Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Text2Vis: A Challenging and Diverse Benchmark for Generating Multimodal Visualizations from Text

T0 review · 3 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 Text2Vis introduces a 1,985-sample text-to-visualization benchmark spanning 20+ chart types and hard data-science queries, and a cross-modal actor-critic loop raises GPT-4o's final pass rate from 26% to 42%.

desk verdict The benchmark is a solid contribution, but the headline agentic gain is likely test-label leakage because the critic prompt includes the gold answer. read the letter →

arxiv 2507.19969 v1 pith:TSEEERIP submitted 2025-07-26 cs.CL cs.CV

Mizanur Rahman , Md Tahmid Rahman Laskar , Shafiq Joty , Enamul Hoque This is my paper · ORCID
classification cs.CLcs.CV
keywords text-to-visualizationbenchmarkLLM-as-judgedatasciencequerieschartgenerationactor-criticrefinementmultimodalfeedbackGPT-4oevaluation
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 argues that existing text-to-visualization benchmarks are too easy to expose how language models actually handle realistic data-analysis questions, and it offers a harder replacement: Text2Vis, with 1,985 samples, over 20 chart types, and queries that require multi-step reasoning, prediction, outlier detection, conversational context, and sometimes web data retrieval. In direct prompting, all eleven tested models fall short, with GPT-4o passing only 26% of samples and the best open-source model at 13%. The paper proposes a cross-modal actor-critic inference loop in which a critic inspects the generated answer, the code, and the rendered chart and gives structured feedback, and one round of that refinement raises GPT-4o's pass rate to 42% while improving readability and chart correctness. It also introduces a fully automatic LLM-based judge for answer correctness, code execution, readability, and chart correctness, reporting that the judge agrees with human ratings on a 236-sample subset and can score all 1,985 samples in about five minutes for roughly two dollars.

What carries the argument

The central objects are the Text2Vis benchmark instance—a data table, natural-language query, short answer, executable Matplotlib/Seaborn code, and annotated chart—and the cross-modal actor-critic inference loop that refines those outputs. In that loop, an actor first emits a JSON pair of answer and visualization code; a critic then inspects the answer, the code, and the rendered chart image and returns structured feedback on answer correctness, code syntax and semantics, and visual clarity; the actor uses that feedback for exactly one refinement pass. The benchmark's hard queries (multi-step calculations, trend prediction, outlier detection, retrieval, conversation, and unanswerable cases) are what make the combined feedback necessary: removing any single feedback modality degrades the final pass rate, and code-only feedback reaches high execution rates but poor final pass rates.

What would settle it

Have independent human annotators grade the 1,985 agentic-refined outputs with the paper's own rubric; if the human pass rate no longer shows a 16-point gain over direct inference, the reported improvement is an artifact of using GPT-4o as both critic and judge.

Watch

Extended reading notes

Core claim

The paper's central claim is that text-to-visualization is currently under-benchmarked, and that a harder benchmark plus a refinement loop is enough to substantially improve state-of-the-art performance. On the paper's own setup, direct inference lets GPT-4o pass only 26% of 1,985 samples, while the best open-source model (Qwen2.5-7B) passes 13%. The paper further claims that a cross-modal actor-critic loop, where a critic examines the generated answer, the code, and the rendered chart and the actor revises once, raises GPT-4o's pass rate to 42% while improving readability and chart correctness. It also asserts that its automatic GPT-4o judge, checked against human labels on 236 samples, is reliable enough to replace manual annotation for large-scale benchmarking.

Load-bearing premise

The headline gain rests on trusting GPT-4o to grade its own agentic refinements, since human validation covered only 236 direct-inference outputs and no human check was run on the refined outputs.

