Pith. sign in

REVIEW 3 major objections 5 minor 46 references

Spatial Information Integration in Small Language Models for Document Layout Generation and Classification

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A locally runnable small language model can generate receipt layouts that beat LayoutTransformer on overlap and area fidelity.

desk verdict The BERT coordinate-concatenation result is real and worth attention; the layout-generation superiority claim over LayoutTransformer is not established because the comparison is conditional versus unconditional. read the letter →

arxiv 2501.05497 v1 pith:BE2WYRAQ submitted 2025-01-09 cs.CL cs.AIcs.IR

Pablo Melendez , Clemens Havas This is my paper · ORCID
classification cs.CLcs.AIcs.IR
keywords layoutgenerationsmalllanguagemodelsspatialinformationdocumenttextclassificationboundingboxessemi-structureddocumentsreceipts
verification ladder T0 review T1 audit T2 compute T3 formal

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 argues that a small, locally runnable language model can create synthetic document layouts for semi-structured documents like receipts, where public training data is scarce. By fine-tuning Llama3-8B on fewer than one hundred annotated receipts, the authors obtain a conditional layout generator that takes a list of section labels in natural language and returns bounding-box coordinates. Measured against LayoutTransformer, a pretrained transformer baseline fine-tuned on the same data, the proposed method produces far fewer intra-label overlapping boxes (1 versus 23) and comes closer to ground-truth area sizes for six of eight labels. The paper also reports that fine-tuning BERT for text classification improves from 62% to 74% accuracy when bounding-box coordinates are concatenated to the input strings. If these results hold, document-processing models could be trained on locally generated synthetic layouts rather than scarce and privacy-sensitive real documents.

What carries the argument

The load-bearing mechanism is the prompt format that casts layout generation as a text-completion task: 'Provide bounding box coordinates x1, y1, x2, y2 for these sections of a receipt document: <labels>', with the model answering one line per label in the form 'Label: x1, y1, x2, y2'. This natural-language conditioning, combined with LoRA fine-tuning of Llama3-8B in 4-bit precision, lets a model trained on fewer than 100 documents generalize to new label configurations. The evaluation machinery is cluster-based: for each label, origin points and closing points of generated boxes are clustered, then compared with ground-truth clusters using Mahalanobis distance, average-area difference, and intra-label overlap counts.

What would settle it

Train or prompt a label-conditioned baseline such as LayoutGPT on the same 107 receipts and rerun the overlap and area-difference metrics; if that conditional baseline produces one or fewer overlaps and area errors comparable to Llama3-8B, the paper's claimed superiority over LayoutTransformer would not distinguish the method from other conditional generators.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that a 4-bit quantized Llama3-8B fine-tuned with LoRA on 87% of a 107-receipt collection can generate plausible section layouts (Logo, Header, VAT_Table_Summary, PaymentInformation, LineItemTable, Footer, Contact, InvoiceDetails) from a prompt listing the desired labels, and that this conditional generator outperforms LayoutTransformer—which had to be pretrained on PubLayNet before fine-tuning—on the evaluation metrics used. The proposed method produced only one intra-label overlapping bounding box across all labels compared with 23 for LayoutTransformer, approached the ground-truth average area size more closely for six of eight labels, and matched or came close on Mahalanobis distances to cluster centroids for origin and closing points. In the classification part, the paper claims that adding bounding-box coordinates to plain strings raises fine-tuned BERT accuracy from 0.62 to 0.74, while few-shot prompting of larger SLMs gains nothing from the same coordinates.

Load-bearing premise

The main comparison assumes that the evaluation metrics—cluster distances, area differences, and overlap counts—give a fair picture when one method is forced to emit exactly one box per requested label while the other emits however many boxes it chooses.

