Spatial Reasoning on Controlled Synthetic DR
100Binary Robust AccuracyLoRA-DPO
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen2-VL-7B2026.06 | 100 | 99.9 | 100 | 0.1 | 1.5 | 2.2 | |
| LoRA-DPOModel Family=LLaVA family, Backbone=LLaVA-v1.6-Vicuna-7B2026.06 | 99.5 | 97.8 | 92.2 | 16.4 | 32.4 | 47.4 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen3-VL-2B2026.06 | 99.4 | 94.6 | 88.9 | 13 | 18.1 | 14 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen2-VL-2B2026.06 | 98.8 | 90.1 | 64.3 | 14.6 | 33 | 46.5 | |
| LoRA-DPOModel Family=InternVL family, Backbone=InternVL2.5-1B2026.06 | 98.8 | 96 | 91.4 | 12.8 | 41 | 42.8 | |
| LoRA-DPOModel Family=LLaVA family, Backbone=LLaVA-v1.6-Mistral-7B2026.06 | 98.5 | 100 | 100 | 56.5 | 80.9 | 100 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen3-VL-8B2026.06 | 97.3 | 93.2 | 87.9 | 5.2 | 9.4 | 12.4 | |
| LoRA-DPOModel Family=InternVL family, Backbone=InternVL2-1B2026.06 | 91.5 | 74.4 | 52.3 | 24.6 | 40.3 | 27.5 | |
| LoRA-DPOModel Family=LLaVA family, Backbone=LLaVA-1.5-7B2026.06 | 90.3 | 83.9 | 76.5 | 54.6 | 82.4 | 76.5 |