Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Benchmarking Vision-Language Models on Chinese Ancient Documents: From OCR to Knowledge Reasoning

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read AncientDoc is the first benchmark that evaluates vision-language models on Chinese ancient documents from OCR to knowledge reasoning.

desk verdict AncientDoc is a genuinely useful first benchmark for Chinese ancient documents; the main caveat is that the QA gold labels start from Qwen-generated content and the paper doesn't yet show how much human review changed them. read the letter →

arxiv 2509.09731 v1 pith:J2ZPN4CE submitted 2025-09-10 cs.CL

Haiyang Yu , Yuchuan Wu , Fan Shi , Lei Liao , Jinghui Lu , Xiaodong Ge , Han Wang , Minghan Zhuo
show 13 more authors
Xuecheng Wu Xiang Fei Hao Feng Guozhi Tang An-Lan Wang Hanshen Zhu Yangfan He Quanhuan Liang Liyuan Meng Chao Feng Can Huang Jingqun Tang Bin Li
This is my paper · ORCID
classification cs.CL
keywords Chineseancientdocumentsvision-languagemodelsdocumentunderstandingbenchmarkOCRvernaculartranslationquestionansweringtraditionalcharactershistorical
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

AncientDoc is presented as the first benchmark built specifically for evaluating vision-language models (VLMs) on Chinese ancient documents. It assembles about 2,973 page images from over 100 books spanning 14 traditional genres, and defines five tasks that climb from raw text extraction to deep understanding: page-level OCR, vernacular translation, reasoning-based question answering, knowledge-based QA, and linguistic-variant QA. The authors argue that existing document benchmarks, which focus on English printed text or simplified Chinese, cannot expose how models handle vertical typesetting, traditional and obsolete characters, missing punctuation, and historical background knowledge. If the benchmark holds up, it gives the field a shared yardstick for measuring and improving AI systems on a culturally significant class of documents that is currently underserved. The paper also reports a first sweep of mainstream open and closed VLMs, showing the gap is still large.

What carries the argument

The load-bearing object is AncientDoc itself: a multi-task benchmark of 2,973 scanned pages from roughly 100 ancient Chinese books in 14 categories, annotated by a pipeline that pre-generates questions and answers with Qwen2.5-VL-72B and then applies human revision. The five tasks are the mechanism that carries the argument, because they make the suite progress from low-level recognition (page-level OCR) through intralingual translation to three kinds of QA (reasoning, knowledge, linguistic variant), forcing any evaluated model to engage with visual layout, classical language, and historical context at once. The scoring machinery pairs hard n-gram and embedding metrics with an LLM judge cali

What would settle it

Take a random sample of the benchmark's gold QA pairs, have independent specialists in classical Chinese re-answer the questions from the page images without ever seeing the Qwen-generated pre-annotations, and compare the two sets. If agreement is low, or if the human-only answers systematically differ from the Qwen-then-proofread gold labels in ways that correlate with Qwen's known output patterns, the benchmark's claim to measure ancient-document understanding fails. Equivalently, a leakage check—does any of the 2,973 page images appear in a VLM's training data—would show whether reported sc

Watch

Extended reading notes

Core claim

The central claim is that a single benchmark can span the full pipeline of Chinese ancient-document understanding, from recognizing characters on a page to reasoning about allusions, and that no prior suite covers this range. AncientDoc operationalizes that claim with five tasks: page-level OCR, vernacular translation, reasoning-based QA, knowledge-based QA, and linguistic variant QA. Its construction uses images from digitized holdings, selected to cover 14 genres across dynasties, with annotations produced by pre-generating QA pairs with Qwen2.5-VL-72B and then revising them manually. Evaluation combines hard lexical metrics (CER/CHRF++/BERTScore) with an LLM judge chosen for agreement wit

Load-bearing premise

The benchmark's validity rests on the assumption that the human review of Qwen2.5-VL-72B's pre-generated questions and answers was thorough enough to remove subtle mistakes; if not, the gold standard carries one model's biases rather than a neutral expert reading.

