Geometry Problem Solving on PGPS9K N=1000
94.5Top-1 AccuracyDraw2Think BL
Evaluation Results
| Method | Links | |
|---|---|---|
| Draw2Think BLBackbone=gemini-3-flash-preview, Execution Environment=direct VLM2026.05 | 94.5 | |
| Draw2Think CTBackbone=gemini-3-flash-preview, Execution Environment=executable environment2026.05 | 90.5 | |
| Pi-GPSProtocol=Neural-symbolic, Input Format=Parsed diagram annotations2026.05 | 69.8 | |
| PGPSNet-v2Protocol=Neural-symbolic, Input Format=Parsed diagram annotations2026.05 | 69.2 | |
| GeoGen-SFTBackbone=qwen2.5-vl-7b, Protocol=Fine-tuned, Mechanism=SFT2026.05 | 54.3 | |
| Qwen2.5-VL-7BProtocol=Zero-shot2026.05 | 43 |