Physical Reasoning Visual Question Answering on CRAFT VQA
39.4Descriptive Accuracy+EG+VAORA
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| +EG+VAORAbase model=Qwen3VL-8B-SFT, Expert-Guided reward=true, VAORA=true2026.07 | 39.4 | 48.2 | 46.3 | 45.9 | 50.5 | 46.06 | |
| +EGbase model=Qwen3VL-8B-SFT, Expert-Guided reward=true2026.07 | 38.5 | 44.5 | 46 | 42.3 | 51.5 | 44.56 | |
| SFTbase model=Qwen3VL-8B2026.07 | 37.7 | 46.9 | 43.9 | 44.7 | 47.5 | 44.14 | |
| Qwen3-VL-8B-Instruct2026.07 | 26.6 | 41.7 | 40.7 | 42 | 40.7 | 38.34 |