Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Multi-Agent Interactive Question Generation Framework for Long Document Understanding

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A fully automated multi-agent pipeline can generate long-context English and Arabic document QA data at scale, and the resulting benchmark is hard for current vision-language models.

desk verdict Useful automated QA generation pipeline for long docs, but the benchmark's difficulty claims rest on unvalidated ground truth. read the letter →

arxiv 2507.20145 v1 pith:OUKSLNE5 submitted 2025-07-27 cs.CL cs.AI

Kesen Wang , Daulet Toibazar , Abdulrahman Alfulayt , Abdulaziz S. Albadawi , Ranya A. Alkahtani , Asma A. Ibrahim , Haneen A. Alhomoud , Sherif Mohamed
show 1 more author
Pedro J. Moreno
This is my paper
classification cs.CLcs.AI
keywords multi-agentquestiongenerationlongdocumentunderstandingArabicNLPbenchmarkvision-languagemodelsautomateddataannotationunanswerablequestionscross-pagereasoninglow-resourcelanguages
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 a fully automated, multi-agent pipeline can replace expensive human annotation for long-document question generation, producing 6,732 English and Arabic question-answer pairs (AraEngLongBench) from hundreds of pages without manual labeling. The authors further claim that this benchmark is hard: the best tested model, Gemini-1.5 Pro, reaches only 49.8% overall accuracy, and every evaluated model scores below 10% on unanswerable questions in both languages. If the pipeline works as described, it offers a scalable route to long-context training and evaluation data for low-resource languages such as Arabic, and a stress test that current vision-language models are far from passing.

What carries the argument

The load-bearing mechanism is the multi-agent interactive chain: Agent 1 drafts and later refines questions from OCR and layout output, Agent 2 filters irrelevant or redundant questions, Agent 3 grounds candidate answers in the page images, Agent 4 produces reference answers plus difficulty estimates and feeds corrections back into Agent 1, and Agent 5 performs an end-of-process consistency check tying each answer to evidence. Two properties carry the argument: the iterative feedback loop (with its 40% accuracy threshold that triggers harder questions) and the automatic evidence-validation step that is meant to prevent the mismatches seen in prior benchmarks. The preprocessing stack—LVLM-based OCR instead of a generic engine, YOLO-style layout detection, and 10-page overlapping chunks—is what lets the agents reason across page boundaries.

What would settle it

Sample 200 question-answer-evidence tuples from AraEngLongBench and have independent human annotators verify each answer against the cited source pages; if a substantial fraction of reference answers are wrong or unsupported, the agent pipeline has contaminated the ground truth and the model scores in Table 1 would need to be recomputed on a human-verified subset.

Watch

Extended reading notes

Core claim

The paper's central assertion is that an agentic chain of five specialized roles—question generation, question extraction, answer generation, assessment with feedback, and evidence validation—can generate benchmark-quality long-document questions without human annotation. Starting from PDFs, the pipeline uses LVLM-based OCR and layout analysis to build structured chunks with a 10-page overlap, then iterates the agents until the question set reaches a target difficulty (Agent 4 raises complexity if answer accuracy exceeds 40%). On the resulting AraEngLongBench, the best overall score is Gemini-1.5 Pro at 49.8%, followed by Gemini-2.0 Flash at 49.6%; among open models, Qwen-2.5-VL leads at 44.5% and Llama-3.2-VL trails at 32.4%. The authors read the uniformly low unanswerable-question scores (roughly 3–18%) as evidence that current LVLMs cannot reliably abstain when a question has no answer in the document, and the lower Arabic scores as evidence of a cross-lingual gap.

Load-bearing premise

The reference answers and evidence in AraEngLongBench are generated and validated only by agents, so the entire accuracy comparison is trustworthy only if the agent-produced automated text extraction, layout parsing, and question-answer pairs are correct.

