Pith. sign in

REVIEW 4 major objections 4 minor 70 references

SlideAgent: Hierarchical Agentic Framework for Multi-Page Visual Document Understanding

T0 review · 4 major / 4 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 By pre-building a three-level knowledge map of a slide deck, SlideAgent improves multi-page visual QA by 7.9 to 9.8 accuracy points over the same base LLM.

desk verdict A plausible and well-engineered agentic pipeline for slide QA, but the reported gains are not yet cleanly attributed to the hierarchy because there is no flat-OCR control. read the letter →

arxiv 2510.26615 v4 pith:5QJC4YEW submitted 2025-10-30 cs.CL

Yiqiao Jin , Rachneet Kaur , Zhen Zeng , Sumitra Ganesh , Srijan Kumar This is my paper
classification cs.CL
keywords slidedecksmulti-pagevisualdocumentsquestionansweringhierarchicalagenticframeworkretrieval-augmentedgenerationlayoutunderstandingelementgroundingmultimodallargelanguagemodels
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

SlideAgent makes a direct claim: a multimodal LLM answers questions about multi-page slide decks substantially better if the document is first decomposed into a structured, query-agnostic knowledge base at three levels—deck, page, and element. That decomposition is done once, before any question, using specialized agents that write a global summary, a page-by-page narrative that carries context forward, and element-level annotations tied to detected bounding boxes. At query time an orchestrator chooses which levels are needed, generates subqueries for retrieval, and lets the agents reason before a synthesizer reconciles any disagreements. The paper reports accuracy gains of 7.9 points on a proprietary model and 9.8 points on an open-source model across slide QA benchmarks, with the strongest effects in multi-hop and layout-sensitive questions. If true, this means fine-grained document reasoning does not require retraining or a stronger model—only a better way to present the page to the model.

What carries the argument

The central mechanism is the hierarchical knowledge construction stage—global, page, element—where page knowledge is generated sequentially so each slide's description is conditioned on the previous slide, and element knowledge is anchored to detected bounding boxes so the synthesizer can ground answers spatially. A query orchestrator then activates only the levels a question needs and generates subqueries for retrieval.

What would settle it

Give SlideAgent a set of test decks whose slides contain deliberately low-contrast charts and dense overlapping text, run it once with its normal detector and once with human-verified element boxes, and compare accuracy; if accuracy does not improve when the boxes are corrected, or the element agent's contribution vanishes, the parsing pipeline rather than the hierarchical reasoning is carrying the reported gains.

Watch

Extended reading notes

Core claim

The central discovery is that fine-grained, layout-aware reasoning over multi-page documents is largely a representation problem. When a large multimodal model is handed a full slide or page image, it misses chart segments, footnotes, and spatial cues that carry the answer; when the same model is given the same visual content after it has been decomposed into typed, coordinate-annotated elements with semantic descriptions, it answers correctly far more often. SlideAgent operationalizes this by constructing, before any question arrives, a query-agnostic knowledge base with three levels—global (deck summary and narrative), page (sequential summaries that carry context forward), and element (pe

Load-bearing premise

The load-bearing premise is that the off-the-shelf OCR and layout-detection pipeline reliably splits each slide into semantically coherent elements with correct bounding boxes; on visually complex or low-contrast slides it can fail, and the paper's own future-work section concedes this.

Editorial extensions

If this is right

  • No fine-tuning is required: the same base model gains roughly 8–10 accuracy points on slide-deck QA by routing questions through the three-level knowledge base.
  • The largest gains are in multi-hop reasoning and visual/layout questions, where whole-page processing is weakest; the paper reports about a 10-point improvement in multi-hop reasoning and 7.7 points in layout/visual reasoning.
  • Structured page knowledge improves page retrieval for sparse lexical and dense text retrievers, not just multimodal retrieval, making the pipeline cheaper at scale.
  • The gains hold even when the correct pages are provided to the model, implying that element-level grounding, not retrieval alone, drives most of the improvement.
  • The framework is model-agnostic: quality gains appear with both proprietary and open-source backbones and across technical and financial slide domains.

