Pith. sign in

REVIEW 4 major objections 6 minor 8 cited by

A 0.1B-parameter model trained on 40 million samples claims to recognize text and formulas from character to document level as accurately as billion-parameter vision-language parsers, while running 2–9 times faster.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A 0.1B-parameter text/formula recognition model trained on a new 40M-sample dataset matches or beats much larger OCR models and runs 2-9× faster.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Solid engineering contribution, but the accuracy claim needs a train/test contamination check before I'd trust the numbers. the 4 major comments →

arxiv 2512.21095 v2 pith:NCXJGK6H submitted 2025-12-24 cs.CV

UniRec-0.1B: Unified Text and Formula Recognition with 0.1B Parameters

Yongkun Du , Zhineng Chen , Yazhen Xie , Weikang Bai , Hao Feng , Wei Shi , Yuchen Su , Can Huang
show 1 more author
Yu-Gang Jiang
This is my paper
classification cs.CV
keywords unified text and formula recognitiondocument parsinglightweight vision-language modelhierarchical supervisionsemantic-decoupled tokenizermulti-level OCRLaTeX formula recognitiontraining dataset construction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper sets out to prove that unified document recognition does not need a giant multimodal language model. It introduces a 0.1B-parameter recognizer, UniRec-0.1B, trained from scratch on a new 40-million-sample dataset of Chinese and English text, formula, and mixed content. The claim is that with the right supervision signals and tokenizer design, this small model matches or beats much larger document-parsing systems in accuracy and is several times faster. If correct, the result matters because text and formulas make up the majority of document parsing cost, so shrinking this module alone could accelerate real pipelines without adding error.

Core claim

The paper claims that a 0.1B encoder-decoder can jointly recognize text and formulas across five granularities—character, word, line, paragraph, and multi-paragraph—when trained from scratch on 40M labeled samples. Two design choices carry the result. First, hierarchical supervision inserts explicit line-break and paragraph-end tokens into the label sequence, teaching the decoder spatial structure. Second, a semantic-decoupled tokenizer trains separate vocabularies for plain text and formulas so that tokens shared in surface form (like "sum" or "frac") receive distinct embeddings. On the paper's own benchmark and on a public page-level benchmark, the model reports lower edit distances than l

What carries the argument

Two mechanisms carry the argument. Hierarchical Supervision Training (HST) injects structural tokens into every training label: <|ln|> marks a line break, <|pn|> marks a paragraph end, forcing the decoder to represent layout rather than treating content as a flat string. The Semantics-Decoupled Tokenizer (SDT) trains one tokenizer on plain text and another on LaTeX formulas, then imports the formula tokens into the text vocabulary as distinct special tokens, so surface-identical tokens keep separate embeddings. A third supporting mechanism is the data pipeline: the 40M-sample corpus is partly generated by assigning unique colors to LaTeX tokens, rendering to PDF, and aligning by color to yie

Load-bearing premise

The load-bearing premise is that the benchmark pages are not near-duplicates of training pages; the paper does not report a deduplication check between the evaluation benchmark and its 40M-sample corpus.

What would settle it

Compute image or token-level similarity between every evaluation block and every training sample; if a noticeable share of benchmark blocks are near-duplicates of training content, the claimed edit-distance gains are not evidence of generalization. Simpler: run the released model on freshly collected documents from the same nine domains that were never rendered during training and compare edit distances to the reported numbers.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Substituting a 0.1B recognition module for the text/formula stages of two-stage parsers could cut page-level inference time by an order of magnitude while holding or lowering edit distance.
  • A single model producing character- through paragraph-level outputs may remove the need for separate text-line and formula detection stages in OCR pipelines.
  • Training a capable recognizer from scratch on 40M carefully labeled samples suggests data engineering can compensate for parameter count in narrow multimodal tasks.
  • The large formula-recognition gain attributed to the decoupled tokenizer implies tokenizer design is a major axis for compact multimodal models, not just architecture or data volume.

Where Pith is reading between the lines

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

  • Editorial inference: if the clean-generalization check passes, the small-expert/large-layout split could become the efficient default architecture for document parsing, since layout analysis remains heavy but content recognition becomes cheap.
  • Editorial inference: the decoupling idea is testable beyond OCR — small multimodal models for chart understanding, handwriting, or screenshot-to-code might see similar gains from separate token spaces for symbolic and natural-language content.
  • Editorial inference: the reported margins likely overstate the gap because the evaluation benchmark shares source families (online TeX documents, industry reports, newspapers, K-12 papers) with training; a fully disjoint benchmark is the natural next test.
  • Editorial inference: one could probe memorization directly by rendering held-out scientific pages and measuring whether the model's advantage persists when the test distribution is shifted to newer or less common document styles.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces UniRec-0.1B, a 0.1B-parameter encoder-decoder model for unified text and formula recognition at character, word, line, paragraph, and multi-paragraph levels. To support it, the authors construct UniRec40M, a 40M-sample Chinese/English dataset assembled from TeX sources, digital-born PDFs, and public datasets, and propose Hierarchical Supervision Training (HST) and a Semantics-Decoupled Tokenizer (SDT). The model is evaluated on a newly built UniRec-Bench derived from OmniDocBench and on OmniDocBench itself, where it is reported to be competitive with or better than much larger document parsing models while achieving a 2-9x speedup. The paper includes ablations of HST and SDT, speed measurements, and a code/dataset release.

