Pith. sign in

REVIEW 4 major objections 5 minor 41 references

Evaluating LLMs for Visualization Generation and Understanding

T0 review · 4 major / 5 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 LLMs can generate correct code for simple charts like bar and pie and answer basic chart questions, but they fail on complex visualizations and fine-grained perceptual judgments; GPT-4o led the evaluated models.

desk verdict A useful but thin snapshot of four LLMs on chart generation and FigureQA; the manual 'understanding' analysis is too under-specified to carry the abstract's specific perceptual-failure claims. read the letter →

arxiv 2507.22890 v1 pith:OQRFODS4 submitted 2025-06-16 cs.HC cs.AI

Saadiq Rauf Khan , Vinit Chandak , Sougata Mukherjea This is my paper
classification cs.HCcs.AI
keywords largelanguagemodelsvisualizationgenerationunderstandingFigureQAchartquestionansweringVega-LiteMatplotlibzero-shotprompting
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 asks whether current LLMs are ready to produce and interpret data visualizations from natural language. The authors evaluated four models for chart-code generation across 24 chart types in Python and Vega-Lite, and three models for chart understanding on the FigureQA dataset plus author-created manual questions. They found that models generate correct code for common charts such as bar, pie, and line charts, and answer simple yes/no chart questions with moderate accuracy. They also found consistent failures on uncommon charts, Vega-Lite scripting, dotted lines, close bar lengths, and fine-grained color and length judgments. The paper's contribution is a mapping of where these models currently succeed and fail, with GPT-4o as the strongest overall performer in this sample.

What carries the argument

The evaluation rests on two instruments: a 24-chart generation suite covering common and uncommon visualization types, and a question-answering battery built from the FigureQA dataset (synthetic bar, line, and pie charts with binary questions) plus the authors' own non-binary questions about bar counts, colors, intersections, and areas. The central mechanism is zero-shot prompting: each model receives only the prompt text, with or without a short system prompt, and the chart image, and the output is compared against the specified requirements or ground-truth answers. This design makes the comparison a direct measure of what the models can do without fine-tuning or few-shot examples.

What would settle it

Re-score the manual analysis with an independent answer key and two or more annotators who did not write the questions; if agreement falls below about 80%, or if a model given the question text without the chart image matches the reported accuracy on the binary FigureQA items, then the specific claims about perceptual failures are not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that LLM performance on visualization tasks splits by task difficulty: generation of standard chart types in Python with default settings is mostly reliable, while generation of specialist charts (bullet, pictogram, graduated symbol maps), customized visual variables, and Vega-Lite scripts degrades sharply; understanding of charts is likewise reliable for simple comparisons and pie and donut layouts but unreliable for line charts, dotted lines, close boundaries, and exact lengths. The quantitative anchor is that GPT-4o succeeded on about 95% of default Python chart prompts and 66% of FigureQA questions, with Gemini at 64% and Claude at 55%, and manual testing showed all models dropping to well below 50% image-level accuracy on line charts. The paper treats these numbers as evidence that current LLMs are useful but not yet dependable for visualization generation and understanding.

Load-bearing premise

The load-bearing premise is that the authors' manual questions and ground-truth answers are a valid and reliable measure of chart understanding; the paper does not release the full question set or answer key, reports no inter-rater validation, and invokes a human baseline that was never actually measured.

Editorial extensions

If this is right

  • If correct, applications that rely on LLMs to generate standard charts such as bar, column, line, and scatter plots can treat those as near-solved for Python, while bullet charts, pictograms, and Vega-Lite output need human review.
  • If correct, chart-question answering systems should avoid trusting LLMs for fine-grained perceptual judgments such as near-equal bar lengths, dotted-line counts, and exact values read from axes.
  • If correct, adding a simple system prompt such as 'Analyse the following chart carefully and answer the following questions correctly' is a low-cost intervention that improves accuracy for most models and chart types.
  • If correct, the large gap between Python and Vega-Lite generation means LLM-based visualization tools that accept Vega-Lite specifications should not assume parity across target languages.

