Mathematical Reasoning on GSM-symbolic P2 (Accuracy and Length Metrics)
88.4AccuracyAdaSR-HRPO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| AdaSR-HRPOModel=Qwen3-4B2026.06 | 88.4 | 196.39 | 131.162 | 327.552 | |
| AdaSR-GRPOModel=Qwen3-4B2026.06 | 87.4 | 209.866 | 126.268 | 336.134 | |
| StreamingThinker (SFT)Model=Qwen3-4B2026.06 | 81.6 | 227.21 | 130.374 | 357.584 | |
| AdaSR-HRPOModel=Qwen3-1.7B2026.06 | 78.8 | 210.046 | 160.21 | 370.256 | |
| HRPO-tokenModel=Qwen3-1.7B2026.06 | 77 | 213.528 | 145.432 | 358.96 | |
| AdaSR-GRPOModel=Qwen3-1.7B2026.06 | 75.8 | 235.784 | 148.704 | 384.488 | |
| HRPO-sentenceModel=Qwen3-1.7B2026.06 | 75.4 | 250.946 | 131.024 | 381.97 | |
| StreamingThinker (SFT)Model=Qwen3-1.7B2026.06 | 64.2 | 229.29 | 132.616 | 361.906 | |
| Read-then-thinkModel=Qwen3-4B2026.06 | 46.2 | — | 1,937.396 | 1,937.396 | |
| Read-then-thinkModel=Qwen3-1.7B2026.06 | 42.4 | — | 1,866.474 | 1,866.474 |