Mathematical Reasoning on AIME 2025 (Accuracy, CLC)
60.1AccuracySOLAR
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SOLARBackbone=Qwen3-8B, Fine-tuning strategy=SOLAR2026.06 | 60.1 | 53.02 | |
| SFTBackbone=Qwen3-8B, Fine-tuning strategy=SFT2026.06 | 58.38 | 50.13 | |
| Qwen3-8BBackbone=Qwen3-8B, Fine-tuning strategy=Base2026.06 | 54.91 | 45.06 | |
| Soft ThinkingBackbone=Qwen3-8B, Fine-tuning strategy=Soft Thinking2026.06 | 54.71 | 44.22 | |
| SOLARBackbone=Qwen3-4B, Fine-tuning strategy=SOLAR2026.06 | 30.95 | 24.51 | |
| MidAlignBackbone=Qwen3-4B, Fine-tuning strategy=MidAlign2026.06 | 29.05 | 23.22 | |
| SFTBackbone=Qwen3-4B, Fine-tuning strategy=SFT2026.06 | 27.9 | 20.44 | |
| AlignXBackbone=Qwen3-4B, Fine-tuning strategy=AlignX2026.06 | 27.76 | 20.1 | |
| MAPOBackbone=Qwen3-4B, Fine-tuning strategy=MAPO2026.06 | 25.81 | 18.94 | |
| Soft ThinkingBackbone=Qwen3-4B, Fine-tuning strategy=Soft Thinking2026.06 | 23.33 | 18.22 | |
| Qwen3-4BBackbone=Qwen3-4B, Fine-tuning strategy=Base2026.06 | 23.05 | 17.61 |