Pith. sign in

REVIEW 5 major objections 5 minor 65 references

The OCR Quest for Generalization: Learning to recognize low-resource alphabets with model editing

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper establishes that averaging the per-domain fine-tuning differences of OCR expert models, then briefly fine-tuning the merged model, transfers to low-resource alphabets and historical ciphers substantially better than centralized…

desk verdict A useful new application of task-vector averaging to low-resource OCR, with a strong benchmark and honest framing; the main claim is plausible but under-verified without error bars, code, and a partition-sensitivity test. read the letter →

arxiv 2506.06761 v2 pith:XVHG52IZ submitted 2025-06-07 cs.LG cs.CV

classification cs.LGcs.CV
keywords transferlearningtaskarithmeticmodeleditinglow-resourceOCRhistoricalciphersmergingmeta-learningopticalcharacterrecognition
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

This paper tries to establish that an OCR system can learn new low-resource alphabets and historical ciphered scripts by training several expert models on separate source domains, averaging their task vectors, and then fine-tuning briefly on the new script. The claim is that this distributed route beats classical centralized fine-tuning even when both routes use the same seed and exactly the same target training data: for example, Arabic accuracy rises from 17.5% to 42.2%, Bangla from 11.5% to 46.2%, and the Borg cipher from 0.0% to 50.5%. The authors care because under-represented scripts are usually excluded from large pretraining, and because training on separate domains allows documents to stay where they are instead of being centralized. The intended upshot is a practical recipe for rapidly adding a new alphabet to a reading system.

What carries the argument

Load-bearing is the task vector, the difference between a per-domain fine-tuned expert and the shared pretrained seed, together with the averaging rule θ_{t+1} = θ_t + (1/N) Σ τ_t^n (Equation 2). This is the same task-arithmetic operation used in model editing; iterated, it coincides with the FedAvg and Reptile updates. The mechanism works because each expert moves only a short distance from a common HierText-pretrained seed, so their averaged displacement stays in a near-linear region and produces a generalist rather than a cancelling mixture. That generalist is then used as the initialization for short fine-tuning on the unseen script.

What would settle it

Re-run the same seven source domains from the same seed but with a larger per-domain training budget so the experts drift far apart; if the averaged model then transfers to Arabic or Borg worse than the centralized fine-tune under an identical target-data budget, the advantage depends on staying near the seed rather than on averaging itself. A more direct check is to compute pairwise cosine similarity between the task vectors: the claim predicts the two excluded domains are the least orthogonal, and that any new set of domains with uniformly high pairwise similarity reproduces the transfer gains.

Watch

Extended reading notes

Core claim

The central claim is that the model obtained by averaging per-domain task vectors, θ_ind = θ_0 + (1/N) Σ (θ_dn − θ_0), is a better starting point for low-resource transfer than a model fine-tuned centrally on the union of those same domains. Starting from that averaged seed, the same brief fine-tuning on each unseen script outperforms the centralized seed in all nine tested targets, and turns near-zero performance into usable recognition in the hardest cases. The paper also reports that the averaged model generalizes out-of-domain better than the centralized model on most held-out printed, map, form, meter-reading, and curved-text benchmarks, while staying competitive in-domain. The authors interpret the averaging as a single meta-learning step in which independent experts, kept close to a common high-resource seed, contribute a generalist displacement that preserves transferable features.

Load-bearing premise

The recipe assumes that models fine-tuned from the same starting point stay close enough in their internal parameters that averaging their differences yields a useful generalist instead of destructive cancellation; the paper inherits this assumption from earlier task-arithmetic work and had to exclude two of its own source domains because they violated it.

Editorial extensions

If this is right

  • A low-resource script can be added with a small amount of target data and a short fine-tuning run, since the averaged seed already carries transferable features.
  • With the same target data and training budget, transfer accuracy roughly doubles or triples compared with centralized fine-tuning on several scripts, including Arabic, Bangla, and Borg.
  • Out-of-domain robustness on unseen document types improves without sacrificing in-domain accuracy, so the averaging step can replace centralized pretraining as a general-purpose seed.
  • Source domains can be fine-tuned independently, which means institutions can keep historical documents local and still contribute to a shared reading model.