Reading between the lines

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

  • A natural extension, not run in the paper, is to measure how sensitive the gains are to the merging threshold (set to 15 pixels in the appendix); sweeping it over varied slide layouts would show whether the framework can be tuned per domain or whether a single threshold is a hidden liability.
  • The paper's query-agnostic knowledge base is built once per deck, which suggests the cost could be amortized over many questions on the same deck; a follow-up that reports construction time and token cost relative to QA savings would clarify when this pipeline pays for itself.
  • Because the framework treats elements independently, a graph-based extension that links a chart to its legend or an arrow to its node—already flagged as future work—could be tested against the current element descriptions to see whether explicit relation modeling closes the remaining gap on multi-hop and visual 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 / 4 minor

Summary. The paper proposes SlideAgent, a hierarchical agentic framework for multi-page visual document understanding. It constructs query-agnostic knowledge at three levels—global, page, and element—using LLM-based agents plus an OCR/layout parsing pipeline, then retrieves relevant knowledge and synthesizes answers. Experiments on SlideVQA, TechSlides, and FinSlides with GPT-4o and InternVL3-8B report consistent gains over base models (+7.9 and +9.8 overall, respectively), with ablation studies, query-type breakdowns, and retrieval improvements over BM25, BGE, SFR, COLPALI, and VisRAG. The central claim is that the hierarchical decomposition and selective agent activation yield substantial, model-agnostic improvements for slide QA.

Significance. If the reported gains are correct, SlideAgent provides a training-free, model-agnostic recipe for improving MLLMs on multi-page visual documents. The paper's strengths include evaluation across three datasets and two backbone families, ablations of major components, a ground-truth-page oracle setting, and retrieval experiments with multiple text and multimodal retrievers. However, the attribution of the gains to the hierarchical agentic design is weakened by the absence of a control that provides baselines with the same OCR/layout text, and by an internal contradiction about whether element-level retrieval was actually evaluated. Table 4 also contains identical base scores for two different retrievers, which raises a data-integrity concern. These issues need to be resolved before the central claim is fully supported.

major comments (4)
  1. [§3.1, Tables 1/3; Eq. (2)] There is no control condition that gives the baselines the same EasyOCR/Docling text, bounding boxes, and element types that SlideAgent injects into the prompt. The reported gains could therefore be driven by the extra OCR/layout signal rather than by the hierarchical agentic design. Please add a flat condition (e.g., SlideAgent with the parsed text/boxes concatenated without global/page/element decomposition) or otherwise feed comparable low-level signals to the baselines.
  2. [§3.2 vs §6] §3.2 claims that the ground-truth-page oracle results 'demonstrate the effectiveness of element-level retrieval,' but §6 explicitly states 'We did not evaluate element-level retrieval.' In the ground-truth-page setting retrieval is removed entirely, so the claim is unsupported. This is a direct internal contradiction that must be fixed, either by removing the claim or by adding an actual element-level retrieval experiment.
  3. [Table 4] The base scores for BGE and SFR are identical across all six metrics (MRR 70.1, Recall@1 56.1, nDCG@1 60.9, Recall@3 75.8, Hit@3 78.1, nDCG@3 69.2). Two distinct retrievers cannot produce exactly these numbers; this is likely a copy error. Since SFR is chosen as the base retriever partly on the basis of this table, the numbers must be corrected and the conclusions re-checked.
  4. [Appendix A.1, §2.1] The element parsing pipeline is load-bearing: element merging with threshold τ=15 px and EasyOCR/Docling detection are not validated on SlideVQA/TechSlides/FinSlides. The paper itself notes in Future Work that the pipeline 'may struggle with visually complex or low-contrast designs.' Without basic detection-quality statistics, a sensitivity analysis on τ, or an error analysis on the benchmark data, it is hard to assess how much of the reported improvement depends on the assumption that off-the-shelf layout parsing is accurate for these datasets.
minor comments (4)
  1. [Throughout] The model name 'LLaVA' is typeset as 'LLaV A' in several places. Please fix the spacing.
  2. [Table 3 caption] The caption says 'all baseline methods (Type 2/3) sharing the same base model (GPT-4o),' but the table contains open-source raw models that do not use GPT-4o. Clarify the caption to avoid confusion.
  3. [Abstract] 'achieves significant improvement' would read better as 'achieves significant improvements' or 'achieves a significant improvement.'
  4. [§3.4, Fig. 3] The 'w/o P' ablation removes the page agent, but it also removes a text-compression/summarization step. The steep drop does not cleanly isolate the value of hierarchy versus prompt-length reduction. A brief discussion of this confound would be helpful.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central claim is an empirical benchmark comparison; minor self-citations are not load-bearing, and the main flagged issue is an unsupported attribution, not a circular derivation.