Significance. If the empirical claims hold, this is a valuable result: a 0.1B model trained from scratch that can replace 0.9B-3B recognition modules in document parsing pipelines, with substantial speedups and open data/code. The ablation design is informative, and using OmniDocBench as an external check plus measuring speed under a fixed protocol are strengths. However, the central evidence is weakened by the lack of any train/test disjointness analysis between UniRec40M and UniRec-Bench/OmniDocBench, and by small formula-test samples with only point estimates. These issues must be resolved before the efficiency-accuracy trade-off claim can be accepted.

major comments (4)
  1. [§3.1, §5.1] UniRec-Bench is derived from OmniDocBench (Sec. 5.1), while UniRec40M is built from the same document families: arXiv/Wikipedia TeX sources, industry research reports/newspapers extracted via PyMuPDF, K-12 exam papers, and handwritten notes (Sec. 3.1). The paper reports no deduplication, no URL/TeX-source/PDF-hash overlap check, and no exclusion of OmniDocBench pages from training. Because Tables 2-4 all rely on this benchmark or OmniDocBench itself, the headline accuracy advantage (Abstract) could be inflated by memorization. The Appendix (Fig. 9 caption) even shows a training-specific label artifact (PyMuPDF sub/superscript flattening) appearing in predictions, demonstrating that training-pipeline idiosyncrasies can propagate to evaluation. Please add a quantitative disjointness analysis (hash/TeX-source/block-level near-duplicates) and re-run or confirm the comparisons on a verified d
  2. [§5.3.1, Table 2] The formula-level comparison rests on only 620 formula test blocks (header row of Table 2). The claimed formula gain over PaddleOCR-VL is about 0.9 percentage points (0.134 vs 0.125 edit distance), and even the larger gains over Dolphin are point estimates without confidence intervals or repeated seeds. With 620 samples, this margin is within plausible sampling noise. Report bootstrap confidence intervals, per-domain formula counts, or evaluate on a larger public formula set before concluding that SDT provides a robust formula advantage.
  3. [§5.3.2, Table 4] The 'new SOTA' claim is based on a 0.2-point overall edit-distance improvement (PaddleOCR-VL 0.115 vs wUniRec-0.1B 0.113) on OmniDocBench, with no repeated runs, no error bars, and no significance test. Since the integration replaces only the text/formula recognizer and keeps layout and table modules fixed, the aggregate metric may also dilute recognition gains. Report per-sample paired differences and significance, and clarify whether OmniDocBench pages were excluded from UniRec40M; otherwise Table 4 cannot validate the replacement claim.
  4. [§4.2] The Semantics-Decoupled Tokenizer is a central contribution, but the description ('train two independent tokenizers on plain text and mathematical formulas, then integrate formula tokens as special tokens') omits the tokenizer training corpus, tokenization algorithm, and merge/conflict handling. Final vocabulary size is given but not the formula/text split or how OOV tokens are mapped. Without these details, SDT cannot be implemented or ablated by others. Please provide the tokenizer recipe or release the tokenizers.
minor comments (6)
  1. [Abstract, Table 2] The abstract says 'outperforms both general-purpose VLMs and leading document parsing expert models,' but Table 2 shows UniRec-0.1B and PaddleOCR-VL tied on the overall average (0.100) and UniRec slightly worse on formula (0.134 vs 0.125). Suggest 'outperforms or is on par with.'
  2. [Table 1] The table sums to 39.6M across the listed entries, while the text repeatedly states 40M. Please reconcile the total and the rounding.
  3. [§5.1, Tables 2-3] The edit-distance metric is used without a precise definition. Clarify whether it is normalized, how paragraph-level ground truth is aligned with model output, and how structural tokens (<|ln|>, <|pn|>) are counted before or after post-processing.
  4. [Fig. 1] The top chart labels are extremely small and the '1-Edit' metric is not defined. Please enlarge the figure and define the metric in the caption or text.
  5. [§5.3.1] The sentence 'UniRec-0.1B outperforms them by 18.8%, 20.3%, and 10.4%' should state that these are absolute edit-distance differences, not relative improvements.
  6. [Conclusion] Typo: 'exceling' should be 'excelling.'

