DocAtlas: Multilingual Document Understanding Across 80+ Languages
Summary
DocAtlas is a framework that creates high-fidelity OCR datasets and benchmarks across 82 languages, using differential rendering and synthetic generation. It demonstrates that Direct Preference Optimization improves multilingual model adaptation without degrading base-language performance.
View Cached Full Text
Cached at: 05/20/26, 02:35 AM
Paper page - DocAtlas: Multilingual Document Understanding Across 80+ Languages
Source: https://huggingface.co/papers/2605.12623
Abstract
DocAtlas framework creates high-fidelity OCR datasets across 82 languages using differential rendering and synthetic generation, demonstrating improved multilingual model adaptation through Direct Preference Optimization.
Multilingual document understandingremains limited forlow-resource languagesdue to scarce training data and model-based annotation pipelines that perpetuate existing biases. We introduceDocAtlas, a framework that constructs high-fidelityOCR datasetsand benchmarks covering 82 languages and 9 evaluation tasks. Our dual pipelines,differential renderingof native DOCX documents and synthetic LaTeX-based generation for right-to-left scripts produce precise structural annotations in a unifiedDocTag formatencoding layout, text, and component types, without learned models for core annotation. Evaluating 16 state-of-the-art models reveals persistent gaps in low-resource scripts. We show thatDirect Preference Optimization(DPO) using rendering-derived ground truth as positive signal achieves stablemultilingual adaptation, improving both in-domain (+1.9%) and out-of-domain (+1.8%) accuracy without measurable base-language degradation, wheresupervised fine-tuningdegrades out-of-domain performance by up to 21%. Our best variant,DocAtlas-DeepSeek, improves +1.7% over the strongest baseline.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2605\.12623
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2605.12623 in a model README.md to link it from this page.
Datasets citing this paper1
#### ahmedheakl/docatlas_instruct Viewer• Updatedabout 2 hours ago • 181k
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2605.12623 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
DocAtlas: Long-Document Understanding as Mutable-State Interaction
DocAtlas is a research system that frames long-document understanding as a mutable-state interaction process, using an external document harness with search, reading, note-taking, and review tools. It achieves state-of-the-art results on MMLongBench-Doc with GPT-5.4 and substantially improves compact VLM agents via end-to-end reinforcement learning.
dots.ocr: Multilingual Document Layout Parsing in a Single Vision-Language Model
This paper presents dots.ocr, a unified Vision-Language Model that jointly learns layout detection, text recognition, and relational understanding for multilingual document layout parsing. It achieves state-of-the-art results on OmniDocBench and introduces the XDocParse benchmark spanning 126 languages.
DocOCR-Eval: A Correction-Based Framework for OCR Tool Selection Without Ground Truth
DocOCR-Eval proposes an annotation-free framework that uses a correction and ranking strategy to evaluate and select OCR tools without ground truth labels, showing that aggregating multiple multimodal large language models improves alignment with human rankings.
DocAnnot -- Accelerating the Creation of Key Information Extraction Datasets with GenAI-Powered Auto-annotation
Introduces DocAnnot, a framework that uses a large vision-language model, OCR, and a spatially informed contextual matching algorithm to automatically generate training datasets for key information extraction from documents, reducing manual annotation effort. Evaluated on CORD and SROIE benchmarks, it achieves reasonable F1-scores and enables efficient human verification.
ForMaT: Dataset for Visually-Grounded Multilingual PDF Translation
This paper introduces ForMaT, a parallel corpus of 3,956 PDFs across 15 language pairs designed for visually-grounded multilingual translation, preserving layout metadata to benchmark layout-aware MT systems.