Editorial extensions

If this is right

  • Text2Vis gives the field a harder, more realistic test: GPT-4o passes only 26% of samples directly, and the best open-source model passes 13%, so there is clear room for improvement in both closed and open models.
  • The agentic loop is the main lever: with one round of answer-plus-code feedback, GPT-4o's answer match rises from 42% to 53%, readability from 3.45 to 3.99, chart correctness from 3.15 to 4.02, and final pass rate from 26% to 42%, a change reported as statistically significant (McNemar's test, p<0.01).
  • Each feedback modality contributes: removing answer, code, or visual feedback individually degrades performance, and code execution success alone is not enough, since code-only feedback reaches 94% execution but only a 32% final pass rate.
  • The automatic LLM-based evaluation protocol, validated against human judgments on 236 samples with a Cohen's kappa of 0.78, makes it feasible to benchmark large model suites cheaply and repeatedly.
  • Models struggle most on retrieval-based and multi-chart queries and on unanswerable queries, identifying those as the next targets for visualization-capable LLMs.

Reading between the lines

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

  • If GPT-4o's dual role as critic and judge creates any preference for its own corrected outputs, the true gain of the agentic loop could be smaller than the reported 16 percentage points; a human evaluation of the 1,985 agentic-refined outputs would settle this.
  • The 11% unanswerable-query subset could be used as a targeted hallucination probe, since models scored near zero pass rate there, and improving refusal behavior may be the cheapest path to higher overall pass rates.
  • Because open-ended queries allow multiple valid visualizations, pass-rate comparisons on the closed-ended subset alone may be a more stable basis for ranking models.
  • The framework's 42% pass rate still means most hard queries fail, and the bottleneck is answer correctness rather than code execution, suggesting that coupling the refinement loop with retrieval tools or a calculator could yield larger gains than further code tuning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces Text2Vis, a benchmark of 1,985 text-to-visualization samples spanning 20+ chart types, with each sample containing a data table, a natural language query, a short answer, visualization code, and an annotated chart. The authors benchmark 11 open- and closed-source models under direct inference, report GPT-4o as the strongest model with a 26% final pass rate, and propose a cross-modal actor-critic agentic framework that jointly refines textual answers and visualization code, claiming it raises GPT-4o's pass rate to 42% (Table 4B). They also introduce a fully automated GPT-4o-based evaluation framework for answer match, code execution, readability, and chart correctness, and validate it against human annotations on 236 samples (Section 5.2, Table 5). The central contributions are the benchmark itself, the agentic framework, and the scalable LLM-based evaluation suite.

Significance. If the benchmark data are released cleanly and the evaluation protocol is sound, Text2Vis is a useful addition to text-to-visualization evaluation: it covers a broader set of chart types, query types (multi-turn, unanswerable, retrieval-augmented, multi-chart), and data science reasoning tasks than prior benchmarks such as nvBench and VisEval. A reproducible, code-based benchmark with human-validated automated scoring at roughly two dollars for 1,985 samples is a practical contribution, and the reported human-judge correlations (Pearson r around 0.85-0.92 on direct inference) support the automatic judge for that setting. However, the headline agentic gain is not yet credible because the published agentic prompt appears to give the critic the ground-truth answer, and no human evaluation covers the agentic-refined outputs. The benchmark contribution is significant, but the framework and evaluation claims need stronger validation.

