Geometric Reasoning on GeoQA
72.52AccuracyGeoSketch-Qwen2.5VL-7B
Evaluation Results
| Method | Links | |
|---|---|---|
| GeoSketch-Qwen2.5VL-7BTraining Stage=RL2025.09 | 72.52 | |
| Theorem-SFTBackbone=Qwen2.5-VL-7B, Supervision Layers=MLP2026.05 | 69.73 | |
| Theorem-SFTBackbone=Qwen2.5-VL-7B, Supervision Layers=All2026.05 | 66.89 | |
| GeoSketch-Qwen2.5VL-7BTraining Stage=SFT2025.09 | 65.09 | |
| BaseBackbone=Qwen2.5-VL-7B2026.05 | 49.46 | |
| Qwen2.5VL-7BMode=Base2025.09 | 47.28 | |
| Theorem-SFTBackbone=Qwen2.5-VL-3B, Supervision Layers=MLP2026.05 | 41.89 | |
| Vanilla SFTBackbone=Qwen2.5-VL-7B2026.05 | 38.78 | |
| Theorem-SFTBackbone=Gemma-3-4B-IT, Supervision Layers=MLP2026.05 | 36.72 | |
| BaseBackbone=Qwen2.5-VL-3B2026.05 | 35.95 | |
| Theorem-SFTBackbone=Gemma-3-4B-IT, Supervision Layers=All2026.05 | 35.72 | |
| Theorem-SFTBackbone=Qwen2.5-VL-3B, Supervision Layers=All2026.05 | 35.54 | |
| Vanilla SFTBackbone=Qwen2.5-VL-3B2026.05 | 34.59 | |
| Vanilla SFTBackbone=Gemma-3-4B-IT2026.05 | 31.89 | |
| BaseBackbone=Gemma-3-4B-IT2026.05 | 23.24 |