Spatial Reasoning on Controlled Synthetic - CS
100Binary Robust AccuracyLoRA-DPO
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| LoRA-DPOModel Family=LLaVA family, Backbone=LLaVA-v1.6-Vicuna-7B2026.06 | 100 | 98.6 | 69.4 | 27.7 | 40.5 | 19.4 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen2-VL-2B2026.06 | 100 | 100 | 100 | 25.8 | 38.8 | 48.5 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen2-VL-7B2026.06 | 100 | 100 | 100 | 0 | 0 | -0.1 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen3-VL-2B2026.06 | 100 | 100 | 100 | 16.1 | 26.7 | 35.6 | |
| LoRA-DPOModel Family=LLaVA family, Backbone=LLaVA-v1.6-Vicuna-7B2026.06 | 100 | 99.8 | 98.9 | 16.2 | 41.8 | 71 | |
| LoRA-DPOModel Family=LLaVA family, Backbone=LLaVA-v1.6-Mistral-7B2026.06 | 99.8 | 90.9 | 90.3 | 52.4 | 74.2 | 90.3 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen3-VL-8B2026.06 | 99.8 | 99.3 | 96.5 | 7 | 15.5 | 21.5 | |
| LoRA-DPOModel Family=InternVL family, Backbone=InternVL2-1B2026.06 | 99.6 | 92.5 | 67.8 | 30.4 | 57.9 | 42.8 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen2-VL-7B2026.06 | 99.1 | 91.9 | 87.8 | 3.1 | 8.3 | 9.3 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen3-VL-2B2026.06 | 99 | 94.6 | 89.6 | 14.7 | 17.8 | 14.6 | |
| LoRA-DPOModel Family=InternVL family, Backbone=InternVL2.5-1B2026.06 | 98.9 | 90.4 | 82.5 | 2.8 | 34.7 | 50.4 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen3-VL-8B2026.06 | 98.6 | 96.2 | 94.3 | 3.9 | 10.3 | 23.6 | |
| LoRA-DPOModel Family=LLaVA family, Backbone=LLaVA-v1.6-Mistral-7B2026.06 | 97.7 | 100 | 100 | 43.4 | 67.3 | 97.4 | |
| LoRA-DPOModel Family=InternVL family, Backbone=InternVL2.5-1B2026.06 | 97.2 | 90 | 75.1 | 24.9 | 32.8 | 31.6 | |
| LoRA-DPOModel Family=LLaVA family, Backbone=LLaVA-1.5-7B2026.06 | 93.8 | 90.3 | 80.4 | 72.5 | 90 | 80.4 | |
| LoRA-DPOModel Family=InternVL family, Backbone=InternVL2-1B2026.06 | 89.6 | 65.7 | 49.6 | 33.2 | 30.7 | 24.6 | |
| LoRA-DPOModel Family=LLaVA family, Backbone=LLaVA-1.5-7B2026.06 | 88.8 | 81.2 | 68.9 | 34.2 | 58.2 | 66.1 | |
| LoRA-DPOModel Family=Qwen family, Backbone=Qwen2-VL-2B2026.06 | 88.1 | 69.6 | 35.6 | 21.7 | 40.3 | 25.2 |