Mathematical Reasoning on GSM (Pass@1, # L)
82.18Pass@1 AccuracySFT-CoT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFT-CoTBackbone=Qwen3-4B2026.05 | 82.18 | 85.2 | |
| Ours-RLBackbone=Qwen3-4B2026.05 | 75.28 | 38.36 | |
| Ours-SFTBackbone=Qwen3-4B2026.05 | 63.84 | 38.84 | |
| SFT-CoTBackbone=Llama-3.2-1B2026.05 | 53.49 | 77.89 | |
| Ours-RLBackbone=Llama-3.2-1B2026.05 | 46.47 | 32.09 | |
| Ours-SFTBackbone=Llama-3.2-1B2026.05 | 40.76 | 31.25 | |
| CoLaR-2Backbone=Qwen3-4B2026.05 | 40 | 39.6 | |
| SFT-w/o CoTBackbone=Qwen3-4B2026.05 | 38.44 | 0 | |
| RoTBackbone=Qwen3-4B2026.05 | 37.8 | 32 | |
| CoconutBackbone=Llama-3.2-1B2026.05 | 24.24 | 6 | |
| SFT-w/o CoTBackbone=Llama-3.2-1B2026.05 | 21.23 | 0 | |
| CoLaR-2Backbone=Llama-3.2-1B2026.05 | 17.23 | 37.85 | |
| CoconutBackbone=Qwen3-4B2026.05 | 16.9 | 6 |