PaDoc: Layout-Grounded Parallel Decoding for Document Parsing

A recent scholarly article presents PaDoc, a layout-based parser designed for comprehensive document parsing. This research, found on arXiv (2608.06146), tackles the limitations of existing autoregressive document parsers that convert page layouts and content into a single sequence, leading to increased decoding lengths with more content. While crop-based two-stage parsers allow for region-level parallel processing, they necessitate repeated visual prefills and fragmented context. In contrast, PaDoc views the predicted layout as a branching framework over a unified page representation, facilitating simultaneous decoding of layout and content branches. This is accomplished through a prefix-conditioned factorization based on a region-sufficiency assumption, minimizing decoding depth to the longest layout-content path. The implementation occurs within a single multimodal large language model (MLLM) utilizing packed variable-length ancestor attention to maintain visibility during standard next-token prediction. The authors of the paper remain unnamed, but the research aims to enhance document parsing efficiency while preserving full-page context.