Document parser reads hundreds of pages without growing per-step compute

The model attends only to relevant pages and token buckets, giving 7x faster decoding than full attention at hundred-page scale.

Independent

Shu Wei · Jingjing Wu · Qunyi Xie · Lingshu Zhang · Xiangcheng Du · Yannian Fu · +6 more

Research Digest·30 October 2026·2 min read

Illustration: Document parser reads hundreds of pages without growing per-step compute

In one line

Unlimited OCR 2 achieves constant computational cost across arbitrarily long documents, enabling one-shot hundred-page parsing.

Before you read this

Two ideas this paper builds on

sparse attentionSelectively attending to only the most relevant positions or tokens to reduce computational overhead.34 papers →vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.214 papers →

Read it as

Reading a very long document normally overwhelms a computer's memory and makes processing slower with every page. Parsing a hundred-page book is impractical.

This work introduces a method that mimics human transcription: it focuses closely only on the page currently being copied. A "Page Gate" limits the visual field, and a "Token Sieve" selects the most useful parts, fixing the memory footprint to a constant size.

The result is that processing speed stays constant no matter the length. On new hundred-page benchmarks, the system achieves top accuracy while running seven times faster, enabling one-shot parsing of ultra-long documents.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

sparse attention unlimited-horizon decoding long-context processing structured layout reasoning kv cache reuse vision-language model