Pith. sign in

REVIEW 5 major objections 7 minor 33 references

On the Comprehensibility of Multi-structured Financial Documents using LLMs and Pre-processing Tools

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper shows that converting complex financial PDF tables and charts into structured text before retrieval lets text-only LLMs outperform a multimodal model fed raw pages: GPT-4 reaches 76% versus GPT-4o's 56%, at lower cost.

desk verdict A useful, honest engineering comparison of preprocessing-plus-text RAG against direct MLLM input for financial PDFs, but the small private sample makes the headline GPT-4o gain statistically fragile. read the letter →

arxiv 2506.05182 v2 pith:4N3H5WA7 submitted 2025-06-05 cs.IR

Shivani Upadhyay , Messiah Ataey , Syed Shariyar Murtaza , Yifan Nie , Jimmy Lin This is my paper
classification cs.IR
keywords documentpre-processingretrieval-augmentedgenerationmultimodalLLMfinancialQAtablequestionansweringchartGPT-4ostructuredcomprehension
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 claims that multimodal models like GPT-4o are not reliable at answering questions from complex financial PDFs when the pages are fed directly as images: it reports 56% accuracy on a 75-question expert-written set spanning nested tables and charts. The proposed remedy is a document pre-processor that converts text, tables, and charts into a structured text format before retrieval, so a text-only LLM in a RAG pipeline can do the reasoning. With that pre-processing, accuracy rises to 61.3% for GPT-4o and 76% for GPT-4, and the per-call cost drops substantially compared with image input. If this holds, financial question-answering systems can become both more accurate and cheaper by normalizing structure into text rather than relying on raw multimodal input.

What carries the argument

The load-bearing mechanism is the document pre-processor, a three-part pipeline that turns mixed financial PDF content into JSON text before any retrieval or generation. A commercial document intelligence service extracts textual and tabular content, merging hierarchical column headers so relational information survives flattening; an open-source PDF library uses the service's bounding boxes to crop embedded images; and ChartVLM, a Pix2Struct-based model fine-tuned on the ChartX benchmark, reads chart images and emits their underlying data as CSV, converted to JSON. The resulting JSON is chunked at 600 tokens, embedded, indexed, and searched; the top three chunks are placed into a prompt that asks a text-only LLM for a one-line answer. The argument is that this normalization removes structural failure modes such as nested tables, merged cells, and multi-dimensional plots that trip up direct multimodal input.

What would settle it

Re-run the same pre-processing-and-RAG pipeline on a public financial document QA set with fixed ground truth, such as a sample of annual reports with nested tables and charts, using the same models and prompts, and check whether text-only GPT-4 still beats direct multimodal GPT-4o by the reported margin; a result at or below parity would falsify the claim that the pre-processing, rather than the model or the retrieval setup, produces the gain.

Watch

Extended reading notes

Core claim

The central claim is that converting document structure into text is a stronger route to comprehension than giving a multimodal model the original page images. The paper reports that GPT-4o, a leading multimodal model, scores 56% on a private set of 75 financial QA pairs (53 table questions, 22 chart questions) when pages are fed as images plus text, while text-only GPT-4 in a retrieval-augmented pipeline over pre-processed content scores 76%, GPT-4o scores 61.3%, and GPT-3.5 Turbo scores 60.0%. The pre-processor extracts text and nested tables with a commercial OCR/layout service, crops figure images with an open-source PDF library, and converts chart images to JSON tables with ChartVLM, a Pix2Struct-based chart model. The paper also claims a per-page cost of $0.00231 for the pipeline, making it 74% cheaper than the least expensive commercial alternative it compares against.

Load-bearing premise

The claimed advantage depends entirely on the private 75-question evaluation set with human-scored answers, where the decisive gaps are only 4 questions (56% to 61.3%) and 15 questions (56% to 76%); the paper does not report the number of raters, the scoring instructions, or inter-rater agreement.

