Medical Question Answering on PubMedQA (Accuracy and Length Metrics)
0.807AccuracyAdaSR-HRPO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| AdaSR-HRPOModel=Qwen3-4B2026.06 | 0.807 | 356.728 | 114.512 | 471.24 | |
| AdaSR-HRPOModel=Qwen3-1.7B2026.06 | 0.793 | 414.206 | 136.642 | 550.848 | |
| AdaSR-GRPOModel=Qwen3-4B2026.06 | 0.792 | 378.214 | 111.842 | 490.056 | |
| AdaSR-GRPOModel=Qwen3-1.7B2026.06 | 0.776 | 433.573 | 143.314 | 576.887 | |
| StreamingThinker (SFT)Model=Qwen3-4B2026.06 | 0.756 | 402.877 | 116.106 | 518.983 | |
| StreamingThinker (SFT)Model=Qwen3-1.7B2026.06 | 0.749 | 414.737 | 120.534 | 535.271 | |
| Read-then-thinkModel=Qwen3-4B2026.06 | 0.615 | — | 331.835 | 331.835 | |
| Read-then-thinkModel=Qwen3-1.7B2026.06 | 0.588 | — | 434.629 | 434.629 |