full rationale

The paper's central claim—that SlideAgent improves multi-page visual QA over its base MLLM (+7.9 proprietary, +9.8 open-source)—is established empirically against external baselines (Tables 1-3, 11-12), not derived from premises that already contain the result. The hierarchical knowledge K_g/K_p/K_e is constructed before queries are seen, and its contribution to page-level retrieval is tested externally in Table 4 against raw retrievers using ranking labels. Free parameters such as the merging threshold τ=15 and the answer-agreement threshold NLS=0.75 are fixed constants, not fitted to test answers. Self-citations (Jin et al. 2024, 2025) are incidental and do not carry the argument: they support retriever/QA choices but those choices are independently evaluated. The one notable flaw is an over-attribution rather than circularity: §3.2 states that oracle-setting gains 'demonstrat[e] the effectiveness of element-level retrieval,' while §6 admits 'We did not evaluate element-level retrieval due to the lack of corresponding annotations in existing datasets.' Also, no baseline receives SlideAgent's injected OCR/layout text, so part of the +7.9/+9.8 gain could be due to extra textual signal rather than the hierarchy; this is a confound/ablation gap, not a circular reduction. Thus the paper's derivation chain is not circular, and only a low score is warranted.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method is an empirical pipeline built on LLMs and off-the-shelf OCR/layout tools; no new physical or mathematical entities are introduced. The numeric free parameters are hand-set thresholds without sensitivity analysis, and the axioms are practical assumptions about tool reliability and benchmark representativeness.

free parameters (2)
  • Element merging distance threshold τ = 15 pixels
    Chosen by hand in Appendix A.1 / Algorithm 1; controls which OCR boxes are merged into elements. No sensitivity analysis is reported.
  • Agent agreement Levenshtein threshold = 0.75 (Normalized Levenshtein Similarity)
    Appendix B.2; used to decide when agent answers agree so the synthesizer is bypassed. Chosen by hand.
assumptions (4)
  • domain assumption EasyOCR and Docling layout parsing accurately detect and classify slide elements
    The element agent (Section 2.1) and Appendix A.1 rely on these tools; the paper's own Future Work admits they may struggle with low-contrast or visually complex designs.
  • domain assumption Text-based dense retrieval (SFR) can represent slide pages when queries are expanded with generated subqueries and page knowledge
    Section 2.2 and Table 4: the method depends on textual retrieval of inherently visual content.
  • domain assumption LLM-generated global/page/element knowledge is faithful and errors do not propagate
    Knowledge construction (Section 2.1) feeds model-written summaries back as context; hallucination in Kg/Kp/Ke would bias final answers.
  • domain assumption SlideVQA, TechSlides, and FinSlides are representative of multi-page visual document QA
    All conclusions are drawn from these public benchmarks; no real-world deployment evaluation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SlideAgent: Hierarchical Agentic Framework for Multi-Page Visual Document Understanding." pith.science (2026). https://pith.science/paper/5QJC4YEW

@misc{pith2026251026615,
  author       = {Pith},
  title        = {Pith review of: SlideAgent: Hierarchical Agentic Framework for Multi-Page Visual Document Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5QJC4YEW}},
  note         = {Machine review of arXiv:2510.26615}
}
read the original abstract

Multi-page visual documents such as manuals, brochures, presentations, and posters convey key information through layout, colors, icons, and cross-slide references. While multimodal large language models (MLLMs) offer opportunities in document understanding, current systems struggle with complex, multi-page visual documents, particularly in fine-grained reasoning over elements and pages. We introduce SlideAgent, a versatile agentic framework for understanding multi-modal, multi-page, and multi-layout documents, especially slide decks. SlideAgent employs specialized agents and decomposes reasoning into three specialized levels--global, page, and element--to construct a structured, query-agnostic representation that captures both overarching themes and detailed visual or textual cues. During inference, SlideAgent selectively activates specialized agents for multi-level reasoning and integrates their outputs into coherent, context-aware answers. Extensive experiments show that SlideAgent significantly improves accuracy over both proprietary (+7.9%) and open-source models (+9.8%).

Figures

Figures reproduced from arXiv: 2510.26615 by the authors.

