pith. sign in

arxiv: 2606.23050 · v1 · pith:65UZ4NNGnew · submitted 2026-06-22 · 💻 cs.CV · cs.CL

Unlimited OCR Works

Youyang Yin , Huanhuan Liu , YY , Qunyi Xie , Chaorun Liu , Shiqi Yang , Shaohua Wang , Zhanlong Liu
show 9 more authors
Hao Zou Jinyue Chen Shu Wei Jingjing Wu Mingxin Huang Zhen Wu Guibin Wang Tengyu Du Lei Jia
This is my paper

Pith reviewed 2026-06-26 09:01 UTC · model grok-4.3

classification 💻 cs.CV cs.CL
keywords OCRattention mechanismKV cachelong contextsequence modelingdecoder design
0
0 comments X

The pith

Replacing attention with R-SWA keeps the KV cache constant so OCR can process dozens of pages in one pass under a 32K limit.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Unlimited OCR to handle long documents without the memory and speed penalties that come from growing KV caches in standard LLM-based OCR decoders. It starts from DeepSeek OCR and swaps every attention layer for Reference Sliding Window Attention, which the authors designed to hold KV cache size fixed no matter how long the output becomes. This combination of the encoder's high compression and the new constant-cache decoder lets the model transcribe many pages at once while staying inside normal context windows. The same attention change is presented as a general tool for any parsing task that needs to copy or transcribe long sequences.

Core claim

By replacing all attention layers in the decoder with Reference Sliding Window Attention, Unlimited OCR maintains a constant KV cache size throughout decoding, allowing transcription of dozens of pages in a single forward pass under a 32K maximum length.

What carries the argument

Reference Sliding Window Attention (R-SWA), an attention mechanism that reduces computation costs while enforcing constant KV cache size for the entire decoding process.

If this is right

  • OCR models can now handle multi-page documents without splitting or repeated passes.
  • The constant cache removes the progressive slowdown that normally appears as output length grows.
  • R-SWA can be swapped into other sequence-to-sequence tasks that require long output sequences.
  • The design emulates human working memory by avoiding ever-growing state during copying tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Tasks like automatic speech recognition or machine translation could adopt the same attention change to handle long inputs or outputs efficiently.
  • If R-SWA truly preserves accuracy, it offers a drop-in replacement for standard attention in any decoder that copies or transcribes text.
  • Future work could test whether the constant cache also reduces peak memory enough to run larger models on the same hardware.

Load-bearing premise

That swapping in R-SWA keeps the original OCR accuracy and language-modeling benefits intact even though no accuracy numbers or baseline comparisons are shown.

What would settle it

Run Unlimited OCR and the baseline DeepSeek OCR on the same set of multi-page documents and measure character error rate; if error rate rises sharply with R-SWA, the claim that accuracy is preserved fails.

read the original abstract

Recently, end-to-end OCR models, exemplified by DeepSeek OCR, have once again thrust OCR into the spotlight. A widely held view is that employing a large language model (LLM) as the decoder allows the model to leverage the prior distribution of language, leading to improved OCR performance. However, the downside is equally evident: as the output sequence lengthens, the accumulated KV cache drives up memory consumption and progressively slows down generation. This stands in stark contrast to humans, who exhibit no such decline in efficiency during long-horizon copying tasks. In this technical report, we propose Unlimited OCR, a model designed to emulate human parsing working memory. Taking DeepSeek OCR as the baseline, we replace all attention layers in the decoder with our proposed Reference Sliding Window Attention (R-SWA), which reduces attention computation costs while maintaining a constant KV cache throughout the entire decoding process. By combining the high compression rate of DeepSeek OCR's encoder with our constant KV cache design, Unlimited OCR can transcribe dozens of pages of documents in a single forward pass under a standard maximum length of 32K. More importantly, R-SWA is a general-purpose parsing attention mechanism - beyond OCR, it is equally applicable to tasks such as ASR, translation, etc. Codes and model weights are publicly available at http://github.com/baidu/Unlimited-OCR.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes Unlimited OCR, an extension of DeepSeek OCR that replaces all decoder attention layers with Reference Sliding Window Attention (R-SWA). This is claimed to enforce a constant KV cache size while preserving the LLM decoder's language prior, enabling transcription of dozens of pages in one forward pass under a 32K context limit. R-SWA is further positioned as a general-purpose mechanism applicable beyond OCR to tasks such as ASR and translation.

Significance. If the accuracy-preservation claim holds with supporting measurements, the work would address a practical bottleneck in long-context LLM-based OCR and parsing models by decoupling memory usage from output length. The public release of code and weights would further strengthen its potential impact as a reusable attention variant.

major comments (2)
  1. [Abstract] Abstract: The central claim that R-SWA 'maintains' the language prior benefits and overall OCR accuracy of the DeepSeek OCR baseline is asserted without any CER/WER numbers, ablation tables, or direct comparisons, leaving the accuracy-preservation step as an unevaluated assumption rather than a demonstrated result.
  2. [Abstract] Abstract: No quantitative results, ablation studies, or error analysis are supplied to support the performance claims (dozens of pages in a single 32K pass) or the generality claim for ASR/translation, which are load-bearing for the paper's contribution.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback on the need for empirical support. We respond point-by-point below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that R-SWA 'maintains' the language prior benefits and overall OCR accuracy of the DeepSeek OCR baseline is asserted without any CER/WER numbers, ablation tables, or direct comparisons, leaving the accuracy-preservation step as an unevaluated assumption rather than a demonstrated result.

