Pith. sign in

REVIEW 3 major objections 4 minor 42 references

A routed multi-expert system reusing fine-tuning checkpoints matches the domain-label oracle on three Manchu script styles, reaching 0.30% CER on regular script, 1.57% on memorials, and 4.83% on running script, with a 99.3% accurate page-st

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 02:55 UTC pith:L5KHVCED

load-bearing objection Honest proof-of-concept for OCR expert routing, but the headline oracle-match is statistically fragile due to retrospective selection and tiny test sets. the 3 major comments →

arxiv 2607.14041 v1 pith:L5KHVCED submitted 2026-07-15 cs.CV cs.AIcs.LG

Multi-Expert Routing for Multi-Domain Low-Resource OCR: A Manchu Case Study

Zhan Chen , Jiqiao Ma , Chih-wen Kuo This is my paper
classification cs.CV cs.AIcs.LG
keywords Manchu OCRmulti-expert routingdomain classificationlow-resource OCRhistorical document recognitioncheckpoint reusepage-level image classifierNougat
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.

For low-resource OCR where a single language appears in several visually distinct scripts, training one recognizer forces a compromise. This paper shows a better use of an iterative fine-tuning history: keep every checkpoint as a potential domain specialist, add a lightweight page-level image classifier that decides which checkpoint reads each page, and the routed system can match the performance of an oracle that already knows each page's script style. On Manchu test sets, the router achieves 99.3% page-level domain accuracy and the routed system reaches 0.30% character error on regular script, 1.57% on memorials, and 4.83% on running script—numbers no single checkpoint attains. The result matters because it converts ordinary training byproducts into value, and because two of the three chosen experts were never trained specifically for the domain they end up serving.

Core claim

On three frozen, source-separated Manchu test sets, a routed system composed of checkpoints from a fine-tuning version stream plus a small page-level image classifier matches the domain-label oracle to two decimal places: 0.30% CER on regular script (125 pages), 1.57% on memorials (10 pages), and 4.83% on running script (13 pages). The router places 147 of 148 pages in the correct domain, and the one misrouted page changes the regular-script CER only at the third decimal place. The selected expert pool combines checkpoints from different training stages: v3.0 for regular script and v3.0-s2 for memorials were chosen after training, while the running-script expert, v5.0, was trained specifical

What carries the argument

The central mechanism is a two-part system: a pool of page-to-text experts, all sharing a Nougat encoder–decoder architecture but trained at different points in the version stream with different data snapshots and augmentation recipes, and a small residual image classifier (ResNet-18, 224×224 input, ImageNet-initialized) that maps a page image to a three-way domain label (regular / memorial / running). The classifier is trained on source-metadata-derived labels with source-balanced sampling, and it dispatches each page to the checkpoint that historically performed best on that domain. The domain-label oracle—using true domain labels and the fixed expert mapping—serves as the reference policy

Load-bearing premise

The 10-page memorial and 13-page running-script test sets are large enough that the CER rankings that select the domain experts reflect genuine differences in skill rather than page-to-page noise.

What would settle it

Run a bootstrap or leave-one-page-out analysis on the frozen 10 memorial and 13 running-script pages to obtain confidence intervals for each expert's mean CER; if the best-expert intervals overlap or the chosen specialist changes under resampling, the routed system's two-decimal match to the oracle is not a stable result.

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

If this is right

  • If the routed system matches the oracle, the remaining CER is recognizer error rather than routing error, so future gains lie in stronger specialists, especially for running script.
  • A new visual domain can be added by training or designating one checkpoint and adding a router class, without retraining the existing expert pool.
  • The extra inference cost is one tiny classifier forward pass per page, making the approach practical for large archival corpora.
  • Version histories in long-running OCR projects have reuse value beyond the final model, turning discarded checkpoints into complementary specialists.
  • The published protocol fixes the test sets, router design, and per-page predictions, so the routed numbers can be independently reproduced and extended.

