Document Visual Question Answering: A Comprehensive Analysis of Methods, Datasets, and Challenges

Document Visual Question Answering (DocVQA) is a key task within Visual Document Understanding (VDU), requiring models to jointly interpret textual content, layout structure, and visual elements within complex document images. The field has grown rapidly with the introduction of OCR-based architectures, multimodal transformers, and end-to-end OCR-free models. However, the literature remains fragmented, lacking a unified perspective on model design, dataset properties, and evaluation practices. This study provides a overview of the current DocVQA paradigms, including OCR-based encoder-only models, generative encoder-decoder models, OCR-free pixel-to-text models, and auxiliary techniques based on contrastive learning and knowledge distillation. We review key models and datasets related to VQA—specifically DocVQA, InfographicVQA, ChartQA, and VisualMRC—emphasizing their reasoning requirements, dataset distinctions, and inherent limitations. The survey further consolidates evaluation metrics across extractive, numerical, and generative settings and examines performance trends that arise from architectural choices. Finally, we outline the key challenges related to OCR brittleness, visual–text fusion, numerical reasoning, multi-page inference, and multilingual generalization, and discuss emerging directions toward unified, robust, and scalable DocVQA systems. This survey aims to provide a coherent technical foundation for advancing research in reliable and general-purpose document question answering.