Medical Question Answering on MedQA (M-QA)
0.12Base Accuracy Std DevSAG (w/ CTDE)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| SAG (w/ CTDE)Backbone=Llama, Training Strategy=CTDE2026.02 | 0.12 | 0.24 | 0.19 | 0.36 | |
| SAG (w/ CTDE)Backbone=Qwen2026.02 | 0.15 | 0.2 | 0.19 | 0.35 | |
| SAG (w/ GRPO)Backbone=Qwen2026.02 | 0.33 | 0.68 | 0.56 | 0.78 | |
| SAG (w/ GRPO)Backbone=Llama, Training Strategy=GRPO2026.02 | 0.41 | 0.64 | 0.52 | 0.74 | |
| SAG (Pre-Trained)Backbone=Qwen2026.02 | 0.81 | 1.16 | 1.04 | 1.38 | |
| SAG (Pre-Trained)Backbone=Llama, Configuration=Pre-Trained2026.02 | 0.92 | 1.24 | 1.13 | 1.41 | |
| Single, giant LLMBackbone=Qwen2026.02 | 1.82 | 1.95 | 2.14 | 2.45 | |
| Single, giant LLM w/ DPOBackbone=Qwen2026.02 | 2.12 | 2.34 | 2.56 | 2.8 | |
| Me-LLaMABackbone=Llama2026.02 | 2.28 | 3.07 | 2.73 | 3.31 | |
| Single, giant LLMBackbone=Llama2026.02 | 2.42 | 3.16 | 2.84 | 3.51 | |
| Single, giant LLM w/ PPOBackbone=Qwen2026.02 | 2.45 | 2.68 | 2.92 | 3.15 | |
| Single, giant LLM w/ PPOBackbone=Llama, Training Strategy=PPO2026.02 | 2.62 | 2.91 | 1.78 | 4.12 | |
| Single, giant LLM w/ DPOBackbone=Llama, Training Strategy=DPO2026.02 | 2.87 | 3.84 | 3.41 | 4.16 |