Survey Examines Multimodal AI Advances in Visually Rich Document Understanding

A comprehensive survey published on arXiv (ID: 2507.09861v2) analyzes recent progress in Visually Rich Document Understanding (VRDU) using Multimodal Large Language Models (MLLMs). The research highlights two primary technical approaches: OCR-based and OCR-free methods for extracting information from document images. Key challenges identified include data scarcity, processing multi-page documents, and handling multilingual content. The survey focuses on techniques for integrating textual, visual, and layout features within these models. Training paradigms such as pretraining and instruction tuning are examined as critical components. Emerging trends like Retrieval-Augmented Generation and agentic frameworks are noted as promising directions for future development. The analysis underscores VRDU's growing importance in automating the interpretation of complex documents containing intricate visual and structural elements.