Reading between the lines

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

  • An implication the authors leave implicit is that the specific accuracy numbers are time-sensitive, since the models were evaluated at one snapshot; the task categories where all models fail, such as dotted lines and close boundaries, are more likely to persist across model generations.
  • A testable extension would be to separate chart understanding from chart reading: asking the same questions with the numeric data table available would show whether failures come from visual perception or from the underlying reasoning about the data.
  • The close-boundary and dotted-line failures suggest a visual-acuity floor in current vision-language models; connecting these qualitative results to standard perception tests, such as length discrimination thresholds, could turn them into a quantitative benchmark.
  • The manual results could become a reusable benchmark if the question set and answer key were released; without them, other labs cannot independently verify which models fail at which perceptual tasks.
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 manuscript reports a small-scale empirical evaluation of four LLMs (GPT-3.5, GPT-4o, Gemini-1.5-pro, and Claude 3 Opus) for visualization generation in Python and Vega-Lite, and of three of these models (GPT-4o, Gemini-1.5-pro, Claude 3 Opus) for visualization understanding on the public FigureQA benchmark plus author-authored manual questions. The headline findings are that GPT-4o leads code generation (95% of 24 Python chart prompts), all models handle common chart types but fail on less common ones, and that on FigureQA GPT-4o (66.0%) slightly exceeds Gemini (64.3%) and Claude (54.6%). The manual analysis in Sections 4.4 and 4.5 is used to claim that LLMs struggle with close boundaries, dotted lines, and exact lengths and that GPT-4o is the best overall performer for chart understanding. The paper positions these observations as insights for improving both LLMs and information visualization systems.

Significance. If the results are taken at face value, the paper provides a useful, comparative snapshot of LLM capabilities for visualization tasks in a relatively under-explored area. Its strengths include using the external FigureQA benchmark for the automated part, covering a broad set of 24 chart types for generation, and reporting per-model success counts explicitly. However, the manual analysis, which carries the paper's most distinctive qualitative conclusions, is not currently reproducible, the automated comparison lacks any statistical grounding, and the generation results rely on an underspecified correctness criterion. Until these issues are addressed, the paper's significance is limited to a preliminary descriptive study that would need substantial strengthening to support its central qualitative claims.

major comments (4)
  1. [§4.4–4.5, Tables 6–9] The manual analysis that supports the abstract's specific failure claims is not reproducible and has no validation. The authors state that for each of 20 randomly chosen charts per type they created a group of questions and then "selected a subset" for evaluation, without giving the selection rule, the exact questions, the per-chart denominators, or an answer key. All answers were checked by the authors against their own expected answers, with no second annotator, inter-rater reliability, or scoring rubric. The sentence at the end of §4.5 that inter-LLM comparison "will inherently compare the LLMs' performances with a human baseline" is unsupported because no human participants were recruited and no human accuracy was measured. Since the abstract's headline limitations (e.g., "identifying relationships between close boundaries and determining lengths of shapes") rest entirely on Tables 6–9 and the associated qualitative observations, these conclusions are not yet supported. Please release the full question set and answer key, specify the selection rule, define scoring criteria (including how color-name mismatches such as Claude's are treated), and provide either an independent-annotation agreement measure or a measured human baseline.
  2. [§4.2, Table 5] The ranking GPT-4o > Gemini-1.5-pro > Claude 3 Opus on FigureQA is presented without any statistical support. The paper reports only raw counts and percentages over 1,342 binary questions; there are no confidence intervals, standard errors, or significance tests. The GPT-4o–Gemini gap is 23 correct answers (1.7 percentage points), and a 95% Wilson interval for 66.0% accuracy on 1,342 trials is approximately ±2.5%, so the observed difference is within sampling noise. The paper should report confidence intervals and a paired significance test (e.g., McNemar's test) for model comparisons, and should state how the 100 images were sampled and whether the chart-type balance is known. Without this, the claim that GPT-4o is the best performer on FigureQA is not quantitatively established.
  3. [§3.1.6, Tables 1–3] The correctness criterion for chart generation is not operationalized. The paper says the authors "examined the generated charts to determine whether they satisfied all the requirements specified in the prompts" but does not define what counts as correct when code fails to run, when the output is semantically correct but visually different, or when the model produces a different chart type (e.g., Gemini's violin plot for a range plot in §3.2). The claim "GPT-4o produced 95% of charts" therefore depends on an unstated scoring rubric. The authors should provide a concrete rubric, ideally with executable checks for code validity and with chart-type classification performed by more than one rater or against objective criteria.
  4. [§3.1.7 and §4.2] The experimental procedure is under-specified in ways that affect reproducibility. The statement "each prompt is repeated three times, and we accept the output only if they remain the same" is ambiguous: it is unclear whether a chart was counted as correct only when all three repetitions were identical, or when at least two agreed, and what was recorded when repetitions differed. Additionally, the paper does not report the exact model versions or access dates for GPT-4o, Gemini-1.5-pro, and Claude 3 Opus, despite these being rapidly evolving deployed systems. For an LLM evaluation to be reproducible, the snapshot dates and API settings (e.g., temperature and other sampling parameters) must be reported. Please clarify the repetition protocol and provide model version and configuration details.
minor comments (5)
  1. [Throughout] There are multiple typos and inconsistencies: "Vega-lite" versus "Vega-Lite", "Anthropics's" versus "Anthropic's", and section headings such as "T esting" and "Data Set".
  2. [References] Reference [7] contains a malformed URL ("https://https://openai.com/index/gpt-4"), and several model references (GPT-4 vs. GPT-4o, Gemini-1.5-pro, Claude 3 Opus) lack version identifiers and access dates.
  3. [Tables 6–9] The percentages in Tables 6–9 would be much easier to interpret if the paper reported the exact number of questions per chart type and per condition, since the denominators are not stated in the tables.
  4. [§4.4] The manual questions include open-ended items such as "What are their colors?" and "Which color has the minimum value?"; the scoring of these questions needs an explicit policy for accepting synonyms or approximate color names, especially because the paper criticizes Claude for not using the given color names.
  5. [§4.3] The claim that binary questions allow models to achieve "approximately 50% accuracy by random guessing" is correct for balanced classes, but the paper should verify or state the class balance in the selected FigureQA subset.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the study is an empirical evaluation against external benchmarks and chart-type specifications, with no derived result reducing to its own inputs.