Editorial extensions

If this is right

  • Financial report QA systems can use cheaper text-only LLM inference instead of image input while improving accuracy.
  • The pipeline should transfer to other heavily structured document domains, such as insurance filings and regulatory disclosures, where nested tables and plots are common.
  • Image-based RAG that retrieves raw chart images via a vision-language embedding model is likely to underperform text-normalized retrieval, consistent with the paper's 40.9% result for that variant.
  • Because the gains appear with different underlying LLMs (GPT-3.5, GPT-4, GPT-4o), the benefit is tied to the text normalization step rather than to any single model.

Reading between the lines

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

  • The 76% ceiling suggests that after successful preprocessing the remaining errors come from the LLM's reasoning or retrieval, not from structure loss; a per-question error analysis could locate whether the last 24% are retrieval misses, arithmetic errors, or chart-conversion errors.
  • ChartVLM's failure to fully read composite charts, acknowledged in the paper, means that chart-heavy documents may benefit from a dedicated chart-captioning or multi-image parsing step; this could be tested by swapping in a different chart-to-table model and measuring end-to-end accuracy.
  • The cost calculation assumes a fixed 600 tokens per page; pages with dense tables or many charts will produce more tokens after preprocessing, so the claimed 74% cost advantage over commercial alternatives may shrink for chart-heavy filings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 7 minor

Summary. The paper proposes a document pre-processing pipeline for RAG-based QA over multi-structured financial PDFs, combining Azure Document Intelligence, pypdf, and ChartVLM to convert text, tables, and charts into a unified JSON-based textual representation. The authors evaluate this pipeline with GPT-3.5, GPT-4, and GPT-4o on a private 75-question set, reporting 61.3% accuracy for GPT-4o and 76% for GPT-4, versus 56% for direct input of page images to GPT-4o. They also report component-level results on VQAonBD and ComplexChartQA, and a cost analysis claiming the proposed solution is cheaper than commercial alternatives.

Significance. If the claims were fully supported, the paper would provide a useful practical recipe for financial document QA: a text-only RAG pipeline with structure-aware preprocessing can outperform direct multimodal input at lower cost. The paper's strengths are its clear description of the pipeline, the public code repository, and the attempt to isolate component performance on public datasets. However, the central comparative claims rest on a small private evaluation set whose scoring procedure is not reported, and the key GPT-4o difference is not statistically reliable. The paper is therefore a plausible engineering report whose headline results require stronger evidence.