Editorial extensions

If this is right

  • Synthetic receipt layouts can be produced locally on commodity hardware, removing the need to share sensitive document images with external services.
  • The generator is conditional: a user can specify which and how many sections should appear, something LayoutTransformer cannot do directly.
  • Document-processing models such as LayoutLM could be trained or augmented with this synthetic layout data instead of relying on scarce public receipt datasets.
  • Text classification with fine-tuned transformers can be improved by simply concatenating spatial coordinates, although the effect did not appear in few-shot prompting of larger models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the comparison were rerun against a conditional baseline such as LayoutGPT, the overlap advantage might shrink because LayoutGPT also emits one box per label; the paper leaves this comparison for future work.
  • The 12% BERT improvement could partly reflect that coordinates strongly correlate with section identity (e.g., footers near the page bottom); shuffling coordinates while preserving text would test whether the model uses spatial structure or mere position priors.
  • The success on receipts plausibly transfers to other semi-structured documents such as invoices and purchase orders with similar section variability; a cheap test would fine-tune on a mixed corpus and measure overlap counts.
  • Because the generator runs locally, it could be paired with a local OCR engine in a fully on-premise pipeline to create unlimited labeled training data without privacy leakage.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper proposes a method for document layout generation by fine-tuning Llama3-8B with LoRA on a small dataset of 107 receipt images. The model is prompted with a label set and outputs one bounding box per label, making the approach conditional. The authors compare this method against LayoutTransformer using Mahalanobis distance, area difference, and intra-label overlap counts in Table 2, and report that the proposed method 'clearly outperformed' the baseline. The paper also evaluates text classification using fine-tuned BERT and few-shot SLMs (Llama3-8B, Llama3.1-8B, Gemma2-9B) with and without concatenated bounding-box coordinates, finding a 12-point accuracy improvement for BERT when spatial information is added.

Significance. If the comparative claim were supported, the work would offer a valuable demonstration that a locally runnable small language model can generate document layouts from very few examples, with the practical advantage of being fully conditional on user-specified labels. The classification result, showing that concatenating bounding-box coordinates improves fine-tuned BERT accuracy from 0.62 to 0.74, is a suggestive empirical finding, although it is based on a single dataset and lacks significance testing. The paper is transparent about its hyperparameters and acknowledges several limitations in Section 5 and the Conclusion. However, the headline claim that the proposed method outperforms LayoutTransformer is not currently established because the evaluation protocol confounds the comparison and no uncertainty quantification is reported.

major comments (3)
  1. [Section 3.3.2, Table 2] The overlap metric is biased by the conditional/unconditional asymmetry. The proposed method receives the label set and produces exactly one bounding box per label, while LayoutTransformer is unconditional and can emit multiple boxes per label, including duplicates. Consequently, the comparison of 1 overlap versus 23 overlaps largely measures the difference between conditional and unconditional generation protocols rather than layout quality. The paper acknowledges the converse bias in Section 5, but Section 4.1 and the Conclusion still present the overlap count as evidence of clear superiority. To support the claim, the authors should include a conditional baseline (e.g., LayoutGPT, which is cited in Section 2.3) or otherwise control for the number of boxes per label.
  2. [Section 4.1, Table 2] The Mahalanobis distance results are mixed: LayoutTransformer is closer to ground truth for origin distance on 4 of 8 labels and for closing distance on 3 of 8 labels. The superiority claim therefore rests on the area-difference and overlap-count metrics, both of which are confounded by the generation protocol. In addition, no confidence intervals or significance tests are reported for any layout metric, and the number of generated samples per model is not specified, so it is unclear whether the observed area differences (e.g., -866.95 for LayoutTransformer versus -223.82 for the proposed method on label I) are reliable. The statement in Section 4.1 that 'the proposed method clearly outperformed generation with LayoutTransformer' is not supported by the reported evidence.
  3. [Section 5 and Conclusion] The discussion correctly notes that LayoutTransformer's intra-label overlaps 'might give this method an unfair advantage' by reducing Mahalanobis distance. This caveat is not carried through to the interpretation of the results: Section 4.1 and the Conclusion still use the overlap count as decisive evidence of superiority, and the Appendix's claim that the proposed method 'outperforms the LayoutTransformer approach' is stated without the conditional-protocol caveat. The manuscript should either quantify the effect of duplicate boxes on all three metrics (Mahalanobis distance, area difference, overlap count) or restrict the claims to the conditional generation setting and adjust the abstract accordingly.
minor comments (5)
  1. [Section 4.1] In the paragraph describing closing points, the phrase 'while the distance to the origin point for C, L, I, LIT, and F labels was closer to the ground truth when using the proposed method' should read 'closing point' rather than 'origin point'.
  2. [Table 2] The columns labeled 'Overlaps' are not defined in the table caption; a note should clarify that these are intra-label overlaps and how they are counted when a model produces multiple bounding boxes for the same label.
  3. [Section 3.3.2] The sentence 'Once the model has been finetuned, the next step is to prompt it to return layout information based on provided labels by the user' is redundant and could be shortened.
  4. [References] Reference [37] is malformed ('Shumaylov Z. Zhao Y. et al. Shumailov, I. 2024.') and should be corrected to the standard nature citation format.
  5. [Section 4.2, Table 3] The reported 12-point accuracy improvement for BERT with bounding boxes is accompanied by standard deviations (4.51% without, 4.13% with) but no significance test; given the overlap in standard deviations, a paired test would strengthen the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the layout and classification experiments are supervised evaluations on held-out subsets, with no self-citations and no fitted parameters presented as predictions.

