Mathematical Reasoning on Math500 (Acc, #L)
53.2AccuracySFT-CoT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFT-CoTBackbone=Qwen3-4B2026.05 | 53.2 | 248.2 | |
| Ours-RL (gsm)Backbone=Qwen3-4B2026.05 | 42.6 | 107.77 | |
| Ours-SFT (gsm)Backbone=Qwen3-4B2026.05 | 41.2 | 132.29 | |
| Ours-SFT (math)Backbone=Qwen3-4B2026.05 | 34.4 | 100.03 | |
| SFT-w/o CoTBackbone=Qwen3-4B2026.05 | 29.4 | 0 | |
| Ours-RL (gsm)Backbone=Llama-3.1-1B2026.05 | 14.6 | 90.02 | |
| SFT-CoTBackbone=Llama-3.1-1B2026.05 | 14 | 205.83 | |
| Ours-SFT (math)Backbone=Llama-3.1-1B2026.05 | 11.6 | 112.32 | |
| Ours-SFT (gsm)Backbone=Llama-3.1-1B2026.05 | 10.8 | 97.8 | |
| SFT-w/o CoTBackbone=Llama-3.1-1B2026.05 | 8.22 | 0 |