Editorial extensions

If this is right

  • A fixed, reusable suite now exists for tracking progress on Chinese ancient documents; future VLMs can be ranked on the same 2,973 pages and 14 genres.
  • Current systems are far from fluent: the best page-level OCR reaches only about 18 character F1 and the best vernacular translation earns 4.72/10 from the human-aligned judge.
  • No single model dominates all five tasks, so the benchmark exposes separate capabilities—recognition, translation, reasoning, factual knowledge, and stylistic analysis—rather than one general document-understanding score.
  • The methodology of calibrating an LLM judge against human raters on a subset makes open-ended ancient-language evaluation reproducible where n-gram metrics alone are unreliable.

Reading between the lines

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

  • Because the gold QA pairs were drafted by one VLM (Qwen2.5-VL-72B) before human review, model rankings may partly measure agreement with that VLM's reading; a fully human-authored gold subset would reveal the size of that effect.
  • The paired five-task design invites error-cascade analysis: on the same page, a model's OCR output can be compared with its translation and QA answers to trace whether recognition failures drive reasoning failures.
  • The suite could be adapted, with permission, to other premodern Sinitic corpora (classical Japanese, Korean and Vietnamese Literary Sinitic) to test whether document-understanding skills transfer across historical languages.
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. The paper introduces AncientDoc, a benchmark for evaluating vision-language models (VLMs) on Chinese ancient documents. It comprises 2,973 page images from about 100 books held by Harvard Library, covering 14 document categories, and defines five tasks: page-level OCR, vernacular translation to modern Chinese, reasoning-based QA, knowledge-based QA, and linguistic-variant QA. The QA and translation labels were pre-generated by Qwen2.5-VL-72B and then manually reviewed; OCR references were also manually corrected. The authors evaluate 25+ open- and closed-source VLMs with CER/character-based metrics for OCR and BERTScore/CHRF++ plus a GPT-4o-based 0-10 score for the understanding tasks, after selecting GPT-4o as the LLM scorer most aligned with human ratings on a 250-sample study. The central claim is that AncientDoc is the first benchmark covering the full range from OCR to knowledge reasoning for Chinese ancient documents.

Significance. AncientDoc addresses a real gap: existing document benchmarks are predominantly English or simplified Chinese, and none systematically covers vernacular translation, knowledge-based QA, or linguistic-variant reasoning on traditional vertical Chinese texts. The public release, the multi-task design, and the inclusion of a human-calibrated LLM scorer are positive features; the authors also report a self-bias check for GPT-4o's own predictions. If the label quality can be demonstrated, this would be a valuable community resource. However, the current manuscript provides insufficient annotation-quality evidence for the QA/translation labels, and the reported OCR metrics are internally inconsistent, so the validity of the resulting model rankings is not yet established.