full rationale

The paper's layout generation claim rests on fine-tuning Llama3-8B with LoRA on 87% of the receipt dataset and comparing the generated bounding boxes against the withheld 13% using Mahalanobis distances, area differences, and intra-label overlaps. The classification experiments split strings into 85% training/validation and 15% testing. No evaluation metric is defined in terms of the method's own outputs in a way that would make the reported results true by construction, and no fitted parameter is renamed as a prediction. The only notable concern is the asymmetric comparison between the conditional proposed method and the unconditional LayoutTransformer, especially for the overlap metric; however, the paper itself acknowledges this asymmetry in Section 5, noting that LayoutTransformer's overlaps 'might give this method an unfair advantage' for the Mahalanobis distance. That is an experimental-validity issue, not circular reasoning: the overlap count is not a mathematical consequence of the proposed method's definition, and the paper does not assert a derivation from the metric. There are also no self-citations, no imported uniqueness theorems, and no ansatz smuggled in through prior work by the same authors. Therefore the derivation chain is self-contained and no circularity is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claims rest on a proprietary 107-document dataset, author-chosen hyperparameters, an unvalidated evaluation metric, and small held-out splits. No new theoretical entities are introduced. Ground truth quality and test set representativeness are the most fragile inputs.

free parameters (4)
  • LoRA fine-tuning hyperparameters = rank=32, scale=64, dropout=0.05, lr=1.5e-4, epochs=4
    Chosen by the authors without a search or sensitivity analysis; the layout generation results depend on these settings.
  • Layout generation train/test split = 87% train, 13% test (about 93 and 14 documents)
    Arbitrary split; the small test set drives the evaluation of the central layout claim.
  • BERT fine-tuning hyperparameters = epochs=3, lr=5e-5, cross-entropy loss
    Standard choices, but the 12% accuracy gain is shown only for this configuration.
  • Classification train/validation/test split = 85% train/validation, 15% test
    Arbitrary split for the text classification task.
assumptions (4)
  • domain assumption Manual annotations and section definitions for the 107 receipts are correct.
    Section 3.1 describes annotation with LabelStudio but provides no quality metrics or inter-annotator agreement.
  • domain assumption The proprietary OCR returns correct strings and bounding boxes.
    Section 3.1 states strings and boxes come from a proprietary OCR; errors would propagate into the classification labels.
  • ad hoc to paper Mahalanobis distance and area difference are valid proxies for layout generation quality.
    Section 3.3.2 introduces these metrics specifically to compare conditional and unconditional models; their validity for downstream document generation is not established.
  • domain assumption The small held-out test sets are representative of the receipt distribution.
    With roughly 14 test documents for layout and 15% of 5731 strings for classification, the conclusions may not generalize to other receipt samples.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatial Information Integration in Small Language Models for Document Layout Generation and Classification." pith.science (2026). https://pith.science/paper/BE2WYRAQ

@misc{pith2026250105497,
  author       = {Pith},
  title        = {Pith review of: Spatial Information Integration in Small Language Models for Document Layout Generation and Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BE2WYRAQ}},
  note         = {Machine review of arXiv:2501.05497}
}
read the original abstract

Document layout understanding is a field of study that analyzes the spatial arrangement of information in a document hoping to understand its structure and layout. Models such as LayoutLM (and its subsequent iterations) can understand semi-structured documents with SotA results; however, the lack of open semi-structured data is a limitation in itself. While semi-structured data is common in everyday life (balance sheets, purchase orders, receipts), there is a lack of public datasets for training machine learning models for this type of document. In this investigation we propose a method to generate new, synthetic, layout information that can help overcoming this data shortage. According to our results, the proposed method performs better than LayoutTransformer, another popular layout generation method. We also show that, in some scenarios, text classification can improve when supported by bounding box information.

