Reproduce “TexOCR: Advancing Document OCR Models for Compilable Page-to-LaTeX Reconstruction” 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
Chengye Wang, Lin Fu, Zexi Kuang, Yilun Zhao
Researchers from Zhejiang and Yale Universities introduce TEXOCR, a framework designed to convert scientific PDFs into compilable LaTeX, addressing the limitation of existing OCR models that often produce brittle or unexecutable LaTeX. The approach combines a new benchmark, a large training corpus, and a two-stage training pipeline incorporating reinforcement learning with verifiable rewards to improve structural faithfulness and compilation success.
Key insights
- Compilability and structural faithfulness are critical for practical LaTeX reconstruction, revealing limitations in current MLLMs that achieve high surface transcription but fail on global syntactic requirements.
- Reinforcement Learning with Verifiable Rewards (RLVR), using unit tests derived from LaTeX invariants, significantly improves structural faithfulness and end-to-end usability beyond token-level likelihood objectives.
- Single-page inference consistently outperforms multi-image or merged multi-page strategies, preserving visual clarity and preventing resolution loss, which is crucial for stable LaTeX generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.