Medical Question Answering on CMExam
79.27AccuracyModelMerge
Evaluation results
| Method | Links | |
|---|---|---|
| ModelMergeBackbone Model=HuatuoGPT-Vision-7B2025.12 | 79.27 | |
| BiRG-LoRASetting=Qwen, CMExam train2026.06 | 77.62 | |
| MoELoRASetting=Qwen, CMExam train2026.06 | 77.62 | |
| DoRASetting=Qwen, CMExam train2026.06 | 77.55 | |
| Arrow LoRA-weight routerFamily=Adapter library, Params=N/A2026.06 | 77.3 | |
| MedAdapter-style candidate verifierFamily=Adapter library, Params=N/A2026.06 | 77.07 | |
| BiRG-LoRA r16/top4Family=Single adapter, Params=66.59M2026.06 | 76.27 | |
| LoRA r16/alpha32Family=Fixed LoRA, Params=43.65M2026.06 | 76.2 | |
| DoRA r24/alpha48Family=PEFT, Params=66.87M2026.06 | 75.94 | |
| LoRA r4/alpha8Family=Fixed LoRA, Params=10.91M2026.06 | 75.82 | |
| LoRA r24/alpha48Family=Fixed LoRA, Params=65.47M2026.06 | 75.82 | |
| Rank-wise gate without clinical priorsFamily=Rank-wise control, Params=64.88M2026.06 | 75.82 | |
| LoRA-Mixer-style hard-soft routerFamily=Adapter library, Params=N/A2026.06 | 75.52 | |
| MeteoRA-style soft MoEFamily=Adapter library, Params=N/A2026.06 | 75.33 | |
| MoELoRA 4e/top2/r8Family=LoRA-MoE, Params=92.60M2026.06 | 75.32 | |
| OriginalBackbone Model=HuatuoGPT-Vision-7B2025.12 | 69.75 | |
| Qwen3-8B, no SFTFamily=Base model, Params=02026.06 | 68.95 | |
| ModelMergeBackbone Model=Lingshu-7B2025.12 | 68.82 | |
| PSIBackbone Model=HuatuoGPT-Vision-7B2025.12 | 68.45 | |
| OriginalBackbone Model=Lingshu-7B2025.12 | 67.23 | |
| PSIBackbone Model=Lingshu-7B2025.12 | 66.67 | |
| RESTABackbone Model=HuatuoGPT-Vision-7B2025.12 | 58.73 | |
| MoELoRASetting=Llama, CMB train2026.06 | 55.97 | |
| BiRG-LoRASetting=Llama, CMB train2026.06 | 55.63 | |
| DoRASetting=Llama, CMB train2026.06 | 54.97 | |
| RESTABackbone Model=Lingshu-7B2025.12 | 48.3 |