Reproduce “ParseFixer: An Agentic Framework for Document Parsing via Selective Multimodal Correction” 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

LeKai Yu, Hao Liu, Kun Wang, Zhiran Li, Ruping Cao, Fan Liu, Yupeng Hu

ParseFixer, an agentic framework, improves document parsing by combining a backbone parser with selective multimodal correction to accurately preserve both content and structure. The system achieved an Overall score of 61.78 in the DataMFM Challenge Track 1, securing third place, and obtained the best Text ED (0.10) and Reading Order (75.48) scores among all participating teams.

Key insights

  • Decoupling initial parsing from targeted, verified correction using an agentic framework enhances overall accuracy and reliability for complex document structures.
  • Selective application of powerful multimodal models, guided by explicit failure diagnosis and rule-based verification, effectively mitigates the risk of introducing new errors (e.g., hallucinations or content loss) often associated with unconstrained generative models.
  • A hybrid approach that combines a stable backbone parser with advanced multimodal large language models for specific failure modes proves more effective than relying on a single model for all parsing tasks.

Summary by alphaXiv.

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