major comments (3)
  1. [Sec. 3.4, Table 3, Eqs. (1)-(4)] The CER and character-level metrics are mutually inconsistent. For Gemini2.5-Pro, CER=32.03 implies that the correct-character rate is at least about 68% (since recall >= 1 - CER when insertions are non-negative), but Char Recall is reported as 18.53. Similar contradictions appear for every model, e.g., Qwen2.5-VL-7B has CER=35.47 but Char Recall=12.75. Unless the CER and character metrics use different alignment or normalization procedures (not described in Eqs. 1-4), these numbers cannot coexist. This directly undermines the OCR comparison, which is a core component of the benchmark.
  2. [Sec. 3.2, Data Annotation] The gold QA and translation labels are pre-generated by Qwen2.5-VL-72B and only described as receiving 'comprehensive manual review and revision.' The paper reports no annotation statistics: number of annotators, their classical-Chinese expertise, per-task workload, proportion of items filtered or modified, or inter-annotator agreement. For knowledge-based and linguistic-variant QA, subtle errors in historical facts or classical allusions are easy to miss in a proofreading pass. If the reference answers carry systematic biases of the generating model, then model rankings may partly measure agreement with Qwen2.5-VL-72B rather than genuine ancient-document understanding. The authors should provide independent evidence of label quality, such as expert re-annotation of a random sample, revision/filtering statistics, or detailed annotation guidelines.
  3. [Sec. 3.4, Tables 2-7] The human-alignment study uses only 250 QA pairs (50 per task) and reports no number of human raters, no score distributions, and no confidence intervals. GPT-4o scores are then used as a primary evaluation metric, yet many ranking differences are tiny (e.g., Table 5: Qwen-VL-Max 7.46 vs. Qwen2.5-VL-72B 7.43; Table 6: Doubao-V2 7.36 vs. Gemini2.5-Pro 7.36). Without variance estimates or significance tests, it is impossible to tell whether these differences are signal or noise. Please provide bootstrap confidence intervals, per-model standard deviations, or at least a clear statement of score variance for the main results.
minor comments (5)
  1. [Abstract / Sec. 3.3] The paper alternates between 'about 3,000 pages' and the exact figure 2,973 pages; use the exact count consistently throughout.
  2. [Table 1 / Table 4 captions] There are typos in the captions: 'bemchmarks' should be 'benchmarks' and 'transoaltion' should be 'translation'.
  3. [Sec. 3.2] The sentence 'we construct one QA pair for page-level OCR and vernacular translation' is imprecise; page-level OCR and translation are not QA pairs. Use a term such as 'annotation item' or 'task instance.'
  4. [Sec. 3.4] GPT-4o is selected as 'closer to human ratings across several metrics,' but it has the largest absolute bias (-0.896) among the five scorers. The selection criterion (e.g., weighting of correlation vs. error vs. bias) should be stated explicitly.
  5. [References and model tables] Several model names in the tables (e.g., InternVL2.5, Qwen2.5-Omni, LLaVA-OneVision) lack explicit citation entries. Please add or clarify references for all evaluated models.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: benchmark construction is empirical, not a derivation that reduces to its inputs.

full rationale

AncientDoc is presented as a new benchmark, not as a derived prediction. The QA labels are pre-generated by Qwen2.5-VL-72B and then manually reviewed (Sec 3.2), but this is a data-collection procedure rather than a formal derivation whose conclusion is equivalent to its premise. The evaluation uses standard metrics plus GPT-4o as a scorer, with the scorer explicitly selected based on agreement with human ratings on 250 QA pairs (Sec 3.4, Tab. 2), and the paper checks self-bias by noting GPT-4o scores its own translations poorly (Sec 4.3). No load-bearing step reduces to a fitted parameter, a self-citation, or an imported uniqueness theorem. The lack of inter-annotator agreement statistics is a validity/quality concern, but it does not make the benchmark's claim circular. The minor citation of a ByteDance technical report for the Doubao baseline is not load-bearing.

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

The benchmark relies on annotation quality assumptions and on the choice of evaluators. No numerical parameters are fitted; design choices like two QA pairs per task for three tasks are fixed heuristics.

assumptions (4)
  • domain assumption Human proofreading of LLM-generated QA pairs yields reliable ground truth
    Sec 3.2: 'comprehensive manual review and revision' assumes annotator correctness.
  • domain assumption BERTScore and GPT-4o scoring are valid proxies for answer quality
    Sec 3.4 and Tab. 2: GPT-4o chosen by alignment on 250 samples; BERTScore used without task-specific calibration.
  • domain assumption The Harvard Library digitized collection is representative of Chinese ancient documents
    Sec 3.2: data sourced exclusively from Harvard Library.
  • domain assumption GPT-4o is unbiased as a scorer of its own outputs
    Sec 4.3: single example of GPT-4o translation score 0.92 used as evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking Vision-Language Models on Chinese Ancient Documents: From OCR to Knowledge Reasoning." pith.science (2026). https://pith.science/paper/J2ZPN4CE

@misc{pith2026250909731,
  author       = {Pith},
  title        = {Pith review of: Benchmarking Vision-Language Models on Chinese Ancient Documents: From OCR to Knowledge Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J2ZPN4CE}},
  note         = {Machine review of arXiv:2509.09731}
}
read the original abstract