Editorial extensions

If this is right

  • Following the described pipeline should let a research group generate long-document QA data for new domains and languages without hiring annotators.
  • AraEngLongBench, if its ground truth holds up, gives the community a benchmark where current LVLMs score below 50% and near chance on unanswerable items, so it can serve as a difficulty probe for long-context understanding.
  • The Arabic sub-benchmark exposes a cross-lingual gap: every model scores lower on Arabic than on English, which should push work on low-resource document understanding.
  • The 40% complexity threshold provides a controllable dial for generating harder or easier benchmark splits.
  • The uniformly low unanswerable-question performance indicates a concrete failure mode—inability to abstain—that future LVLMs will need to address.

Reading between the lines

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

  • If the agent-generated reference answers contain hidden OCR or hallucination errors, the reported model scores would be an underestimate of true ability; a human audit of a random sample would settle this.
  • Because the pipeline steps are language-agnostic apart from OCR, it should transfer to other right-to-left or low-resource scripts without redesign.
  • The near-universal failure on unanswerable questions may be partly a generation artifact: agent-written 'unanswerable' items may be hard to distinguish from ordinary hard questions, making the UA score a test of calibration rather than of abstention ability.
  • The authors leave open how iteration count affects question depth; a natural next experiment is to compare one-loop vs. five-loop outputs on multi-hop reasoning questions.
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

4 major / 5 minor

Summary. The paper proposes a fully automated multi-agent framework for generating long-context question-answer pairs from long English and Arabic documents, and introduces the resulting benchmark, AraEngLongBench, containing 6,732 questions across 12 question types. The pipeline uses LVLM-based OCR, layout analysis, a five-agent generation/refinement loop, and an internal evidence-validation step, with no human annotation. The authors evaluate six LVLMs on the benchmark and report that the best model, Gemini 1.5 Pro, reaches only 49.8% overall accuracy, with all models scoring below 10% on unanswerable questions. The central claim is that AraEngLongBench is a challenging benchmark for long-context document understanding, especially for Arabic.

Significance. If the benchmark's ground truth is reliable, the paper makes a useful contribution: it offers a scalable, agent-driven alternative to costly human annotation, addresses the scarcity of Arabic long-document QA data, and releases code and data. The experiments provide a first indication that current LVLMs, including Gemini-1.5 Pro and Qwen 2.5 VL, struggle with cross-page and unanswerable questions. However, the central difficulty claim depends on the correctness of automatically generated reference answers, and the paper does not report any human validation, error analysis, or audit of the 6,732 QA pairs. This is a load-bearing weakness: without evidence that the reference answers and supporting evidence are correct, the reported accuracy numbers could reflect benchmark noise as much as model competence.

major comments (4)
  1. [Section 3.2.2, Table 1] The central claim that AraEngLongBench is challenging rests on the correctness of the reference answers and evidence pages generated by the pipeline, but the paper reports no human validation, inter-annotator agreement, or sampled audit of the 6,732 QA pairs. Agent 5's 'evidence validation' is performed by an LVLM, which is the same class of system being evaluated later, and no validation metric or pass rate is reported. I request a human-validated sample (for example, 100-200 QA pairs per language) with agreement statistics and an error taxonomy, and a re-evaluation or correction of the contaminated subset.
  2. [Section 3.2.2 (Agent 4)] The 40% accuracy trigger is a load-bearing design choice: whenever the answer-generation agent answers correctly in more than 40% of cases, Agent 1 is instructed to increase question complexity. This explicitly selects for questions that are hard for the generating LVLM, not for questions that are hard because they require genuine long-document understanding. Without external validation, this mechanism may bias the benchmark toward underspecified, ambiguous, or ungrounded questions. Please provide an analysis of how many questions were modified by this trigger, examples of original versus modified questions, and results that separate trigger-modified questions from the rest.
  3. [Section 4, Table 1 (UA column)] All models score below 10% on unanswerable questions, but the paper does not specify how 'unanswerable' questions are constructed, what reference answer is stored for them, or how model responses are scored as correct. The low UA accuracy may reflect a mismatch between the model output format and the scoring rule rather than an inability to abstain. I ask for the UA generation protocol, scoring criteria, and several worked examples showing what counts as a correct answer on these items.
  4. [Section 4 (answer-type distribution)] The dataset is reported to contain 6,732 questions, but the answer-type distribution is described as covering 5,085 answers, leaving 1,647 questions with no explanation. It is unclear whether all 6,732 questions have reference answers used in Table 1, or whether the evaluation is on a subset. Please clarify the relationship between the 6,732 and 5,085 numbers, and state exactly which question set was used for the experiments in Table 1.
minor comments (5)
  1. [Table 1] The 'Average' column is not defined; it should state whether it is an unweighted mean over the listed categories, over languages, or over all questions. The category definitions (SC, MC, LC, SP, CP, UA) appear only in the text and should be spelled out in the caption for clarity.
  2. [Figure 1] The axes are labeled only as 'Average Page Numbers' and 'Average Text Tokens' with no units or scale, and the claim that AraEngLongBench 'exceeds other datasets' would be stronger with quantitative values for each dataset or a comparison table.
  3. [References] Reference [31] is listed with the placeholder 'Author(s)' and the full author list and venue are incomplete; several other references use arXiv identifiers without a year format, which should be standardized.
  4. [Throughout] The terminology is inconsistent: the paper uses 'LVLM' in the abstract and 'V-LLM' in the experiments, and there are typographical errors such as 'LLaV A-NeXT' and 'even certain LVLMs even underperform'; a careful copyedit would improve readability.
  5. [Section 2.3] The claim that the proposed framework 'establishes a new standard for large-scale, high-quality annotation' is stronger than the evidence supports, given the lack of validation data; please temper this claim or provide supporting evidence.

