Reproduce “HPD-Parsing: Hierarchical Parallel Document Parsing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shu Wei, Jingjing Wu, Lingshu Zhang, Qunyi Xie, Hao Zou, Le Xiang, Xu Fan, Yangliu Xu and 4 more
HPD-Parsing introduces a Hierarchical Parallel Decoding paradigm that coordinates global layout generation with concurrent localized content decoding, achieving over 3x throughput improvement on document parsing and setting a new state-of-the-art accuracy on the OmniDocBench v1.6 benchmark.
Key insights
- Decoder latency, driven by long sequential execution paths, is the dominant inference cost in autoregressive document parsing.
- Documents possess a hierarchical structure where global layout is coordinated, but detailed content within regions is largely localized and exhibits limited dependence on distant parts.
- Exploiting this hierarchical structure allows for significant parallelization in content decoding, breaking the traditional sequence-length-dependent bottleneck.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.