Circularity Check

0 steps flagged

No circular derivation; the paper's claims are empirical and evaluated against external benchmarks. Train/test overlap is a contamination risk, not a circularity.

full rationale

UniRec-0.1B is an empirical systems paper with no analytic derivation chain whose outputs are fixed by its inputs. The central claims—recognition accuracy on UniRec-Bench and OmniDocBench, ablation gains from HST/SDT, and inference speedups—are established by trained-model comparisons against external baselines and an external benchmark (OmniDocBench). The model is trained from scratch with a cross-entropy loss on UniRec40M; UniRec-Bench is a block-level re-crop of OmniDocBench, but the paper does not fit any parameter to UniRec-Bench, and the full-page OmniDocBench numbers are an external check. The HST and SDT ablations compare trained models with and without the proposed components; although HST tokens appear in the label sequence, the evaluation removes/replaces them before scoring, so the reported gains are not a self-definitional artifact. Self-citations (e.g., SVTR/SVTRv2, Dolphin) are used as related work or baselines, not as the sole justification of a load-bearing premise. The admitted PyMuPDF sub/superscript labeling artifact (Fig. 9 caption) is a disclosed limitation, and the lack of a deduplication check between UniRec40M and the OmniDocBench-derived UniRec-Bench is a genuine contamination risk for the accuracy evidence; however, contamination is an empirical validity concern, not deductive circularity, and it does not make any 'prediction' equivalent to a fitted input by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 3 axioms · 0 invented entities

No new physical entities are introduced. The special tokens <|ln|> and <|pn|> and the separate vocabularies are implementation-level design elements within the model, not new postulated entities with external falsifiable predictions. The ledger therefore focuses on the dataset-labeling and benchmark-disjointness assumptions that the empirical claims depend on.

free parameters (1)
  • Architecture and training hyperparameters = 6 layers, D=768, vocab 56,371, max res 960x1408, lr 1e-4, batch 64, 10 epochs, max len 1024
    These hand-chosen settings determine the accuracy/speed tradeoff of UniRec-0.1B. They are standard for a 0.1B encoder-decoder, but they are not derived from theory and jointly affect every headline number.
axioms (3)
  • domain assumption Color-based TeX/PDF alignment produces correct word-, line-, and paragraph-level labels at scale.
    Section 3.1(1) builds UniRec40M by inserting LaTeX color commands and aligning rendered PDFs with source tokens. No manual quality audit or error rate for this labeling pipeline is reported.
  • domain assumption Test documents in UniRec-Bench and OmniDocBench are disjoint from UniRec40M training examples.
    Section 5.1 derives UniRec-Bench from OmniDocBench, while Section 3.1 lists arXiv/Wikipedia, reports, newspapers, and K-12 exams as training sources. No deduplication is described; this premise is the main threat to validity.
  • domain assumption Edit distance between prediction and ground truth is an adequate accuracy metric for text and formula recognition.
    Section 5.1 defines accuracy via edit distance. It does not discuss LaTeX-equivalent output variants, spacing normalization, or semantic equivalence, all of which can affect edit distance.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of UniRec-0.1B: Unified Text and Formula Recognition with 0.1B Parameters." pith.science (2026). https://pith.science/paper/NCXJGK6H

@misc{pith2026251221095,
  author       = {Pith},
  title        = {Pith review of: UniRec-0.1B: Unified Text and Formula Recognition with 0.1B Parameters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NCXJGK6H}},
  note         = {Machine review of arXiv:2512.21095}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Text and formulas constitute the core informational components of many documents. Accurately and efficiently recognizing both is crucial for developing robust and generalizable document parsing systems. Recently, vision-language models (VLMs) have achieved impressive unified recognition of text and formulas. However, they are large-sized and computationally demanding, restricting their usage in many applications. In this paper, we propose UniRec-0.1B, a unified recognition model with only 0.1B parameters. It is capable of performing text and formula recognition at multiple levels, including characters, words, lines, paragraphs, and documents. To implement this task, we first establish UniRec40M, a large-scale dataset comprises 40 million text, formula and mixed samples, enabling the training of a powerful yet lightweight model. Secondly, we identify two challenges when building such a lightweight but unified expert model. They are: structural variability across levels and semantic entanglement between textual and formulaic content. To tackle these, we introduce a hierarchical supervision training that explicitly guides structural comprehension, and a semantic-decoupled tokenizer that separates text and formula representations. Finally, we develop a comprehensive evaluation benchmark covering Chinese and English documents from multiple domains and with multiple levels. Experimental results on this and public benchmarks demonstrate that UniRec-0.1B outperforms both general-purpose VLMs and leading document parsing expert models, while achieving 2-9x speedup, validating its effectiveness and efficiency. Codebase and Dataset: https://github.com/Topdu/OpenOCR.

discussion (0)

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

Forward citations

Cited by 8 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. How Far Is Document Parsing from Solved? PureDocBench: A Source-TraceableBenchmark across Clean, Degraded, and Real-World Settings

    cs.CV 2026-05 conditional novelty 8.0

    PureDocBench shows document parsing is far from solved, with top models at ~74/100, small specialists competing with large VLMs, and ranking reversals under real degradation.

  2. MinerU2.5-Pro: Pushing the Limits of Data-Centric Document Parsing at Scale

    cs.CV 2026-04 unverdicted novelty 7.0

    A fixed 1.2B model trained via diversity-aware sampling, cross-model verification, annotation refinement, and progressive stages achieves new state-of-the-art document parsing accuracy of 95.69 on OmniDocBench v1.6.

  3. MonkeyOCRv2: A Visual-Text Foundation Model for Document AI

    cs.CV 2026-07 conditional novelty 6.5

    A document-oriented ViT family pretrained with text generation plus pixel reconstruction on 113M images transfers across recognition, detection, parsing, and understanding, setting open-source SOTA on MDPBench with a ...