major comments (5)
  1. [Section 5, Table 1] The headline comparison between direct GPT-4o (56%) and pre-processed text-mode GPT-4o (61.3%) corresponds to a difference of only 4 correct answers out of 75 (42 vs 46 correct). Under a two-proportion test, the standard error of this difference is roughly 8 percentage points, so the observed 5.3-point gap is within sampling noise. The paper reports no significance tests, confidence intervals, or per-question results, and Appendix A gives only four example queries. The claim in the abstract that 'integrating pre-processing tools raises the accuracy of LLMs to 61.3% for GPT-4o' is therefore not supported by the evidence as presented. Please provide a statistical analysis, the full question-level outcome table, or a larger evaluation set.
  2. [Section 5, human evaluation paragraph] The evaluation protocol for the private 75-question set is not specified: the paper does not state the number of raters, their instructions or rubric, whether answers were judged blindly, or the inter-rater agreement. Since the accuracy differences at issue are small (4–15 questions), rater variability could easily change the ranking of systems. Please describe the human evaluation in sufficient detail to be replicable, and report agreement statistics.
  3. [Section 5, Cost Effectiveness; Appendix D] The cost comparisons rest on the assumption that a typical document page contains approximately 600 tokens (Appendix D). This assumption is stated but not justified with data from the evaluation documents; page token counts for dense financial tables and charts are likely to vary widely. Because the claimed cost savings (e.g., the 74.33% reduction versus Claude 3 Opus in Table 2) are sensitive to this figure, please provide the page-level token distribution and a sensitivity analysis. In addition, the 'Cost per call' column in Table 1 does not specify whether it is per 600-token page or per actual query, and the currency units are omitted.
  4. [Section 4.1.3 and Section 5, row 2(a)] The direct GPT-4o baseline is underspecified. The text says that PDF pages were converted to images and fed to GPT-4o, and that GPT-4o has a maximum input length of 20 images. The paper does not describe how documents longer than 20 pages are handled, how the relevant pages are selected, or whether all pages of the document were passed or only a subset. This is a load-bearing detail because the baseline accuracy of 56% depends on the fairness and completeness of this input construction. Please specify the page-selection mechanism and the exact input format (e.g., concatenated images with or without text).
  5. [Section 4.1.3 and Section 6] The private evaluation set contains 53 table questions and 22 chart questions, but Table 1 reports only aggregate accuracy. Section 6 acknowledges that ChartVLM often misinterprets composite charts, and Appendix C.4 shows a case where it entirely overlooks one data series. Because the aggregate numbers may mask particularly poor chart performance, please report the accuracy separately for table questions and chart questions. This breakdown is necessary to assess whether the pre-processor genuinely improves chart comprehension or whether the improvement comes almost entirely from the table component.
minor comments (7)
  1. [Table 3] The column header 'VGAonBD' is a typo for 'VQAonBD'.
  2. [Section 2] The model name 'LlaV A' is misspelled; it should be 'LLaVA'.
  3. [Table 5] The caption says 'Direct input analysis with the private evaluation set (tables)', but the reported sample sizes are not tied to the 53 table questions; please clarify how many examples underlie each accuracy value.
  4. [Appendix D] The URLs for Vertex AI and Anthropic both point to the same Google Cloud pricing page; please verify the links so that each product's pricing is correctly attributed.
  5. [Section 4.2.1] The description of VQAonBD says the validation set has 'almost 50 questions partitioned into 5 categories for each image'; please clarify whether this means per-image question templates or a fixed set of question types, and how the 4535 images were used.
  6. [Figure 4] The prompt instructs the model to 'answer the questions in one line', while Section 5 says outputs were assessed by human evaluators. Please state whether all compared systems used the same prompt and whether the one-line constraint was enforced in the evaluation.
  7. [Appendix A] The example queries use abbreviations such as 'ROE' and 'YoY' without expansion; adding a footnote with definitions would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper is an empirical evaluation with external tools and datasets, and its central claims are measurements rather than derivations from the methods' definitions.

full rationale

The paper makes empirical claims about the accuracy of LLMs and MLLMs on financial document QA, comparing a direct multimodal approach with a RAG pipeline that uses document pre-processing tools. All components (Azure Document Intelligence, pypdf, ChartVLM, embeddings, LLMs) are external or standard tools, and the reported accuracies are measured on a private evaluation set assembled by financial experts. There is no derivation chain in which a prediction is equivalent to an input by construction. The choice of retrieval parameter k=3 and the JSON context format are engineering decisions informed by experiments, but the reported accuracies are not computed from those choices; they are observed outcomes of running the system. The paper does not fit any parameter to the test set and then rename that fit as a prediction. No load-bearing self-citation appears, since none of the listed references are authored by the present paper's authors. The skeptical concerns about statistical significance, small sample size, and missing inter-rater agreement pertain to the reliability or validity of the measurements, not to circularity. Therefore, under the defined criteria, no significant circularity is present.

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

No new theoretical constructs are introduced; the paper combines existing commercial and open-source tools. The main dependencies are the private evaluation set, the unspecified human scoring process, and the fairness of the direct-model baseline, which are listed as assumptions.

free parameters (3)
  • k (number of retrieved chunks) = 3
    Section 4.1.2 states k=3 was 'empirically chosen' as a trade-off between accuracy and context size; no ablation or held-out tuning set is reported, so this choice could depend on the evaluation data.
  • Chunk size = 600 tokens
    Section 4.1.1 truncates extracted content into 600-token chunks with LangChain's TokenTextSplitter; this affects retrieval and prompt context, and no sensitivity analysis is given.
  • Assumed tokens per page = 600
    Appendix D uses 600 tokens per document page for all cost-per-page calculations; if the true average differs, the advertised cost advantage changes.
assumptions (4)
  • domain assumption The private 75-pair dataset is representative of real financial document QA
    Section 4.1.3 says the set was assembled by financial experts from Canadian company documents, but it is not released and no comparison to a public benchmark is made for the headline result.
  • domain assumption Human evaluator scores are consistent and unbiased
    Section 5 says generated results were assessed by human evaluators, yet the number of evaluators, instructions, and inter-rater agreement are not reported.
  • ad hoc to paper The direct GPT-4o baseline is a fair comparison
    Section 5 describes converting PDF pages to images and feeding up to 20 images to GPT-4o, but does not explain how pages were selected for multi-page documents; an easier or harder page selection changes the baseline.
  • domain assumption The proprietary services (Azure Document Intelligence, OpenAI models, ChartVLM) behave consistently with the paper's descriptions
    The pipeline treats these as black boxes; no model version pins or deterministic seeds are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Comprehensibility of Multi-structured Financial Documents using LLMs and Pre-processing Tools." pith.science (2026). https://pith.science/paper/4N3H5WA7