Figures

Figures reproduced from arXiv: 2501.05497 by the authors.

Figure 1
Figure 1. Dataset samples to the left (a) vs Llama3-8B syn [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. LayoutTransformer samples [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Labels by bounding box total area any impact in accuracy for text classification, different observations were made; for example, classification precision was particularly low for L, H and F labels, all of which are present mostly at the top or the bottom of the page which suggests that the models are struggling to understand these sections of the document; however, it is important to consider other potential causes,… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 28 canonical work pages

  1. [1]

    Abayomi-Alli, Robertas Damaševičius, Atika Qazi, Mariam Adedoyin- Olowe, and Sanjay Misra

    Olusola O. Abayomi-Alli, Robertas Damaševičius, Atika Qazi, Mariam Adedoyin- Olowe, and Sanjay Misra. 2022. Data Augmentation and Deep Learning Methods in Sound Classification: A Systematic Review. Electronics 11, 22 (2022). https: //doi.org/10.3390/electronics11223795

    doi:10.3390/electronics11223795 2022

  2. [2]

    Serge Abiteboul. 1997. Querying semi-structured data. In Database Theory — ICDT ’97, Foto Afrati and Phokion Kolaitis (Eds.). Springer Berlin Heidelberg, Berlin, Heidelberg, 1–18

    work page 1997

  3. [3]

    Chan, Biao Zhang, Aleksandra Faust, and Hugo Larochelle

    Rishabh Agarwal, Avi Singh, Lei M Zhang, Bernd Bohnet, Luis Rosias, Stephanie C.Y. Chan, Biao Zhang, Aleksandra Faust, and Hugo Larochelle. 2024. Many-shot In-Context Learning. In ICML 2024 Workshop on In-Context Learning . https://openreview.net/forum?id=goi7DFHlqS

    work page 2024

  4. [4]

    Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Hu- mayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard Baraniuk

    work page

  5. [5]

    Markus Bayer, Marc-André Kaufhold, and Christian Reuter. 2022. A Survey on Data Augmentation for Text Classification. ACM Comput. Surv. 55, 7, Article 146 (dec 2022), 39 pages. https://doi.org/10.1145/3544558

    doi:10.1145/3544558 2022

  6. [6]

    Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. 2023. AudioLM: A Language Modeling Ap- proach to Audio Generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing 31 (2023), 2523–2533. https://doi.org/1...

    doi:10.1109/taslp.2023 2023

  7. [7]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    2020

  8. [8]

    Arindam Chaudhuri, Krupa Mandaviya, Pratixa Badelia, and Soumya K. Ghosh

    work page

Show all 46 references
  1. [9]

    C. H. Chen. 2005. Handbook Of Pattern Recognition And Computer Vision . World Scientific Publishing Co., Inc., USA

    2005

  2. [10]

    Google DeepMind. 2024. Gemma 2: Improving Open Language Models at a Practical Size. Google DeepMind Research Technical Report. Google

    2024

  3. [11]

    Li Deng. 2012. The MNIST Database of Handwritten Digit Images for Machine Learning Research [Best of the Web]. IEEE Signal Processing Magazine 29, 6 (2012), 141–142. https://doi.org/10.1109/MSP.2012.2211477

    2012

  4. [12]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In North American Chapter of the Association for Computational Linguistics . https: //api.semanticscholar.org/CorpusID:52967399

    2019

  5. [13]

    Esser, R

    P. Esser, R. Rombach, and B. Ommer. 2021. Taming Transformers for High- Resolution Image Synthesis. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE Computer Society, Los Alamitos, CA, USA, 12868–12878. https://doi.org/10.1109/CVPR46437.2021.01268

    2021

  6. [14]

    Weixi Feng, Wanrong Zhu, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Xuehai He, S Basu, Xin Eric Wang, and William Yang Wang. 2023. LayoutGPT: Com- positional Visual Planning and Generation with Large Language Models. In Advances in Neural Information Processing Systems, A. Oh, T....

