Geometry Problem Solving on UniGeo calculation (test)
96.9Top-1 AccuracyDraw2Think CT
Evaluation Results
| Method | Links | |
|---|---|---|
| Draw2Think CTBackbone=gemini-3-flash-preview, Execution Environment=executable environment2026.05 | 96.9 | |
| Draw2Think BLBackbone=gemini-3-flash-preview, Execution Environment=direct VLM2026.05 | 93.6 | |
| Gemini-2.5-ProProtocol=Zero-shot, Reported by=TrustGeoGen [9]2026.05 | 79.6 | |
| OpenAI-o3Protocol=Zero-shot, Reported by=TrustGeoGen [9]2026.05 | 79.3 | |
| GeoGen-SFTBackbone=qwen2.5-vl-7b, Protocol=Fine-tuned, Mechanism=SFT2026.05 | 78 | |
| GeoSketch-RLBackbone=qwen2.5-vl-7b, Protocol=Fine-tuned, Mechanism=SFT+RL2026.05 | 72.5 | |
| NeSyGeoBackbone=qwen2.5-vl-7b, Protocol=Fine-tuned, Mechanism=SFT2026.05 | 71.8 | |
| GeoDRLProtocol=Neural-symbolic, Input Format=Parsed diagram annotations2026.05 | 66.7 | |
| Qwen2.5-VL-7BProtocol=Zero-shot, Reported by=GeoGen [35]2026.05 | 64.2 | |
| SocraticGeoBackbone=qwen2.5-vl-7b, Protocol=Fine-tuned, Mechanism=GRPO2026.05 | 49.2 | |
| GPT-4oProtocol=Zero-shot, Reported by=TrustGeoGen [9]2026.05 | 42.3 |