Circularity Check

1 steps flagged · score 3.0 of 10

Benchmark-difficulty claim is partly construction-driven by an internal 40% accuracy threshold, but no derivational or self-citation circularity is present.

  1. fitted input called prediction [Section 3.2.2, Multi-Agent Interactive Chain (Agent 4 / Agent 1 feedback loop)]
    "Notably, if Agent 4 detects an accuracy rate exceeding 40%, it triggers Agent 1 to increase question complexity, challenging the answer generation agent."

    Agent 4's 40% accuracy check is used as a construction criterion, not just an evaluation: whenever the agentic chain can answer its own generated questions above 40%, Agent 1 is forced to increase question complexity. Thus the final dataset is filtered and iterated until the generating system finds its questions hard. The Abstract then reports that 'our generated English and Arabic questions (AraEngLongBench) are quite challenging to major open- and close-source LVLMs' as an empirical discovery. Table 1's low scores are therefore partly engineered by the difficulty threshold rather than independently observed. Because the benchmark's test-time models are held out and can score above or below 40% by category, the circularity is partial, not complete.

full rationale

This paper is an empirical benchmark-generation paper rather than a derivation, so most circularity patterns (self-definitional equations, imported uniqueness theorems, self-citation chains) do not apply. There are no load-bearing self-citations. The only circularity-adjacent mechanism is in Section 3.2.2: the pipeline uses an internal 40% accuracy threshold to increase question complexity, so the released benchmark is deliberately filtered to be hard for the very LVLM class being evaluated. That makes the Abstract's conclusion that the benchmark is 'quite challenging' partially a construction artifact, not a fully independent measurement. However, the threshold is a generation policy rather than a fitted predictor, the final Table 1 scores come from held-out models with varied context windows, and not every reported accuracy is forced below 40% (e.g., Gemini SP/CP scores exceed 70%), so this is partial bias, not equivalence. The lack of human validation of reference answers and of Agent 5's evidence validation (Sections 3.2.1-3.2.2) is a serious validity and correctness risk, but it is not in itself an instance of derivation circularity; it would require a separate audit of benchmark quality.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No new physical or formal entities are introduced. The paper's claims rest on three domain assumptions about the reliability of OCR, layout analysis, and LLM-agent-generated QA content, plus two hand-set engineering parameters. Exhaustive enumeration: the ledger captures all explicit parameters and assumptions that the central claim depends on.