full rationale

The paper is an empirical evaluation, not a derivation chain. Chart-generation success is scored against externally specified chart types and the visual requirements stated in each prompt, with outputs executed and inspected. The visualization-understanding results are measured against the external FigureQA benchmark, which supplies fixed chart images, question templates, and ground-truth answers. No parameter is fitted to a subset of data and then renamed as a prediction, and no mathematical or statistical claim is derived from an equation that contains its own conclusion. The author-generated component is the manual non-binary question set in Sections 4.4 and 4.5, and the answers were checked by the authors against their own expected answers. This is a legitimate validity and reproducibility concern, but it is not circularity: the reported observations are measurements taken with that instrument, not a claim that the instrument's design logically forces the results. Similarly, the unsupported statement that the inter-LLM comparison 'will inherently compare the LLMs’ performances with a human baseline' is an incorrect methodological assertion, but it does not make any result equivalent to its inputs by construction. No load-bearing self-citation appears, and no prior work by these authors is invoked to justify a central premise. Accordingly, no specific circular step can be exhibited under the required standard.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper relies on the validity and representativeness of the FigureQA benchmark, on the reliability of the authors' visual inspection of generated charts, on the author-generated manual questions as a measure of understanding, and on the assumption that the small random samples are adequate. None of these are fitted parameters or invented entities; they are domain and methodological assumptions. The main risk is that the subjective and small-sample assumptions weaken the generality of the conclusions.

assumptions (5)
  • domain assumption FigureQA provides valid and reliable ground truth for visualization understanding.
    The automated evaluation uses FigureQA's synthesized QA pairs as the correctness standard; if the benchmark contains ambiguities or errors, the reported accuracies would be affected. The paper cites the dataset but does not validate its labels.
  • domain assumption The 100 randomly selected images and 1,342 questions are representative of visualization understanding tasks.
    The selection is random but small, with no seed reported, and the authors note API cost limited the sample size. The paper does not analyze how this sample compares with the full FigureQA distribution.
  • domain assumption Visual inspection by the authors is a reliable measure of whether generated code is correct.
    The generation success determinations are made by examining charts with no stated rubric, no independent raters, and no inter-rater agreement statistics.
  • ad hoc to paper The author-created manual questions and their answers validly measure chart understanding.
    The questions were designed and answered by the authors specifically for this paper, without validation against an external standard, so the correctness judgments are internal to the paper.
  • domain assumption LLM outputs are stable enough that three repetitions with agreement indicate reliable performance.
    The paper tunes parameters to reduce randomness and repeats each prompt three times, accepting output only if the same. If outputs vary beyond this control, the reported results may not be representative. The paper itself notes that responses differed 15% of the time for one counting question type.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating LLMs for Visualization Generation and Understanding." pith.science (2026). https://pith.science/paper/OQRFODS4

@misc{pith2026250722890,
  author       = {Pith},
  title        = {Pith review of: Evaluating LLMs for Visualization Generation and Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OQRFODS4}},
  note         = {Machine review of arXiv:2507.22890}
}
read the original abstract

