Multimodal Reasoning on Multi-Task Suite VSP, Maze, TSP, Sudoku, Jigsaw, VisPuzzle
88.5Average ScoreDiffThinker++
Evaluation Results
| Method | Links | |
|---|---|---|
| DiffThinker++Setting=Flow Matching2025.12 | 88.5 | |
| DiffThinkerSetting=Flow Matching2025.12 | 87.4 | |
| Qwen3-VL-32BSetting=SFT2025.12 | 62.9 | |
| Qwen3-VL-8BSetting=SFT2025.12 | 51.6 | |
| Gemini-3-FlashSetting=N/A2025.12 | 41.3 | |
| Qwen3-VL-32BSetting=GRPO2025.12 | 26.9 | |
| GPT-5Setting=N/A2025.12 | 21.1 | |
| Qwen3-VL-8BSetting=GRPO2025.12 | 11.9 | |
| Qwen3-VL-32BSetting=N/A2025.12 | 10.5 | |
| Qwen3-VL-8BSetting=N/A2025.12 | 9.1 | |
| Qwen-Image-Edit-2511Setting=N/A2025.12 | 6.7 | |
| Qwen-Image-Edit-2509Setting=N/A2025.12 | 4 |