    Authors: We agree the comment is correct: the abstract asserts preservation of the language prior without supporting measurements or comparisons. The technical report centers on the R-SWA design for constant KV cache while retaining the original decoder structure, but this does not constitute a demonstration. We will revise the abstract to describe R-SWA as intended to preserve the prior rather than claiming it maintains accuracy. revision: yes

  2. Referee: [Abstract] Abstract: No quantitative results, ablation studies, or error analysis are supplied to support the performance claims (dozens of pages in a single 32K pass) or the generality claim for ASR/translation, which are load-bearing for the paper's contribution.

    Authors: We agree the comment is correct: the abstract states the ability to transcribe dozens of pages and positions R-SWA as general-purpose without quantitative results, ablations, or error analysis. The constant-KV-cache property is a direct consequence of the sliding-window design, but the specific performance numbers and cross-task applicability are not demonstrated. We will revise the abstract and add a limitations paragraph to qualify these statements as design implications rather than evaluated outcomes. revision: yes

standing simulated objections not resolved
  • Supplying CER/WER numbers, ablation tables, error analysis, or results on ASR/translation tasks, as the current manuscript is a method-focused technical report without experimental evaluations.

Circularity Check

0 steps flagged

No derivation chain present; architectural proposal only

full rationale

The manuscript describes an engineering modification: replace decoder attention layers in an external baseline (DeepSeek OCR) with a new mechanism called R-SWA to enforce constant KV cache size. No equations, no fitted parameters, no derived predictions, and no self-citations appear in the provided text. The central claim is a design assertion rather than a mathematical reduction, so no load-bearing step reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, no free parameters, axioms, or invented entities can be identified with any specificity. The R-SWA mechanism is described at a high level without implementation details or assumptions listed.

pith-pipeline@v0.9.1-grok · 5818 in / 1046 out tokens · 32416 ms · 2026-06-26T09:01:27.459363+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 15 linked inside Pith

  1. [1]

    URLhttps://huggingface.co/nanonets/Nanonets-OCR-s

    Nanonets-ocr-s, 2025. URLhttps://huggingface.co/nanonets/Nanonets-OCR-s

    2025

  2. [2]

    URLhttps://github.com/DocTron-hub/OCRVerse

    Ocrverse, 2025. URLhttps://github.com/DocTron-hub/OCRVerse

    2025

  3. [3]

    URLhttps://github.com/chatdoc-com/OCRFlux

    Ocrflux, 2025. URLhttps://github.com/chatdoc-com/OCRFlux

    2025

  4. [4]

    G. AI. Gemini 2.5-pro, 2025. URLhttps://gemini.google.com/

    2025

  5. [5]

    URLhttps://github.com/alibaba/Logics-Parsing

    alibaba, 2026. URLhttps://github.com/alibaba/Logics-Parsing

    2026

  6. [6]

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P . Wang, J. Lin, C. Zhou, and J. Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023

    Pith/arXiv arXiv 2023

  7. [8]

    URLhttps://arxiv.org/abs/2511.21631

    Pith/arXiv arXiv

  8. [9]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P . Wang, S. Wang, J. Tang, H. Zhong, Y. Zhu, M. Yang, Z. Li, J. Wan, P . Wang, W. Ding, Z. Fu, Y. Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  9. [10]

    Blecher, G

    L. Blecher, G. Cucurull, T. Scialom, and R. Stojnic. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418, 2023

    Pith/arXiv arXiv 2023

  10. [11]

    C. Cui, T. Sun, S. Liang, et al. Paddleocr-vl: Boosting multilingual document parsing via a 0.9 b ultra-compact vision-language model. arXiv preprint arXiv:2510.14528, 2025

    arXiv 2025

  11. [12]

    C. Cui, T. Sun, M. Lin, T. Gao, Y. Zhang, J. Liu, X. Wang, Z. Zhang, C. Zhou, H. Liu, et al. Paddleocr 3.0 technical report. arXiv preprint arXiv:2507.05595, 2025

    Pith/arXiv arXiv 2025

  12. [13]

    D. Dong, M. Zheng, D. Xu, C. Luo, B. Zhuang, Y. Li, R. He, H. Wang, W. Zhang, W. Wang, et al. Qianfan-ocr: A unified end-to-end model for document intelligence. arXiv preprint arXiv:2603.13398, 2026

    arXiv 2026

  13. [14]

    H. Feng, S. Wei, X. Fei, W. Shi, Y. Han, L. Liao, J. Lu, B. Wu, Q. Liu, C. Lin, et al. Dol- phin: Document image parsing via heterogeneous anchor prompting. arXiv preprint arXiv:2505.14059, 2025

    arXiv 2025

  14. [15]

