Baidu’s OCR breakthrough reads dozens of pages in one pass

A single scan now unlocks an entire folder. Baidu’s latest optical-character-recognition system, dubbed “Unlimited OCR,” can ingest dozens of document pages in one go while keeping its memory footprint flat—no more stalling after a handful of pages. The trick lies in a modified attention mechanism that mimics how humans forget irrelevant details, letting the model scale efficiently without choking on long documents.

Traditional OCR pipelines hit a wall after about ten pages because their memory demands rise linearly with input length. Baidu researchers adjusted the transformer’s attention so that only the most pertinent context is retained, allowing the system to process substantially longer sequences without a proportional jump in compute. On the leading OCR benchmark, the approach currently sits at the top, underscoring both its speed and accuracy gains.

For anyone drowning in stacks of scanned contracts, research papers, or archival records, this could translate into real-world time savings. Instead of splitting files into smaller batches and stitching results, users can feed entire folders straight into the model and receive a unified text output. The technique is still fresh—expect it to trickle into commercial tools once the underlying method is open-sourced or licensed.

---

Source: The Decoder. AI-assisted editorial synthesis — TechnoExpress.