@misc{pith2026250605182,
  author       = {Pith},
  title        = {Pith review of: On the Comprehensibility of Multi-structured Financial Documents using LLMs and Pre-processing Tools},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4N3H5WA7}},
  note         = {Machine review of arXiv:2506.05182}
}
read the original abstract

The proliferation of complex structured data in hybrid sources, such as PDF documents and web pages, presents unique challenges for current Large Language Models (LLMs) and Multi-modal Large Language Models (MLLMs) in providing accurate answers. Despite the recent advancements of MLLMs, they still often falter when interpreting intricately structured information, such as nested tables and multi-dimensional plots, leading to hallucinations and erroneous outputs. This paper explores the capabilities of LLMs and MLLMs in understanding and answering questions from complex data structures found in PDF documents by leveraging industrial and open-source tools as part of a pre-processing pipeline. Our findings indicate that GPT-4o, a popular MLLM, achieves an accuracy of 56% on multi-structured documents when fed documents directly, and that integrating pre-processing tools raises the accuracy of LLMs to 61.3% for GPT-4o and 76% for GPT-4, and with lower overall cost. The code is publicly available at https://github.com/OGCDS/FinancialQA.

Figures

Figures reproduced from arXiv: 2506.05182 by the authors.

Figure 1
Figure 1. Flowchart of a RAG framework utilizing docu [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The internal workings of the document pre [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. RAG ingestion pipeline used for performing PDF evaluations. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Prompt utilized within RAG framework. 4.2.1 Dataset The following datasets are used for evaluating LLM table and chart comprehension: • VQAonBD dataset: Tabular images from the dataset (i.e., the validation set of 4535 images) are used as input for the evaluations. The…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 6 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  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]

    Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. 2021. F in QA : A dataset of numerical reasoning over financial data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3697--3711, Online and Pun...

    work page 2021

  4. [4]

    Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. 2023. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807

    arXiv 2023

  5. [5]

    Smith, and Matt Gardner

    Pradeep Dasigi, Kyle Lo, Iz Beltagy, Arman Cohan, Noah A. Smith, and Matt Gardner. 2021. A dataset of information-seeking questions and answers anchored in research papers. arXiv preprint arXiv:2105.03011

    arXiv 2021

  6. [6]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997

    arXiv 2023

  7. [7]

    Enamul Hoque, Parsa Kavehzadeh, and Ahmed Masry. 2022. Chart question answering: State of the art and future directions. arXiv preprint arXiv:2205.03966

    work page Pith review arXiv 2022

  8. [8]

    Wonseok Hwang, Jinyeong Yim, Seunghyun Park, and Minjoon Seo. 2019. A comprehensive exploration on WikiSQL with table-aware word contextualization. arXiv preprint arXiv:1902.01069

    arXiv 2019

Show all 33 references
  1. [9]

    Sujay Kumar Jauhar, Peter Turney, and Eduard Hovy. 2016. TabMCQ : A dataset of general knowledge tables and multiple-choice questions. arXiv preprint arXiv:1602.03960

    2016 arXiv

  2. [10]

    Nengzheng Jin, Joanna Siebert, Dongfang Li, and Qingcai Chen. 2022. A survey on table question answering: Recent advances. arXiv preprint arXiv:2207.05270

    2022 arXiv

  3. [11]

    Yang Jin, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaoqiang Lei, et al. 2023. Unified language-vision pretraining with dynamic discrete visual tokenization. arXiv preprint arXiv:2309.04669

    2023

  4. [12]

    Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Akos Kadar, Adam Trischler, and Yoshua Bengio. 2018. FigureQA : An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300

    2018 arXiv

  5. [13]

    Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. 2023. Pix2Struct : Screenshot parsing as pretraining for visual language understanding. arXiv preprint arXiv:2210.03347

    2023 arXiv

  6. [14]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2021. Retrieval-augmented generation for knowledge-intensive nlp tasks. arXiv preprint ...