    Huang, C

    A. Huang, C. Yao, C. Han, F. Wan, H. Guo, H. Lv, H. Zhou, J. Wang, J. Zhou, J. Sun, et al. Step3-vl-10b technical report. arXiv preprint arXiv:2601.09668, 2026

    arXiv 2026

  15. [16]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 12

    Pith/arXiv arXiv 2023

  16. [17]

    J. Li, D. Li, S. Savarese, and S. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  17. [18]

    Z. Li, Y. Liu, Q. Liu, Z. Ma, Z. Zhang, S. Zhang, Z. Guo, J. Zhang, X. Wang, and X. Bai. Monkeyocr: Document parsing with a structure-recognition-relation triplet paradigm. arXiv preprint arXiv:2506.05218, 2025

    arXiv 2025

  18. [19]

    A. Liu, A. Mei, B. Lin, B. Xue, B. Wang, B. Xu, B. Wu, B. Zhang, C. Lin, C. Dong, et al. Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556, 2025

    Pith/arXiv arXiv 2025

  19. [20]

    Y. Liu, Z. Zhao, L. Tian, et al. Points-reader: Distillation-free adaptation of vision-language models for document conversion. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 1576–1601, November 2025

    2025

  20. [21]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016

    Pith/arXiv arXiv 2016

  21. [22]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In ICLR, 2019

    2019

  22. [23]

    Gpt-4 technical report, 2023

    OpenAI. Gpt-4 technical report, 2023

    2023

  23. [24]

    Ouyang, Y

    L. Ouyang, Y. Qu, H. Zhou, J. Zhu, R. Zhang, Q. Lin, B. Wang, Z. Zhao, M. Jiang, X. Zhao, et al. Omnidocbench: Benchmarking diverse pdf document parsing with comprehensive annotations. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24838–24848, 2025

    2025

  24. [25]

    Poznanski, A

    J. Poznanski, A. Rangapur, J. Borchardt, J. Dunkelberger, R. Huff, D. Lin, C. Wilhelm, K. Lo, and L. Soldaini. olmocr: Unlocking trillions of tokens in pdfs with vision language models. arXiv preprint arXiv:2502.18443, 2025

    arXiv 2025

  25. [26]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P . Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. PMLR, 2021

    2021

  26. [27]

    dots.ocr, 2025

    Rednote. dots.ocr, 2025. URLhttps://github.com/rednote-hilab/dots.ocr

    2025

  27. [28]

    Shoeybi, M

    M. Shoeybi, M. Patwary, R. Puri, P . LeGresley, J. Casper, and B. Catanzaro. Megatron-lm: Training multi-billion parameter language models using model parallelism. arXiv preprint arXiv:1909.08053, 2019

    Pith/arXiv arXiv 1909

  28. [29]

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023

    Pith/arXiv arXiv 2023

  29. [30]

    H. V . Team, P . Lyu, X. Wan, G. Li, S. Peng, W. Wang, L. Wu, H. Shen, Y. Zhou, C. Tang, et al. Hunyuanocr technical report. arXiv preprint arXiv:2511.19575, 2025

    arXiv 2025

  30. [31]

    B. Wang, C. Xu, X. Zhao, L. Ouyang, F. Wu, Z. Zhao, R. Xu, K. Liu, Y. Qu, F. Shang, et al. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024. 13

    Pith/arXiv arXiv 2024

  31. [32]

    W. Wang, Z. Gao, L. Gu, et al. Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.arXiv preprint arXiv:2508.18265, 2025

    Pith/arXiv arXiv 2025

  32. [33]

    H. Wei, L. Kong, J. Chen, L. Zhao, Z. Ge, J. Yang, J. Sun, C. Han, and X. Zhang. Vary: Scaling up the vision vocabulary for large vision-language model. In European Conference on Computer Vision, pages 408–424. Springer, 2024

    2024

  33. [34]

    H. Wei, C. Liu, J. Chen, J. Wang, L. Kong, Y. Xu, Z. Ge, L. Zhao, J. Sun, Y. Peng, et al. General ocr theory: Towards ocr-2.0 via a unified end-to-end model. arXiv preprint arXiv:2409.01704, 2024

    Pith/arXiv arXiv 2024

  34. [35]

    H. Wei, Y. Sun, and Y. Li. Deepseek-ocr: Contexts optical compression. arXiv preprint arXiv:2510.18234, 2025

    Pith/arXiv arXiv 2025

  35. [36]

    H. Wei, Y. Sun, and Y. Li. Deepseek-ocr 2: Visual causal flow. arXiv preprint arXiv:2601.20552, 2026

    arXiv 2026

  36. [37]

    H. Wu, H. Lou, X. Li, Z. Zhong, Z. Sun, P . Chen, X. Zhou, K. Zuo, Y. Chen, X. Tang, et al. Firered-ocr technical report. arXiv preprint arXiv:2603.01840, 2026

    arXiv 2026

  37. [38]

    J. Zhu, W. Wang, Z. Chen, Z. Liu, S. Ye, L. Gu, H. Tian, Y. Duan, W. Su, J. Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025. 14

    Pith/arXiv arXiv 2025