Mathematical Reasoning on In-Domain Reasoning Suite
91.4MATH ScoreLIE
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| LIEModel Size=8B, Training Recipe=GSPO + LIE2026.02 | 91.4 | 60.4 | 73.4 | 37.2 | 31.6 | 58.8 | |
| GSPOModel Size=8B, Training Recipe=GSPO2026.02 | 89.8 | 58.8 | 72.9 | 34.4 | 25.8 | 56.3 | |
| LIEModel Size=4B, Training Recipe=GSPO + LIE2026.02 | 88.4 | 57.2 | 66.2 | 30.5 | 26.7 | 53.8 | |
| GSPOModel Size=4B, Training Recipe=GSPO2026.02 | 85.2 | 51.7 | 62.7 | 26.7 | 20.5 | 49.4 | |
| LIEModel Size=1.7B, Training Recipe=GSPO + LIE2026.02 | 77 | 39.6 | 45.2 | 17.5 | 11.5 | 38.2 | |
| GSPOModel Size=1.7B, Training Recipe=GSPO2026.02 | 70.6 | 32.3 | 38.4 | 8.1 | 3.8 | 30.6 | |
| Qwen3-BaseModel Size=8B, Training Recipe=Base2026.02 | 67.8 | 35.3 | 38.9 | 10.3 | 8.5 | 32.2 | |
| Qwen3-BaseModel Size=4B, Training Recipe=Base2026.02 | 66 | 33.2 | 36.6 | 8.5 | 6.9 | 30.2 | |
| Qwen3-BaseModel Size=1.7B, Training Recipe=Base2026.02 | 51.2 | 20.7 | 25.8 | 3.4 | 1.7 | 20.6 |