Mathematical Reasoning on AIME 2025 (pass@1, pass@16)
62.01Pass@1POPE (+ hard + guide + 1K easy problems)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| POPE (+ hard + guide + 1K easy problems)Training Mixture=Hard (256) + Easy (1K)2026.01 | 62.01 | 84.31 | |
| SFT (+ hard + 1K easy problems)Training Mixture=Hard (256) + Easy (1K)2026.01 | 61.88 | 83.79 | |
| POPE (+ hard + guide + easy)Training Mixture=Hard (256) + Easy (256)2026.01 | 58.75 | 83.87 | |
| SFT (+ hard + easy)Training Mixture=Hard (256) + Easy (256)2026.01 | 57.19 | 82.5 | |
| POPE (+ hard + guide)Training Mixture=Hard problems only2026.01 | 53.12 | 82.61 | |
| SFT (+ hard)Training Mixture=Hard problems only2026.01 | 49.58 | 81.43 | |
| Qwen3-4B-InstructType=Base model2026.01 | 48.13 | 77.29 | |
| Prefix + rejection-sampled SFTTraining Mixture=Hard problems only2026.01 | 38.12 | 77.62 | |
| Full-oracle SFTTraining Mixture=Hard problems only2026.01 | 33.89 | 64.12 |