List your company

Geometry Problem Solving on PGPS9K N=1000

94.5Top-1 AccuracyDraw2Think BL
40.9454.84568.7582.655May 20, 2026

Evaluation results

MethodLinks
Draw2Think BLBackbone=gemini-3-flash-preview, Execution Environment=direct VLM2026.0594.5
Draw2Think CTBackbone=gemini-3-flash-preview, Execution Environment=executable environment2026.0590.5
Pi-GPSProtocol=Neural-symbolic, Input Format=Parsed diagram annotations2026.0569.8
PGPSNet-v2Protocol=Neural-symbolic, Input Format=Parsed diagram annotations2026.0569.2
GeoGen-SFTBackbone=qwen2.5-vl-7b, Protocol=Fine-tuned, Mechanism=SFT2026.0554.3
Qwen2.5-VL-7BProtocol=Zero-shot2026.0543