    2021 arXiv

  7. [15]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023 a . Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744

    2023 arXiv

  8. [16]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023 b . Visual instruction tuning. arXiv preprint arXiv:2304.08485

    2023 arXiv

  9. [17]

    Rodriques, and Andrew D

    Jakub Lála, Odhran O'Donoghue, Aleksandar Shtedritski, Sam Cox, Samuel G. Rodriques, and Andrew D. White. 2023. PaperQA : Retrieval-augmented generative agent for scientific research. arXiv preprint arXiv:2312.07559

    2023 arXiv

  10. [18]

    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

  11. [19]

    Minesh Mathew, Dimosthenis Karatzas, and C. V. Jawahar. 2021. DocVQA : A dataset for vqa on document images. arXiv preprint arXiv:2007.00398

    2021 arXiv

  12. [20]

    Khapra, and Pratyush Kumar

    Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. 2020. PlotQA : Reasoning over scientific plots. arXiv preprint arXiv:1909.00997

    2020 arXiv

  13. [21]

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen Krueger, Kevin Button, Matthew Knight, Benjamin Chess, and John Schulman. 2022...

    2022 arXiv

  14. [22]

    Braza, Ryan-Rhys Griffiths, Manu Ponnapati, Albert Bou, Jon Laurent, Ori Kabeli, Geemi Wellawatte, Sam Cox, Samuel G

    Siddharth Narayanan, James D. Braza, Ryan-Rhys Griffiths, Manu Ponnapati, Albert Bou, Jon Laurent, Ori Kabeli, Geemi Wellawatte, Sam Cox, Samuel G. Rodriques, and Andrew D. White. 2024. Aviary: training language agents on challenging scientific tasks. arXiv preprent arXiv:2412.21154

    2024 arXiv

  15. [23]

    Phuc Nguyen, Nam Tuan Ly, Hideaki Takeda, and Atsuhiro Takasu. 2023. TabIQA : Table questions answering on business document images. arXiv preprint arXiv:2303.14935

    2023 arXiv

  16. [24]

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, et al. 2024. GPT-4 technical report. arXiv preprint arXiv:2303.08774

    2024 arXiv

  17. [25]

    Panupong Pasupat and Percy Liang. 2015. Compositional semantic parsing on semi-structured tables. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Lon...

    2015

  18. [26]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings...

    2021

  19. [27]

    Skarlinski, Sam Cox, Jon M

    Michael D. Skarlinski, Sam Cox, Jon M. Laurent, James D. Braza, Michaela Hinks, Michael J. Hammerling, Manvitha Ponnapati, Samuel G. Rodriques, and Andrew D. White. 2024. Language agents achieve superhuman synthesis of scientific knowledge. arXiv preprent arXiv:2409.13740

    2024 arXiv

  20. [28]

    Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, et al. 2023. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286

    2023 arXiv

  21. [29]

    Dai, Anja Hauth, Katie Millican, David Silver, et al

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, Katie Millican, David Silver, et al. 2024. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805

    2024 arXiv

  22. [30]

    Rubèn Tito, Dimosthenis Karatzas, and Ernest Valveny. 2021. Document Collection Visual Question Answering, page 778–792. Springer International Publishing

    2021

  23. [31]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. LLaMA : Open and efficient foun...

    2023 arXiv

  24. [32]

    Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, and Yu Qiao. 2024. ChartX & ChartVLM : A versatile benchmark and foundation model for complicated chart reasoning. arXiv preprint arXiv:2402.12185

    2024 arXiv

  25. [33]

    Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. 2021. TAT-QA : A question answering benchmark on a hybrid of tabular and textual content in finance. arXiv preprint arXiv:2105.07624

    2021 arXiv

Pith tools

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