major comments (3)
  1. [A.8.2, Table 17] The agentic-refinement prompt printed in Table 17 explicitly supplies the critic with the 'Ground truth answer' and instructs it to 'validate and refine' the initial model response. This means the critic, and through its feedback the actor, has direct access to the gold label during the agentic loop. The example in Figure 3 is consistent with this reading: the critic tells the actor the exact values (2009, 50023.8). If the prompt was used as written, the answer-match improvements in Table 4(B) (42% to 53%) and the final-pass-rate improvement (26% to 42%) are label-leakage artifacts rather than evidence for the framework. The main text in Section 4.3 does not state that ground truth was withheld from the critic, and the appendix prompt suggests it was not. The authors must either confirm that the actual inference-time prompt did not include the ground-truth answer and provide that prompt verbatim, or re-run the agentic experiments with the gold answer withheld. This is load-bearing for the paper's central claim.
  2. [Section 5.2, Tables 4 and 5] The human evaluation validates the automated judge only on direct-inference outputs from three models (GPT-4o, Llama-3.1-8B, Qwen2.5-7B). No human evaluation is reported for the agentic-refined outputs in Table 4(B). Because GPT-4o serves as the actor, the critic in the main configuration, and the judge in Table 4(B), the demonstrated high agreement between human judgments and GPT-4o on direct inference does not automatically transfer to agentic-refined outputs. Even without label leakage, the reported agentic gain could reflect a judge-critic preference bias. A human validation of at least a stratified sample of agentic outputs is needed before the framework improvement can be claimed.
  3. [Section 4.4, Table 18] The answer-match rubric is underspecified: the rule 'Match if numbers are close (e.g., 48.77 vs 48.73)' does not define a tolerance, and the examples in parentheses suggest exact decimal equality or very small differences, but the boundary is left to the judge model. This is particularly important because the same GPT-4o model generates answers and evaluates them. Additionally, the pass-rate thresholds (readability and chart correctness scores of at least 3.5) are arbitrary and no sensitivity analysis is reported. Please operationalize the numeric-similarity rule (e.g., relative or absolute tolerance, rounding convention) and report how pass rates and model rankings shift under reasonable threshold variations.
minor comments (5)
  1. [Section 4.3 vs. A.8.2] The main text describes the critic as analyzing 'the initial response' without mentioning ground-truth input, while the appendix prompt in Table 17 lists 'Ground truth answer' as an input. This inconsistency must be resolved in revision, and the authors should explicitly state whether the gold answer was available to the critic at inference time.
  2. [Figure 5] In the current PDF, Figure 5 appears as a series of escaped Unicode tokens rather than a readable plot; the displayed text is illegible and should be replaced with a proper rendered figure.
  3. [Abstract] The abstract contains 'GPT-4o`s' with a backtick instead of an apostrophe; please correct the punctuation throughout.
  4. [Table 2] The row for question complexity lists '343 245 1173 224' with no column headers in the rendered table; add labels for Easy, Medium, Hard, and Extra Hard to match the text.
  5. [Section 3.1/3.2] The text reports 2001 curated tables, 173 synthetic tables, and 239 removed table-query pairs, but the final split of 1,935 curated and 50 web-retrieval samples is not fully reconciled; please clarify the construction arithmetic.

Circularity Check

1 steps flagged · score 6.0 of 10

The agentic critic is given the ground-truth answer (Table 17), so part of the reported 26% to 42% pass-rate gain is by construction; the Text2Vis benchmark itself remains independently grounded.

  1. self definitional [Appendix A.8.2, Table 17 (Agentic Framework prompt); used in Section 4.3 and reported in Table 4]
    "Given a structured data table, Ground truth answer, a user-generated question, and an initial model response, your task is to validate and refine the model output for accuracy, correctness, and completeness."

    The critic's refinement input includes the ground-truth answer. Since the critic's task is to validate the answer and refine the output, the actor's revised 'Answer' can simply be the gold value; Figure 3 shows the critic supplying 'The highest deviation is 50023.8 million euros...in 2009' and the final answer becoming '2009, 50023.8'. Table 4's pass rate requires Answer Match, and Table 18's judge compares the generated answer with the ground truth. Thus the Answer Match gain (42% to 53%) and much of the pass-rate gain (26% to 42%) are forced by label leakage, not by the framework. Section 5.2's human validation covers only direct-inference outputs, so it cannot rule this out. As written, the answer-refinement claim is circular.

full rationale

The Text2Vis benchmark construction and the direct-inference evaluations are not circular: queries and answers were manually verified, and Table 5 (Section 5.2) provides human validation of the automated judge on 236 direct-inference samples with high correlation. The circularity is confined to the agentic inference claim. The Table 17 prompt explicitly provides 'Ground truth answer' to the critic, whose job is to validate and refine the initial model response. With the gold answer in the prompt, the actor's refined answer trivially matches the ground truth, and the same GPT-4o-based evaluator (Table 18) then scores it as a match. This affects the headline agentic contribution (Table 4), including the claimed improvement in answer match and final pass rate. The code-execution, readability, and chart-correctness gains may still have independent content, but the answer-match portion of the claimed improvement reduces to feeding the label into the refiner. If the actual deployed prompt withheld the ground truth despite this template, the authors need to state and show that; the manuscript does not. This is a partial, but concrete, circularity in the paper's strongest contribution.

