Spatial reasoning on satellite imagery on SQuID (fragmentation)
81.63AccuracyQVLM
Evaluation Results
| Method | Links | |
|---|---|---|
| QVLMCode Generator=GPT-5, Segmentation Model=ConvNeXt2026.01 | 81.63 | |
| QVLMCode Generator=gpt-oss-120B, Segmentation Model=ConvNeXt2026.01 | 68.92 | |
| QVLMCode Generator=GPT-5, Segmentation Model=DINOv3 + Mask2Former2026.01 | 64.1 | |
| QVLMCode Generator=Llama-3.1-8B, Segmentation Model=ConvNeXt2026.01 | 50 | |
| GPT-52026.01 | 26.53 | |
| QWEN 30B A3B thinking2026.01 | 26.53 |