Cross-viewpoint spatial reasoning on ViewSpatial-Bench
46.3Camera-Relative Direction AccuracyViSRA
Evaluation results
| Method | Links | |
|---|---|---|
| ViSRABackbone=LLaVA-OneVision-7B2026.05 | 46.3 | |
| ViSRABackbone=InternVL3-8B2026.05 | 45.2 | |
| ViSRABackbone=Qwen2.5-VL-7B2026.05 | 43.2 | |
| Qwen2.5-VL-7BType=Base Model2026.05 | 28.8 | |
| InternVL3-8BType=Base Model2026.05 | 28.5 | |
| LLaVA-OneVision-7BType=Base Model2026.05 | 26.9 |