Where Pith is reading between the lines

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

  • If the pattern holds on larger handwriting test sets, retrospective checkpoint selection could become a standard lifecycle practice: keep all checkpoints, pick per-domain bests on a style-annotated held-out set, and dispatch with a cheap visual classifier.
  • The approach likely transfers to other multi-style historical scripts where domain labels can be derived from source metadata (e.g., handwritten vs. print, different hands or periods), not just Manchu.
  • Because the router matches the oracle at the reported precision, the practical bottleneck is the absence of a strong running-script recognizer; investing in handwriting-specific training data or architectures should be the next step.
  • The small handwriting test sets (10 and 13 pages) mean the expert-selection step is potentially noisy; a prospective evaluation with a separate model-selection set and larger probes would test whether the two-decimal match survives.

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

3 major / 4 minor

Summary. The paper proposes a multi-expert OCR system for historical Manchu documents, where a lightweight ResNet-18 domain router dispatches each page to one of several pre-existing Nougat checkpoints drawn from an iterative fine-tuning history. On three frozen test sets (125 regular-script pages, 10 memorial pages, 13 running-script pages), the routed system reports CERs of 0.30%, 1.57%, and 4.83%, matching a “domain-label oracle” that uses true page labels and one designated expert per domain. The router achieves 99.3% page-level domain accuracy. The authors are explicit that the expert pool was designated retrospectively from the completed test-set results, and they frame the handwriting sets as small domain-shift probes. The paper also reports ablations on source-balanced router training and compares three backbone architectures.

Significance. The core idea—reusing checkpoints from a training history as domain specialists and selecting among them at inference with a lightweight classifier—is practical and potentially valuable for low-resource historical scripts. If the reported headline result were statistically well founded, it would be a useful demonstration that model-level routing can combine complementary strengths without retraining a generalist. The paper is also unusually transparent: frozen source-separated test sets, per-page predictions, a leakage-safe router training protocol, and explicit disclosure of the retrospective expert-pool designation are all strengths. The router's 99.3% accuracy on 148 pages is a genuine measurement. However, the central claim that the routed system “matches the domain-label oracle to two decimal places” is weakened by the fact that the oracle's chosen experts are the column minima of the same frozen test sets, and the 10- and 13-page probes cannot support the reported precision.

major comments (3)
  1. [§5–§6, Table 1] The headline result is close to being an artifact of selection on the evaluation data. The designated experts (v3.0 for regular, v3.0-s2 for memorials, v5.0 for running) are the column minima of Table 1 computed on the same frozen test sets later used in Table 2. The reported “match” with the domain-label oracle is therefore not an independent estimate of expert-pool selection; it is a restatement of the selection rule. The instability is concrete: on the 10-page memorial set, v3.0-s2 (1.57% CER) and v5.0-mem-tuned (1.60%) differ by roughly one misrecognized character across all pages. This is far below page-level noise. Please either designate experts on a validation split that is disjoint from P125/Z10/C13, or, at minimum, provide per-page bootstrap or permutation-based confidence intervals demonstrating that the expert ordering is stable. Without this, the two-decimal match should be
  2. [§7, Table 2] All CER numbers in Table 2 are point estimates with no uncertainty quantification, despite sample sizes of 10 and 13 pages. On a 10-page set, a single page's CER can change the mean by roughly 0.1–1 percentage points, so the claim of matching “to two decimal places” is not meaningful without associated confidence intervals. The paper even reports routed and oracle CERs that differ in the third decimal (0.299% vs 0.296% for regular script) and then rounds them to the same value. I request bootstrap CIs for every cell in Table 2 and Figure 4, or a paired comparison (e.g., sign test or Wilcoxon) for the expert differences that motivate selection. This is essential for the 1.57% vs 1.60% memorial comparison.
  3. [§7 “Effect of domain-specific training”] The claim that the v5.0 running-script campaign “reduces” running-script CER from 26.68% to 4.83% is presented as evidence for adding a purpose-trained expert, but the v5.0 checkpoint differs from v4.1-s2 in data snapshot, augmentation recipe, and training schedule simultaneously. The paper acknowledges this confound in §10, but the running-text interpretation remains causal. Please either add a controlled comparison (same data and recipe, only the target-domain composition varied) or explicitly relabel this observation as a confounded, retrospective comparison. This does not invalidate the routing architecture, but it does affect the strength of the “new domain can be added as an expert” claim.
minor comments (4)
  1. [§2] Typo: “the manc.humanuscript database” should be “the manc.hu manuscript database.”
  2. [§6, Figure 3] The confusion matrix in Figure 3 is informative but the caption and text use “rows: true domain; columns: expert routed to.” A small schematic labeling the axes directly on the figure would improve readability.
  3. [§10, Appendix C] The parser showcase in Appendix C and the “10/10” match in Figure 6, while labeled as an interface example, are tangential to the OCR benchmark. Consider moving this material to a separate project webpage or adding an explicit sentence that this is not an evaluation result. The provisional ensemble targets (0.28%, 1.39%, 3.81%) in §10 are also potentially confusing; they are already marked as not part of the present benchmark, but a short “these are future targets, not results” note immediately before the list would help.
  4. [Table 3] The sentence in §7 reporting Donut's running-script result says “reaches 5.89% on running script,” which matches Table 3, but the earlier text says “Donut is weaker on regular script” while its running-script CER is essentially tied with Nougat. Please clarify that Donut is weak on regular script but competitive on running script in this table.