Reading between the lines

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

  • If the locality assumption holds for other backbones, the same recipe could let speech or translation systems adopt low-resource languages by averaging per-domain task vectors instead of central fine-tuning.
  • A practical caveat implicit in the paper is that the merge is only safe when the expert models are close in parameter space; the authors had to exclude two source domains from one averaging variant, so a deployment would need an automatic near-orthogonality check before merging.
  • The paper's claim could be probed further by testing whether the benefit survives when source domains are deliberately chosen to be very dissimilar to each other, which would stress the locality condition and reveal how general the recipe really is.
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

5 major / 5 minor

Summary. The paper proposes a distributed, model-merging approach to OCR generalization and low-resource alphabet transfer. Starting from a ViT-CTC model pretrained on HierText, the authors fine-tune separate expert models on seven OCR sub-domains (handwritten, scene, and printed), form a merged model by averaging the corresponding task vectors (Eq. 2), and use this merged model both for out-of-domain evaluation and as an initialization for fine-tuning on nine low-resource target scripts (Arabic, Chinese, Japanese, Korean, Bangla, Hindi, Vatican, Borg, Copiale). The central empirical claim is that fine-tuning this merged initialization on target data substantially outperforms fine-tuning the plain HierText seed under the same target data and hyperparameters, e.g., Arabic 17.5% to 42.2%, Bangla 11.5% to 46.2%, and Borg 0.0% to 50.5% (Table 3). The paper also studies the effect of the number of aggregation rounds T and sub-sampling of source domains, and releases model weights via a public model card.

Significance. If the central claim holds, the paper offers a practical, decentralized route to low-resource OCR with modest compute requirements, and it provides a broad evaluation across 20 datasets with a unified architecture and hyperparameter regime. The strengths are concrete: public datasets, a fixed architecture, identical target data for the compared routes, consistent large gains across nine targets, and publicly released model weights. The stress-test concern that the transfer advantage depends on hand-selecting near-orthogonal source domains does not, on my reading, land for the main transfer claim: Section 5.1 says the transfer experiments use θ_ind, the average of all seven sub-domain experts, while the orthogonal subset (Avg(Orth)) appears only as an ablation in Table 2. However, the paper should state this explicitly, because the row label 'Distributed (from Z)' in Table 3 is ambiguous. The main weakness is that the comparison in Table 3 is not fully controlled, as the distributed route consumes additional source-domain data; this is detailed in Major Comment 1.