  4. HPD-Parsing: Hierarchical Parallel Document Parsing

    cs.CL 2026-07 conditional novelty 6.0

    Hierarchical parallel decoding — a global layout branch plus concurrent content branches with multi-token prediction — reaches 4,752 tokens/sec (≈3× a vanilla autoregressive baseline) at competitive accuracy on OmniDocBench.

  5. Boosting Document Parsing Efficiency and Performance with Coarse-to-Fine Visual Processing

    cs.CV 2026-03 conditional novelty 6.0

    PaddleOCR-VL uses a Valid Region Focus Module to select key visual tokens and a 0.9B model for guided recognition, delivering SOTA document parsing with far fewer tokens and parameters.

  6. ABot-OCR Technical Report

    cs.CV 2026-05 unverdicted novelty 5.0

    ABot-OCR is a new end-to-end VLM for direct image-to-Markdown transcription using a custom data engine and structure-constrained RL optimization, reporting SOTA scores of 92.81/93.30 on OmniDocBench v1.5/v1.6.

  7. OvisOCR2 Technical Report

    cs.CV 2026-07 conditional novelty 4.0

    A 0.8B end-to-end document-to-Markdown model reportedly tops OmniDocBench v1.6 (96.58) and PureDocBench (75.06), outperforming pipeline-based OCR systems.

  8. PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training

    cs.CV 2026-06 unverdicted novelty 3.0

    PaddleOCR-VL-1.6 improves on PaddleOCR-VL-1.5 via region-aware data optimization and progressive post-training to reach 96.33% on OmniDocBench v1.6.

Reference graph

Works this paper leans on

87 extracted references · 20 linked inside Pith · cited by 8 Pith papers

  1. [1]

    Gpt-4 technical report.arXivpreprintarXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXivpreprintarXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXivpreprintarXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Scene text recognition with permuted autoregressive sequence models

    Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In ECCV, pages 178–196. Springer, 2022

    2022

  4. [4]

    Nougat: Neural optical understanding for academic documents.arXivpreprint arXiv:2308.13418, 2023

    Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents.arXivpreprint arXiv:2308.13418, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Pix2text.https://github.com/breezedeus/Pix2Text, 2022

    breezedeus. Pix2text.https://github.com/breezedeus/Pix2Text, 2022. Accessed: 2025-06-23

    2022

  6. [6]

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

    chatdoc com. Ocrflux.https://github.com/chatdoc-com/OCRFlux, 2025. Accessed:2025-09-25

    2025

  7. [7]

    Ocean-ocr: Towards general ocr application via a vision-language model.arXiv preprint arXiv:2501.15558, 2025

    Song Chen, Xinyu Guo, Yadong Li, Tao Zhang, Mingan Lin, Dongdong Kuang, Youwei Zhang, Lingfeng Ming, Fengyu Zhang, Yuran Wang, et al. Ocean-ocr: Towards general ocr application via a vision-language model.arXiv preprint arXiv:2501.15558, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024

    2024

  9. [9]

    Paddleocr-vl: Boosting multilingual document parsing via a 0.9 b ultra-compact vision-language model.arXivpreprintarXiv:2510.14528, 2025

    Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, Zelun Zhang, Jiaxuan Liu, Xueqing Wang, Changda Zhou, Hongen Liu, Manhui Lin, et al. Paddleocr-vl: Boosting multilingual document parsing via a 0.9 b ultra-compact vision-language model.arXivpreprintarXiv:2510.14528, 2025

    arXiv 2025

  10. [10]

    Paddleocr 3.0 technical report.arXivpreprintarXiv:2507.05595, 2025

    Cheng Cui, Ting Sun, Manhui Lin, Tingquan Gao, Yubo Zhang, Jiaxuan Liu, Xueqing Wang, Zelun Zhang, Changda Zhou, Hongen Liu, et al. Paddleocr 3.0 technical report.arXivpreprintarXiv:2507.05595, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Image-to-markup generation with coarse-to- fine attention

    Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M Rush. Image-to-markup generation with coarse-to- fine attention. InICML, pages 980–989. PMLR, 2017

    2017

  12. [12]

    SVTR: Scene text recognition with a single visual model

    Yongkun Du, Zhineng Chen, Caiyan Jia, Xiaoting Yin, Tianlun Zheng, Chenxia Li, Yuning Du, and Yu-Gang Jiang. SVTR: Scene text recognition with a single visual model. InIJCAI, pages 884–890, 2022

    2022

  13. [13]

    Out of length text recognition with sub-string matching

    Yongkun Du, Zhineng Chen, Caiyan Jia, Xieping Gao, and Yu-Gang Jiang. Out of length text recognition with sub-string matching. InAAAI, pages 2798–2806, 2025

    2025

  14. [14]