Circularity Check

0 steps flagged

No significant circularity; the routed-vs-oracle comparison is an explicitly scoped benchmark evaluation with leakage-safe router training and full disclosure of retrospective expert selection.

full rationale

The paper's main result is that a trained page-level router can reproduce the domain-label oracle on three frozen test sets. The expert pool is chosen retrospectively from Table 1, and the paper says so: 'Because the expert pool was designated retrospectively from the completed test-set results, the routed number is a frozen-benchmark comparison rather than an independently held-out estimate of expert-pool selection.' This is a disclosed limitation, not a concealed circularity: the routers are trained only on training-split pages ('the frozen test sets ... are not used for router fitting or router checkpoint selection'), and the match to the oracle is a definitional consequence of 99.3% domain accuracy, not a fitted parameter renamed as a prediction. The CER values 0.30/1.57/4.83 are the column minima of Table 1, but the paper never claims to have predicted them from the router; it claims only that routing selects the already-known best checkpoints, which is exactly the benchmark-scoped claim. No load-bearing self-citation or imported uniqueness theorem appears. The small 10- and 13-page handwriting sets raise statistical precision concerns, but those are correctness risks, not circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim rests on a post-hoc expert selection that uses the same frozen test sets whose error rates are then reported as the system's performance. The router itself is trained on source labels, and the small test sets make the two-decimal precision a fragile assumption. No new physical or conceptual entities are introduced.

free parameters (2)
  • expert-pool designation per domain = regular: v3.0; memorial: v3.0 s2; running: v5.0
    The best checkpoint per domain was selected after the fact using frozen test-set results (Section 6: 'the expert pool was designated retrospectively from the completed test-set results'). The reported routed CERs therefore incorporate test-set information and are not held-out estimates.
  • router architecture and training recipe = ResNet-18, 224x224, ImageNet init, inverse-frequency weighting, source-balanced sampling
    Model-selection choices made by the authors; no sensitivity analysis is reported for these choices.
axioms (4)
  • domain assumption Domain labels are recoverable from source metadata (memorial source -> memorial; running source -> running; all others -> regular).
    Used to train the router and define the oracle. Section 6: 'Domain labels are derived from source metadata.' If source metadata does not align with visual style, the router and oracle are mis-specified.
  • domain assumption The three-way division (regular/memorial/running) captures the visually relevant variation in the corpus.
    Section 2 groups the corpus into three visual styles; finer distinctions (seals, marginalia) are acknowledged as future work in Section 8.
  • domain assumption Test sets of 10 (memorial) and 13 (running) pages are sufficient to estimate the reported two-decimal CER differences.
    The paper itself calls the handwriting sets 'domain-shift probes rather than large-sample estimates' (Section 10), yet the headline match is reported to two decimals.
  • domain assumption Nougat is the appropriate shared backbone for all experts.
    Chosen after a common-recipe comparison of three backbones (Appendix A); it is an empirical model-selection decision, not a formal proof.