Information Visualization has been utilized to gain insights from complex data. In recent times, Large Language models (LLMs) have performed very well in many tasks. In this paper, we showcase the capabilities of different popular LLMs to generate code for visualization based on simple prompts. We also analyze the power of LLMs to understand some common visualizations by answering questions. Our study shows that LLMs could generate code for some simpler visualizations such as bar and pie charts. Moreover, they could answer simple questions about visualizations. However, LLMs also have several limitations. For example, some of them had difficulty generating complex visualizations, such as violin plot. LLMs also made errors in answering some questions about visualizations, for example, identifying relationships between close boundaries and determining lengths of shapes. We believe that our insights can be used to improve both LLMs and Information Visualization systems.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 35 canonical work pages

  1. [1]

    https://www.tableau.com/

    Tableau. https://www.tableau.com/

    work page

  2. [2]

    https://matplotlib.org/

    Matplotlib. https://matplotlib.org/

    work page

  3. [3]

    https://vega.github.io/vega-lite/

    Vega-Lite. https://vega.github.io/vega-lite/

    work page

  4. [4]

    IEEE Transactions on Visualization and Computer Graphics 28(12), 5049–5070 (2022)

    Wu, A., Wang, Y., Shu, X., Moritz, D., Cui, W., Zhang, H., Zhang, D., Qu, H.: AI4VIS: Survey on Artificial Intelligence Approaches for Data Visualization. IEEE Transactions on Visualization and Computer Graphics 28(12), 5049–5070 (2022)

    work page 2022

  5. [5]

    : Language mod- els are few-shot learners

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., et al. : Language mod- els are few-shot learners. In: Proceedings of the 34th International Conference on Neural Information Processing Systems (NIPS), pp. 1877–1901 (2020)

    work page 2020

  6. [6]

    https: //arxiv.org/abs/2107.03374

    Chen, M., Tworek, J., Jun, H., Yuan, Q., Ponde, H., et al.: Evaluating Large Lan- guage Models Trained on Code (2021). https: //arxiv.org/abs/2107.03374

    arXiv 2021

  7. [7]

    https://https://openai.com/index/ gpt-4

    GPT-4. https://https://openai.com/index/ gpt-4

    work page

  8. [8]

    https://gemini.google.com

    Gemini. https://gemini.google.com

    work page

Show all 41 references
  1. [9]

    https://claude.ai

    Claude. https://claude.ai

  2. [10]

    ACM Computing Surveys 55(9), 1–35 (2023)

    Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, Prompt, and Pre- dict: A Systematic Survey of Prompting Methods in Natural Language Processing. ACM Computing Surveys 55(9), 1–35 (2023)

    2023

  3. [11]

    IEEE Transactions on Visualization & Computer Graphics 26(1), 895–905 (2020)

    Wang, Y., Sun, Z., Zhang, H., Cui, W., Xu, K., Ma, X., Zhang, D.: DataShot: Auto- matic Generation of Fact Sheets from Tabular Data. IEEE Transactions on Visualization & Computer Graphics 26(1), 895–905 (2020)

    2020

  4. [12]

    IEEE Computer Graphics and Applications 39(5), 33–46 (2019)

    Dibia, V., Demiralp, C.: Data2Vis: Auto- matic Generation of Data Visualizations Using Sequence-to-Sequence Recurrent Neu- ral Networks. IEEE Computer Graphics and Applications 39(5), 33–46 (2019)

    2019

  5. [13]

    IEEE Transactions on Visual- ization & Computer Graphics 30(1), 284–294 (2024)

    Xiao, S., Huang, S., Lin, Y., Ye, Y., Zeng, W.: Let the Chart Spark: Embedding Semantic Context into Chart with Text-to-Image Gen- erative Model. IEEE Transactions on Visual- ization & Computer Graphics 30(1), 284–294 (2024)

    2024

  6. [14]

    In: International ACM Conference on Management of Data (SIGMOD) (2018)

    Luo, Y., Qin, X., Tang, N., Li, G., Wang, X.: DeepEye: Creating Good Data Visualiza- tions by Keyword Search. In: International ACM Conference on Management of Data (SIGMOD) (2018)

    2018

  7. [15]

    http://opennlp.apache

    Apache OpenNLP. http://opennlp.apache. org

  8. [16]

    IEEE Transac- tions on Visualization and Computer Graph- ics 27(2) (2021)