    Contextperception parallel decoder for scene text recognition.IEEE Trans

    YongkunDu,ZhinengChen,CaiyanJia,XiaotingYin,ChenxiaLi,YuningDu,andYu-GangJiang. Contextperception parallel decoder for scene text recognition.IEEE Trans. Pattern Anal. Mach. Intell., 47(6):4668–4683, 2025. doi: 10.1109/TPAMI.2025.3545453

    arXiv 2025

  15. [15]

    Instruction-guided scene text recognition

    Yongkun Du, Zhineng Chen, Yuchen Su, Caiyan Jia, and Yu-Gang Jiang. Instruction-guided scene text recognition. IEEE Trans.PatternAnal. Mach.Intell., 47(4):2723–2738, 2025. doi: 10.1109/TPAMI.2025.3525526

    arXiv 2025

  16. [16]

    SVTRv2: Ctc beats encoder-decoder models in scene text recognition

    Yongkun Du, Zhineng Chen, Hongtao Xie, Caiyan Jia, and Yu-Gang Jiang. SVTRv2: Ctc beats encoder-decoder models in scene text recognition. InICCV, pages 20147–20156, 2025

    2025

  17. [17]

    LaTeX rainbow: Universal LaTeX to PDF document semantic & layout annotation framework

    Changxu Duan, Zhiyin Tan, and Sabine Bartsch. LaTeX rainbow: Universal LaTeX to PDF document semantic & layout annotation framework. InProceedings of the Second Workshopon Information Extraction from Scientific Publications, pages 56–67, 2023

    2023

  18. [18]

    Readlikehumans: Autonomous, bidirectional and iterative language modeling for scene text recognition

    ShanchengFang,HongtaoXie,YuxinWang,ZhendongMao,andYongdongZhang. Readlikehumans: Autonomous, bidirectional and iterative language modeling for scene text recognition. InCVPR, pages 7098–7107, 2021

    2021

  19. [19]

    ABINet++: Autonomous, bidirectional and iterative language modeling for scene text spotting.IEEETrans.PatternAnal.Mach

    Shancheng Fang, Zhendong Mao, Hongtao Xie, Yuxin Wang, Chenggang Yan, and Yongdong Zhang. ABINet++: Autonomous, bidirectional and iterative language modeling for scene text spotting.IEEETrans.PatternAnal.Mach. Intell., 45(6):7123–7141, 2023. 11

    2023

  20. [20]

    Dolphin: Document image parsing via heterogeneous anchor prompting.arXivpreprintarXiv:2505.14059, 2025

    HaoFeng,ShuWei,XiangFei,WeiShi,YingdongHan,LeiLiao,JinghuiLu,BinghongWu,QiLiu,ChunhuiLin,etal. Dolphin: Document image parsing via heterogeneous anchor prompting.arXivpreprintarXiv:2505.14059, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    open-parse.https://github.com/Filimoa/open-parse, 2024

    Filimoa. open-parse.https://github.com/Filimoa/open-parse, 2024. Accessed: 2025-06-23

    2024

  22. [22]

    Gemini 2.5

    Google DeepMind. Gemini 2.5. https://blog.google/technology/google-deepmind/ gemini-model-thinking-updates-march-2025/, 2025

    2025

  23. [23]

    Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks

    Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks. InICML, pages 369–376, 2006

    2006

  24. [24]

    Self-Supervised Character-to- Character distillation for text recognition

    Tongkun Guan, Wei Shen, Xue Yang, Qi Feng, Zekun Jiang, and Xiaokang Yang. Self-Supervised Character-to- Character distillation for text recognition. InICCV, pages 19473–19484, 2023

    2023

  25. [25]

    Dong Guo, Faming Wu, Feida Zhu, Fuxing Leng, Guang Shi, Haobin Chen, Haoqi Fan, Jian Wang, Jianyu Jiang, Jiawei Wang, et al. Seed1. 5-vl technical report.arXivpreprintarXiv:2505.07062, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    ICPR2018 contest on robust reading for multi-type web images

    Mengchao He, Yuliang Liu, Zhibo Yang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Yongpan Wang, Xin Zhang, and Lianwen Jin. ICPR2018 contest on robust reading for multi-type web images. InICPR, pages 7–12, 2018

    2018

  27. [27]

    Gtc: Guided training of ctc towards efficient and accurate scene text recognition

    Wenyang Hu, Xiaocong Cai, Jun Hou, Shuai Yi, and Zhiping Lin. Gtc: Guided training of ctc towards efficient and accurate scene text recognition. InAAAI, volume 34, pages 11005–11012, 2020

    2020

  28. [28]

    Ocr-free document understanding transformer

    Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. InECCV, pages 498–517. Springer, 2022

    2022

  29. [29]

    Show, attend and read: A simple and strong baseline for irregular text recognition

    Hui Li, Peng Wang, Chunhua Shen, and Guyu Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. InAAAI, pages 8610–8617, 2019

    2019

  30. [30]

    dots.ocr: Multilingual document layout parsing in a single vision-language model, 2025

    Yumeng Li, Guang Yang, Hao Liu, Bowen Wang, and Colin Zhang. dots.ocr: Multilingual document layout parsing in a single vision-language model, 2025

    2025

  31. [31]

