Baidu's 'Unlimited OCR' Processes Multi-Page Documents with Constant Memory

Baidu researchers have developed 'Unlimited OCR,' an optical character recognition model capable of processing dozens of document pages in a single inference pass. The key innovation is an attention mechanism called Reference Sliding Window Attention (R-SWA), which keeps memory use constant regardless of document length. This is achieved by limiting the look-back window for previously generated tokens in the KV cache, preventing unbounded memory growth.