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

The benchmark uses no fitted physical parameters or invented entities. The load-bearing assumptions are about data provenance (ChartQA tables), annotation reliability (LLM-generated queries and answers manually reviewed without quantified agreement), and the validity of GPT-4o as a judge, especially for agentic outputs that were not human-validated. Free parameters include the 3.5 pass threshold and the unspecified numeric-tolerance rule for answer matching.

free parameters (2)
  • Pass quality threshold = readability >= 3.5 and chart correctness >= 3.5
    Used to compute final pass rate; chosen by authors with no sensitivity analysis reported.
  • Answer-match closeness tolerance = unspecified (e.g., 48.77 vs 48.73 accepted)
    Vague matching rule delegated to GPT-4o; affects answer match rates and pass rates.
assumptions (4)
  • domain assumption ChartQA-derived tables are representative of real-world visualization tasks
    Section 3.1 selects 2,001 tables from ChartQA sources (Statista, Pew, OWID, OECD) as the base; no independent representativeness analysis is provided.
  • domain assumption LLM-generated queries, answers, and code are correct after manual review
    Section 3.2 says outputs were 'manually reviewed, corrected, and refined', but no inter-annotator agreement or error rate is reported.
  • domain assumption GPT-4o can accurately judge answer correctness, readability, and chart correctness
    Section 4.4 and Section 5.2 rely on GPT-4o as judge with human agreement on 236 samples; this is a general assumption about LLM-as-judge validity.
  • ad hoc to paper GPT-4o judge agreement with humans on direct inference transfers to agentic-refined outputs
    Section 5.2 human evaluation covers direct inference only; agentic pass rates in Table 4 are not human-validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Text2Vis: A Challenging and Diverse Benchmark for Generating Multimodal Visualizations from Text." pith.science (2026). https://pith.science/paper/TSEEERIP

@misc{pith2026250719969,
  author       = {Pith},
  title        = {Pith review of: Text2Vis: A Challenging and Diverse Benchmark for Generating Multimodal Visualizations from Text},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TSEEERIP}},
  note         = {Machine review of arXiv:2507.19969}
}
read the original abstract

Automated data visualization plays a crucial role in simplifying data interpretation, enhancing decision-making, and improving efficiency. While large language models (LLMs) have shown promise in generating visualizations from natural language, the absence of comprehensive benchmarks limits the rigorous evaluation of their capabilities. We introduce Text2Vis, a benchmark designed to assess text-to-visualization models, covering 20+ chart types and diverse data science queries, including trend analysis, correlation, outlier detection, and predictive analytics. It comprises 1,985 samples, each with a data table, natural language query, short answer, visualization code, and annotated charts. The queries involve complex reasoning, conversational turns, and dynamic data retrieval. We benchmark 11 open-source and closed-source models, revealing significant performance gaps, highlighting key challenges, and offering insights for future advancements. To close this gap, we propose the first cross-modal actor-critic agentic framework that jointly refines the textual answer and visualization code, increasing GPT-4o`s pass rate from 26% to 42% over the direct approach and improving chart quality. We also introduce an automated LLM-based evaluation framework that enables scalable assessment across thousands of samples without human annotation, measuring answer correctness, code execution success, visualization readability, and chart accuracy. We release Text2Vis at https://github.com/vis-nlp/Text2Vis.

Figures

Figures reproduced from arXiv: 2507.19969 by the authors.