pith-pipeline@v1.3.0-alltime-deepseek · 11750 in / 11401 out tokens · 107241 ms · 2026-08-02T02:55:06.687753+00:00 · methodology

0 comments
read the original abstract

Historical Manchu OCR must accommodate various visually distinct writing styles, including regular script, running script, and the semi-cursive chancery hand used in palace memorials, despite limited labeled data. We study a multi-expert system that reuses checkpoints from an iterative fine-tuning process as domain specialists and uses a lightweight page-level image classifier to dispatch pages by visual style. When the checkpoint pool lacks a suitable specialist, we train an additional expert for that domain. On three frozen test sets, the routed system matches the selected specialist for each style at two-decimal precision: 0.30 percent CER on regular script, 1.57 percent on memorials, and 4.83 percent on running script. The router achieves 99.3 percent page-level domain accuracy and matches the domain-label oracle at the same precision. Two of the three selected specialists were not trained specifically for their final domain; only the running-script expert was trained with that domain as its target. We report the evaluation protocol, router design, and per-page predictions to make the comparison reproducible.

Figures

Figures reproduced from arXiv: 2607.14041 by Chih-wen Kuo, Jiqiao Ma, Zhan Chen.

Figure 1
Figure 1. Figure 1: A two-dimensional projection of the checkpoint trade-off: regular-script CER versus running-script CER (from [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The routed system. A lightweight domain router (a small image classifier) reads each page [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Source-balanced router training (abla￾tion) on the frozen test sets. Within the regular￾script class, a naive router underfits the minority manc.hu woodblock print and misroutes most of it to the running-script expert (blue, 7/12), inflating its routed CER to 4.1% against a domain-label ora￾cle of 0.5%; balancing the regular-script sources (orange) recovers domain-label-oracle-level rout￾ing (1/12, 0.6%). … view at source ↗
Figure 3
Figure 3. Figure 3: Domain-router confusion matrix on the frozen test sets (rows: true domain; columns: ex￾pert routed to). The learned router places 147 of 148 pages correctly (99.3%); the single error is one atypical manc.hu regular-script page sent to the running-script expert. Its regular-script CER is 0.299%, versus 0.296% for the domain-label or￾acle, which rounds to 0.30% for both. does not replace the selected checkpo… view at source ↗
Figure 5
Figure 5. Figure 5: Model output on one held-out page per visual style (regular, memorial, running), with the [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example output from the public Manchu-language assistance parser. The interface shows [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

42 extracted references · 12 linked inside Pith

  1. [1]

    Qwen-vl: A frontier large vision-language model with versatile abilities.arXiv preprint arXiv:2308.12966, 2023

    Jinze Bai, Shuai Bai, Shusheng Yang, Shi- jie Wang, Sinan Tan, Peng Wang, Jun- yang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities.arXiv preprint arXiv:2308.12966, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    HHD-Ethiopic: A histor- ical handwritten dataset for Ethiopic OCR with baseline models and human-level perfor- mance

    Birhanu Hailu Belay, Isabelle Guyon, Tadele Mengiste, Bezawork Tilahun, Marcus Li- wicki, Tesfa Tegegne, Romain Egele, and Tsiyon Worku. HHD-Ethiopic: A histor- ical handwritten dataset for Ethiopic OCR with baseline models and human-level perfor- mance. InDocument Analysis and Recogni- tion – ICDAR 2024, pages 23–38. Springer, 2024

    2024

  3. [3]