free parameters (2)
  • 10-page overlap chunk size = 10 pages
    Chunk size for splitting documents into processible pieces; chosen by hand, no ablation or sensitivity analysis is reported, and it directly affects cross-page question generation.
  • 40% accuracy complexity trigger = 40%
    If Agent 4 detects accuracy exceeding 40%, Agent 1 is triggered to increase question complexity. This threshold is a hand-set heuristic that shapes the difficulty and distribution of the final benchmark, with no ablation.
assumptions (3)
  • domain assumption LVLM-based OCR and layout analysis produce accurate text and structural representations of document pages.
    Invoked in Section 3.2.1; the entire pipeline and the correctness of generated QA pairs depend on this, but no OCR accuracy metrics are reported.
  • domain assumption A multi-agent LLM chain with feedback produces factually grounded questions and answers that align with document evidence.
    Invoked in Section 3.2.2; Agent 5 is claimed to validate alignment, but no human evaluation or error analysis supports this.
  • domain assumption The scraped web documents are representative and license-compliant after exclusion.
    Section 3.1 says documents with restrictive licenses or defects were excluded, but the exact sources, volumes, and filtering criteria are not given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Agent Interactive Question Generation Framework for Long Document Understanding." pith.science (2026). https://pith.science/paper/OUKSLNE5

@misc{pith2026250720145,
  author       = {Pith},
  title        = {Pith review of: Multi-Agent Interactive Question Generation Framework for Long Document Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OUKSLNE5}},
  note         = {Machine review of arXiv:2507.20145}
}
read the original abstract

Document Understanding (DU) in long-contextual scenarios with complex layouts remains a significant challenge in vision-language research. Although Large Vision-Language Models (LVLMs) excel at short-context DU tasks, their performance declines in long-context settings. A key limitation is the scarcity of fine-grained training data, particularly for low-resource languages such as Arabic. Existing state-of-the-art techniques rely heavily on human annotation, which is costly and inefficient. We propose a fully automated, multi-agent interactive framework to generate long-context questions efficiently. Our approach efficiently generates high-quality single- and multi-page questions for extensive English and Arabic documents, covering hundreds of pages across diverse domains. This facilitates the development of LVLMs with enhanced long-context understanding ability. Experimental results in this work have shown that our generated English and Arabic questions (\textbf{AraEngLongBench}) are quite challenging to major open- and close-source LVLMs. The code and data proposed in this work can be found in https://github.com/wangk0b/Multi_Agentic_QA_Long_Doc.git. Sample Question and Answer (QA) pairs and structured system prompts can be found in the Appendix.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 17 canonical work pages

  1. [1]

    Gpt-4 technical report,

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

    arXiv 2023

  2. [2]

    Gemini 1.5: Unlocking multimodal un- derstanding across millions of tokens of context,

    Gemini Team, “Gemini 1.5: Unlocking multimodal un- derstanding across millions of tokens of context,” 2024, Accessed: 2025-02-11

    work page 2024

  3. [3]

    Claude 3 haiku: Our fastest model yet,

    Anthropic, “Claude 3 haiku: Our fastest model yet,” 2024, Accessed: 2025-02-11

    work page 2024

  4. [4]

    Internlm- xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd,

    Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al., “Internlm- xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd,” arXiv preprint arXiv:2404.06512, 2024

    arXiv 2024

  5. [5]

    Llava-next: Stronger llms supercharge multi- modal capabilities in the wild,

    Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chun- yuan Li, “Llava-next: Stronger llms supercharge multi- modal capabilities in the wild,” 2024

    work page 2024

  6. [6]

    Cogvlm: Visual ex- pert for pretrained language models,

    Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al., “Cogvlm: Visual ex- pert for pretrained language models,”arXiv preprint arXiv:2311.03079, 2023

    arXiv 2023

  7. [7]

    Docvqa: A dataset for vqa on document images,

    Minesh Mathew, Dimosthenis Karatzas, and CV Jawa- har, “Docvqa: A dataset for vqa on document images,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2021, pp. 2200–2209

    work page 2021

  8. [8]

    Chartqa: A benchmark for question answering about charts with visual and logical reasoning,

    Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque, “Chartqa: A benchmark for question answering about charts with visual and logical reasoning,”arXiv preprint arXiv:2203.10244, 2022

    arXiv 2022

Show all 31 references
  1. [9]

    Infograph- icvqa,

    Minesh Mathew, Viraj Bagal, Rub `en Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar, “Infograph- icvqa,” inProceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision, 2022, pp. 1697–1706

    2022

  2. [10]

    Towards com- plex document understanding by discrete reasoning,

    Fengbin Zhu, Wenqiang Lei, Fuli Feng, Chao Wang, Haozhou Zhang, and Tat-Seng Chua, “Towards com- plex document understanding by discrete reasoning,” in Proceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 4857–4866

    2022

  3. [11]

    Document understand- ing dataset and evaluation (dude),

    Jordy Van Landeghem, Rub `en Tito, Łukasz Borch- mann, Michał Pietruszka, Pawel Joziak, Rafal Powal- ski, Dawid Jurkiewicz, Micka¨el Coustaty, Bertrand An- ckaert, Ernest Valveny, et al., “Document understand- ing dataset and evaluation (dude),” inProceedings of the IEEE/CVF I...

    2023

  4. [12]

    Mmlongbench-doc: Benchmark- ing long-context document understanding with visual- izations,

    Yubo Ma, Yuhang Zang, Liangyu Chen, Meiqi Chen, Yizhu Jiao, Xinze Li, Xinyuan Lu, Ziyu Liu, Yan Ma, Xiaoyi Dong, et al., “Mmlongbench-doc: Benchmark- ing long-context document understanding with visual- izations,”arXiv preprint arXiv:2407.01523, 2024

    2024 arXiv

  5. [13]

    Longdocurl: a comprehensive multimodal long document benchmark integrating un- derstanding, reasoning, and locating,

    Chao Deng, Jiale Yuan, Pi Bu, Peijie Wang, Zhong- Zhi Li, Jian Xu, Xiao-Hui Li, Yuan Gao, Jun Song, Bo Zheng, et al., “Longdocurl: a comprehensive multimodal long document benchmark integrating un- derstanding, reasoning, and locating,”arXiv preprint arXiv:2412.18424, 2024

    2024 arXiv

  6. [14]

    M-longdoc: A bench- mark for multimodal super-long document understand- ing and a retrieval-aware tuning framework,

    Yew Ken Chia, Liying Cheng, Hou Pong Chan, Chao- qun Liu, Maojia Song, Sharifah Mahani Aljunied, Sou- janya Poria, and Lidong Bing, “M-longdoc: A bench- mark for multimodal super-long document understand- ing and a retrieval-aware tuning framework,”arXiv preprint arXiv:2411.06...

    2024 arXiv

  7. [15]

    Camel-bench: A comprehensive arabic lmm bench- mark,

    Sara Ghaboura, Ahmed Heakl, Omkar Thawakar, Ali Alharthi, Ines Riahi, Abduljalil Saif, Jorma Laakso- nen, Fahad S Khan, Salman Khan, and Rao M Anwer, “Camel-bench: A comprehensive arabic lmm bench- mark,”arXiv preprint arXiv:2410.18976, 2024

    arXiv

  8. [16]

    Doclaynet: A large human-annotated dataset for document-layout segmen- tation,

    Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S Nassar, and Peter Staar, “Doclaynet: A large human-annotated dataset for document-layout segmen- tation,” inProceedings of the 28th ACM SIGKDD con- ference on knowledge discovery and data mining, 2022, pp. 3743–3751

    2022

  9. [17]

    A large-scale multi-document summarization dataset from the wikipedia current events portal,

    Demian Gholipour Ghalandari, Chris Hokamp, Nghia The Pham, John Glover, and Georgiana Ifrim, “A large-scale multi-document summarization dataset from the wikipedia current events portal,”arXiv preprint arXiv:2005.10070, 2020

    2005 arXiv

  10. [18]

    Docile benchmark for document information localiza- tion and extraction,

    ˇStˇep´an ˇSimsa, Milan ˇSulc, Michal U ˇriˇc´aˇr, Yash Pa- tel, Ahmed Hamdi, Mat ˇej Koci ´an, Maty ´aˇs Skalick `y, Jiˇr´ı Matas, Antoine Doucet, Micka ¨el Coustaty, et al., “Docile benchmark for document information localiza- tion and extraction,” inInternational Conference...

