Elamur Research⌕
☰
home / papers / 2026 / PAR: Training-Free Positional Perturbation and Attention Recycling for Faithful OCR

PAR: Training-Free Positional Perturbation and Attention Recycling for Faithful OCR

Paper · 2026 · Yao Yao, Manwen Liao, Weitian Zhang et al.

Summary

The paper addresses Linguistic Priors Hallucination in vision-language models used for OCR. When a model sees familiar text, it tends to output the memorized version instead of transcribing what the image actually shows. To measure this, the authors build the GlitchText Probing Dataset. It takes highly familiar Chinese poems and English quotes and plants deliberate glyph-level or morphological alterations in them. A shuffled control variant removes linguistic context. Using last-layer attention, the authors also show that models rely less on image tokens as generation gets longer.

Their remedy, PAR, is training-free and applied only at inference. It has two parts:

  • Positional Perturbation (PP) adds structured phase noise to rotary positional embeddings, which breaks the momentum of long-range language dependencies.
  • Foveal Attention Recycling (FAR) detects when the model is overconfident in its language prior. It then shifts attention back to key visual regions using a Gaussian kernel, modeled on foveal vision.

PAR lowers CER by about 12%, with the largest gains on long outputs, and gives modest gains on OmniDocBench. This matters for tasks that need exact transcription, such as archival digitization and legal or scholarly work.

Key points

  • GlitchText construction: Texts were drawn from the Chinese Poetry Dataset (73,281 poems) and the English Quotes Dataset (2,510 Goodreads quotes). DeepSeek-OCR perplexity was used to pick the most familiar texts. The final set has 500 Chinese and 200 English items, each with 20 injected glitches.
  • Models tested: DeepSeek-OCR, Qwen3-VL-2B/8B, InternVL3.5-8B, GPT-4o and Claude-3.5-Sonnet. Metrics were CER, Identification Rate and Correction Rate.
  • Effect of shuffling on DeepSeek-OCR (Chinese): Glitch identification rose from 44.81% to 82.39% and correction fell from 49.57% to 11.52%. This indicates the model can see the glitches but overrides them.
  • GPT-4o behaves differently: Its CER worsened on shuffled text (0.2283 to 0.5587), suggesting it uses context to make up for weaker fine-grained visual grounding.
  • OmniDocBench results: Overall scores rose for every model tested, for example Qwen3-VL-8B from 53.57 to 54.93. Code is released at github.com/Zoeyyao27/PAR-for-Faithful-OCR.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Shanghai Jiao Tong UniversityShanghai · China212 works · 19,544 citations🎓University of Hong KongPok Fu Lam · China118 works · 13,496 citations🎓Wuhan UniversityWuhan · China47 works · 1,246 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

Cite

Cite as Yao Yao, Manwen Liao, Weitian Zhang, Zuchao Li and Hai Zhao. 2026. PAR: Training-Free Positional Perturbation and Attention Recycling for Faithful OCR. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1065

@inproceedings{yao2026par,
  title = {{PAR: Training-Free Positional Perturbation and Attention Recycling for Faithful OCR}},
  author = {Yao Yao, Manwen Liao, Weitian Zhang, Zuchao Li and Hai Zhao},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1065},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1065}
}