    Monkeyocr: Documentparsingwithastructure-recognition-relationtripletparadigm

    Zhang Li, Yuliang Liu, Qiang Liu, Zhiyin Ma, Ziyang Zhang, Shuo Zhang, Zidun Guo, Jiarui Zhang, Xinyu Wang, andXiangBai. Monkeyocr: Documentparsingwithastructure-recognition-relationtripletparadigm. arXivpreprint arXiv:2506.05218, 2025

    arXiv 2025

  32. [32]

    Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model.arXiv preprint arXiv:2405.04434, 2024

    Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model.arXiv preprint arXiv:2405.04434, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Casia online and offline chinese handwriting databases

    Cheng-Lin Liu, Fei Yin, Da-Han Wang, and Qiu-Feng Wang. Casia online and offline chinese handwriting databases. In ICDAR, pages 37–41, 2011

    2011

  34. [34]

    Textmonkey: An ocr-free large multimodal model for understanding document.arXivpreprintarXiv:2403.04473, 2024

    Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: An ocr-free large multimodal model for understanding document.arXivpreprintarXiv:2403.04473, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Towards end-to-end unified scene text detection and layout analysis

    Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. InCVPR, pages 1049–1059, 2022

    2022

  36. [36]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InICLR, 2019

    2019

  37. [37]

    pix2tex - latex ocr.https://github.com/lukas-blecher/LaTeX-OCR, 2022

    Lukas Blecher. pix2tex - latex ocr.https://github.com/lukas-blecher/LaTeX-OCR, 2022. Accessed: 2025-06- 23

    2022

  38. [38]

    MORAN: A multi-object rectified attention network for scene text recognition

    Canjie Luo, Lianwen Jin, and Zenghui Sun. MORAN: A multi-object rectified attention network for scene text recognition. PatternRecognit., 90:109–118, 2019

    2019

  39. [39]

    Nanonets-ocr-s: A model for transforming documents into structured markdown with intelligent content recognition and semantic tagging, 2025

    Souvik Mandal, Ashish Talewar, Paras Ahuja, and Prathamesh Juvatkar. Nanonets-ocr-s: A model for transforming documents into structured markdown with intelligent content recognition and semantic tagging, 2025

    2025

  40. [40]

    Mathpix snip: Convert images and pdfs to latex, docx, and more.https://mathpix.com/, 2025

    Mathpix. Mathpix snip: Convert images and pdfs to latex, docx, and more.https://mathpix.com/, 2025

    2025

  41. [41]

    Smoldocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion.arXivpreprintarXiv:2503.11576, 2025

    Ahmed Nassar, Andres Marafioti, Matteo Omenetti, Maksym Lysak, Nikolaos Livathinos, Christoph Auer, Lucas Morin, Rafael Teixeira de Lima, Yusik Kim, A Said Gurbuz, et al. Smoldocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion.arXivpreprintarXiv:2503.11576, 2025. 12

    Pith/arXiv arXiv 2025

  42. [42]

    Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, Linke Ouyang, Zhiyuan Zhao, Tao Chu, Tianyao He, Fan Wu, Qintong Zhang, et al. Mineru2. 5: A decoupled vision-language model for efficient high-resolution document parsing. arXiv preprintarXiv:2509.22186, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Mineru2.0-2505-0.9b.https://huggingface.co/opendatalab/MinerU2.0-2505-0.9B, 2025

    opendatalab. Mineru2.0-2505-0.9b.https://huggingface.co/opendatalab/MinerU2.0-2505-0.9B, 2025

    2025

  44. [44]

    Omnidocbench: Benchmarking diverse pdf document parsing with comprehensive annotations

    Linke Ouyang, Yuan Qu, Hongbin Zhou, Jiawei Zhu, Rui Zhang, Qunshu Lin, Bin Wang, Zhiyuan Zhao, Man Jiang, Xiaomeng Zhao, et al. Omnidocbench: Benchmarking diverse pdf document parsing with comprehensive annotations. InCVPR, pages 24838–24848, 2025

    2025

  45. [45]

    Marker.https://github.com/datalab-to/marker, 2025

    Vik Paruchuri. Marker.https://github.com/datalab-to/marker, 2025. Accessed: 2025-09-25

    2025

  46. [46]

    olmocr: Unlocking trillions of tokens in pdfs with vision language models.arXiv preprint arXiv:2502.18443, 2025

    Jake Poznanski, Jon Borchardt, Jason Dunkelberger, Regan Huff, Daniel Lin, Aman Rangapur, Christopher Wilhelm, Kyle Lo, and Luca Soldaini. olmocr: Unlocking trillions of tokens in pdfs with vision language models.arXiv preprint arXiv:2502.18443, 2025

    arXiv 2025

  47. [47]

    NRTR: A no-recurrence sequence-to-sequence model for scene text recognition

    Fenfen Sheng, Zhineng Chen, and Bo Xu. NRTR: A no-recurrence sequence-to-sequence model for scene text recognition. InICDAR, pages 781–786, 2019

    2019

  48. [48]

    An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition.IEEE Trans.PatternAnal

    Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition.IEEE Trans.PatternAnal. Mach. Intell., 39(11):2298–2304,