    Nougat: Neu- ral optical understanding for academic doc- uments.arXiv preprint arXiv:2308.13418, 2023

    Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neu- ral optical understanding for academic doc- uments.arXiv preprint arXiv:2308.13418, 2023

    Pith/arXiv arXiv 2023

  4. [4]

    Automated sign detection across the Elec- tronic Babylonian Library: A large-scale dataset and end-to-end cuneiform OCR pipeline.arXiv preprint arXiv:2606.22608, 2026

    Wentao Che, Esteban Garc ´es Arias, Asim Niaz, Andreas Bender, and Enrique Jim´enez. Automated sign detection across the Elec- tronic Babylonian Library: A large-scale dataset and end-to-end cuneiform OCR pipeline.arXiv preprint arXiv:2606.22608, 2026

    Pith/arXiv arXiv 2026

  5. [5]

    Finetuning vision-language models as ocr systems for low-resource languages: A case study of manchu.arXiv preprint arXiv:2507.06761, 2025

    Yan Hon Michael Chung and Donghyeok Choi. Finetuning vision-language models as ocr systems for low-resource languages: A case study of manchu.arXiv preprint arXiv:2507.06761, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Pamela Kyle Crossley and Evelyn S. Rawski. A profile of the Manchu language in ch’ing history.Harvard Journal of Asiatic Studies, 53(1):63–102, 1993

    1993

  7. [7]

    Damai Dai, Chengqi Deng, Chenggang Zhao, R. X. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y . Wu, et al. DeepSeekMoE: Towards ultimate expert spe- cialization in mixture-of-experts language models.arXiv preprint arXiv:2401.06066, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Old Uyghur OCR: The first work-in-progress via reproducing fine-tuning of VLMs.Journal of Old Turkic Studies, 9(2):313–318, 2025

    Mehmet Oguz Derin and Erdem Uc ¸ar. Old Uyghur OCR: The first work-in-progress via reproducing fine-tuning of VLMs.Journal of Old Turkic Studies, 9(2):313–318, 2025

    2025

  9. [9]

    Mark C. Elliott. The Manchu-language archives of the Qing dynasty and the origins of the palace memorial system.Late Imperial China, 22(1):1–70, 2001

    2001

  10. [10]

    Switch transformers: Scaling to tril- lion parameter models with simple and effi- cient sparsity.Journal of Machine Learning Research, 23(120):1–39, 2022

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to tril- lion parameter models with simple and effi- cient sparsity.Journal of Machine Learning Research, 23(120):1–39, 2022

    2022

  11. [11]

    Deep residual learning for im- age recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im- age recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016

    2016

  12. [12]

    Hopcroft, and Kilian Q

    Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Wein- berger. Snapshot ensembles: Train 1, get M for free. InInternational Conference on Learning Representations (ICLR), 2017

    2017

  13. [13]

    Averaging weights leads to wider optima and better generalization

    Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. InConfer- ence on Uncertainty in Artificial Intelligence (UAI), 2018

    2018

  14. [14]

    Adaptive mixtures of local experts.Neural Computa- tion, 3(1):79–87, 1991

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts.Neural Computa- tion, 3(1):79–87, 1991

    1991

  15. [15]

    Jiang, Alexandre Sablayrolles, An- toine Roux, et al

    Albert Q. Jiang, Alexandre Sablayrolles, An- toine Roux, et al. Mixtral of experts.arXiv preprint arXiv:2401.04088, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    9 Transkribus—a service platform for tran- scription, recognition and retrieval of historical documents

    Philip Kahle, Sebastian Colutto, G¨unter Hackl, and G ¨unter M ¨uhlberger. 9 Transkribus—a service platform for tran- scription, recognition and retrieval of historical documents. In14th IAPR Inter- national Conference on Document Analysis and Recognition (ICDAR), volume 4, pages 19–24. IEEE, 2017

    2017

  17. [17]