Figure 1
Figure 1. Example from the Text2Vis benchmark. Input: A data table containing historical stock prices and a query. Output: Python code for visualization, the predicted answer, and an annotated textual explanation. The chart is generated from the code. used for exploratory data analysis, outlier detec￾tion, pattern recognition, and feature identification. However, creating accurate and intuitive visualiza￾tions is challenging … view at source ↗
Figure 2
Figure 2. Examples of different question types used in data analysis, including trend prediction, reasoning, outlier [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Our Agentic Inference Framework where the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Common chart types in our Text2Vis. A Appendices A.1 Common Chart Types Text2Vis includes a wide range of chart types that reflect the diversity of real-world data analysis tasks 4. Line charts are essential for visualizing trends over time, making them ideal for time-…
Figure 5
Figure 5. Figure 5: Error type distribution with square root trans [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Common errors in Data Visualization generation. [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Most Frequent Words in Error Messages Across All Evaluated Models. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages

  1. [1]

    - Clearly specify the selected domain in the output

    Generate a Synthetic Data Table: - Randomly choose one domain from: Healthcare, Technology, Finance, Marketing, Retail, Education, Sports, Energy, Logistics, or any other relevant field. - Clearly specify the selected domain in the output. - Ensure domain-specific numerical scales (e.g., revenue in 100k+, sales in thousands, ratings 1.0-5.0, percentages 0...

    work page

  2. [2]

    arXiv preprint arXiv:2411.15594

    A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shi- rong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. Yucheng Han, Chi Zhang, Xin Chen, Xu Y...

    arXiv 2025

  3. [3]

    Provide a Short Answer: - The answer must be exactly one word or one number

    work page

  4. [4]

    arXiv preprint arXiv:2408.05346

    Datanarrative: Automated data-driven story- telling with visualizations and texts. arXiv preprint arXiv:2408.05346. Albert Q Jiang, Alexandre Sablayrolles, Arthur Men- sch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guil- laume Lample, Lucile Saulnier, and 1 others. 2023. Mistral 7b. arXiv preprint arXiv:2...

    arXiv 2023

  5. [5]

    - Highlight the main insight derived from the visualization

    Include a Text Summary: - Provide a concise summary. - Highlight the main insight derived from the visualization

    work page

  6. [6]

    arXiv preprint arXiv:2402.02167

    Vi (e) va llm! a conceptual stack for evaluating and interpreting generative ai-based visualizations. arXiv preprint arXiv:2402.02167. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas- try, Amanda Askell, Pamela Mishkin, Jack Clark, and 1 others. 2021. Learning transferable visual models from natural lan...

    arXiv 2021

  7. [10]

    - The question should involve detecting patterns, finding skewness, assessing data spread, or identifying peaks and outliers

    Generate a Single, Very Complex Data Science Question: - The question must require multi-step reasoning and deep analysis related to data distributions, frequency analysis, or variability. - The question should involve detecting patterns, finding skewness, assessing data spread, or identifying peaks and outliers. - Ensure the question requires multi-step ...

    work page

  8. [12]

    - Data should be binned appropriately to represent the distribution

    Output Python Code Using Matplotlib: - The code should create a Histogram that effectively visualizes the dataset and addresses the generated question. - Data should be binned appropriately to represent the distribution. - Ensure clear labeling of axes and meaningful annotations to highlight key insights. - Use colors, bin adjustments, and density plots i...

    work page

Show all 29 references
  1. [14]

    Histogram

    Provide Metadata: - Domain: The selected domain. - ChartType: Must be "Histogram". - xlabel: The numerical variable representing the bins. - ylabel: The frequency count of values. Output Requirements: - Return all the above information in a valid JSON format without any additi...

  2. [15]

    • Design the question specifically for a scatter plot visualization

    Generate a Single, Very Complex Data Science Question: • The question must require multi-step reasoning and deep analysis. • Design the question specifically for a scatter plot visualization. For example, it may ask to analyze the relationship, correlation, or pattern between ...

  3. [16]

    Provide a Short Answer: • The answer must be precise

  4. [17]

    • Ensure the code annotates key insights on the plot

    Output Python Code for a Scatter Plot Visualization: • Use matplotlib to generate a scatter plot. • Ensure the code annotates key insights on the plot

  5. [18]

    Include a Text Summary: • Provide a concise explanation of the reasoning behind the answer, highlighting the main insight derived from the scatter plot

  6. [19]

    Scatter". • xlabel: The variable used for the X-axis. • ylabel: The variable used for the Y-axis (if not applicable, use

    Provide Metadata: • ChartType: Set this to "Scatter". • xlabel: The variable used for the X-axis. • ylabel: The variable used for the Y-axis (if not applicable, use "N/A"). To ensure clarity, two examples with scatterplot are included to demonstrate the expected structure. Ple...

