Spatial Reasoning on VSI-Bench Extra
72RDB (Back)ViSRA
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| ViSRABackbone=Qwen2.5-VL-7B2026.05 | 72 | 56 | 49.7 | 61.8 | |
| ViSRABackbone=Qwen2.5-VL-3B2026.05 | 71.5 | 48.6 | 47.3 | 59 | |
| ViSRABackbone=InternVL3-8B2026.05 | 62.4 | 49.5 | 47.3 | 55 | |
| InternVL3-8Bmodel_type=Base Model2026.05 | 32.5 | 47 | 33.3 | 36.4 | |
| Qwen3-VL-8Bmodel_type=Post-trained Model2026.05 | 32.3 | 45 | 40.1 | 37.7 | |
| Qwen2.5-VL-7Bmodel_type=Base Model2026.05 | 30.1 | 44 | 27.8 | 32.9 | |
| Spatial-MLLM-4Bmodel_type=Post-trained Model2026.05 | 30 | 35.7 | 14.2 | 27 | |
| Qwen2.5-VL-3Bmodel_type=Base Model2026.05 | 29.3 | 43.8 | 31.3 | 33.5 | |
| InternVL3.5-8Bmodel_type=Post-trained Model2026.05 | 28.8 | 38.5 | 34.5 | 32.8 |