    GlotOCR Bench: OCR models still struggle beyond a handful of Unicode scripts

    Amir Hossein Kargaran, Nafiseh Nikeghbal, Jana Diesner, Franc ¸ois Yvon, and Hinrich Sch¨utze. GlotOCR Bench: OCR models still struggle beyond a handful of Unicode scripts. arXiv preprint arXiv:2604.12978, 2026

    Pith/arXiv arXiv 2026

  18. [18]

    AMADI LontarSet: The first handwritten Balinese palm leaf manuscripts dataset

    Made Windu Antara Kesiman, Sophea Prum, Jean-Christophe Burie, and Jean-Marc Ogier. AMADI LontarSet: The first handwritten Balinese palm leaf manuscripts dataset. In 15th International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 168–173. IEEE, 2016

    2016

  19. [19]

    Kraken—an universal text recognizer for the humanities

    Benjamin Kiessling. Kraken—an universal text recognizer for the humanities. InDigital Humanities Conference (DH2019), 2019

    2019

  20. [20]

    Ocr-free document understanding transformer

    Geewook Kim, Teakgyu Hong, Moon- bin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang- doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. InEuropean Conference on Computer Vision, pages 498–517. Springer, 2022

    2022

  21. [21]

    Pix2Struct: Screenshot parsing as pretraining for visual language understand- ing

    Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Mar- tin Eisenschlos, Urvashi Khandelwal, Pe- ter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2Struct: Screenshot parsing as pretraining for visual language understand- ing. InInternational Conference on Machine Learning (ICML), pages 18893–18912, 2023

    2023

  22. [22]

    TrOCR: Transformer-based optical character recogni- tion with pre-trained models

    Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. TrOCR: Transformer-based optical character recogni- tion with pre-trained models. InProceed- ings of the AAAI Conference on Artificial In- telligence, volume 37, pages 13094–13102, 2023

    2023

  23. [23]

    Om- niOCR: Generalist OCR for ethnic minority languages.arXiv preprint arXiv:2602.21042, 2026

    Bonan Liu, Zeyu Zhang, Bingbing Meng, Han Wang, Hanshuo Zhang, Chengping Wang, Daji Ergu, and Ying Cai. Om- niOCR: Generalist OCR for ethnic minority languages.arXiv preprint arXiv:2602.21042, 2026

    arXiv 2026

  24. [24]

    Visual instruction tun- ing.Advances in neural information process- ing systems, 36, 2024

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tun- ing.Advances in neural information process- ing systems, 36, 2024

    2024

  25. [25]

    Kosmos-2.5: A mul- timodal literate model.arXiv preprint arXiv:2309.11419, 2023

    Tengchao Lv, Yupan Huang, Jingye Chen, Lei Cui, et al. Kosmos-2.5: A mul- timodal literate model.arXiv preprint arXiv:2309.11419, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Ancient but digitized: Developing handwritten optical character recognition for East Syriac script through creating KHAMIS dataset.arXiv preprint arXiv:2408.13631, 2024

    Ameer Majeed and Hossein Hassani. Ancient but digitized: Developing handwritten optical character recognition for East Syriac script through creating KHAMIS dataset.arXiv preprint arXiv:2408.13631, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    A survey of historical document image datasets

    Konstantina Nikolaidou, Mathias Seuret, Hamam Mokayed, and Marcus Liwicki. A survey of historical document image datasets. International Journal on Document Analy- sis and Recognition (IJDAR), 25(4):305–338, 2022

    2022

  28. [28]

    OpenITI MAKHZAN: An open annotated dataset of Arabic, Persian, Ottoman Turkish, and Urdu print and manuscript data.Zenodo, 2025

    Open Islamicate Texts Initiative. OpenITI MAKHZAN: An open annotated dataset of Arabic, Persian, Ottoman Turkish, and Urdu print and manuscript data.Zenodo, 2025. dataset

    2025

  29. [29]