major comments (5)
  1. [Table 3, Section 5.1] The claim of 'same exact training data' is not supported by the experiment as reported. The Distributed (from Z) route uses the seven source sub-domains D to build θ_ind before target fine-tuning, whereas Centralized (from Z) goes directly from the HierText seed θ0 to the target g. These two routes therefore do not consume the same training data: the distributed route has an additional training stage on D. To support the headline, the authors should add a centralized baseline that also consumes D before the same target fine-tuning (e.g., starting from the 'FT (HW, Scene)' or 'All Above' model in Table 1), or alternatively restrict the distributed route to sub-domain models trained only on data available to the centralized route. Without this control, the reported gains may be attributable to additional source-domain training rather than to the distributed averaging mechanism.
  2. [Table 3, ×Δ column] The summary statistic '×Δ' is never defined in the paper. The values 0.69, 1.62, and 1.04 are used to support the claim that the distributed strategy yields large improvements, but a reader cannot determine whether ×Δ is an average ratio, an average absolute gain, or some other aggregate. Please define it explicitly and, if it is used to quantify the 'almost three times' improvement, report the underlying per-target values and their uncertainty.
  3. [Section 4.2, Tables 1-4] All main results are single runs without reported seeds, error bars, or significance tests. Figure 5 reports standard deviations, but it is unclear whether these are across independent training runs or across sub-sampling proportions. For a claim of consistent superiority across nine targets, the authors should report mean±std over at least three independent seeds for the key comparisons in Table 3, or justify why single runs are sufficient given the effect size.
  4. [Section 5.1, Table 3] It is not stated explicitly which merged model feeds the 'Distributed (from Z)' row. The text says θ_ind is used, but the row label and the earlier discussion of Avg(Orth) leave room for ambiguity. Please state in the table caption and in Section 5.1 whether the transfer results use Avg(Ind.) (all seven experts) or Avg(Orth.) (excluding MLT and Parzival). This matters because if Avg(Orth.) is used, the orthogonal-subset selection becomes load-bearing; if Avg(Ind.) is used, the skeptic's concern about hand-selecting orthogonal domains is addressed for the transfer claim. Either way, the ambiguity should be resolved.
  5. [Section 4.2, Table 3] The fixed 30-epoch transfer budget is justified as 'ensuring convergence in all cases', but Table 3 shows near-zero accuracy for Borg (0.0%) and Chinese (1.03%) under Centralized (from Z), which suggests that the centralized baseline may be severely undertrained or overfit under this budget. Please show learning curves or a sweep of transfer epochs for both routes, or qualify the claim as 'under a fixed optimization budget'. This would rule out the possibility that the distributed advantage is an artifact of the chosen epoch count rather than a property of the merged initialization.
minor comments (5)
  1. [Throughout] There are several typos, including 'valiate' (Section 5.1), 'tunning', and 'finetunning'. A careful proofread is recommended.
  2. [Figure 7] The claim that performance 'decays due to over-specialization' after many meta-learning steps is not quantitatively supported in the text; please report the actual accuracies or a statistical comparison, or soften the claim.
  3. [Related Work / Conclusions] The paper contrasts itself with 'state-of-the-art meta-learning' but does not run any meta-learning baseline such as Reptile or MAML. If the distinction is important, please add such a baseline or reframe the claim to avoid implying a direct comparison.
  4. [Table 4] The ablation table has dense formatting with abbreviated task-vector names and no explicitly labeled row/column groups; adding clear group labels and a note on how 'Best' and 'Worst' are defined would improve readability.
  5. [Abstract / Introduction] The phrase 'Even when using the same exact training data' is potentially misleading, as discussed in Major Comment 1; consider rephrasing to 'same target training data' or 'identical fine-tuning data' once the controlled comparison is added.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the task-vector averaging comparison is self-contained against external baselines with no target-derived fitting.

full rationale

The paper's central claim is an empirical comparison: starting from a shared HierText seed θ0, averaging task vectors from seven independently fine-tuned Latin OCR experts (Eq. 2) gives an initialization that transfers to low-resource alphabets and ciphers better than centralized fine-tuning from the same seed. No target-derived quantity is fitted: the same k, T, learning rate, and 30 transfer epochs are used for every script, and Table 3's 'Distributed (from Z)' uses all seven domain experts (θind) without per-target selection. The averaged model is defined by Eq. 2 from source-domain task vectors only, and the transfer gain is measured against external baselines on held-out target splits, not against a quantity used to construct the model. The locality and orthogonality assumptions are imported from external prior work ([16], [37], [40]) rather than from the authors' own results; the only author-overlapping citation ([42], the Esposalles dataset) is a data source, not a load-bearing premise. The Avg(Orth) exclusion of MLT and Parzival is a post hoc ablation for the OOD table, not the route used for the main transfer results, so it does not make the headline prediction equivalent to its inputs. No equation in the paper defines a prediction in terms of the target data used to evaluate it, and no fitted parameter is renamed as a prediction; hence no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

One hand-set hyperparameter family and four domain assumptions. No new theoretical entities. The largest burden is the locality assumption behind the averaging in Eq. 2; the paper flags its failure modes by excluding non-orthogonal experts in one ablation.