Chinese ancient documents, invaluable carriers of millennia of Chinese history and culture, hold rich knowledge across diverse fields but face challenges in digitization and understanding, i.e., traditional methods only scan images, while current Vision-Language Models (VLMs) struggle with their visual and linguistic complexity. Existing document benchmarks focus on English printed texts or simplified Chinese, leaving a gap for evaluating VLMs on ancient Chinese documents. To address this, we present AncientDoc, the first benchmark for Chinese ancient documents, designed to assess VLMs from OCR to knowledge reasoning. AncientDoc includes five tasks (page-level OCR, vernacular translation, reasoning-based QA, knowledge-based QA, linguistic variant QA) and covers 14 document types, over 100 books, and about 3,000 pages. Based on AncientDoc, we evaluate mainstream VLMs using multiple metrics, supplemented by a human-aligned large language model for scoring.

Figures

Figures reproduced from arXiv: 2509.09731 by the authors.

Figure 1
Figure 1. Some examples of each task in AncientDoc. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) The distribution of page counts across differ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Comparison of consistency between Doubao scoring and human scoring. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Comparison of consistency between Gemini scoring and human scoring. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Comparison of consistency between GPT-4o scoring and human scoring. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Comparison of consistency between Qwen-Plus scoring and human scoring. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Comparison of consistency between Qwen-VL-72B scoring and human scoring. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. HCSU: A Dataset and Benchmark for Fine-Grained Historical Calligraphy Style Understanding

    cs.CV 2026-07 conditional novelty 6.5 of 10

    HCSU supplies the first large decoupled Tie/Bei calligraphy dataset with expert aesthetic labels and finds SOTA LVLMs remain knowledgeable yet unperceptive on fine-grained style.

  2. LPCAN: Lightweight Pyramid Cross-Attention Network for Rail Surface Defect Detection Using RGB-D Data

    cs.CV 2026-01 reject novelty 4.0 of 10

    A lightweight RGB-D cross-attention network is proposed for rail defect detection, but the SOTA accuracy and generalization claims are internally inconsistent and the implementation is not public.

  3. Knowledge-Embedded and Hypernetwork-Guided Few-Shot Substation Meter Defect Image Generation Method

    cs.CV 2026-01 reject novelty 3.0 of 10

    Fine-tuning Stable Diffusion with DreamBooth-style knowledge and hypernetwork-guided crack control maps can synthesize substation meter defect images that boost a YOLOv8 defect detector's mAP when added to the training set.

Reference graph

Works this paper leans on

41 extracted references · 17 linked inside Pith · cited by 3 Pith papers

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Akoushideh, A.; Ranjkesh Rashtehroudi, A.; and Shahbahrami, A. 2025. Persian/Arabic Scene Text Recognition With Convolutional Recurrent Neural Network. IET Smart Cities, 7(1): e70001

    2025

  4. [4]

    Bai, J.; Bai, S.; Yang, S.; Wang, S.; Tan, S.; Wang, P.; Lin, J.; Zhou, C.; and Zhou, J. 2023. Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond. arXiv preprint arXiv:2308.12966

    arXiv 2023

  5. [5]

    Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; et al. 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923

    arXiv 2025

  6. [6]

    Borchmann, .; Pietruszka, M.; Stanislawek, T.; Jurkiewicz, D.; Turski, M.; Szyndler, K.; and Grali \'n ski, F. 2021. Due: End-to-end document understanding benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)

    2021

  7. [7]

    Chen, Z.; Wang, W.; Cao, Y.; Liu, Y.; Gao, Z.; Cui, E.; Zhu, J.; Ye, S.; Tian, H.; Liu, Z.; et al. 2024. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271

    arXiv 2024

  8. [8]

    Comanici, G.; Bieber, E.; Schaekermann, M.; Pasupat, I.; Sachdeva, N.; Dhillon, I.; Blistein, M.; Ram, O.; Zhang, D.; Rosen, E.; et al. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261

    arXiv 2025

Show all 41 references
  1. [9]

    Cui, R.; Zhu, A.; and Ding, Z. 2025. Enhanced Chinese scene text recognition model base on cross-domain feature fusion. Signal, Image and Video Processing, 19(6): 499

