Mathematical Reasoning on MetaMathQA (Accuracy and Length Analysis)
92.4AccuracyAdaSR-HRPO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| AdaSR-HRPOModel=Qwen3-4B2026.06 | 92.4 | 121.708 | 115.65 | 237.358 | |
| AdaSR-GRPOModel=Qwen3-4B2026.06 | 90.9 | 140.01 | 113.545 | 253.555 | |
| StreamingThinker (SFT)Model=Qwen3-4B2026.06 | 86 | 152.2 | 121.998 | 274.198 | |
| Read-then-thinkModel=Qwen3-1.7B2026.06 | 84.8 | — | 1,281.774 | 1,281.774 | |
| Read-then-thinkModel=Qwen3-4B2026.06 | 82.8 | — | 1,379.088 | 1,379.088 | |
| AdaSR-HRPOModel=Qwen3-1.7B2026.06 | 82.6 | 173.246 | 129.356 | 302.602 | |
| AdaSR-GRPOModel=Qwen3-1.7B2026.06 | 82.3 | 158.712 | 159.285 | 317.997 | |
| StreamingThinker (SFT)Model=Qwen3-1.7B2026.06 | 68.8 | 149.196 | 121.465 | 270.661 |