15.2 | 10.85 | 24.72 | |
| AutoRefine-EmbeddingBackbone=Qwen-2.5-3B, Approach=Clinical Sparse Reward2026.05 | 6 | 11.9 | 8.95 | 22.68 | |
| AutoRefine-HardSearchBackbone=Qwen-2.5-3B, Approach=Clinical Sparse Reward2026.05 | 5.5 | 13.8 | 9.65 | 21.79 | |
| AutoRefine-HardDocBackbone=Qwen-2.5-3B, Approach=Clinical Sparse Reward2026.05 | 5.5 | 11.4 | 8.45 | 20.77 | |
| Search-R1Backbone=Qwen-2.5-3B, Approach=RAG+RL2026.05 | 3.5 | 11.9 | 7.7 | 19.6 | |
| AutoRefine-ICDTreeBackbone=Qwen-2.5-3B, Approach=Clinical Sparse Reward2026.05 | 3 | 12.6 | 7.8 | 22.62 | |
| Baichuan2-13B-ChatModel Type=Medical LLM2026.05 | 2.5 | 6.6 | 4.55 | 7.9 | |
| Qwen2.5-3BDescription=directly inference without any training2026.05 | 0.5 | 1 | 0.75 | 0.84 | |
| HuatuoGPT2-13BModel Type=Medical LLM2026.05 | 0.5 | 1.8 | 1.15 | 2.09 |