Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

ChartGen: Scaling Chart Understanding Via Code-Guided Synthetic Chart Generation

T0 review · 4 major / 6 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 The paper claims that a fully automated two-stage pipeline generates a 222,500-pair synthetic chart-to-code dataset from 13,000 seed images, and that a held-out benchmark built from it shows six open-weight VLMs scoring at most 0.58/1 on…

desk verdict A genuinely useful chart-to-code resource, but the quality filter only checks executability and duplicates, so the benchmark numbers rest on unverified semantic validity. read the letter →

arxiv 2507.19492 v1 pith:CZM3QNOJ submitted 2025-05-31 cs.HC cs.AIcs.CV

Jovana Kondic , Pengyuan Li , Dhiraj Joshi , Zexue He , Shafiq Abedin , Jennifer Sun , Ben Wiesel , Eli Schwartz
show 7 more authors
Ahmed Nassar Bo Wu Assaf Arbelle Aude Oliva Dan Gutfreund Leonid Karlinsky Rogerio Feris
This is my paper
classification cs.HCcs.AIcs.CV
keywords chart-to-codereconstructionsyntheticchartgenerationvision-languagemodelscodeunderstandingredrawingmultimodaldataset
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

The paper tries to establish that chart-to-code reconstruction—turning a chart image back into the Python script that made it—can be studied at scale through fully synthetic data, and that current open-weight vision-language models are far from solving it. It introduces ChartGen, a two-stage pipeline that first asks a VLM to redraw real chart images as plotting code, then asks a code LLM to iteratively mutate that code into new chart types, libraries, and styles. Applied to 13,000 seed charts from ChartQA, the pipeline yields 222,500 unique image-code pairs across 27 chart types and 11 plotting libraries, along with CSVs, summaries, DocTags, and QA pairs. On a curated 4,300-pair held-out subset, the best of six open-weight VLMs reaches only 0.58 out of 1 on data fidelity and 7.48 out of 10 on rendered-image similarity, which the paper reads as evidence that chart redrawing remains unsolved.

What carries the argument

The load-bearing mechanism is the two-stage code-guided generation loop: (1) a VLM (phi-3.5-vision-instruct) is prompted to reconstruct each seed chart image into a Python plotting script; (2) a code-focused LLM (Codestral-22B) iteratively rewrites that script roughly twenty times under prompts that force new chart types, plotting libraries, data distributions, topics, and styles, and each rewritten script is executed to render a new chart image. Sample validity is enforced by two filters—executable scripts and de-duplication via identical image embeddings—so the code itself is the carrier of chart semantics, and the rendered images are the visible output of that code.

What would settle it

Take a random sample of roughly 200 ChartGen-200K images and have independent annotators read the plotted values directly from the pixels; if a large share of images are unreadable or the values disagree with the paired CSV, executability is not a sufficient validity filter and the 4.3K evaluation set cannot support the paper's conclusion that current VLMs are the bottleneck.

Watch

Extended reading notes

Core claim

The central claim is that a purely automated, code-guided pipeline can turn a modest collection of real-world chart images into a chart-understanding corpus large and diverse enough to train and evaluate modern vision-language models, and that using that corpus to benchmark six open-weight VLMs exposes a large gap between syntactic code generation and faithful chart reconstruction. The paper presents ChartGen-200K, a 222.5K-sample dataset generated from 13K ChartQA seeds, covering 27 chart types and 11 plotting libraries, and a held-out 4.3K chart redrawing benchmark. On this benchmark, the strongest model attains 0.58/1 data fidelity, 5.83/10 code-semantic similarity, and 7.48/10 image similarity, despite execution rates mostly above 65%, so the paper concludes that recovering executable, semantically faithful plotting code from chart pixels remains an open problem.

Load-bearing premise

The pipeline's validity filter assumes that any Python script that executes without error and renders an image not already in the corpus is a legitimate chart; it never checks whether the augmented charts have readable axes, truthful data relationships, or meaningful semantics, so if many are meaningless the dataset and the benchmark built on it lose force.

