HPD-Parsing: Hierarchical Parallel Document Parsing

Traditional Vision-Language Model (VLM) document parsers generate output via a single, token-by-token autoregressive trajectory. This creates a sequential bottleneck where inference time increases linearly with document length. The authors investigate whether document parsing can be restructured to decouple global layout coordination from localized content generation to enable parallel processing.

HPD-Parsing introduces a Hierarchical Parallel Decoding (HPD) paradigm. The system uses a main layout branch to establish the document's structure and reading order. When the model identifies a specific region, it dynamically forks concurrent content branches to decode that region's text. To further optimize, the authors integrate Progressive Multi-Token Prediction (P-MTP), which allows each branch to predict multiple tokens per decoding step. The framework also utilizes shared-prefix KV cache reuse to avoid redundant computation across branches and employs a staged adaptation strategy to maintain accuracy during the transition from sequential to parallel decoding.

HPD-Parsing achieves a peak throughput of 4,752 tokens per second, which is 3.06 times faster than the vanilla autoregressive baseline and 2.62 times faster than the previous fastest document parsing model. The authors demonstrate that by isolating content decoding to specific branches, the model avoids the performance degradation typically associated with long-sequence autoregressive generation, all while maintaining competitive parsing accuracy on standard benchmarks.

This work provides a scalable solution for high-throughput document processing, which is critical for applications like Retrieval-Augmented Generation (RAG) and large-scale information extraction. By proving that document parsing does not require a single, rigid sequential generation path, the authors open a new design space for efficient, unified VLM-based document analysis.