Grade-school reasoning on GSM Hard
53.4Pass@1 Success RateSFT-CoT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFT-CoTModel=Qwen3-VL-4B-Instruct2026.01 | 53.4 | 191.1 | |
| SFT-CoTModel=Qwen3-VL-2B-Instruct2026.01 | 33.1 | 207.2 | |
| RoT (Ours)Model=Qwen3-VL-4B-Instruct2026.01 | 14.1 | 32 | |
| SFT-CoTModel=LLaVa-V1.6-Mistral-7B2026.01 | 12.2 | 209.7 | |
| SFT-w/o CoTModel=Qwen3-VL-4B-Instruct2026.01 | 9.48 | 0 | |
| RoT (Ours)Model=Qwen3-VL-2B-Instruct2026.01 | 8.64 | 32 | |
| SFT-w/o CoTModel=Qwen3-VL-2B-Instruct2026.01 | 4.7 | 0 | |
| RoT (Ours)Model=LLaVa-V1.6-Mistral-7B2026.01 | 3.64 | 32 | |
| SFT-w/o CoTModel=LLaVa-V1.6-Mistral-7B2026.01 | 2.27 | 0 |