Spatial reasoning on satellite imagery on SQuID (connectivity)
74.04AccuracyQVLM
Evaluation Results
| Method | Links | |
|---|---|---|
| QVLMCode Generator=GPT-5, Segmentation Model=ConvNeXt2026.01 | 74.04 | |
| QVLMCode Generator=gpt-oss-120B, Segmentation Model=ConvNeXt2026.01 | 53.25 | |
| QVLMCode Generator=Llama-3.1-8B, Segmentation Model=ConvNeXt2026.01 | 49.04 | |
| QVLMCode Generator=GPT-5, Segmentation Model=DINOv3 + Mask2Former2026.01 | 45 | |
| QWEN 30B A3B thinking2026.01 | 44.23 | |
| GPT-52026.01 | 37.5 |