Spatial Relation Classification on EmbSpatial
90.29Left AccuracyGemma4 + FT
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Gemma4 + FTBackbone=Gemma4, Training Strategy=Fine-tuned2026.05 | 90.29 | 77.78 | 86.57 | 85.24 | 63.04 | 58.96 | 77.01 | |
| Gemma4 + GeoWorld-VLMBackbone=Gemma4, Training Strategy=GeoWorld-VLM (Ours)2026.05 | 89.21 | 78.14 | 88.43 | 86.72 | 67.03 | 63.06 | 78.78 | |
| InternVL3.5-2B + DINOBackbone=InternVL3.5-2B, Training Strategy=DINOv3 distillation2026.05 | 87.41 | 78.85 | 90.67 | 84.5 | 67.75 | 65.3 | 79.09 | |
| Gemma4 + DINOBackbone=Gemma4, Training Strategy=DINOv3 distillation2026.05 | 87.05 | 71.68 | 87.69 | 84.5 | 65.58 | 58.96 | 75.91 | |
| InternVL3.5-2B + FTBackbone=InternVL3.5-2B, Training Strategy=Fine-tuned2026.05 | 86.69 | 82.8 | 82.84 | 82.29 | 65.58 | 68.66 | 78.17 | |
| InternVL3.5-2B + GeoWorld-VLMBackbone=InternVL3.5-2B, Training Strategy=GeoWorld-VLM (Ours)2026.05 | 84.89 | 81.36 | 89.18 | 87.82 | 69.93 | 64.93 | 79.7 | |
| Gemma4Backbone=Gemma4, Training Strategy=Base2026.05 | 84.53 | 60.57 | 60.45 | 55.35 | 53.99 | 47.76 | 60.55 | |
| InternVL3.5-2BBackbone=InternVL3.5-2B, Training Strategy=Base2026.05 | 78.42 | 58.42 | 64.55 | 28.78 | 43.12 | 42.16 | 52.68 |