📐Pi-GPS:

Enhancing Geometry Problem Solving by Unleashing the Power of Diagrammatic Information

📄 Paper • 🚀 Quick Start • 📊 Results • 📚 Citation

---

🎉 Announcement

🏆 This paper has been accepted at ICCV 2025! 🎉

We're excited to share our work on geometry problem solving with the community!

---

👥 Authors

Junbo Zhao^1† • Ting Zhang^1†✉ • Jiayu Sun¹ • Mi Tian² • Hua Huang^1✉

¹Beijing Normal University　　²TAL

_{† Equal contribution　　✉ Corresponding author}

---

📖 Overview

We propose Pi-GPS, a novel framework for geometry problem solving that leverages diagrammatic information to resolve textual ambiguities. Our approach combines multi-modal understanding with symbolic reasoning to achieve state-of-the-art performance.

📐 Framework Overview of Pi-GPS

---

🔍 Key Features

Component	Description
🔧 Rectifier	Utilizes multi-modal language models (MLLMs) to disambiguate text by incorporating diagrammatic context
✅ Verifier	Ensures refined text adheres to geometric rules, effectively reducing model hallucinations
🧠 Symbolic Solver	Combines neural parsing with symbolic reasoning for robust problem solving

---

📊 Performance

🏆 Pi-GPS achieves state-of-the-art results, outperforming previous neural-symbolic methods by nearly 10% on standard benchmarks!

Category	Method	Geometry3K		PGPS9K
		Completion	Choice	Completion	Choice
MLLMs	Qwen-VL	22.1	26.7	20.1	23.2
	GPT-4o	34.8	58.6	33.3	51.0
	Claude 3.5 Sonnet	32.0	56.4	27.6	45.9
	Gemini 2	38.9	60.7	38.2	56.8
Neural Methods	NGS	35.3	58.8	34.1	46.1
	Geoformer	36.8	59.3	35.6	47.3
	SCA-GPS	-	76.7	-	-
	GOLD*	-	62.7	-	60.6
	PGPSNet-v2-S*	65.2	76.4	60.3	69.2
	LANS (Diagram GT)*	72.1	82.3	66.7	74.0
Neural-symbolic Methods	Inter-GPS	43.4	57.5	-	-
	GeoDRL	57.9	68.4	55.6	66.7
	E-GPS	-	67.9	-	-
	Pi-GPS (ours)	70.6	77.8	61.4	69.8

---

🚀 Getting Started

📋 Prerequisites

📥 Installation

# Clone the repository
git clone https://github.com/hellozting/Pi-GPS.git
cd Pi-GPS

# Install dependencies
pip install -r requirements.txt

📊 Dataset Preparation

We use two standard geometry datasets:

Dataset	Link	Description
📐 Geometry3K	Download	3K geometry problems with diagrams
📊 PGPS9K	Download	9K plane geometry problems

Download and extract datasets:

# Extract images for unified processing
python data/extract.py

---

⚡ Quick Start

🎯 Run with Pre-processed Data

python Solver/test.py --label final

🛠️ Run from Scratch

Click to expand step-by-step instructions

1️⃣ Text Parser

python Parser/text_parser.py

2️⃣ Diagram Parser

Follow PGDPNet setup instructions.

3️⃣ Disambiguation Module

# Add points to images
python Disambiguation_module/addPointsToImage.py

# Set your MLLM API in Disambiguation_module/align.py
python Disambiguation_module/GuidedAlign.py

4️⃣ Theorem Predictor

# Set your LLM API in Theorem_predictor/LLMPredict.py
python Theorem_predictor/LLMPredict.py

5️⃣ Solver

python Solver/test.py --label final_result \
  --text_logic_form_path Disambiguation_module/disambiguated_text_logic_forms_pred.json \
  --diagram_logic_form_path Parser/diagram_parser/PGDP.json \
  --predict_path Theorem_predictor/LLM_pred_seq.json

---

📚 Citation

If you find our work helpful, please consider citing:

@InProceedings{Zhao_2025_ICCV,
    author    = {Zhao, Junbo and Zhang, Ting and Sun, Jiayu and Tian, Mi and Huang, Hua},
    title     = {Pi-GPS: Enhancing Geometry Problem Solving by Unleashing the Power of Diagrammatic Information},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {1526-1536}
}

---

🤝 Contributing

We welcome contributions! Please feel free to:

• 🐛 Report bugs
• 💡 Suggest features
• 📝 Improve documentation
• 🔧 Submit pull requests

---

🙏 Acknowledgments

Our work builds upon several excellent projects:

---

📞 Contact

Questions or suggestions?

📧 [email protected]
🐙 Open an Issue

---

_{Made with ❤️ by the Pi-GPS Team}