Document AI Translation Moves Beyond OCR Pipelines to End-to-End Systems

Translating scanned contracts, PDFs, and official documents has long required cascading multi-stage pipelines of optical character recognition OCR, translation, and desktop publishing. 

A growing body of research is now targeting document image AI translation — translating text embedded in scanned or rendered documents while preserving layout.

In 2025, the field is stepping into the spotlight with the International Conference on Document Analysis and Recognition (ICDAR), taking place September 16-21 in Wuhan, China, which is launching its first shared task on document image AI translation.

A number of recent papers highlight how researchers are tackling the challenges of document image AI translation.

End-to-End Document Image AI Translation

Researchers from the Chinese Academy of Sciences propose training a small document AI translation model with the help of a multimodal large language model (LLM), then deploying only the small model. This boosts performance on cross-domain, long-context, and complex-layout documents, while avoiding the high cost of running large models at inference, according to the researchers.

Researchers from Zhejiang University, the University of Chinese Academy of Sciences, Xiaohongshu Inc., and East China Normal University propose a framework that uses reinforcement learning to train multimodal translation models end-to-end. It treats document image translation as a three-part task: OCR, layout reasoning, and translation. The model uses a mixed reward system that combines multiple signals; translation accuracy, text recognition, and layout fidelity, into a single objective. This helps the model balance text and layout quality, achieving state-of-the-art results on benchmarks and generalizing well to unseen documents.

In another paper, researchers from the Chinese Academy of Sciences introduce a method where a multimodal LLM self-reviews its own OCR output during translation. This synchronous self-review process enhances the model’s ability to detect and correct OCR errors in real-time, leading to more accurate translations of document images.

Industry Contributions 

Industry teams are also moving quickly. Huawei’s translation service center has developed, and submitted at ICDAR, a system that uses a large vision-language model to translate document images end-to-end. It combines multi-task learning, chain-of-thought reasoning, and post-processing to improve layout-aware translation quality.

New Benchmarks and Datasets

In parallel, new datasets are setting a stronger foundation for evaluation. Johns Hopkins University’s OJ4OCRMT dataset, built from the multilingual Official Journal of the EU, offers aligned document images for benchmarking document image translation pipelines across multiple languages.