    2023

  7. [15]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2020. Generative adversarial networks. Commun. ACM 63, 11 (oct 2020), 139–144. https://doi.org/10.1145/ 3422622

    2020

  8. [16]

    Davis, Vijay Mahade- van, and Abhinav Shrivastava

    Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S. Davis, Vijay Mahade- van, and Abhinav Shrivastava. 2020. LayoutTransformer: Layout Generation and Completion with Self-attention. 2021 IEEE/CVF International Conference on Com- puter Vision (ICCV) (2020), 984–994. https...

    2020

  9. [17]

    Tampakas

    Emmanouil Ikonomakis, Sotiris Kotsiantis, and V. Tampakas. 2005. Text Classifi- cation Using Machine Learning Techniques. WSEAS transactions on computers 4 (08 2005), 966–974

    2005

  10. [18]

    Radu Tudor Ionescu and Marius Popescu. 2016. Spatial Information in Text Categorization. Springer International Publishing, Cham, 229–241. https://doi. org/10.1007/978-3-319-30367-3_9

    2016 doi

  11. [19]

    Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. 2019. FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents. In 2019 Inter- national Conference on Document Analysis and Recognition Workshops (ICDARW) , Vol. 2. 1–6. https://doi.org/10.1109/ICDARW.2019.10029

    2019

  12. [20]

    Yixing Jiang, Jeremy Andrew Irvin, Ji Hun Wang, Muhammad Ahmed Chaudhry, Jonathan H Chen, and Andrew Y. Ng. 2024. Many-Shot In-Context Learning in Multimodal Foundation Models. In ICML 2024 Workshop on In-Context Learning . https://openreview.net/forum?id=j2rKwWXdcz

    2024

  13. [21]

    Jyothi, T

    A. Jyothi, T. Durand, J. He, L. Sigal, and G. Mori. 2019. LayoutVAE: Stochastic Scene Layout Generation From a Label Set. In 2019 IEEE/CVF International Con- ference on Computer Vision (ICCV) . IEEE Computer Society, Los Alamitos, CA, USA, 9894–9903. https://doi.org/10.1109/IC...

    2019

  14. [22]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. 2013. Auto-Encoding Variational Bayes. CoRR abs/1312.6114 (2013). https://api.semanticscholar.org/CorpusID:216078090

    2013 arXiv

  15. [23]

    Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. 2019. LayoutGAN: Generating Graphic Layouts with Wireframe Discriminators. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net. https...

    2019

  16. [24]

    Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. 2020. DocBank: A Benchmark Dataset for Document Layout Analysis. In Proceedings of the 28th International Conference on Computational Linguistics , Donia Scott, Nuria Bel, and Chengqing Zong (E...

    2020 doi

  17. [25]

    Lawrence Zitnick

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. 2014. Microsoft COCO: Common Objects in Context. In Computer Vision – ECCV 2014 , David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars (Eds.)....

    2014

  18. [26]

    M Lopez and DJ Smith. 1997. Information extraction for semi-structured docu- ments. Proc. Workshop on Management of Semi-structured Data ; Conference date: 01-01-1997

    1997

  19. [27]

    Kiran Maharana, Surajit Mondal, and Bhushankumar Nemade. 2022. A review: Data pre-processing and data augmentation techniques. Global Transitions Pro- ceedings 3, 1 (2022), 91–99. https://doi.org/10.1016/j.gltp.2022.04.020 Interna- tional Conference on Intelligent Engineering ...

