DocAtlas: Multilingual Document Understanding Across 80+ Languages

Researchers introduced DocAtlas, a framework for creating high-fidelity OCR datasets and benchmarks across 82 languages using differential rendering and synthetic generation.

Across 16 models and 9 document-understanding tasks, the study found persistent weaknesses in low-resource scripts and complex multilingual layouts.

Training with DPO on rendering-derived ground truth improved both in-domain and out-of-domain accuracy, while supervised fine-tuning could sharply hurt transfer performance.

The work aims to reduce data scarcity and bias in multilingual OCR without sacrificing base-language accuracy.