Mathematical Reasoning on OlympiadBench (Pass@1, Avg@8, Pass@8)
60.9Avg@8MEL
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| MELBackbone=Qwen3-14B-Base2026.02 | 60.9 | 61.87 | 75.82 | |
| GRPOBackbone=Qwen3-14B-Base2026.02 | 58.46 | 58.16 | 74.78 | |
| MELBackbone=Qwen3-8B-Base2026.02 | 55.6 | 54.01 | 73 | |
| GRPOBackbone=Qwen3-8B-Base2026.02 | 54.6 | 53.56 | 73.74 | |
| MELBackbone=Qwen3-4B-Base2026.02 | 49.48 | 48.51 | 69.73 | |
| GRPOBackbone=Qwen3-4B-Base2026.02 | 48.46 | 48.51 | 67.21 | |
| GRPOBase Model=Qwen3-4B-Base, Training Data=DAPO-Math-14k2026.03 | 43.37 | — | 60.74 | |
| EOPDBase Model=Qwen3-4B-Base, Training Data=DAPO-Math-14k2026.03 | 43.24 | — | 63.11 | |
| OPDBase Model=Qwen3-4B-Base, Training Data=DAPO-Math-14k2026.03 | 42.1 | — | 58.8 | |
| BaselineBackbone=Qwen3-14B-Base2026.02 | 40.5 | 45.25 | 65.58 | |
| BaselineBackbone=Qwen3-8B-Base2026.02 | 39.41 | 44.51 | 64.09 | |
| KDBase Model=Qwen3-4B-Base, Training Data=DAPO-Math-14k2026.03 | 37.33 | — | 62.52 | |
| BaselineBackbone=Qwen3-4B-Base2026.02 | 35.37 | 39.17 | 60.38 | |
| EOPDBase Model=Qwen3-1.7B-Base, Training Data=MATH2026.03 | 30.28 | — | 51.11 | |
| OPDBase Model=Qwen3-1.7B-Base, Training Data=MATH2026.03 | 30.09 | — | 51.56 | |
| GRPOBase Model=Qwen3-1.7B-Base, Training Data=MATH2026.03 | 29.89 | — | 50.81 | |
| KDBase Model=Qwen3-1.7B-Base, Training Data=MATH2026.03 | 27.66 | — | 48.4 | |
| OPDBase Model=Qwen3-0.6B-Base, Training Data=MATH2026.03 | 20.15 | — | 37.19 | |
| EOPDBase Model=Qwen3-0.6B-Base, Training Data=MATH2026.03 | 19.52 | — | 39.56 | |
| GRPOBase Model=Qwen3-0.6B-Base, Training Data=MATH2026.03 | 18.91 | — | 34.96 | |
| KDBase Model=Qwen3-0.6B-Base, Training Data=MATH2026.03 | 18.15 | — | 36.59 |