Editorial extensions

If this is right

  • The ChartGen pipeline converts 13,000 seed chart images into 222,500 unique chart image-code pairs spanning 27 chart types and 11 plotting libraries.
  • Each pair ships with aligned CSV, DocTags, text summary, and QA pairs, so the corpus supports tasks beyond chart-to-code, such as chart summarization and question answering.
  • On the held-out 4.3K chart redrawing benchmark, the best of six open-weight VLMs reaches 0.58/1 data fidelity and 7.48/10 image similarity, so faithful chart-to-code reconstruction remains largely unsolved.
  • The pipeline can start from any seed chart collection, including charts whose original source code or plotting library is unknown, because the first stage reconstructs code from pixels alone.

Reading between the lines

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

  • The same code-guided loop could be pointed at other image domains with a canonical code representation—UI screenshots to HTML, diagrams to TikZ, or maps to GIS scripts—to build paired synthetic corpora without hand annotation; ChartGen itself demonstrates only charts.
  • Because the evaluation uses GPT-4o as the judge, the reported scores could be recalibrated against human agreement; if human and GPT-4o rankings diverge, the relative ordering of the six VLMs may change.
  • The dataset's QA and summary modalities, all derived from the same plotting code, make it possible to train one model jointly on chart-to-code, chart-to-text, and chart QA; the paper does not test whether joint training improves chart grounding.
  • The 7.48/10 image-similarity ceiling suggests the remaining gap is partly visual style rather than data semantics; a per-chart-type breakdown across the 27 types would show whether heatmaps, 3D plots, or less common libraries account for most of the failure.
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 / 6 minor

Summary. The paper introduces ChartGen, a fully automated two-stage pipeline for synthetic chart generation: a VLM (phi-3.5-vision-instruct) first reconstructs 13K seed chart images into Python plotting scripts, and a code-specialized LLM (Codestral-22B) iteratively augments these scripts to produce a large corpus of chart image-code pairs. The authors report a dataset of 222.5K (elsewhere stated as 200K) unique chart image-code pairs spanning 27 chart types and 11 plotting libraries, augmented with CSV tables, DocTags, summaries, and QA pairs. A held-out evaluation subset of 4.3K pairs is used to benchmark six open-weight VLMs (3B–26B) using GPT-4o as an automated judge for code data fidelity, code semantics, and rendered-image similarity; the best model reaches 0.58/1 data fidelity and 7.48/10 image similarity, supporting the claim that chart-to-code reconstruction remains unsolved for open-weight models.

Significance. If the synthetic corpus is semantically sound, ChartGen provides a valuable public resource: the scale (200K+ pairs), the breadth of plotting libraries (11) and chart types (27), and the detailed release of pipeline prompts, model configurations, and compute costs are concrete strengths. The proposed evaluation protocol, which separates data fidelity from semantic/style consistency and also compares rendered images, is a useful contribution to chart-to-code benchmarking. The main significance, however, is conditional on whether the automatically generated charts are actually meaningful visualizations, which the paper does not yet verify.