free parameters (3)
  • Default sub-domain training steps and aggregation rounds = k = 75, T = 1
    Hand-chosen operating point for all domain experts; Figure 7 explores k and T, but main transfer tables use this default, so the reported gains are tied to it.
  • Transfer fine-tuning epochs = 30
    Set empirically in Section 4.2 to maintain a common regime across all targets; centralized and distributed comparisons both use 30, but the absolute accuracies depend on this choice.
  • Optimizer, learning rate, batch size = Adam, lr = 1e-5, batch = 128
    Fixed hyperparameters with no per-target tuning; the paper argues this is a fair common baseline, but no sensitivity analysis is provided for the headline comparisons.
assumptions (4)
  • domain assumption Task vectors from fine-tuned models can be linearly averaged without destructive interference (locality/linearity assumption).
    Invoked in Eq. 2 and Section 3.3; inherited from [16] and [40]. The Avg(Orth) variant in Table 2 removes MLT and Parzival because of non-orthogonality, showing the assumption is imperfect for OCR.
  • domain assumption A HierText Latin-pretrained seed is a sufficiently central initialization for all downstream OCR experts.
    Section 3.2 states the seed ensures the experts do not diverge significantly; random-init distributed models are worse in Table 3, so the main result depends on this seed choice.
  • standard math The FedAvg/Reptile/task-arithmetic equivalence justifies interpreting T averaging rounds as a meta-learning procedure.
    Cited from [11, 19, 37] in Section 3.2; the paper does not re-derive the equivalence but relies on it to connect its method to earlier meta-learning.
  • domain assumption The fixed 30-epoch transfer protocol is equally fair to centralized and distributed starting points.
    All target scripts use the same epochs and learning rate; if centralized fine-tuning benefits from earlier stopping or a lower learning rate, the headline gain shrinks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The OCR Quest for Generalization: Learning to recognize low-resource alphabets with model editing." pith.science (2026). https://pith.science/paper/XVHG52IZ

@misc{pith2026250606761,
  author       = {Pith},
  title        = {Pith review of: The OCR Quest for Generalization: Learning to recognize low-resource alphabets with model editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XVHG52IZ}},
  note         = {Machine review of arXiv:2506.06761}
}
read the original abstract

Achieving robustness in recognition systems across diverse domains is crucial for their practical utility. While ample data availability is usually assumed, low-resource languages, such as ancient manuscripts and non-western languages, tend to be kept out of the equations of massive pretraining and foundational techniques due to an under representation. In this work, we aim for building models which can generalize to new distributions of data, such as alphabets, faster than centralized fine-tune strategies. For doing so, we take advantage of the recent advancements in model editing to enhance the incorporation of unseen scripts (low-resource learning). In contrast to state-of-the-art meta-learning, we showcase the effectiveness of domain merging in sparse distributions of data, with agnosticity of its relation to the overall distribution or any other prototyping necessity. Even when using the same exact training data, our experiments showcase significant performance boosts in \textbf{transfer learning} to new alphabets and \textbf{out-of-domain evaluation} in challenging domain shifts, including historical ciphered texts and non-Latin scripts. This research contributes a novel approach into building models that can easily adopt under-represented alphabets and, therefore, enable document recognition to a wider set of contexts and cultures.

Figures

Figures reproduced from arXiv: 2506.06761 by the authors.

Figure 1
Figure 1. Summary of the two main contributions of this work. Given a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Model sampling method for out-of-domain generalization, where models [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Landscape of some of the datasets used in this research: Esposalles (a), George [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Taxonomy (Venn diagram) of the datasets used in this work. We categorize data [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Average accuracy achieved by our approach (red), baseline (green) and tradi [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Performance of the different models trained on multilingual and ciphered datasets [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Accuracy of models pre-trained under a different number of epochs ( [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 55 canonical work pages

  1. [1]

    Albets, R.P.: Els arxius, la llei d’arxius i documents i el sistema d’arxius de catalunya: on som? Lligall 36, 44–67 (2013), https://raco.cat/ index.php/lligall/article/view/340107/431046, accessed: 2025- 02-13

  2. [2]

    https://cl.lingfil.uu.se/~bea/borg/ (Jan 2018)

