BAIDU'S UNLIMITED OCR PROCESSES DOZENS OF PAGES AT ONCE

Traditional OCR models struggle with longer documents because memory usage grows with each additional page. Baidu's approach treats memory management similarly to human forgetting—selectively retaining critical information while discarding less relevant data. The system achieves this through a redesigned attention mechanism that prevents exponential memory growth. By processing multiple pages without degradation, the model handles real-world scanning workflows more efficiently. Baidu's Unlimited OCR currently ranks first on major OCR benchmarks, outperforming existing solutions in both speed and accuracy across extended documents. The technique addresses a practical pain point for document digitization services, which frequently process multi-page forms, contracts, and records. The development suggests attention mechanisms—a core component of modern AI—can be optimized for longer sequences without sacrificing performance. This efficiency gain could extend beyond OCR to other tasks involving lengthy inputs.