Visual Reasoning on VSP (Accuracy by Scale 3-8)
100Accuracy (Scale 3)DiffThinker
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| DiffThinkerTraining Setting=Task-Specific Training2026.03 | 100 | 100 | 100 | 98 | 100 | 100 | |
| EndoCoTTraining Setting=Task-Specific Training2026.03 | 100 | 100 | 100 | 99 | 100 | 99 | |
| DiffThinkerTraining Setting=Unified Training2026.03 | 100 | 99 | 99 | 99 | 100 | 97 | |
| EndoCoTTraining Setting=Unified Training2026.03 | 100 | 98 | 98 | 100 | 100 | 100 | |
| Qwen3-VL-8BTraining Setting=Task-Specific Training, Optimization Method=SFT2026.03 | 99 | 96 | 98 | 96 | 92 | 86 | |
| Qwen3-VL-8BTraining Setting=Task-Specific Training, Optimization Method=GRPO2026.03 | 91 | 70 | 70 | 31 | 34 | 24 | |
| Qwen3-VL-8BTraining Setting=Zero-Shot2026.03 | 64 | 46 | 33 | 21 | 12 | 21 | |
| ChronoEditTraining Setting=Zero-Shot2026.03 | 60 | 20 | 12 | 7 | 16 | 13 | |
| Qwen-Image-Edit-2511Training Setting=Zero-Shot2026.03 | 50 | 55 | 44 | 16 | 23 | 23 | |
| ThinkGenTraining Setting=Zero-Shot2026.03 | 44 | 4 | 11 | 13 | 9 | 10 |