    Aldarrab, N., Knight, K., Megyesi, B.: The Borg.lat.898 Cipher. https://cl.lingfil.uu.se/~bea/borg/ (Jan 2018)

  3. [3]

    Atienza, R.: Vision transformer for fast and efficient scene text recogni- tion (2021)

  4. [4]

    Archivaria pp

    Brothman, B.: The past that archives keep: memory, history, and the preservation of archival records. Archivaria pp. 48–80 (2001)

  5. [5]

    In: Proceedings of the 22nd ACM Symposium on Document Engineering

    Campiotti, I., Lotufo, R.: Optical character recognition with transform- ers and ctc. In: Proceedings of the 22nd ACM Symposium on Document Engineering. pp. 1–4 (2022)

  6. [6]

    In: 32nd USENIX Security Symposium (USENIX Security 23)

    Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffu- sion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253–5270 (2023)

  7. [7]

    ACM Computing Surveys (CSUR)54(2), 1–35 (2021)

    Chen, X., Jin, L., Zhu, Y., Luo, C., Wang, T.: Text recognition in the wild: A survey. ACM Computing Surveys (CSUR)54(2), 1–35 (2021)

  8. [8]

    arXiv preprint arXiv:1908.00262 (2019)

    Choi, J., Jeong, M., Kim, T., Kim, C.: Pseudo-labeling curriculum for unsupervised domain adaptation. arXiv preprint arXiv:1908.00262 (2019)

Show all 65 references
  1. [9]

    International Journal on Document Analysis and Recognition (IJDAR)23(1), 31–52 (2020)

    Ch’ng, C.K., Chan, C.S., Liu, C.L.: Total-text: toward orientation ro- bustness in scene text detection. International Journal on Document Analysis and Recognition (IJDAR)23(1), 31–52 (2020)

  2. [10]

    In: Fink, G.A., Jain, R., Kise, K., Zanibbi, R

    Etter, D., Carpenter, C., King, N.: A hybrid model for multilingual ocr. In: Fink, G.A., Jain, R., Kise, K., Zanibbi, R. (eds.) Document Analysis and Recognition - ICDAR 2023. pp. 467–483. Springer Nature Switzerland, Cham (2023) 22

  3. [11]

    Fallah, A., Mokhtari, A., Ozdaglar, A.: Personalized federated learn- ing: A meta-learning approach (2020),https://arxiv.org/abs/2002. 07948

  4. [12]

    Pattern recognition letters33(7), 934–942 (2012)

    Fischer, A., Keller, A., Frinken, V., Bunke, H.: Lexicon-freehandwritten word spotting using character hmms. Pattern recognition letters33(7), 934–942 (2012)

  5. [13]

    In: Proceedings of the 23rd international conference on Machine learning

    Graves, A., Fernández, S., Gomez, F., Schmidhuber, J.: Connection- ist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In: Proceedings of the 23rd international conference on Machine learning. pp. 369–376 (2006)

  6. [14]

    In: International Conference on Historical Cryptology

    Héder, M., Megyesi, B.: The decode database of historical ciphers and keys: Version 2. In: International Conference on Historical Cryptology. pp. 111–114 (2022)

  7. [15]

    In: Document Analysis and Recognition– ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part III 16

    Huang, Y., Jin, L., Peng, D.: Zero-shot chinese text recognition via matching class embedding. In: Document Analysis and Recognition– ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part III 16. pp. 127–141. Springer (2021)

  8. [16]

    Ilharco, G., Ribeiro, M.T., Wortsman, M., Gururangan, S., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic (2023)

  9. [17]

    AI & society37(3), 823–835 (2022)

    Jaillant, L., Caputo, A.: Unlocking digital archives: cross-disciplinary perspectives on ai and born-digital data. AI & society37(3), 823–835 (2022)

  10. [18]

    In: 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW)

    Jaume, G., Ekenel, H.K., Thiran, J.P.: Funsd: A dataset for form under- standing in noisy scanned documents. In: 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW). IEEE

  11. [19]