  49. [49]

    ASTER: An attentional scene text recognizer with flexible rectification.IEEETrans.PatternAnal

    Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. ASTER: An attentional scene text recognizer with flexible rectification.IEEETrans.PatternAnal. Mach.Intell., 41(9):2035–2048, 2019

    2035

  50. [50]

    ICDAR 2019 competition on large-scale street view text with partial labeling-rrc-lsvt

    Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, Chee Seng Chan, and Lianwen Jin. ICDAR 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. InICDAR, pages 1557–1562, 2019

    2019

  51. [51]

    Tal open dataset.https://ai.100tal.com/dataset, 2023

    TAL. Tal open dataset.https://ai.100tal.com/dataset, 2023

    2023

  52. [52]

    Docling.https://github.com/docling-project/docling, 2024

    Docling Team. Docling.https://github.com/docling-project/docling, 2024. Accessed: 2025-06-23

    2024

  53. [53]

    Hunyuanocr technical report, 2025

    HunyuanVisionTeam,PengyuanLyu,XingyuWan,GengluoLi,ShangpinPeng,WeinongWang,LiangWu,Huawen Shen, Yu Zhou, Canhui Tang, Qi Yang, Qiming Peng, Bin Luo, Hower Yang, Xinsong Zhang, Jinnian Zhang, Houwen Peng, Hongming Yang, Senhao Xie, Longsha Zhou, Ge Pei, Binghong Wu, Rui Yan, Kan Wu, Jieneng Yang, Bochao Wang, Kai Liu, Jianchen Zhu, Jie Jiang, Linus, Han Hu...

    2025

  54. [54]

    Kimi-vl technical report.arXivpreprintarXiv:2504.07491, 2025

    Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chenzhuang Du, Chu Wei, et al. Kimi-vl technical report.arXivpreprintarXiv:2504.07491, 2025

    Pith/arXiv arXiv 2025

  55. [55]

    Hunyuan-0.5b.https://github.com/Tencent-Hunyuan/Hunyuan-0.5B, 2025

    Tencent. Hunyuan-0.5b.https://github.com/Tencent-Hunyuan/Hunyuan-0.5B, 2025

    2025

  56. [56]

    unstructured

    Unstructured-IO. unstructured. https://github.com/Unstructured-IO/unstructured, 2022. Accessed: 2025-06-23

    2022

  57. [57]

    SGDR: stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. InICLR, 2017

    2017

  58. [58]

    UniMERNet: A universal network for real-world mathematical expression recognition.arXivpreprintarXiv:2404.15254, 2024

    Bin Wang, Zhuangcheng Gu, Guang Liang, Chao Xu, Bo Zhang, Botian Shi, and Conghui He. UniMERNet: A universal network for real-world mathematical expression recognition.arXivpreprintarXiv:2404.15254, 2024

    Pith/arXiv arXiv 2024

  59. [59]

    Mineru: An open-source solution for precise document content extraction.arXiv preprint arXiv:2409.18839, 2024

    Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction.arXiv preprint arXiv:2409.18839, 2024

    Pith/arXiv arXiv 2024

  60. [60]

    Multi-Granularity Prediction for scene text recognition

    Peng Wang, Cheng Da, and Cong Yao. Multi-Granularity Prediction for scene text recognition. InECCV, pages 339–355, 2022

    2022

  61. [61]

    PETR: Rethinking the capability of transformer-based language model in scene text recognition.IEEETrans.Image Process., 31:5585–5598, 2022

    Yuxin Wang, Hongtao Xie, Shancheng Fang, Mengting Xing, Jing Wang, Shenggao Zhu, and Yongdong Zhang. PETR: Rethinking the capability of transformer-based language model in scene text recognition.IEEETrans.Image Process., 31:5585–5598, 2022. 13

    2022

  62. [62]

    General ocr theory: Towards ocr-2.0 via a unified end-to-end model.arXiv preprint arXiv:2409.01704, 2024

    Haoran Wei, Chenglong Liu, Jinyue Chen, Jia Wang, Lingyu Kong, Yanming Xu, Zheng Ge, Liang Zhao, Jianjian Sun, Yuang 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

  63. [63]

    Deepseek-ocr: Contexts optical compression

    Haoran Wei, Yaofeng Sun, and Yukun Li. Deepseek-ocr: Contexts optical compression. arXiv preprint arXiv:2510.18234, 2025

    Pith/arXiv arXiv 2025

  64. [64]

    Handwritten mathematical expression recognition via paired adversarial learning.IJCV, 128:2386–2401, 2020

    Jin-Wen Wu, Fei Yin, Yan-Ming Zhang, Xu-Yao Zhang, and Cheng-Lin Liu. Handwritten mathematical expression recognition via paired adversarial learning.IJCV, 128:2386–2401, 2020

    2020

  65. [65]

    Qwen3 technical report.arXivpreprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXivpreprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  66. [66]

