Glonce

PaDoc cuts document-parsing latency with parallel decoding

researchThe Glonce editorsHugging Face🔥80Glonce rating80Impact65Novelty75Relevance100Surprise45Players25Credibility85
PaDoc cuts document-parsing latency with parallel decoding

Document parsers that read an image of a page and output structured text, tables and formulas come in two flavors, and researchers behind a new system called PaDoc argue both have a real cost. End-to-end parsers give a single unified interface but serialize the page's layout and the content of every region into one long autoregressive sequence, so decoding time grows with how much the page contains. Crop-based two-stage parsers process regions in parallel instead, but pay for it with repeated visual prefills for each crop and a loss of full-page context.

PaDoc is proposed as a third path: it treats the predicted layout as a branching structure over a single shared representation of the page. Under what the researchers call a region-sufficiency assumption, they derive a factorization in which the layout stream and the content of each region are decoded concurrently, so the decoding depth is bounded by the longest single layout-content path rather than by the total content on the page. The method is built into one multimodal LLM: packed variable-length ancestor attention preserves the model's visibility into the page under ordinary next-token training, while masked parallel decoding creates branches that the vLLM serving backend runs as concurrent requests sharing a cache-resident prefix.

On the OmniDocBench Full benchmark, PaDoc reaches an Overall layout F1 of 91.1 and, among end-to-end parsers, a top-tier Overall score of 94.24, the best Text Edit score (0.038) and the best Formula CDM score (95.59). On a 384-page subset run on a single A800 GPU, PaDoc is the fastest end-to-end parser at all five concurrency levels tested, improving valid-page throughput by 67.4-118% and cutting P95 latency by 39.2-54.9% relative to a same-backbone Sequential SFT baseline. Code is available on GitHub at Longin-Yu/Padoc.

Key facts

  • PaDoc decodes a page's layout and its regional content concurrently instead of serially, bounding decoding depth to the longest single layout-content path rather than total page content.
  • On OmniDocBench Full, PaDoc scores an Overall layout F1 of 91.1 and, among end-to-end parsers, a top-tier Overall score of 94.24, plus the best Text Edit (0.038) and Formula CDM (95.59).
  • On a 384-page subset with one A800 GPU, PaDoc is the fastest end-to-end parser at all five tested concurrency levels.
  • It improves valid-page throughput by 67.4-118% and cuts P95 latency by 39.2-54.9% versus a same-backbone Sequential SFT baseline.
  • Code is released on GitHub at Longin-Yu/Padoc.

Why it matters

Document parsing is normally slow because end-to-end parsers decode one long sequence per page, so latency grows with how much content the page holds. PaDoc's move is to treat the page's predicted layout as a branching structure: once the layout is known, the content of each region can be decoded alongside it instead of after it, in the same model. That turns a genuinely sequential bottleneck into one bounded by the single longest region, which is why the throughput and latency gains reported are large rather than marginal.

Who it affects

Anyone running document parsing at scale, digitizing archives, feeding PDFs into a retrieval pipeline, or converting scanned forms and papers into machine-readable text stands to gain from lower latency and higher throughput per GPU. Teams currently choosing between end-to-end parsers (accurate but slow) and crop-based two-stage parsers (faster but losing full-page context and paying for repeated visual prefills) get a third option that keeps full-page context without the sequential decoding cost.

How to use it

Code is released on GitHub at Longin-Yu/Padoc. The source does not state a license, pricing or a hosted service; what is public is the code implementing the layout-grounded parallel-decoding parser evaluated in the paper.

How solid is it

The accuracy results come from OmniDocBench Full, a standard document-parsing benchmark, where PaDoc posts the best Text Edit and Formula CDM scores among end-to-end parsers plus a top-tier Overall score of 94.24 and an Overall layout F1 of 91.1. The throughput and latency numbers were measured separately, on a 384-page subset with a single A800 GPU across five concurrency levels, against a same-backbone Sequential SFT baseline, which isolates the effect of the parallel-decoding change rather than mixing in a different base model.

Risks and caveats

The concurrent decoding scheme rests on a region-sufficiency assumption the researchers introduce to justify decoding layout and content branches in parallel; the source does not say how the method behaves on pages that violate it. The speed comparison is against one same-backbone Sequential SFT baseline, not a named field of other end-to-end parsers, so the throughput and latency gains are a specific head-to-head result rather than a market-wide comparison. The source gives no training-data details, no publication date and names no individual authors or institutions.