Geometry problem solving on Geo3K
98.2Top-1 AccuracyDraw2Think BL
Evaluation Results
| Method | Links | |
|---|---|---|
| Draw2Think BLBackbone=gemini-3-flash-preview, Execution Environment=direct VLM, Protocol=Training-free2026.05 | 98.2 | |
| Draw2Think CTBackbone=gemini-3-flash-preview, Execution Environment=executable environment, Protocol=Training-free2026.05 | 95.3 | |
| Human expertIdentity=College students2026.05 | 90.9 | |
| OpenAI-o3Protocol=Zero-shot, Reported by=TrustGeoGen [9]2026.05 | 81 | |
| Gemini-2.5-ProProtocol=Zero-shot, Reported by=TrustGeoGen [9]2026.05 | 80.7 | |
| GPT-5-ThinkingModel Category=Close-source Models2025.09 | 79.9 | |
| Pi-GPSProtocol=Neural-symbolic, Input Format=Parsed diagram annotations, Mechanism=PGDP + o3-mini solver2026.05 | 77.8 | |
| Gemini-2.5-ProModel Category=Close-source Models2025.09 | 77.2 | |
| PGPSNet-v2Protocol=Neural-symbolic, Input Format=Parsed diagram annotations, Mechanism=neural program synthesis2026.05 | 76.4 | |
| GeoDRLProtocol=Neural-symbolic, Input Format=Parsed diagram annotations, Mechanism=RL + formal solver2026.05 | 68.4 | |
| GeoGen-SFTProtocol=Fine-tuned, Backbone=qwen2.5-vl-7b, Mechanism=SFT2026.05 | 58.4 | |
| Gemini 2.0 FlashModel Category=Closed-source LMMs2026.02 | 58.2 | |
| InterGPSProtocol=Neural-symbolic, Input Format=Parsed diagram annotations, Mechanism=formal solver2026.05 | 57.5 | |
| GeoFocus-7BModel Category=Open-source Mathematical Geometry LMMs, Parameters=7B2026.02 | 55.3 | |
| VAPO-Thinker-7BModel Category=Our Models2025.09 | 51.3 | |
| VLAA-Thinker-7BModel Category=Open-source Models2025.09 | 50.6 | |
| GeoFocus-3BModel Category=Open-source Mathematical Geometry LMMs, Parameters=3B2026.02 | 50.4 | |
| GeoUniModel Category=Open-source Mathematical Geometry LMMs2026.02 | 50 | |
| Vision-R1-7BModel Category=Open-source Models2025.09 | 49 | |
| GeoGen-SFT-7BModel Category=Open-source Mathematical Geometry LMMs, Parameters=7B2026.02 | 46.3 | |
| R1-OneVision-7BModel Category=Open-source Models2025.09 | 46.1 | |
| InternVL2.5-8BModel Category=Open-source Models2025.09 | 44.8 | |
| Qwen2.5-VL-7BProtocol=Zero-shot, Input Format=Raw images, Reported by=GeoGen [35]2026.05 | 44.8 | |
| VAPO-Thinker-3BModel Category=Our Models2025.09 | 44.2 | |
| Qwen2.5-VL-7BModel Category=Open-source LMMs, Parameters=7B2026.02 | 39.4 | |
| GeoGen-SFT-3BModel Category=Open-source Mathematical Geometry LMMs, Parameters=3B2026.02 | 38.6 | |
| Qwen2.5-VL-7BModel Category=Open-source Models2025.09 | 38.5 | |
| G-LLaVA-13BModel Category=Open-source Mathematical Geometry LMMs, Parameters=13B2026.02 | 35 | |
| GPT-4oModel Category=Closed-source LMMs2026.02 | 34.6 | |
| InternVL-3-8BModel Category=Open-source LMMs, Parameters=8B2026.02 | 33.9 | |
| GPT-4oProtocol=Zero-shot, Reported by=TrustGeoGen [9]2026.05 | 31.5 | |
| Claude 3.7Model Category=Closed-source LMMs2026.02 | 31.1 | |
| InternVL-3-2BModel Category=Open-source LMMs, Parameters=2B2026.02 | 31.1 | |
| Gemma-3-4BModel Category=Open-source LMMs, Parameters=4B2026.02 | 30.6 | |
| GeoSketch-RLProtocol=Fine-tuned, Backbone=qwen2.5-vl-7b, Mechanism=SFT+RL2026.05 | 28.8 | |
| Qwen2.5-VL-3BModel Category=Open-source LMMs, Parameters=3B2026.02 | 25.3 | |
| InternVL-2.5-8BModel Category=Open-source LMMs, Parameters=8B2026.02 | 16.6 | |
| mPLUG-Owl2-7BModel Category=Open-source LMMs, Parameters=7B2026.02 | 12.7 | |
| Deepseek-VL-7BModel Category=Open-source LMMs, Parameters=7B2026.02 | 5.2 | |
| Monkey-Chat-7BModel Category=Open-source LMMs, Parameters=7B2026.02 | 4.7 | |
| Chameleon-7BModel Category=Open-source LMMs, Parameters=7B2026.02 | 3.9 |