General Multimodal Reasoning on MMK12, Geo3K, MathVerse, MathVista, and We-Math
76.21MMK12 AccuracyPTD-PO
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| PTD-POModel Scale=Qwen3-VL-8B-Thinking2026.06 | 76.21 | 64.52 | 79.81 | 71.39 | 82.31 | 74.85 | |
| PTD-POModel Scale=Qwen3-VL-4B-Thinking2026.06 | 73.84 | 64.79 | 80.45 | 69.81 | 81.67 | 74.11 | |
| PAPOModel Scale=Qwen3-VL-8B-Thinking2026.06 | 71.17 | 62.53 | 77.31 | 66.2 | 79.1 | 71.26 | |
| HDPOModel Scale=Qwen3-VL-8B-Thinking2026.06 | 71.05 | 64.6 | 78.04 | 66.86 | 78.88 | 71.89 | |
| GRPOModel Scale=Qwen3-VL-8B-Thinking2026.06 | 70.75 | 60.17 | 77.19 | 67.97 | 80.17 | 71.25 | |
| OPSDModel Scale=Qwen3-VL-8B-Thinking2026.06 | 68.73 | 58.19 | 69.84 | 59.44 | 71.72 | 65.58 | |
| HDPOModel Scale=Qwen3-VL-4B-Thinking2026.06 | 66.26 | 62.75 | 76.25 | 67.83 | 76.93 | 70 | |
| PAPOModel Scale=Qwen3-VL-4B-Thinking2026.06 | 65.11 | 62.69 | 77.73 | 70.95 | 80.32 | 71.36 | |
| GRPOModel Scale=Qwen3-VL-4B-Thinking2026.06 | 64.53 | 62.62 | 75.25 | 68.38 | 78.3 | 69.82 | |
| OPSDModel Scale=Qwen3-VL-4B-Thinking2026.06 | 60.97 | 58.13 | 67.98 | 55.23 | 68.06 | 62.07 | |
| SFTModel Scale=Qwen3-VL-4B-Thinking2026.06 | 59.59 | 57.38 | 70.85 | 57.07 | 71.24 | 63.23 | |
| PTD-POModel Scale=Qwen3-VL-2B-Thinking2026.06 | 56.86 | 50.52 | 70.53 | 60.38 | 72.64 | 62.19 | |
| SFTModel Scale=Qwen3-VL-8B-Thinking2026.06 | 55.97 | 59.09 | 67.72 | 57.81 | 70.47 | 62.21 | |
| GRPOModel Scale=Qwen3-VL-2B-Thinking2026.06 | 51.3 | 45.36 | 53.49 | 56.14 | 58.36 | 52.93 | |
| HDPOModel Scale=Qwen3-VL-2B-Thinking2026.06 | 50.22 | 49.36 | 66.6 | 58.86 | 69.81 | 58.97 | |
| PAPOModel Scale=Qwen3-VL-2B-Thinking2026.06 | 49.78 | 44.64 | 60 | 56.6 | 64.27 | 55.06 | |
| OPSDModel Scale=Qwen3-VL-2B-Thinking2026.06 | 44.99 | 38.48 | 52.69 | 48.52 | 56.9 | 48.32 | |
| SFTModel Scale=Qwen3-VL-2B-Thinking2026.06 | 38.59 | 41.95 | 55.5 | 49.05 | 59.55 | 48.93 |