  7. [20]

    unanswerable

    Answer: Provide a precise and concise response based on the data. If no clear answer is available, return "unanswerable"

  8. [21]

    Answer":

    Visualization Code: Generate Python Matplotlib code to create a meaningful visualization that accurately represents the data. Ensure annotations and highlights are included. Important Requirement: • The output must be in a valid JSON format without any extra text, markdown for...

  9. [22]

    Answer Validation: Verify correctness and identify errors if any

  10. [23]

    Visualization Code Validation: Check for syntax errors, readability issues, or execution problems

  11. [24]

    Answer":

    Refinement Task: • Based on the feedback, refine the model response to correct errors. • Ensure the response is precise, formatted correctly, and adheres to the required JSON format. Output Requirements: • Ensure the final output is in a valid JSON format without extra text or...

  12. [25]

    Answer Matching: Compare the generated answer with the ground truth using following evaluation criteria

  13. [26]

    Evaluation Criteria:

    Visualization Evaluation: Score the visualization based on following evaluation criteria. Evaluation Criteria:

  14. [27]

    48.77" vs

    Answer Matching (Binary: 1 or 0) • Match if numbers are close (e.g., "48.77" vs "48.73") or equivalent percentage formats (e.g., "100" vs "100 • Match if the ground truth appears within the generated response (e.g., "100" in "The result is 100"). • For long ground truth answer...

  15. [28]

    Readability and Quality Score (0-5) • Labels and Titles: Are they clear, concise, and correctly positioned? • Layout Spacing: Is the layout well-organized with no clutter? • Color Accessibility: Are colors distinct and accessible (colorblind-friendly)? • Axis Scaling: Are axes...

  16. [29]

    Answer Match

    Chart Correctness Score (0-5) • Query Alignment: Does the visualization correctly address the question? • Data Integrity: Are all data points accurately plotted? • Insight Representation: Does the chart effectively communicate its key insights based on its type? • Handling Mis...

  17. [2017]

    employ- ment trends

    Seq2sql: Generating structured queries from natural language using reinforcement learning. arXiv preprint arXiv:1709.00103. Line 19.8% Bar 17.9% Multichart 10.1% Area 6.9% Pie 5.7% Waterfall 5.5% Scatter 4.7% Treemap 4.7% Histogram 4.6% Boxplot 4.5% Dot 4.1% Donut 3.5% 22 Othe...

    arXiv

  18. [2020]

    IEEE Transactions on Visualization and Computer Graphics, 27(2):369–379

    Nl4dv: A toolkit for generating analytic speci- fications for data visualization from natural language queries. IEEE Transactions on Visualization and Computer Graphics, 27(2):369–379. Mohamed Nejjar, Luca Zacharias, Fabian Stiehle, and Ingo Weber. 2025. Llms for science: Usag...

    2025 arXiv

  19. [2022]

    In Computer Graphics Forum, volume 41, pages 555–572

    Chart question answering: State of the art and future directions. In Computer Graphics Forum, volume 41, pages 555–572. Wiley Online Library. Mohammed Saidul Islam, Md Tahmid Rahman Laskar, Md Rizwan Parvez, Enamul Hoque, and Shafiq Joty

  20. [2023]

    Advances in Neural Information Pro- cessing Systems, 36:46595–46623

    Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Pro- cessing Systems, 36:46595–46623. Victor Zhong, Caiming Xiong, and Richard Socher

  21. [2024]

    CoRR, abs/2406.11931

    Deepseek-coder-v2: Breaking the barrier of closed-source models in code intelligence. CoRR, abs/2406.11931. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 oth...

    2024 arXiv

Pith tools

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