References

30 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Maximilian Augustin, Yannic Neuhaus, and Matthias Hein. 2025. Dash: Detection and assessment of systematic hallucinations of vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22748–22759. find
  2. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, and 45 others. 2025. Qwen3-vl technical report. arXiv prepri…
  3. Paul E. Black. 2021. Ratcliff/Obershelp pattern recognition. Dictionary of Algorithms and Data Structures. Accessed: 2025-12-25. find
  4. Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. 2024. Multi-modal hallucination control by visual information grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog… find
  5. Anisha Gunjal, Jihan Yin, and Erhan Bas. 2024. Detecting and preventing hallucinations in large vision language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18135–18143.
  6. Lehan He, Zeren Chen, Zhelun Shi, Tianyu Yu, Lu Sheng, and Jing Shao. Systematic reward gap optimization for mitigating vlm hallucinations. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. find
  7. Yingchen He, Jennifer M Scholz, Rachel Gage, Christopher S Kallie, Tingting Liu, and Gordon E Legge. 2015. Comparing the visual spans for faces and letters. Journal of vision, 15(8):7–7. find
  8. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276.
  9. Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. See what you are told: Visual attention sink in large multimodal models. In The Thirteenth International Conference on Learning Representations. find
  10. Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. 2024. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13872–13…
  11. Wei Li, Zhen Huang, Houqiang Li, Le Lu, Yang Lu, Xinmei Tian, Xu Shen, and Jieping Ye. 2025a. Visual evidence prompting mitigates hallucinations in large vision-language models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages… find
  12. Wei Li, Zhen Huang, Houqiang Li, Le Lu, Yang Lu, Xinmei Tian, Xu Shen, and Jieping Ye. 2025b. Visual evidence prompting mitigates hallucinations in large vision-language models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 20… find
  13. Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. 2023. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 292–305.
  14. Junhong Liang and Bojun Zhang. 2025. Vision language models are not (yet) spelling correctors. arXiv preprint arXiv:2509.17418. — [arXiv:2509.17418](https://arxiv.org/abs/2509.17418) find
  15. Rui Melo, Rui Abreu, and Corina S Pasareanu. 2025. Microsaccade-inspired probing: Positional encoding perturbations reveal llm misbehaviours. arXiv preprint arXiv:2510.01288. — [arXiv:2510.01288](https://arxiv.org/abs/2510.01288) find
  16. Andrew Morris, Viktoria Maier, and Phil Green. 2004. From wer and ril to mer and wil: improved evaluation measures for connected speech recognition. find
  17. Linke Ouyang, Yuan Qu, Hongbin Zhou, Jiawei Zhu, Rui Zhang, Qunshu Lin, Bin Wang, Zhiyuan Zhao, Man Jiang, Xiaomeng Zhao, Jin Shi, Fan Wu, Pei Chu, Minghao Liu, Zhenxiang Li, Chao Xu, Bo Zhang, Botian Shi, Zhongying Tu, and Conghui He. 2024. Omnidocbench: Benchmarking diverse pdf document parsing wi… — [arXiv:2412.07626](https://arxiv.org/abs/2412.07626) find
  18. Eunkyu Park, Minyeong Kim, and Gunhee Kim. 2025. Halloc: Token-level localization of hallucinations for vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29893–29903. find
  19. Vipula Rawte, Aryan Mishra, Amit Sheth, and Amitava Das. 2025. Defining and quantifying visual hallucinations in vision-language models. In Proceedings of the 5th Workshop on Trustworthy NLP (TrustNLP 2025), pages 501–510, Albuquerque, New Mexico. Association for Computational Linguistics. find
  20. Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, Jitao Sang, and Haoyu Tang. 2023. Evaluation and analysis of hallucination in large vision-language models. CoRR, abs/2308.15126. — [arXiv:2308.15126](https://arxiv.org/abs/2308.15126) find
  21. Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, and 1 others. 2025. Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265.
  22. Xintong Wang, Jingheng Pan, Liang Ding, and Chris Biemann. 2024. Mitigating hallucinations in large vision-language models with instruction contrastive decoding. In ACL (Findings). find
  23. Haoran Wei, Yaofeng Sun, and Yukun Li. 2025. Deepseek-ocr: Contexts optical compression. arXiv preprint arXiv:2510.18234.
  24. Xiaofeng Zhang, Yihao Quan, Chen Shen, Chaochen Gu, Xiaosong Yuan, Shaotian Yan, Jiawei Cao, Hao Cheng, Kaijie Wu, and Jieping Ye. 2025. Shallow focus, deep fixes: Enhancing shallow layers vision attention sinks to alleviate hallucination in lvlms. In Proceedings of the 2025 Conference on Empirical…
  25. Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035–4045.
  26. Yan Shu, Hangui Lin, Yexin Liu, Yan Zhang, Gangyan Zeng, Yan Li, Yu Zhou, Ser-Nam Lim, Harry Yang, and Nicu Sebe. 2025. When semantics mislead vision: Mitigating large multimodal models hallucinations in scene text spotting and understanding. arXiv preprint arXiv:2506.05551. — [arXiv:2506.05551](https://arxiv.org/abs/2506.05551) find
  27. Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063.
  28. Hai-Long Sun, Zhun Sun, Houwen Peng, and Han-Jia Ye. 2025. Mitigating visual forgetting via take-along visual conditioning for multi-modal long CoT reasoning. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5158–5171, Vienna,…
  29. S. M. Towhidul Islam Tonmoy, S. M. Mehedi Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. 2024. A comprehensive survey of hallucination mitigation techniques in large language models. CoRR, abs/2401.01313.
  30. William S Tuten and Wolf M Harmening. 2021. Foveal vision. Current Biology, 31(11):R701–R703. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yao-2026-par-training-free-positional-perturbation/