Mathematical Reasoning on AIME26 (avg@16, pass@16)
21.46Avg@16Rollout-level |Ai| replay
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Rollout-level |Ai| replayModel=Qwen3-4B2026.06 | 21.46 | 0.5 | |
| GRPOModel=Qwen3-4B2026.06 | 16.88 | 0.3333 | |
| σgModel=Qwen3-4B2026.06 | 15.83 | 0.3667 | |
| uniformModel=Qwen3-4B2026.06 | 15.8 | 0.3552 | |
| σgModel=Qwen3-1.7B2026.06 | 5 | 0.2 | |
| uniformModel=Qwen3-1.7B2026.06 | 4.71 | 0.1462 | |
| Rollout-level |Ai| replayModel=Qwen3-1.7B2026.06 | 4.58 | 0.2 | |
| GRPOModel=Qwen3-1.7B2026.06 | 2.08 | 0.1333 | |
| σgModel=Qwen3-0.6B2026.06 | 1.04 | 0.0667 | |
| Rollout-level |Ai| replayModel=Qwen3-0.6B2026.06 | 1.04 | 0.0667 | |
| uniformModel=Qwen3-0.6B2026.06 | 0.62 | 0.0667 | |
| GRPOModel=Qwen3-0.6B2026.06 | 0.42 | 0.0333 |