Reproduce “Diagram Formalization Enhanced Multi-Modal Geometry Problem Solver” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zeren Zhang, Jo-Ku Cheng, Jingyang Deng, Lu Tian, Jinwen Ma, Ziran Qin, Xiaokai Zhang, Na Zhu and 1 more

Researchers from Peking University, 01.AI, and other institutions introduce DFE-GPS, a multi-modal framework for geometry problem-solving that effectively integrates geometric diagram formalization. The DFE-GPS-34B model achieves state-of-the-art performance on the formalgeo7k benchmark, reaching 82.38% accuracy for multiple-choice and 75.33% for open-ended questions, surpassing current advanced LLMs and MLLMs.

Key insights

  • Synthetic data generation is crucial for enabling vision encoders to accurately perceive and formalize geometric structures, with SynthGeo228K significantly boosting Diagram Formalizer performance.
  • Explicitly leveraging and calibrating geometric formal language (CDL) within the LLM enhances its understanding of diagrams and improves problem-solving effectiveness.
  • The proposed Process Evaluation Score (PES), using GPT-4o-mini, provides a comprehensive metric to assess the logical coherence, calculation accuracy, and conciseness of generated step-by-step reasoning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.