Spatial Reasoning on Rel3D
72Consistency ScoreOurs
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Ours2026.04 | 72 | 63 | 45.5 | 76.1 | 65.7 | 66.4 | |
| GPT-5.2Model Type=Closed-source Models, Scale=~100B active parameters, Training Data=≥30T tokens2026.04 | 70.5 | 61.9 | 46.4 | 78.8 | 56.4 | 65.4 | |
| Qwen3-VL-8BModel Type=Open-source Models2026.04 | 67.6 | 53 | 32.2 | 72.9 | 51.8 | 59 | |
| Gemini-2.5-FlashModel Type=Closed-source Models2026.04 | 66.9 | 55.6 | 44.9 | 70.8 | 42.5 | 58.9 | |
| InternVL3.5-8BModel Type=Open-source Models2026.04 | 62.4 | 47.6 | 27.4 | 62.5 | 43.4 | 52.5 | |
| LLaVA-OneVision-1.5-8BModel Type=Open-source Models2026.04 | 62.1 | 49.2 | 46.7 | 68.7 | 51.4 | 56.3 | |
| SpaceQwen2.5-VL-3BModel Type=Spatially Specialized Models2026.04 | 58 | 57.4 | 71.1 | 57.4 | 58.6 | 58.7 | |
| LLaVa-1.6-Mistral-7BModel Type=Open-source Models2026.04 | 57.4 | 41.3 | 57.8 | 34.5 | 34.3 | 45.3 | |
| Cambrian-1-8BModel Type=Open-source Models2026.04 | 53.7 | 34.6 | 41.9 | 34.8 | 35.2 | 41 | |
| VLM3R-7BModel Type=Spatially Specialized Models, Evaluation Mode=Exclusively processes video inputs (single-frame videos)2026.04 | 50 | 34.5 | 50.3 | 39.9 | 41.7 | 42.2 | |
| SpatialLLM-8BModel Type=Spatially Specialized Models2026.04 | 49.7 | 53.5 | 44.3 | 48.3 | 38.1 | 48.8 | |
| SpaceMantis-8BModel Type=Spatially Specialized Models2026.04 | 45.1 | 38.2 | 57.5 | 42.3 | 41.3 | 42.7 | |
| Spatial-MLLM-7BModel Type=Spatially Specialized Models, Evaluation Mode=Exclusively processes video inputs (single-frame videos)2026.04 | 37.3 | 29.2 | 33.4 | 37.5 | 29.4 | 33.3 |