    Jiang, Y., Konečný, J., Rush, K., Kannan, S.: Improving feder- ated learning personalization via model agnostic meta learning (2023), https://arxiv.org/abs/1909.12488

  12. [20]

    In: Pattern Recognition: 40th 23 German Conference, GCPR 2018, Stuttgart, Germany, October 9-12, 2018, Proceedings 40

    Kang, L., Toledo, J.I., Riba, P., Villegas, M., Fornés, A., Rusinol, M.: Convolve, attend and spell: An attention-based sequence-to-sequence model for handwritten word recognition. In: Pattern Recognition: 40th 23 German Conference, GCPR 2018, Stuttgart, Germany, October 9-12,...

  13. [21]

    In: 2013 12th international con- ference on document analysis and recognition

    Karatzas, D., Shafait, F., Uchida, S., Iwamura, M., i Bigorda, L.G., Mestre, S.R., Mas, J., Mota, D.F., Almazan, J.A., De Las Heras, L.P.: Icdar 2013 robust reading competition. In: 2013 12th international con- ference on document analysis and recognition. pp. 1484–1493. IEEE (2013)

  14. [22]

    Archives and manuscripts33(1), 44–61 (2005)

    Ketelaar, E.: Sharing, collected memories in communities of records. Archives and manuscripts33(1), 44–61 (2005)

  15. [23]

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization (2017)

  16. [24]

    In: Proceed- ings of the 4th Workshop on Building and Using Comparable Corpora: Comparable Corpora and the Web

    Knight, K., Megyesi, B., Schaefer, C.: The Copiale Cipher. In: Proceed- ings of the 4th Workshop on Building and Using Comparable Corpora: Comparable Corpora and the Web. pp. 2–9. Association for Computa- tional Linguistics, Portland, Oregon (Jun 2011)

  17. [25]

    Science 350(6266), 1332–1338 (2015)

    Lake, B.M., Salakhutdinov, R., Tenenbaum, J.B.: Human-level concept learning through probabilistic program induction. Science 350(6266), 1332–1338 (2015)

  18. [26]

    Journal of Electronic Imaging 28(1), 013023 (2019)

    Laroca, R., Barroso, V., Diniz, M.A., Gonçalves, G.R., Schwartz, W.R., Menotti, D.: Convolutional neural networks for automatic meter reading. Journal of Electronic Imaging 28(1), 013023 (2019). https://doi.org/10.1117/1.JEI.28.1.013023

  19. [27]

    Rutgers University Press (2016)

    Lawrence, S.C.: Privacy and the past: Research, law, archives, ethics. Rutgers University Press (2016)

  20. [28]

    In: Workshop on challenges in representation learning, ICML

    Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896. Atlanta (2013)

  21. [29]

    arXiv preprint arXiv:2410.11005 (2024) 24

    Lin, F., Mao, S., La Malfa, E., Hofmann, V., de Wynter, A., Yao, J., Chen, S.Q., Wooldridge, M., Wei, F.: One language, many gaps: Eval- uating dialect fairness and robustness of large language models in rea- soning tasks. arXiv preprint arXiv:2410.11005 (2024) 24

  22. [30]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)

    Long, S., Qin, S., Panteleev, D., Bissacco, A., Fujii, Y., Raptis, M.: Towards end-to-end unified scene text detection and layout analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)

  23. [31]

    International Journal on Document Analysis and Recognition5, 39–46 (2002)

    Marti, U.V., Bunke, H.: The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition5, 39–46 (2002)

  24. [32]

    McMahan, H.B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data (2023), https://arxiv.org/abs/1602.05629

  25. [33]

    In: 2019 Interna- tional Conference on Document Analysis and Recognition (ICDAR)

    Michael, J., Labahn, R., Grüning, T., Zöllner, J.: Evaluating sequence- to-sequence models for handwritten text recognition. In: 2019 Interna- tional Conference on Document Analysis and Recognition (ICDAR). pp. 1286–1293. IEEE (2019)

  26. [34]

