Mathematical Reasoning on AIME 2025 (Pass@1, Pass@4)
55Pass@1AOPD
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| AOPDBackbone=Qwen3-8B-Base, SFT Warm-up steps=3K2026.05 | 55 | 70.28 | |
| AOPDBackbone=Qwen3-4B-Base, SFT Warm-up steps=3K2026.05 | 54.37 | 69.73 | |
| OPDBackbone=Qwen3-8B-Base, SFT Warm-up steps=3K2026.05 | 52.29 | 70.44 | |
| OPDBackbone=Qwen3-4B-Base, SFT Warm-up steps=3K2026.05 | 51.88 | 68.12 | |
| ExOPDBackbone=Qwen3-4B-Base, SFT Warm-up steps=3K2026.05 | 51.25 | 66.59 | |
| GKDBackbone=Qwen3-4B-Base, SFT Warm-up steps=3K2026.05 | 50.42 | 69.12 | |
| ExOPDBackbone=Qwen3-8B-Base, SFT Warm-up steps=3K2026.05 | 48.75 | 65.83 | |
| GKDBackbone=Qwen3-8B-Base, SFT Warm-up steps=3K2026.05 | 48.54 | 70.56 | |
| SFT CheckpointBackbone=Qwen3-8B-Base, SFT Warm-up steps=3K2026.05 | 47.71 | 66.73 | |
| SFT CheckpointBackbone=Qwen3-4B-Base, SFT Warm-up steps=3K2026.05 | 46.25 | 64.82 | |
| AOPDBackbone=Qwen3-8B-Base, SFT Warm-up steps=1K2026.05 | 45.21 | 66.86 | |
| GKDBackbone=Qwen3-8B-Base, SFT Warm-up steps=1K2026.05 | 42.08 | 63.69 | |
| SeqKDBackbone=Qwen3-4B-Base, SFT Warm-up steps=3K2026.05 | 40 | 60.09 | |
| SeqKDBackbone=Qwen3-8B-Base, SFT Warm-up steps=3K2026.05 | 40 | 56.98 | |
| SFT CheckpointBackbone=Qwen3-8B-Base, SFT Warm-up steps=1K2026.05 | 39.79 | 58.62 | |
| OPDBackbone=Qwen3-8B-Base, SFT Warm-up steps=1K2026.05 | 38.12 | 56.84 | |
| ExOPDBackbone=Qwen3-8B-Base, SFT Warm-up steps=1K2026.05 | 36.04 | 53.86 | |
| SeqKDBackbone=Qwen3-8B-Base, SFT Warm-up steps=1K2026.05 | 30.63 | 50 |