ZICQ
中 Log in / Sign up
Newsroom LLMs & Foundation Models #Document Parsing #AI Framework #Data-Centric #Robustness #Multimodal

WeVisDoc Released: Revolutionizing End-to-End Document Parsing with Enhanced Robustness and Performance

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 7 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:WeVisDoc is a novel two-stage data-centric framework designed to enhance the robustness and accuracy of end-to-end document parsing. It broadens semantic, structural, and appearance coverage through heterogeneous data generation and structure-preserving degradation synthesis, while using diagnostic probing to optimize and refine parser performance. On OmniDocBench v1.6 and PureDocBench benchmarks, the WeVisDoc-4B model achieved impressive scores of 95.38 and 75.54 respectively, ranking first amo


Overview

WeVisDoc is an innovative two-stage data-centric framework designed to address the robustness challenges in end-to-end document parsing. Document parsing tasks require parsers to maintain reliable performance across diverse layouts and acquisition conditions, but traditional training datasets often favor common document types and clean digital pages, limiting the model's generalization capabilities. WeVisDoc tackles this through the following two stages:

Technical Highlights

  1. Stage I: Expanding Coverage

    • Heterogeneous Data Generation: Generates diverse document types and layouts to broaden semantic, structural, and appearance coverage.
    • Structure-Preserving Degradation Synthesis: Simulates real-world degradation phenomena such as blur, noise, and distortion to enhance the model's adaptability to complex scenarios.
  2. Stage II: Diagnostics and Optimization

    • Diagnostic Probing: Uses a held-out probe set to measure residual errors of the Stage I parser and categorizes them into fixed visual-structural clusters.
    • Targeted Data Construction and Budget Allocation: Generates targeted training data based on diagnostic results and reallocates the target-token budget to optimize parser performance.

Performance

In OmniDocBench v1.6 and PureDocBench benchmarks, the WeVisDoc-4B model demonstrated exceptional performance:

  • OmniDocBench v1.6: Achieved an overall score of 95.38.
  • PureDocBench: Achieved an average overall score of 75.54 across three tracks, ranking first among all compared parsers.

Compared to Stage I, Stage II improved the overall scores of both 2B and 4B models on both benchmarks, with more significant gains on the degraded PureDocBench tracks. For instance, the 4B model gained 4.03 points on the Real Degraded track.

Industry Impact

The release of WeVisDoc marks a significant advancement in the field of document parsing. Its data-centric approach and diagnostic optimization mechanisms provide new insights for AI developers, particularly in handling complex document parsing tasks. Its high performance and robustness make it a promising tool for applications in automated document processing, OCR (Optical Character Recognition), and intelligent document management.

Developer Recommendations

  • Data Diversity: Developers should focus on the diversity of training data, simulating real-world degradation phenomena to enhance model robustness.
  • Diagnostic Optimization: Utilize diagnostic probing mechanisms to identify model weaknesses and optimize accordingly.
  • Resource Allocation: Allocate computing resources wisely, especially when handling large datasets and complex models.

Source: Hugging Face Daily Papers (2026-09-17)

— END —

Tags: #Document Parsing #AI Framework #Data-Centric #Robustness #Multimodal

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…