HPD-Parsing Reframes Document Parsing as Hierarchical Parallel Decoding

Introduction

Unified Vision-Language Model document parsers have made document understanding more end-to-end: instead of stitching together separate OCR, layout analysis, and post-processing modules, a model can read a whole page and generate structured output. But many of these systems still share a familiar limitation. Even when the page is processed as a whole, the output is commonly produced through a single autoregressive path, one token after another.

HPD-Parsing, short for Hierarchical Parallel Document Parsing, addresses that decoding bottleneck. Proposed by the PaddleOCR team, it starts from a simple observation: document parsing is not uniformly sequential. The page layout should be understood globally, but once blocks are identified, their contents can often be parsed at the same time.

Key ideas

• Replacing full-page sequential generation: Rather than forcing all page content into one long token-by-token trajectory, HPD-Parsing introduces a hierarchical decoding structure.
• A main layout branch for coordination: The layout branch organizes the overall document structure and determines how block-level parsing tasks should be assigned.
• Concurrent content branches: Individual blocks are decoded by parallel branches, reducing the dependency on a single output stream.
• Progressive multi-token prediction: The method also uses P-MTP to cut down the number of decoding steps within each branch, further improving efficiency.
• Reported throughput gains: On public benchmarks, HPD-Parsing achieves 4,752 tokens per second. The authors report 2.62× the throughput of the fastest existing document parsing model and 3.06× that of a vanilla autoregressive baseline, while maintaining competitive parsing accuracy.

Why it matters

The contribution is not merely an engineering speed-up. HPD-Parsing reframes document parsing around the natural structure of documents: global layout first, local content in parallel. This is a more task-aware decoding strategy than treating an entire page as one long sequence.

For practical deployments, the implications are clear. In workflows involving receipts, contracts, research papers, forms, or reports, throughput and latency can determine whether a parser is usable at scale. Long or complex pages are exactly where a single autoregressive output path becomes costly, so parallel block-level decoding offers a promising alternative.

The provided summary does not include details such as hardware settings, document categories, or failure cases, so the full paper is still needed for a complete assessment. Even so, HPD-Parsing points to an important direction for multimodal document intelligence: unified VLM parsing does not have to mean fully sequential generation. A system can coordinate globally and execute locally in parallel.

Source: Hugging Face Daily Papers