    Kwai keye-vl 1.5 technical report.arXivpreprintarXiv:2509.01563, 2025

    Biao Yang, Bin Wen, Boyang Ding, Changyi Liu, Chenglong Chu, Chengru Song, Chongling Rao, Chuan Yi, Da Li, Dunju Zang, et al. Kwai keye-vl 1.5 technical report.arXivpreprintarXiv:2509.01563, 2025

    Pith/arXiv arXiv 2025

  67. [67]

    Focalmodulationnetworks

    JianweiYang,ChunyuanLi,XiyangDai,andJianfengGao. Focalmodulationnetworks. In NeurIPS,pages4203–4217, 2022

    2022

  68. [68]

    Towardsaccuratescenetextrecognition with semantic reasoning networks

    DeliYu,XuanLi,ChengquanZhang,JunyuHan,JingtuoLiu,andErruiDing. Towardsaccuratescenetextrecognition with semantic reasoning networks. InCVPR, pages 12113–12122, 2020

    2020

  69. [69]

    RobustScanner: Dynamically enhancing positional clues for robust text recognition

    Xiaoyu Yue, Zhanghui Kuang, Chenhao Lin, Hongbin Sun, and Wayne Zhang. RobustScanner: Dynamically enhancing positional clues for robust text recognition. InECCV, pages 135–151, 2020

    2020

  70. [70]

    Watch, attend and parse: An end-to-end neural network based approach to handwritten mathematical expression recognition.PR, 71: 196–206, 2017

    Jianshu Zhang, Jun Du, Shiliang Zhang, Dan Liu, Yulong Hu, Jinshui Hu, Si Wei, and Lirong Dai. Watch, attend and parse: An end-to-end neural network based approach to handwritten mathematical expression recognition.PR, 71: 196–206, 2017

    2017

  71. [71]

    A tree-structured decoder for image-to-markup generation

    Jianshu Zhang, Jun Du, Yongxin Yang, Yi-Zhe Song, Si Wei, and Lirong Dai. A tree-structured decoder for image-to-markup generation. InICML, pages 11076–11085. PMLR, 2020

    2020

  72. [72]

    Document parsing unveiled: Techniques, challenges, and prospects for structured information extraction

    Qintong Zhang, Bin Wang, Victor Shea-Jay Huang, Junyuan Zhang, Zhengren Wang, Hao Liang, Conghui He, and Wentao Zhang. Document parsing unveiled: Techniques, challenges, and prospects for structured information extraction. arXiv preprintarXiv:2410.21169, 2024

    Pith/arXiv arXiv 2024

  73. [73]

    Decoder pre-training with only text for scene text recognition

    Shuai Zhao, Yongkun Du, Zhineng Chen, and Yu-Gang Jiang. Decoder pre-training with only text for scene text recognition. InACMMM, pages 5191–5200, 2024. ISBN 9798400706868

    2024

  74. [74]

    CLIP4STR: A simple baseline for scene text recognition with pre-trained vision-language model.IEEE Trans.ImageProcess., 33:6893–6904, 2024

    Shuai Zhao, Ruijie Quan, Linchao Zhu, and Yi Yang. CLIP4STR: A simple baseline for scene text recognition with pre-trained vision-language model.IEEE Trans.ImageProcess., 33:6893–6904, 2024

    2024

  75. [75]

    Multi-modal in-context learning makes an ego-evolving scene text recognizer

    Zhen Zhao, Jingqun Tang, Chunhui Lin, Binghong Wu, Can Huang, Hao Liu, Xin Tan, Zhizhong Zhang, and Yuan Xie. Multi-modal in-context learning makes an ego-evolving scene text recognizer. InCVPR, pages 15567–15576, 2024

    2024

  76. [76]

    CDistNet: Perceiving multi- domain character distance for robust text recognition.Int

    Tianlun Zheng, Zhineng Chen, Shancheng Fang, Hongtao Xie, and Yu-Gang Jiang. CDistNet: Perceiving multi- domain character distance for robust text recognition.Int. J.Comput.Vis., 132(2):300–318, 2024

    2024

  77. [77]

    Vl-reader: Vision and language reconstructor is an effective scene text recognizer

    Humen Zhong, Zhibo Yang, Zhaohai Li, Peng Wang, Jun Tang, Wenqing Cheng, and Cong Yao. Vl-reader: Vision and language reconstructor is an effective scene text recognizer. InACMMM, pages 4207–4216, 2024

    2024

  78. [78]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprintarXiv:2504.10479, 2025. 14 Appendix More Recognition Result Visualizations To illustrate the recognition cases of our propos...

    Pith/arXiv arXiv 2025

  79. [80]

    In all, what fraction of the number of bills that Bree received for her birthday were $10 or $20 bills? UniRec-0.1B:

    The table shows the number of bills of each value that Bree received for her birthday. In all, what fraction of the number of bills that Bree received for her birthday were $10 or $20 bills? UniRec-0.1B:

  80. [81]

    In all, what fraction of the number of bills that Bree received for her birthday were $10 or $20 bills? Dolphin-1.5-0.3B:

    The table shows the number of bills of each value that Bree received for her birthday. In all, what fraction of the number of bills that Bree received for her birthday were $10 or $20 bills? Dolphin-1.5-0.3B:

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.