Figure 1
Figure 1. When given the full page, the LLM miscounts [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. SlideAgent generates knowledge about input slide decks in a hierarchical manner at 3 levels: [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison among variants of [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Accuracy of SlideAgent and base models (GPT-4o / InternVL3-8B) on different query types. types, as defined in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Example answers generated by SlideAgent. Agents at different levels work together to provide compre [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison among variants of [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Sample global knowledge Kg generated by the global agent, showing document-level summary, objectives, and narrative structure. Notation Description pi , vi A slide page and its visual content P = {p1, p2, . . . } Set of pages in the slide deck q User query Qˆ Set of su…
Figure 8
Figure 8. Figure 8: Sample page-level knowledge representation [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Sample element-level knowledge representation [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 25 linked inside Pith

  1. [1]

    Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, and 1 others. 2024. Phi-3 technical report: A highly capable language model locally on your phone, 2024. arXiv:2404.14219, 2:6

    arXiv 2024

  2. [2]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  3. [3]

    Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, and 1 others. 2023. Gemini: a family of highly capable multimodal models. arXiv:2312.11805

    arXiv 2023

  4. [4]

    Anthropic. 2025. https://claude.ai/ Claude

    2025

  5. [5]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, and 1 others. 2025. Qwen2. 5-vl technical report. arXiv:2502.13923

    arXiv 2025

  6. [6]

    Aniket Bhattacharyya, Anurag Tripathi, Ujjal Das, Archan Karmakar, Amit Pathak, and Maneesh Gupta. 2025. Information extraction from visually rich documents using llm-based organization of documents into independent textual segments. In ACL

    2025

  7. [7]

    Showmik Bhowmik. 2023. Document layout analysis, volume 3. Springer

    2023

  8. [8]

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, and 1 others. 2024. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185--24198

    2024

Show all 70 references
  1. [9]

    Xin Cheng, Xun Wang, Xingxing Zhang, Tao Ge, Si-Qing Chen, Furu Wei, Huishuai Zhang, and Dongyan Zhao. 2024. xrag: Extreme context compression for retrieval-augmented generation with one token. arXiv:2405.13792

    2024 arXiv

  2. [10]

    Jaemin Cho, Debanjan Mahata, Ozan Irsoy, Yujie He, and Mohit Bansal. 2024. M3docrag: Multi-modal retrieval is what you need for multi-page multi-document understanding. arXiv:2411.04952

    2024 arXiv

  3. [11]

    Timo I Denk and Christian Reisswig. 2019. Bertgrid: Contextualized embedding for 2d document representation and understanding. arXiv preprint arXiv:1909.04948

    2019 arXiv

  4. [12]

    Manuel Faysse, Hugues Sibille, Tony Wu, Bilel Omrani, Gautier Viaud, C \'e line Hudelot, and Pierre Colombo. 2024. Colpali: Efficient document retrieval with vision language models. In ICLR

    2024

  5. [13]

    Tsu-Jui Fu, William Yang Wang, Daniel McDuff, and Yale Song. 2022. Doc2ppt: Automatic presentation slides generation from scientific documents. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 634--642

    2022

  6. [14]

    Liangcai Gao, Yilun Huang, Herv \'e D \'e jean, Jean-Luc Meunier, Qinqin Yan, Yu Fang, Florian Kleber, and Eva Lang. 2019. Icdar 2019 competition on table detection and recognition (ctdar). In 2019 International conference on document analysis and recognition (ICDAR), pages 15...

    2019

  7. [15]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  8. [16]

    Monica Haurilet, Ziad Al-Halah, and Rainer Stiefelhagen. 2019. Spase-multi-label page segmentation for presentation slides. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 726--734. IEEE

    2019

  9. [17]

    huridocs. 2025. https://github.com/huridocs/pdf-document-layout-analysis Pdf document layout analysis: A docker-powered microservice for intelligent pdf document layout analysis, ocr, and content extraction

    2025

  10. [18]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  11. [19]

    Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. 2019. Funsd: A dataset for form understanding in noisy scanned documents. In 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW), volume 2, pages 1--6. IEEE

    2019

  12. [20]

    Yiqiao Jin, Kartik Sharma, Vineeth Rakesh, Yingtong Dou, Menghai Pan, Mahashweta Das, and Srijan Kumar. 2025. Sara: Selective and adaptive retrieval-augmented generation with context compression. arXiv:2507.05633

    2025 arXiv

  13. [21]

    Yiqiao Jin, Qinlin Zhao, Yiyang Wang, Hao Chen, Kaijie Zhu, Yijia Xiao, and Jindong Wang. 2024. Agentreview: Exploring peer review dynamics with llm agents. In EMNLP, pages 1208--1226

    2024

  14. [22]

    Thomas G Kieninger. 1998. Table structure recognition based on robust block segmentation. In Document recognition V, volume 3305, pages 22--32. SPIE

    1998

  15. [23]

    Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, Jy-yong Sohn, and Chanyeol Choi. 2024. https://getlinq.com/blog/linq-embed-mistral/ Linq-embed-mistral:elevating text retrieval with improved gpt data through task-specific control and quality refineme...