    Narechania, A., Srinivasan, A., Stasko, J.T.: NL4DV: A Toolkit for Generating Analytic Specifications for Data Visualization from Natural Language Queries. IEEE Transac- tions on Visualization and Computer Graph- ics 27(2) (2021)

    2021

  9. [17]

    In: Proceedings of 52nd Annual Meeting of the Association for Com- putational Linguistics: System Demonstra- tions, pp

    Manning, C.D., Surdeanu, M., Bauer, J., Finkel, J.R., Bethard, S., McClosky, D.: The Stanford CoreNLP Natural Language Processing Toolkit. In: Proceedings of 52nd Annual Meeting of the Association for Com- putational Linguistics: System Demonstra- tions, pp. 55–60 (2014)

    2014

  10. [18]

    IEEE Transactions on 23 Visualization and Computer Graphics 29(6), 3121–3144 (2023)

    Shen, L., Shen, E., Luo, Y., Yang, X., Hu, X., Zhang, X., Tai, Z., Wang, J.: Towards Natural Language Interfaces for Data Visu- alization: A Survey. IEEE Transactions on 23 Visualization and Computer Graphics 29(6), 3121–3144 (2023)

    2023

  11. [19]

    In: IEEE Pacific Visualization Symposium (PacificVis) 2021 (2021)

    Liu, C., Han, Y., Jiang, R., Yuan, X.: ADVI- Sor: Automatic Visualization Answer for Natural-language Question on Tabular Data. In: IEEE Pacific Visualization Symposium (PacificVis) 2021 (2021)

    2021

  12. [20]

    IEEE Transactions on Visualization and Computer Graphics 28(1), 217–226 (2022)

    Luo, Y., Tang, N., Li, G., Tang, J., Chai, C., Qin, X.: Natural Language to Visualiza- tion by Neural Machine Translation. IEEE Transactions on Visualization and Computer Graphics 28(1), 217–226 (2022)

    2022

  13. [21]

    In: Proceedings of SIGMOD (2021)

    Luo, Y., Tang, N., Li, G., Chai, C., Li, W., Qin., X.: Synthesizing Natural Language to Visualization (NL2VIS) Benchmarks from NL2SQL Benchmarks. In: Proceedings of SIGMOD (2021)

    2021

  14. [22]

    IEEE Transactions on Visualization and Computer Graphics 31(1) (2025)

    Chen, N., Zhang, Y., Xu, J., Ren, K., Yang, Y.: VisEval: A Benchmark for Data Visual- ization in the Era of Large Language Mod- els. IEEE Transactions on Visualization and Computer Graphics 31(1) (2025)

    2025

  15. [23]

    343–352 (2024)

    V´ azquez, P.-P.: Are LLMs ready for Visual- ization? In: IEEE PacificVis 2024 Workshop - Vis Meets AI, pp. 343–352 (2024)

    2024

  16. [24]

    https://arxiv

    Li, G., Wang, X., Aodeng, G., Zheng, S., Zhang, Y., Ou, C., Wang, S., Liu, C.H.: Visu- alization Generation with Large Language Models: An Evaluation (2024). https://arxiv. org/abs/2401.11255

    2024

  17. [25]

    In: IEEE Vis 2024 (2024)

    Tian, Y., Cui, W., Deng, D., Yi, X., Yang, Y., Zhang, H., Wu, Y.: ChartGPT: Leverag- ing LLMs to Generate Charts from Abstract Natural Language. In: IEEE Vis 2024 (2024)

    2024

  18. [26]

    IEEE Access 11 (2023)

    Maddigan, P., Susnjak, T.: Chat2Vis: Gener- ating data visualizations via natural language using chatgpt, codex and GPT-3 large lan- guage models. IEEE Access 11 (2023)

    2023

  19. [27]

    In: Proceedings of the 61st Annual Meet- ing of the Association for Computational Lin- guistics (Volume 3: System Demonstrations), pp

    Dibia, V.: LIDA: A Tool for Automatic Gen- eration of Grammar Agnostic Visualizations and Infographics using Large Language Mod- els. In: Proceedings of the 61st Annual Meet- ing of the Association for Computational Lin- guistics (Volume 3: System Demonstrations), pp. 113–126 (2023)

    2023

  20. [28]

    https:// arxiv.org/abs/2311.16483

    Han, Y., Zhang, C., Chen, X., Yang, X., Wang, Z., Yu, G., Fu, B., Zhang, H.: ChartL- lama: A Multimodal LLM for Chart Under- standing and Generation (2023). https:// arxiv.org/abs/2311.16483

