Baidu launched Unlimited-OCR 2: a model for recognizing documents of almost any length It can extract text from scans and PDF documents, processing tables, formulas and multi-page materials. But the main innovation of the work is not so much in the quality of OCR, but in the way it processes very long documents. Modern LLM-based OCR models have a problem: the more pages they process and the longer the generated text, the more calculations and memory are required to keep working. The model is forced to constantly refer to an increasing amount of prior context. The first Unlimited-OCR already partially solved this problem by limiting the amount of generation history saved. But the visual data of all the pages can still create an increasing computational load. In Unlimited-OCR 2, Baidu proposed the attention sparse sliding window mechanism (SR-SWA). Its idea is quite simple: when the next fragment is detected, the model focuses on the current page and accesses only a small set of the most useful fragments from the rest of the document. Instead of constantly reviewing all accumulated information, it uses a limited working memory - for example, 512 reference tokens. As a result, the computational cost of each successive step remains almost constant, even when the document becomes significantly longer. Everything is like skin: when writing a long book, we first look at the current page, and do not re-read all the previous ones after each sentence. what is the result According to Baidu's own tests, on documents of about 100 pages, the new mechanism accelerates individual decoding steps up to 7 times in relation to full attention and up to 5.6 times in relation to the previous mechanism. In addition, the model uses the processing of pages in overlapping blocks, which allows memory to be freed after the next part of the document is completed. The model contains about 3 billion parameters.

Baidu launched Unlimited-OCR 2: a model for recognizing documents of almost any length It can extract text from scans and PDF documents, processing tables, formulas and multi-page materials. But the main innovation of the work is not so much in the quality of OCR, but in the way it processes very long documents. Modern LLM-based OCR models have a problem: the more pages they process and the longer the generated text, the more calculations and memory are required to keep working. The model is forced to constantly refer to an increasing amount of prior context. The first Unlimited-OCR already partially solved this problem by limiting the amount of generation history saved. But the visual data of all the pages can still create an increasing computational load. In Unlimited-OCR 2, Baidu proposed the attention sparse sliding window mechanism (SR-SWA). Its idea is quite simple: when the next fragment is detected, the model focuses on the current page and accesses only a small set of the most useful fragments from the rest of the document. Instead of constantly reviewing all accumulated information, it uses a limited working memory - for example, 512 reference tokens. As a result, the computational cost of each successive step remains almost constant, even when the document becomes significantly longer. Everything is like skin: when writing a long book, we first look at the current page, and do not re-read all the previous ones after each sentence. what is the result According to Baidu's own tests, on documents of about 100 pages, the new mechanism accelerates individual decoding steps up to 7 times in relation to full attention and up to 5.6 times in relation to the previous mechanism. In addition, the model uses the processing of pages in overlapping blocks, which allows memory to be freed after the next part of the document is completed. The model contains about 3 billion parameters.