    2025

  2. [10]

    Ding, Y.; Luo, S.; Chung, H.; and Han, S. C. 2023. VQA: A new dataset for real-world VQA on PDF documents. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 585--601. Springer

    2023

  3. [11]

    Ding, Y.; Ren, K.; Huang, J.; Luo, S.; and Han, S. C. 2024. MVQA: A dataset for multimodal information retrieval in PDF-based visual question answering. arXiv preprint arXiv:2404.12720

    2024 arXiv

  4. [12]

    Fu, L.; Kuang, Z.; Song, J.; Huang, M.; Yang, B.; Li, Y.; Zhu, L.; Luo, Q.; Wang, X.; Lu, H.; et al. 2024. Ocrbench v2: An improved benchmark for evaluating large multimodal models on visual text localization and reasoning. arXiv preprint arXiv:2501.00321

    2024 arXiv

  5. [13]

    He, P.; Huang, W.; Qiao, Y.; Loy, C.; and Tang, X. 2016. Reading scene text in deep convolutional sequences. In Proceedings of the AAAI conference on artificial intelligence, volume 30

    2016

  6. [14]

    Huang, Y.; Lv, T.; Cui, L.; Lu, Y.; and Wei, F. 2022. Layoutlmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM international conference on multimedia, 4083--4091

    2022

  7. [15]

    P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Radford, A.; et al

    Hurst, A.; Lerer, A.; Goucher, A. P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Radford, A.; et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  8. [16]

    R.; Fujii, Y.; Deselaers, T.; Baccash, J.; and Popat, A

    Ingle, R. R.; Fujii, Y.; Deselaers, T.; Baccash, J.; and Popat, A. C. 2019. A scalable handwritten text recognition system. In 2019 International conference on document analysis and recognition (ICDAR), 17--24. IEEE

    2019

  9. [17]

    K.; and Thiran, J.-P

    Jaume, G.; Ekenel, H. K.; and Thiran, J.-P. 2019. Funsd: A dataset for form understanding in noisy scanned documents. In 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW), volume 2, 1--6. IEEE

    2019

  10. [18]

    Kang, L.; Riba, P.; Rusi \ n ol, M.; Forn \'e s, A.; and Villegas, M. 2022. Pay attention to what you read: non-recurrent handwritten text-line recognition. Pattern Recognition, 129: 108766

    2022

  11. [19]

    Kang, L.; Tito, R.; Valveny, E.; and Karatzas, D. 2024. Multi-page document visual question answering using self-attention scoring mechanism. In International Conference on Document Analysis and Recognition, 219--232. Springer

    2024

  12. [20]

    Li, B.; Zhang, Y.; Guo, D.; Zhang, R.; Li, F.; Zhang, H.; Zhang, K.; Zhang, P.; Li, Y.; Liu, Z.; et al. 2024. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326

    2024 arXiv

  13. [21]

    Liu, H.; Li, C.; Li, Y.; and Lee, Y. J. 2024 a . Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 26296--26306

    2024

  14. [22]

    Liu, Y.; Li, Z.; Huang, M.; Yang, B.; Yu, W.; Li, C.; Yin, X.-C.; Liu, C.-L.; Jin, L.; and Bai, X. 2024 b . Ocrbench: on the hidden mystery of ocr in large multimodal models. Science China Information Sciences, 67(12): 220102

    2024

  15. [23]

    Ma, Y.; Zang, Y.; Chen, L.; Chen, M.; Jiao, Y.; Li, X.; Lu, X.; Liu, Z.; Ma, Y.; Dong, X.; et al. 2024. Mmlongbench-doc: Benchmarking long-context document understanding with visualizations. Advances in Neural Information Processing Systems, 37: 95963--96010

    2024

  16. [24]

    Mathew, M.; Karatzas, D.; and Jawahar, C. 2021. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2200--2209

    2021

  17. [25]

    K.; and Chakraborty, A

    Mishra, A.; Shekhar, S.; Singh, A. K.; and Chakraborty, A. 2019. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), 947--952. IEEE

    2019

  18. [26]

    Popovi \'c , M. 2017. chrF++: words helping character n-grams. In Proceedings of the second conference on machine translation, 612--618

    2017

  19. [27]

    Qwen; :; Yang, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Li, C.; Liu, D.; Huang, F.; Wei, H.; Lin, H.; Yang, J.; Tu, J.; Zhang, J.; Yang, J.; Yang, J.; Zhou, J.; Lin, J.; Dang, K.; Lu, K.; Bao, K.; Yang, K.; Yu, L.; Li, M.; Xue, M.; Zhang, P.; Zhu, Q.; Men, R.; Lin,...