    2024

  16. [24]

    Annie Lang. 2000. The limited capacity model of mediated message processing. Journal of communication, 50(1):46--70

    2000

  17. [25]

    VI Lcvenshtcin. 1966. Binary coors capable or ‘correcting deletions, insertions, and reversals. In Soviet physics-doklady, volume 10

    1966

  18. [26]

    Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. 2020. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038

    2020 arXiv

  19. [27]

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning. arXiv:2308.03281

    2023 arXiv

  20. [28]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. NeurIPS, 36:34892--34916

    2023

  21. [29]

    Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang, Chunfeng Yuan, Bing Li, and 1 others. 2024. Mibench: Evaluating multimodal large language models over multiple images. arXiv preprint arXiv:2407.15272

    2024 arXiv

  22. [30]

    Yubo Ma, Yuhang Zang, Liangyu Chen, Meiqi Chen, Yizhu Jiao, Xinze Li, Xinyuan Lu, Ziyu Liu, Yan Ma, Xiaoyi Dong, and 1 others. 2024. Mmlongbench-doc: Benchmarking long-context document understanding with visualizations. Advances in Neural Information Processing Systems, 37:959...

    2024

  23. [31]

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

    2022 arXiv

  24. [32]

    Minesh Mathew, Viraj Bagal, Rub \`e n Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. 2022. Infographicvqa. In CVPR, pages 1697--1706

    2022

  25. [33]

    Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. 2021. Docvqa: A dataset for vqa on document images. In CVPR, pages 2200--2209

    2021

  26. [34]

    Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. 2024. https://www.salesforce.com/blog/sfr-embedding/ Sfr-embedding-mistral:enhance text retrieval with transfer learning . Salesforce AI Research Blog

    2024

  27. [35]

    Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. 2019. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947--952. IEEE

    2019

  28. [36]

    Laura F Naysmith, Veena Kumari, and Steven CR Williams. 2021. Neural mapping of prepulse-induced startle reflex modulation as indices of sensory information processing in healthy and clinical populations: A systematic review. Human Brain Mapping, 42(16):5495--5518

    2021

  29. [37]

    OpenAI. 2025. https://chat.openai.com/ Gpt-4o

    2025

  30. [38]

    Maciej P Polak and Dane Morgan. 2025. Leveraging vision capabilities of multimodal llms for automated data extraction from plots. arXiv:2503.12326

    2025 arXiv

  31. [39]

    Rafa Powalski, ukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Micha Pietruszka, and Gabriela Pa ka. 2021. Going full-tilt boogie on document understanding with text-image-layout transformer. In International Conference on Document Analysis and Recognition, pages 732--747. Springer

    2021

  32. [40]

    Johannes Rausch, Octavio Martinez, Fabian Bissig, Ce Zhang, and Stefan Feuerriegel. 2021. Docparser: Hierarchical document structure parsing from renderings. In AAAI, volume 35, pages 4328--4338

    2021

  33. [41]

    Stephen Robertson, Hugo Zaragoza, and Michael Taylor. 2004. Simple bm25 extension to multiple weighted fields. In CIKM, pages 42--49

    2004

  34. [42]

    Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. 2024. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. In The Thirty-eight Conference on Neural Info...

    2024

  35. [43]

    Karun Sharma and Vidushee Vats. 2025. Think to ground: Improving spatial reasoning in llms for better visual grounding. In Workshop on Reasoning and Planning for Large Language Models

    2025

  36. [44]

    Michael Shilman, Percy Liang, and Paul Viola. 2005. Learning nongenerative grammatical models for document analysis. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 962--969. IEEE

    2005

  37. [45]

    Jeremy Singer-Vine. 2025. https://github.com/jsvine/pdfplumber pdfplumber

    2025

  38. [46]

    Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. 2019. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317--8326

    2019

  39. [47]

    Ray Smith. 2007. An overview of the tesseract ocr engine. In Ninth international conference on document analysis and recognition (ICDAR 2007), volume 2, pages 629--633. IEEE