    In: BMVC-British machine vision con- ference

    Mishra, A., Alahari, K., Jawahar, C.: Scene text recognition using higher order language priors. In: BMVC-British machine vision con- ference. BMVA (2012)

  27. [35]

    In: ComputerVision–ECCV2020: 16thEuropeanConference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXV 16

    Musgrave, K., Belongie, S., Lim, S.N.: A metric learning reality check. In: ComputerVision–ECCV2020: 16thEuropeanConference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXV 16

  28. [36]

    Nayef, N., Patel, Y., Busta, M., Chowdhury, P.N., Karatzas, D., Khlif, W., Matas, J., Pal, U., Burie, J.C., lin Liu, C., Ogier, J.M.: Icdar2019 robust reading challenge on multi-lingual scene text detection and recog- nition – rrc-mlt-2019 (2019)

  29. [37]

    Nichol, A., Achiam, J., Schulman, J.: On first-order meta-learning al- gorithms (2018), https://arxiv.org/abs/1803.02999

  30. [38]

    Nikolaidou, K., Seuret, M., Mokayed, H., Liwicki, M.: A survey of his- torical document image datasets (2022)

  31. [39]

    Archival Is- sues 41(1) (2021) 25

    Olliff, M.T., Dill, E.: The distributed archives model: a strategy for sharing authority with partners to document communities. Archival Is- sues 41(1) (2021) 25

  32. [40]

    Ortiz-Jimenez, G., Favero, A., Frossard, P.: Task arithmetic in the tan- gent space: Improved editing of pre-trained models (2023)

  33. [41]

    http:// rrc.cvc.uab.es, last visited on November 11

    Robust Reading Competition: Robust reading competition. http:// rrc.cvc.uab.es, last visited on November 11

  34. [42]

    Pat- tern Recognition46(6), 1658–1669 (2013)

    Romero, V., Fornés, A., Serrano, N., Sánchez, J.A., Toselli, A.H., Frinken, V., Vidal, E., Lladós, J.: The esposalles database: An an- cient marriage license corpus for off-line handwriting recognition. Pat- tern Recognition46(6), 1658–1669 (2013)

  35. [43]

    In: 2011 international conference on document analysis and recognition

    Roy, P.P., Ramel, J.Y., Ragot, N.: Word retrieval in historical docu- ment using character-primitives. In: 2011 international conference on document analysis and recognition. pp. 678–682. IEEE (2011)

  36. [44]

    arXiv preprint arXiv:1511.04176 (2015)

    Sahu, D.K., Sukhwani, M.: Sequence to sequence learning for optical character recognition. arXiv preprint arXiv:1511.04176 (2015)

  37. [45]

    In: International Conference on Document Analysis and Recognition

    Seuret, M., van der Loop, J., Weichselbaumer, N., Mayr, M., Molnar, J., Hass, T., Christlein, V.: Combining ocr models for reading early modern books. In: International Conference on Document Analysis and Recognition. pp. 342–357. Springer (2023)

  38. [46]

    arXiv preprint arXiv:2411.11295 (2024)

    Shu, P., Chen, J., Liu, Z., Wang, H., Wu, Z., Zhong, T., Li, Y., Zhao, H., Jiang, H., Pan, Y., et al.: Transcending language boundaries: Harnessing llms for low-resource language translation. arXiv preprint arXiv:2411.11295 (2024)

  39. [47]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Singh, A., Pang, G., Toh, M., Huang, J., Galuba, W., Hassner, T.: Textocr: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

  40. [48]

    In: Document Analysis and Recognition–ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part III 16

    Singh, S.S., Karayev, S.: Full page handwriting recognition via image to sequence extraction. In: Document Analysis and Recognition–ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part III 16. pp. 55–69. Springer (2021)

  41. [49]

    arXiv preprint arXiv:2406.08726 (2024) 26

    Smith, G., Fleisig, E., Bossi, M., Rustagi, I., Yin, X.: Standard language ideology in ai-generated language. arXiv preprint arXiv:2406.08726 (2024) 26