    2024 arXiv

  20. [28]

    Tanaka, R.; Iki, T.; Nishida, K.; Saito, K.; and Suzuki, J. 2024. Instructdoc: A dataset for zero-shot generalization of visual document understanding with instructions. In Proceedings of the AAAI conference on artificial intelligence, volume 38, 19071--19079

    2024

  21. [29]

    Team, B. S. 2025. Seed1.5-VL Technical Report. arXiv preprint arXiv:2505.07062

    2025 arXiv

  22. [30]

    Tian, Z.; Huang, W.; He, T.; He, P.; and Qiao, Y. 2016. Detecting text in natural image with connectionist text proposal network. In European conference on computer vision, 56--72. Springer

    2016

  23. [31]

    Wang, L.; Lyu, C.; Ji, T.; Zhang, Z.; Yu, D.; Shi, S.; and Tu, Z. 2023. Document-level machine translation with large language models. arXiv preprint arXiv:2304.02210

    2023 arXiv

  24. [32]

    Wu, Z.; Chen, X.; Pan, Z.; Liu, X.; Liu, W.; Dai, D.; Gao, H.; Ma, Y.; Wu, C.; Wang, B.; et al. 2024. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302

    2024 arXiv

  25. [33]

    Xu, J.; Guo, Z.; He, J.; Hu, H.; He, T.; Bai, S.; Chen, K.; Wang, J.; Fan, Y.; Dang, K.; et al. 2025. Qwen2. 5-omni technical report. arXiv preprint arXiv:2503.20215

    2025 arXiv

  26. [34]

    Xu, Z.; and Xiang, Y. 2025. Improving Chinese word segmentation with character--lexicon class attention. Neural Computing and Applications, 37(5): 3857--3867

    2025

  27. [35]

    Yang, H.; Jin, L.; Huang, W.; Yang, Z.; Lai, S.; and Sun, J. 2018. Dense and tight detection of Chinese characters in historical documents: Datasets and a recognition guided detector. IEEE Access, 6: 30174--30183

    2018

  28. [36]

    Yin, F.; Wu, Y.-C.; Zhang, X.-Y.; and Liu, C.-L. 2017. Scene text recognition with sliding convolutional character models. arXiv preprint arXiv:1709.01727

    2017 arXiv

  29. [37]

    Zhang, J.; Luan, H.; Sun, M.; Zhai, F.; Xu, J.; Zhang, M.; and Liu, Y. 2018. Improving the transformer translation model with document-level context. arXiv preprint arXiv:1810.03581

    2018 arXiv

  30. [38]

    Zhang, J.; Yu, Y.; and Zhang, Y. 2024. CREAM: coarse-to-fine retrieval and multi-modal efficient tuning for document VQA. In Proceedings of the 32nd ACM International Conference on Multimedia, 925--934

    2024

  31. [39]

    Q.; and Artzi, Y

    Zhang, T.; Kishore, V.; Wu, F.; Weinberger, K. Q.; and Artzi, Y. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675

    2019 arXiv

  32. [40]

    Zhou, X.; Yao, C.; Wen, H.; Wang, Y.; Zhou, S.; He, W.; and Liang, J. 2017. East: an efficient and accurate scene text detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 5551--5560

    2017

  33. [41]

    Zhu, J.; Wang, W.; Chen, Z.; Liu, Z.; Ye, S.; Gu, L.; Tian, H.; Duan, Y.; Su, W.; Shao, J.; et al. 2025. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479

    2025 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.