Mathematical Reasoning on Average GSM, GSM-Hard, SVAMP, MultiArith
79.15Pass@1SFT-CoT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SFT-CoTBackbone=Qwen3-4B2026.05 | 79.15 | 70.05 | 0.348 | |
| Ours-RLBackbone=Qwen3-4B2026.05 | 72.73 | 33.17 | 0.543 | |
| Ours-SFTBackbone=Qwen3-4B2026.05 | 66.19 | 34.23 | 0.335 | |
| SFT-CoTBackbone=Llama-3.2-1B2026.05 | 56.87 | 62.13 | 0.515 | |
| RoTBackbone=Qwen3-4B2026.05 | 55.4 | 32 | 0.022 | |
| SFT-w/o CoTBackbone=Qwen3-4B2026.05 | 54.71 | 0 | — | |
| Ours-RLBackbone=Llama-3.2-1B2026.05 | 54.07 | 25.87 | 1.021 | |
| Ours-SFTBackbone=Llama-3.2-1B2026.05 | 50.82 | 25.42 | 0.911 | |
| CoLaR-2Backbone=Qwen3-4B2026.05 | 47.3 | 31.8 | -0.233 | |
| CoconutBackbone=Qwen3-4B2026.05 | 31.6 | 6 | -3.852 | |
| CoLaR-2Backbone=Llama-3.2-1B2026.05 | 31.4 | 29.84 | 0.126 | |
| CoconutBackbone=Llama-3.2-1B2026.05 | 30.74 | 6 | 0.468 | |
| SFT-w/o CoTBackbone=Llama-3.2-1B2026.05 | 27.65 | 0 | — |