    2007

  40. [48]

    Edward Sun, Yufang Hou, Dakuo Wang, Yunfeng Zhang, and Nancy XR Wang. 2021. D2s: Document-to-slide generation via query-based text summarization. arXiv preprint arXiv:2105.03664

    2021 arXiv

  41. [49]

    Ryota Tanaka, Taichi Iki, Taku Hasegawa, Kyosuke Nishida, Kuniko Saito, and Jun Suzuki. 2025. Vdocrag: Retrieval-augmented generation over visually-rich documents. arXiv:2504.09795

    2025 arXiv

  42. [50]

    Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. 2023. Slidevqa: A dataset for document visual question answering on multiple images. In AAAI, volume 37, pages 13636--13645

    2023

  43. [51]

    Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. 2021. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13878--13888

    2021

  44. [52]

    Deep Search Team. 2024. https://doi.org/10.48550/arXiv.2408.09869 Docling technical report . Technical report

    2024 doi arXiv

  45. [53]

    Rub \`e n Tito, Dimosthenis Karatzas, and Ernest Valveny. 2021. Document collection visual question answering. In International Conference on Document Analysis and Recognition, pages 778--792. Springer

    2021

  46. [54]

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, and 1 others. 2025. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, local...

    2025 arXiv

  47. [55]

    Gaurav Verma, Rachneet Kaur, Nishan Srishankar, Zhen Zeng, Tucker Balch, and Manuela Veloso. 2024. Adaptagent: Adapting multimodal web agents with few-shot learning from human demonstrations. In ACL

    2024

  48. [56]

    Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. 2024 a . Docllm: A layout-aware generative language model for multimodal document understanding. In ACL, pages 8529--8548

    2024

  49. [57]

    Jiayu Wang, Yifei Ming, Zhenmei Shi, Vibhav Vineet, Xin Wang, Sharon Li, and Neel Joshi. 2024 b . Is a picture worth a thousand words? delving into spatial reasoning for vision language models. NeurIPS, 37:75392--75421

    2024

  50. [58]

    Qiuchen Wang, Ruixue Ding, Zehui Chen, Weiqi Wu, Shihang Wang, Pengjun Xie, and Feng Zhao. 2025 a . Vidorag: Visual document retrieval-augmented generation via dynamic iterative reasoning agents. arXiv:2502.18017

    2025 arXiv

  51. [59]

    Yiyang Wang, Rishabh Goel, Sheraz Hassan, Taegen J Doscher, Shilin Wang, Lexington Allen Whalen, Aditya S Gandhi, Yaman S Sangar, Alex Cabral, Xuhai Xu, and 1 others. 2025 b . Puffem: An e-cigarette sleeve for estimating user nicotine intake. In CHASE, pages 129--133. IEEE

    2025

  52. [60]

    Navve Wasserman, Roi Pony, Oshri Naparstek, Adi Raz Goldfarb, Eli Schwartz, Udi Barzelay, and Leonid Karlinsky. 2025. Real-mm-rag: A real-world multi-modal retrieval benchmark. In ACL

    2025

  53. [61]

    Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. 2024. Next-gpt: Any-to-any multimodal llm. In ICML

    2024

  54. [62]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. https://arxiv.org/abs/2309.07597 C-pack: Packaged resources to advance general chinese embedding . Preprint, arXiv:2309.07597

    2023 arXiv

  55. [63]

    Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, and 1 others. 2020 a . Layoutlmv2: Multi-modal pre-training for visually-rich document understanding. arXiv preprint arXiv:2012.14740

    2020 arXiv

  56. [64]

    Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. 2020 b . Layoutlm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192--1200

    2020

  57. [65]

    Yunqiu Xu, Linchao Zhu, and Yi Yang. 2024. Mc-bench: A benchmark for multi-context visual grounding in the era of mllms. arXiv preprint arXiv:2410.12332

    2024 arXiv

  58. [66]

    Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. 2023. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441

    2023 arXiv

  59. [67]

    Shi Yu, Chaoyue Tang, Bokai Xu, Junbo Cui, Junhao Ran, Yukun Yan, Zhenghao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, and 1 others. 2025. Visrag: Vision-based retrieval-augmented generation on multi-modality documents. In ICLR

    2025

  60. [68]

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, and 32 others. 2025. https://arxiv.org/abs/250...

    2025 arXiv

  61. [69]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  62. [70]

    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 gl...

Pith tools

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