    2023 arXiv

  21. [29]

    In: Findings of the Association for Computational Lin- guistics: EMNLP 2023 (2023)

    Ye, J., Hu, A., Xu, H., Ye, Q., Yan, M., et al.: UReader: Universal OCR-free Visually- situated Language Understanding with Mul- timodal Large Language Model. In: Findings of the Association for Computational Lin- guistics: EMNLP 2023 (2023)

    2023

  22. [30]

    In: Findings of the Asso- ciation for Computational Linguistics: ACL 2024 (2024)

    Meng, F., Shao, W., Lu, Q., Gao, P., Zhang, K., Qiao, Y., Luo, P.: ChartAssisstant: A Uni- versal Chart Multimodal Language Model via Chart-to-Table Pre-training and Multitask Instruction Tuning. In: Findings of the Asso- ciation for Computational Linguistics: ACL 2024 (2024)

    2024

  23. [31]

    In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp

    Kanthara, S., Leong, R.T.K., Lin, X., Masry, A., Thakkar, M., Hoque, E., Joty, S.R.: Chart-to-Text: A Large-Scale Benchmark for Chart Summarization. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 4005–40...

    2022

  24. [32]

    In: Findings of the Asso- ciation for Computational Linguistics: ACL, Dublin, Ireland (2022)

    Masry, A., Do, X.L., Tan, J.Q., Joty, S., Hoque, E.: ChartQA: A Benchmark for Ques- tion Answering about Charts with Visual and Logical Reasoning. In: Findings of the Asso- ciation for Computational Linguistics: ACL, Dublin, Ireland (2022)

    2022

  25. [33]

    https://arxiv.org/abs/2308.03349

    Li, S., Tajbakhsh, N.: SciGraphQA: A Large-Scale Synthetic Multi-Turn Question- Answering Dataset for Scientific Graphs (2023). https://arxiv.org/abs/2308.03349

    2023 arXiv

  26. [34]

    In: Computer Vision and Pattern 24 Recognition (CVPR) (2024)

    Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., et al.: HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models. In: Computer Vision and Pattern 24 Recognition (CVPR) (2024)

    2024

  27. [35]

    https:// arxiv.org/abs/2312.15915

    Xu, Z., Du, S., Qi, Y., Xu, C., Yuan, C., Guo, J.: ChartBench: A Benchmark for Complex Visual Reasoning in Charts (2024). https:// arxiv.org/abs/2312.15915

    2024 arXiv

  28. [36]

    IEEE Transactions on Visualization and Computer Graphics 31(1), 1105–1115 (2025)

    Bendeck, A., Stasko, J.: An Empirical Eval- uation of the GPT-4 Multimodal Language Model on Visualization Literacy Tasks. IEEE Transactions on Visualization and Computer Graphics 31(1), 1105–1115 (2025)

    2025

  29. [37]

    IEEE Transactions on Visualization and Computer Graphics 31(1), 1116–1125 (2025)

    Lo, L.Y.-H., Qu, H.: How Good (Or Bad) Are LLMs in Detecting Misleading Visual- izations. IEEE Transactions on Visualization and Computer Graphics 31(1), 1116–1125 (2025)

    2025

  30. [38]

    In: IEEE Vis 2024 (2024)

    Choe, K., Lee, C., Lee, S., Song, J., Cho, A., Kim, N.W., Seo, J.: Enhancing Data Liter- acy On-demand: LLMs as Guides for Novices in Chart Interpretation. In: IEEE Vis 2024 (2024)

    2024

  31. [39]

    https://arxiv.org/ abs/1710.07300

    Kahou, S.E., Michalski, V., Atkinson, A., Kadar, A., Trischler, A., Bengio, Y.: Fig- ureQA: An Annotated Figure Dataset for Visual Reasoning (2018). https://arxiv.org/ abs/1710.07300

    2018 arXiv

  32. [40]

    ACM Computing Surveys 55(12), 1–38 (2023)

    Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y.J., Madotto, A., Fung, P.: Survey of Hallucination in Natural Lan- guage Generation. ACM Computing Surveys 55(12), 1–38 (2023)

    2023

  33. [41]

    Neural Information Processing Systems (2022) 25

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Chi, E.H.-h., Xia, F., Le, Q., Zhou, D.: Chain of Thought Prompting Elicits Rea- soning in Large Language Models. Neural Information Processing Systems (2022) 25

    2022

Pith tools

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