    Dhali, and Lam- bert Schomaker

    Mladen Popovi ´c, Maruf A. Dhali, and Lam- bert Schomaker. Artificial intelligence based writer identification generates new evidence for the unknown scribes of the Dead Sea Scrolls exemplified by the great Isaiah scroll (1qisaa).PLOS ONE, 16(4):e0249769, 2021

    2021

  30. [30]

    Indiscapes: Instance segmentation net- works for layout parsing of historical Indic manuscripts

    Abhishek Prusty, Sowmya Aitha, Abhishek Trivedi, and Ravi Kiran Sarvadevabhatla. Indiscapes: Instance segmentation net- works for layout parsing of historical Indic manuscripts. InInternational Conference on Document Analysis and Recognition (IC- DAR), pages 999–1006. IEEE, 2019

    2019

  31. [31]

    A text recognition dataset from Sahidic Coptic ancient manuscripts

    Fabio Quattrini, Carmine Zaccagnino, Costanza Bianchi, Silvia Cascianelli, and Rita Cucchiara. A text recognition dataset from Sahidic Coptic ancient manuscripts. arXiv preprint arXiv:2606.15987, 2026. 10

    arXiv 2026

  32. [32]

    Ahmed, Polla Fattah, Bryar A

    Yassin Hussein Rassul, Aram M. Ahmed, Polla Fattah, Bryar A. Hassan, Arwaa W. Ab- dulkareem, Tarik A. Rashid, and Joan Lu. Advancing offline handwritten text recog- nition: A systematic review of data aug- mentation and generation techniques.arXiv preprint arXiv:2507.06275, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Scaling vision with sparse mixture of experts

    Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenat- ton, Andr´e Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  34. [34]

    Outrageously large neural networks: The sparsely-gated mixture- of-experts layer

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture- of-experts layer. InInternational Conference on Learning Representations (ICLR), 2017

    2017

  35. [35]

    Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its ap- plication to scene text recognition.IEEE transactions on pattern analysis and machine intelligence, 39(11):2298–2304, 2016

    2016

  36. [36]

    An overview of the Tesseract OCR engine

    Ray Smith. An overview of the Tesseract OCR engine. InNinth International Confer- ence on Document Analysis and Recognition (ICDAR), volume 2, pages 629–633. IEEE, 2007

    2007

  37. [37]

    Few shots are all you need: A progressive learn- ing approach for low resource handwritten text recognition.Pattern Recognition Letters, 160:43–49, 2022

    Mohamed Ali Souibgui, Alicia Forn ´es, Yousri Kessentini, and Be ´ata Megyesi. Few shots are all you need: A progressive learn- ing approach for low resource handwritten text recognition.Pattern Recognition Letters, 160:43–49, 2022

    2022

  38. [38]

    ManchuOCR: An OCR system for the Manchu script.https://github

    tyotakuki. ManchuOCR: An OCR system for the Manchu script.https://github. com/tyotakuki/ManchuOCR. GitHub repository, accessed 2026-07-14

    2026

  39. [39]

    QARI- OCR: High-fidelity Arabic text recognition through multimodal large language model adaptation.arXiv preprint arXiv:2506.02295, 2025

    Ahmed Wasfy, Omer Nacar, Abdelakreem Elkhateb, Mahmoud Reda, Omar Elshehy, Adel Ammar, and Wadii Boulila. QARI- OCR: High-fidelity Arabic text recognition through multimodal large language model adaptation.arXiv preprint arXiv:2506.02295, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    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, 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

  41. [41]

    Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. InInternational Conference on Machine Learning (ICML),...

    2022

  42. [42]

    OCR with the deep CNN model for ligature script-based lan- guages like Manchu.Scientific Programming, 2021:5520338, 2021

    Diandian Zhang, Yan Liu, Zhuowei Wang, and Depei Wang. OCR with the deep CNN model for ligature script-based lan- guages like Manchu.Scientific Programming, 2021:5520338, 2021. 11 A Architecture comparison Table 3 compares three page-level backbones trained with the same source-balanced (v5.0) recipe on the frozen test sets. Nougat is the most balanced of...

    2021