Mathematical Problem Solving on MathVista GPS (N=208)
97.6Top-1 AccuracyDraw2Think CT
Evaluation Results
| Method | Links | |
|---|---|---|
| Draw2Think CTBackbone=gemini-3-flash-preview, Execution Environment=executable environment2026.05 | 97.6 | |
| Draw2Think BLBackbone=gemini-3-flash-preview, Execution Environment=direct VLM2026.05 | 97.1 | |
| MathCanvasBackbone=bagel-7b, Protocol=Fine-tuned, Mechanism=SFT2026.05 | 79.3 | |
| GeoGen-SFTBackbone=qwen2.5-vl-7b, Protocol=Fine-tuned, Mechanism=SFT2026.05 | 74 | |
| SocraticGeoBackbone=qwen2.5-vl-7b, Protocol=Fine-tuned, Mechanism=GRPO2026.05 | 63.6 | |
| Qwen2.5-VL-7BProtocol=Zero-shot2026.05 | 54.3 |