    2023

  11. [19]

    Document visual question answering challenge 2020,

    Minesh Mathew, Ruben Tito, Dimosthenis Karatzas, R Manmatha, and CV Jawahar, “Document visual question answering challenge 2020,”arXiv preprint arXiv:2008.08899, 2020

    2020 arXiv

  12. [20]

    No- vachart: A large-scale dataset towards chart understand- ing and generation of multimodal large language mod- els,

    Linmei Hu, Duokang Wang, Yiming Pan, Jifan Yu, Yingxia Shao, Chong Feng, and Liqiang Nie, “No- vachart: A large-scale dataset towards chart understand- ing and generation of multimodal large language mod- els,” inProceedings of the 32nd ACM International Conference on Multimed...

    2024

  13. [21]

    An overview of the tesseract ocr engine,

    Ray Smith, “An overview of the tesseract ocr engine,” inNinth international conference on document analysis and recognition (ICDAR 2007). IEEE, 2007, vol. 2, pp. 629–633

    2007

  14. [22]

    Learn- ing transferable visual models from natural language su- pervision,

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, et al., “Learn- ing transferable visual models from natural language su- pervision,”arXiv preprint arXiv:2103.00020, 2021

    2021 arXiv

  15. [23]

    Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models,

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi, “Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models,”arXiv preprint arXiv:2301.12597, 2023

    2023 arXiv

  16. [24]

    Brave: Broadening the visual encoding of vision- language models,

    Omid Faraji Kar, Alessio Tonioni, Petra Poklukar, An- shul Kulshrestha, Amir Zamir, and Federico Tombari, “Brave: Broadening the visual encoding of vision- language models,”arXiv preprint arXiv:2404.07204, 2024

    2024 arXiv

  17. [25]

    Automated annotation with generative ai requires validation,

    N. Pangakis, S. Wolken, and N. Fasching, “Automated annotation with generative ai requires validation,”arXiv preprint, 2023

    2023

  18. [26]

    Labelvizier: Error profiling and interactive data annotation for long- document understanding,

    Xinyi Yang, Lei Zhang, Saurabh Trivedi, Qian Liao, Byron C. Wallace, and Matthew Lease, “Labelvizier: Error profiling and interactive data annotation for long- document understanding,”arXiv preprint, 2023

    2023

  19. [27]

    Meganno+: A human-llm collaborative an- notation system,

    H. Kim, K. Mitra, R. L. Chen, S. Rahman, and D. Zhang, “Meganno+: A human-llm collaborative an- notation system,”arXiv preprint, 2024

    2024

  20. [28]

    pdf2image: A python library to convert pdf pages to images using poppler,

    Jeremy Belval, “pdf2image: A python library to convert pdf pages to images using poppler,” GitHub repository, 2018

    2018

  21. [29]

    Tex- thawk2: A large vision-language model excels in bilin- gual ocr and grounding with 16x fewer tokens,

    Xinyu Zhang, Lin Zhang, Jiaxin Wang, et al., “Tex- thawk2: A large vision-language model excels in bilin- gual ocr and grounding with 16x fewer tokens,”arXiv preprint arXiv:2410.05261, 2024

    2024 arXiv

  22. [30]

    Towards gen- eral ocr application via a vision-language model,

    Zhaoyi Liu, Yifan Li, Qi Wang, et al., “Towards gen- eral ocr application via a vision-language model,”arXiv preprint arXiv:2501.15558, 2024

    2024 arXiv

  23. [31]

    The yolo framework: A comprehensive re- view of evolution and applications,

    Author(s), “The yolo framework: A comprehensive re- view of evolution and applications,”Computers, vol. 13, no. 12, pp. 336, 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.