跳转至

文章背景与核心概要

传统的文档解析方法主要面临两大瓶颈：端到端解析器通常采用冗长且串行的解码路径，导致处理效率低下；而基于切片（Crop-based）的两阶段解析器则难以处理碎片化的上下文信息。为了解决这些痛点，本文介绍了 PaDoc 这一创新框架。

PaDoc 将预测出的版面布局视为共享页面表示之上的分支结构，从而实现了基于版面基础的并行解码（Layout-Grounded Parallel Decoding）。该方法允许版面流与区域内容分支并发推进，大幅降低了解码深度。实验表明，该模型在 OmniDocBench 数据集上取得了 SOTA（最先进）的性能，与标准的串行基线相比，具备更高的吞吐量和更低的延迟。

PaDoc: Layout-Grounded Parallel Decoding for Document Parsing

arXiv: 2608.06146
Date: August 6, 2026
Authors: Hao Yu, Jiabo Zhan, Kang Liu, Linnan Zhao, Dongxu Yue, Rui Chen, Jinglin Wang, Chong Sun, Chen Li, Jing Lyu, Chun Yuan

Summary

PaDoc 是一个新颖的文档解析框架，旨在克服传统端到端解析器（常受制于漫长且串行的解码路径）以及基于切片的两阶段解析器（在处理碎片化上下文时表现不佳）的局限性。

通过将预测的版面布局视为共享页面表示之上的分支结构，PaDoc 实现了基于版面基础的并行解码。这种方法使版面流和区域内容分支能够同时推进，从而显著降低了解码深度。该模型在 OmniDocBench 数据集上取得了最先进的性能，与标准串行基线相比，提供了更高的吞吐量并降低了延迟。

PaDoc is a novel document parsing framework designed to overcome the limitations of traditional end-to-end parsers, which often suffer from long, serialized decoding paths, and crop-based two-stage parsers, which struggle with fragmented context.

By treating the predicted layout as a branching structure over a shared page representation, PaDoc enables layout-grounded parallel decoding. This approach allows the layout stream and regional content branches to advance concurrently, significantly reducing decoding depth. The model achieves state-of-the-art performance on the OmniDocBench dataset, offering superior throughput and reduced latency compared to standard sequential baselines.

Key Features & Contributions

  • 并行解码架构： 实现了前缀条件因式分解，允许同时处理区域内容分支。
  • 高效率： 将解码深度降低至最长版面-内容路径的长度。
  • 卓越性能：
    • 整体版面 F1 值达到 91.1，综合得分达到 94.24。
    • 在文本编辑（0.038）和公式 CDM（95.59）指标上实现了同类最佳结果。
  • 吞吐量提升： 与串行 SFT 基线相比，有效页面吞吐量提高了 67.4–118%，P95 延迟降低了 39.2–54.9%。
  • 工程实现： 利用打包的变长祖先注意力（packed variable-length ancestor attention）和掩码并行解码，并与标准的 vLLM 后端兼容。
  • Parallel Decoding Architecture: Implements a prefix-conditioned factorization that allows regional content branches to be processed simultaneously.
  • Efficiency: Reduces decoding depth to the length of the longest layout-content path.
  • Performance:
    • Attains an Overall layout F1 of 91.1 and an Overall score of 94.24.
    • Achieves best-in-class results for Text Edit (0.038) and Formula CDM (95.59).
  • Throughput Gains: Improves valid-page throughput by 67.4–118% and reduces P95 latency by 39.2–54.9% compared to sequential SFT baselines.
  • Implementation: Leverages packed variable-length ancestor attention and masked parallel decoding, compatible with standard vLLM backends.

Access & Resources

Metadata

license icon