Multimodal Reasoning on GeoBench-VLM
35.4Aerial ScoreInternVL3-8B + Ours
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| InternVL3-8B + Oursrefinement=parameter-efficient refinement2026.02 | 35.4 | 61.4 | 42.1 | 45.7 | 24.3 | 48.2 | |
| Qwen2.5-VL-7B + Oursrefinement=parameter-efficient refinement2026.02 | 26.7 | 52.3 | 34 | 36.7 | 14.2 | 39.3 | |
| InternVL3-8Brefinement=none (frozen baseline)2026.02 | 26 | 46.1 | 38 | 35.2 | 20 | 37.4 | |
| LLaVA-1.5-13B + Oursrefinement=parameter-efficient refinement2026.02 | 23.6 | 51.2 | 17.4 | 32.5 | 14.1 | 34.9 | |
| LLaVA-1.5-7Bfinetuned on GeoBench-VLM=true2026.02 | 22.4 | 46.7 | 27.3 | 32.1 | 18.3 | 34.7 | |
| LLaVA-1.5-7B + Oursrefinement=parameter-efficient refinement2026.02 | 21.5 | 49.4 | 12 | 34.5 | 10 | 33.2 | |
| ControlMLLM++training-free=true2026.02 | 18.1 | 27.8 | 16.7 | 22.8 | 14.3 | 22.5 | |
| Qwen2.5-VL-7Brefinement=none (frozen baseline)2026.02 | 18 | 33.5 | 26.5 | 27.9 | 13.3 | 27.4 | |
| Jiang et al.training-free=true2026.02 | 17.4 | 28.1 | 15.8 | 18.9 | 12.8 | 21.4 | |
| LLaVA-1.5-7Brefinement=none (frozen baseline)2026.02 | 16.5 | 29.8 | 14 | 15.5 | 12.2 | 20.9 | |
| LLaVA-1.5-13Brefinement=none (frozen baseline)2026.02 | 13.5 | 37.1 | 19.5 | 13.4 | 13.3 | 23.7 |