    2022 doi

  20. [28]

    Mclachlan

    G. Mclachlan. 1999. Mahalanobis Distance. Resonance 4 (06 1999), 20–26. https: //doi.org/10.1007/BF02834632

    1999 doi

  21. [29]

    Meta. 2023. Llama 2: Open Foundation and Fine-Tuned Chat Models. Meta Research Lab Technical Report. Meta

    2023

  22. [30]

    Meta. 2024. The Llama 3 Herd of Models . Meta Research Lab Technical Report. Meta

    2024

  23. [31]

    MistralAI. 2023. Mistral 7B. Mistral Research Lab Technical Report. MistralAI

    2023

  24. [32]

    Shunji Mori, Hirobumi Nishida, and Hiromitsu Yamada. 1999. Optical character recognition. John Wiley & Sons, Inc., USA

    1999

  25. [33]

    OpenAI. 2023. GPT-4 Technical Report . OpenAI Research Technical Report. OpenAI

    2023

  26. [34]

    Raymond Ptucha, Felipe Petroski Such, Suhas Pillai, Frank Brockler, Vatsala Singh, and Paul Hutkowski. 2019. Intelligent character recognition using fully convolutional neural networks. Pattern Recognition 88 (2019), 604–613. https: //doi.org/10.1016/j.patcog.2018.12.017

    2019 doi

  27. [35]

    Alec Radford and Karthik Narasimhan. 2018. Improving Language Understanding by Generative Pre-Training. (2018). https://api.semanticscholar.org/CorpusID: 49313245

    2018

  28. [36]

    Khoshgoftaar

    Connor Shorten and Taghi M. Khoshgoftaar. 2019. A survey on Image Data Augmentation for Deep Learning. Journal of Big Data 6, 1 (06 Jul 2019), 60. https://doi.org/10.1186/s40537-019-0197-0

    2019 doi

  29. [37]

    Shumaylov Z. Zhao Y. et al. Shumailov, I. 2024. AI models collapse when trained on recursively generated data. Nature 631 (2024), 755–759. https://doi.org/10. 1038/s41586-024-07566-y

    2024

  30. [38]

    Claus Smitt, Michael Halstead, Alireza Ahmadi, and Chris McCool. 2022. Ex- plicitly Incorporating Spatial Information to Recurrent Networks for Agricul- ture. IEEE Robotics and Automation Letters 7, 4 (2022), 10017–10024. https: //doi.org/10.1109/LRA.2022.3188105

    2022

  31. [39]

    Tang, Seong-Whan Lee, and Ching Y

    Yuan Y. Tang, Seong-Whan Lee, and Ching Y. Suen. 1996. Automatic document processing: A survey. Pattern Recognition 29, 12 (1996), 1931–1952. https: //doi.org/10.1016/S0031-3203(96)00044-1

    1996 doi

  32. [40]

    Gemma Team. 2024. Gemma 2: Improving Open Language Models at a Practical Size. ArXiv abs/2408.00118 (2024). https://api.semanticscholar.org/CorpusID: 270843326

    2024 arXiv

  33. [41]

    2020-2022

    Maxim Tkachenko, Mikhail Malyuk, Andrey Holmanyuk, and Nikolai Li- ubimov. 2020-2022. Label Studio: Data labeling software. https:// github.com/heartexlabs/label-studio Open source software available from https://github.com/heartexlabs/label-studio

    2020

  34. [42]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In Advances in Neural Information Processing Systems , I. Guyon, U. Von Luxburg, SAC ’25, March 31–April 4, 2025, Catania...

    2017

  35. [43]

    Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. 2020. LayoutLM: Pre-training of Text and Layout for Document Image Understanding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery amp; Data Mining (KDD ’20) . ACM. htt...

    2020 doi

  36. [44]

    Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. 2019. PubLayNet: largest dataset ever for document layout analysis. In 2019 International Conference on Document Analysis and Recognition (ICDAR) . IEEE, 1015–1022. https://doi.org/ 10.1109/ICDAR.2019.00166

    2019

  37. [2017]

    Springer International Publishing, Cham, 9–41

    Optical Character Recognition Systems . Springer International Publishing, Cham, 9–41. https://doi.org/10.1007/978-3-319-50252-6_2

    doi

  38. [2024]

    In The Twelfth Interna- tional Conference on Learning Representations

    Self-Consuming Generative Models Go MAD. In The Twelfth Interna- tional Conference on Learning Representations . https://openreview.net/forum?id= ShjMHfmPs0

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.