Spatial reasoning on satellite imagery on SQuID Overall
42AccuracyQVLM
Evaluation Results
| Method | Links | |
|---|---|---|
| QVLMCode Generator=GPT-5, Segmentation Model=ConvNeXt2026.01 | 42 | |
| QVLMCode Generator=gpt-oss-120B, Segmentation Model=ConvNeXt2026.01 | 32.14 | |
| QVLMCode Generator=GPT-5, Segmentation Model=DINOv3 + Mask2Former2026.01 | 30.83 | |
| QVLMCode Generator=Llama-3.1-8B, Segmentation Model=ConvNeXt2026.01 | 29 | |
| GPT-52026.01 | 28.1 | |
| QWEN 30B A3B thinking2026.01 | 26.14 |