  42. [50]

    Soltero, R.: Decentralized technologies & the future of historical narra- tives

  43. [51]

    Pattern Recognition Letters160, 43–49 (2022)

    Souibgui, M.A., Fornés, A., Kessentini, Y., Megyesi, B.: Few shots are all you need: a progressive learning approach for low resource handwrit- ten text recognition. Pattern Recognition Letters160, 43–49 (2022)

  44. [52]

    Advances in neural information processing systems14 (2001)

    Szummer, M., Jaakkola, T.: Partially labeled classification with markov random walks. Advances in neural information processing systems14 (2001)

  45. [53]

    In: Document Recognition

    Taghva, K., Borsack, J., Condit, A.: Expert system for automatically correcting ocr output. In: Document Recognition. vol. 2181, pp. 270–

  46. [54]

    Australian Review of Applied Linguistics (2025)

    Tran, H., Stell, A.: Beyond borders or building new walls? the potential for generative ai in recolonising the learning of vietnamese dialects and mandarin varieties. Australian Review of Applied Linguistics (2025)

  47. [55]

    In: arXiv preprint arXiv:1601.07140 (2016), http://vision

    Veit, A., Matera, T., Neumann, L., Matas, J., Belongie, S.: Coco-text: Dataset and benchmark for text detection and recognition in natural images. In: arXiv preprint arXiv:1601.07140 (2016), http://vision. cornell.edu/se3/wp-content/uploads/2016/01/1601.07140v1.pdf

  48. [56]

    Advances in neural information process- ing systems29 (2016)

    Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.: Matching networks for one shot learning. Advances in neural information process- ing systems29 (2016)

  49. [57]

    In: International Conference on Document Analysis and Recogni- tion

    Vlachou-Efstathiou, M., Siglidis, I., Stutzmann, D., Aubry, M.: An interpretable deep learning approach for morphological script type anal- ysis. In: International Conference on Document Analysis and Recogni- tion. pp. 3–21. Springer (2024)

  50. [58]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wan, Z., Zhang, J., Zhang, L., Luo, J., Yao, C.: On vocabulary reliance in scene text recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11425–11434 (2020)

  51. [59]

    In: 2011 International conference on computer vision

    Wang, K., Babenko, B., Belongie, S.: End-to-end scene text recognition. In: 2011 International conference on computer vision. pp. 1457–1464. IEEE (2011) 27

  52. [60]

    In: 2019 International Conference on Document Anal- ysis and Recognition (ICDAR)

    Weinman, J., Chen, Z., Gafford, B., Gifford, N., Lamsal, A., Niehus- Staab, L.: Deep neural networks for text detection and recognition in historical maps. In: 2019 International Conference on Document Anal- ysis and Recognition (ICDAR). pp. 902–909. IEEE (2019)

  53. [61]

    In: Proceedings of the 25th international conference on Ma- chine learning

    Weston, J., Ratle, F., Collobert, R.: Deep learning via semi-supervised embedding. In: Proceedings of the 25th international conference on Ma- chine learning. pp. 1168–1175 (2008)

  54. [62]

    In: International Conference on Document Analysis and Recognition

    Wick, C., Reul, C.: One-model ensemble-learning for text recognition of historical printings. In: International Conference on Document Analysis and Recognition. pp. 385–399. Springer (2021)

  55. [63]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yan, R., Peng, L., Xiao, S., Yao, G.: Primitive representation learning for scene text recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 284–293 (2021)

  56. [64]

    In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Zhang, C., Tao, M., Huang, Q., Lin, J., Chen, Z., Feng, Y.: Mc2: Towards transparent and culturally-aware nlp for minority languages in china. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 8832–8850 (2024)

  57. [65]

    arXiv preprint arXiv:2007.11462 (2020) 28

    Zhang, W., Qiu, Y., Bai, S., Zhang, R., Wei, X., Bai, X.: Fedocr: Communication-efficient federated learning for scene text recognition. arXiv preprint arXiv:2007.11462 (2020) 28

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.