major comments (4)
  1. [§3.1] The dataset-quality filter retains only scripts that execute without error and whose rendered images are not near-duplicates by embedding. This is load-bearing for the central claim that ChartGen-200K is a usable chart-to-code resource and that the 4.3K benchmark measures chart reconstruction. No check verifies that an augmented chart has readable axes, that the plotted data values match the code, or that the intended chart type was actually produced. Since the manuscript asserts in §3.1 that this filtering ensures 'high functional consistency and dataset reliability,' the authors should either add a semantic validation step (e.g., a human audit of a random sample, or automated checks that axis labels, tick values, and data extents are consistent) or explicitly report the failure rates of such checks. Without this, the diversity statistics in §3.3 and the GPT-4o-based scores in Table 2 may be computed on a corrupted resource.
  2. [Abstract / §3.3 / Contribution 2] The dataset size is stated inconsistently: the abstract, §3.1, and §3.4 say 222.5K pairs, while Contribution 2, Table 1, and §3.3 say 200K pairs. The paper also refers to 'over 200 K' in §2.2. Since the scale of the dataset is one of the three headline contributions, these numbers must be reconciled, and the exact number of pairs after filtering should be reported with a breakdown of how many seeds survived each stage.
  3. [§4.2 / Table 2] The evaluation relies entirely on GPT-4o as an automated judge for both code comparison and image comparison, with no human validation of the judge's scores. The reported numbers (e.g., 0.58/1 data fidelity, 7.48/10 image similarity) are only meaningful if GPT-4o's ratings correlate with human judgments on this task. The authors should report a human-annotated subset (e.g., 50–100 samples scored by 2–3 raters) and compute inter-annotator agreement and GPT-4o–human correlation, or at minimum provide qualitative examples of GPT-4o's scoring explanations to demonstrate calibration.
  4. [§2.2] The text claims that transforming the code 'preserves the chart's complete semantics,' but the augmentation prompt explicitly instructs the LLM to choose a new chart type and modify data distributions, topics, and styles. Changing chart type and data distribution necessarily alters semantic content. This claim is internally inconsistent and should be revised to state precisely what is preserved (e.g., the underlying data-generation logic or the code template) rather than 'complete semantics.' This is relevant because the later evaluation treats the reference code and image as ground truth for reconstruction.
minor comments (6)
  1. [Table 1] The table caption cites 'ChartQA [Wei et al., 2024]', but the ChartQA dataset is originally from Masry et al. (2022), which is also in the reference list; the citation should be corrected.
  2. [§2.2 and Figure 2] The augmentation count is given as '∼ 20 iterations' in prose, but Figure 2 shows 19 augmentation steps; please specify the exact number of iterations used and whether all seeds received the same number.
  3. [Appendix A.1 and A.2] Please clarify the decoding configuration: §A.1 says temperature τ = 0.0 for reconstruction and 'sampling disabled' for augmentation, while §A.2 says summaries use 'mildly stochastic decoding' with τ = 0.2. Explain how temperature is applied when sampling is disabled, and report any other sampling parameters (top-p, etc.).
  4. [Figure 3] The figure shows 'Chart types (19)' in one place and 'Chart types (27)' elsewhere; clarify which count refers to the held-out evaluation set and which to the full corpus.
  5. [§3.3] The diversity analysis uses color-entropy and CLIP embedding spread. The reported color entropy change (2.03 to 2.06) is very small and may not be a meaningful measure of visual diversity; please discuss the limitations of this metric or report additional diversity statistics (e.g., number of unique colors, layout variation).
  6. [§1] The phrase '13 K unique chart images from ChartQA' should specify how uniqueness was defined (e.g., exact duplicate removal, near-duplicate clustering), since ChartQA contains repeated or similar charts across questions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ChartGen's scale and benchmark claims are built from external seed data and an external GPT-4o judge; self-citations are limited to auxiliary tooling and are not load-bearing.

full rationale

ChartGen contains no derivation chain whose outputs are equivalent to its inputs. The pipeline claim (13K seed images to 222.5K pairs) is a production count: Stage 1 uses phi-3.5-vision to reconstruct each ChartQA seed image into one script (Sec. 2.1), Stage 2 augments each script roughly 20 times with Codestral-22B (Sec. 2.2), and Sec. 3.1 filters by code executability and duplicate-image embeddings. The benchmark (Sec. 4) draws 4.3K pairs from the corpus and scores six open-weight VLMs with GPT-4o; no parameter is fitted to the evaluation outcome, and the reference code is the exact code that rendered each held-out image, so the chart-to-code task is well-posed. The closest candidate for circularity is that the evaluation set is drawn from the same synthetic distribution the paper produces; that is an external-validity limitation, not a logical reduction, because the pairs are generated before evaluation and the reference is the code that produced the image. Self-citations appear only as auxiliary tools or evaluated models: DocTags (Nassar et al. 2025) and Docling (Livathinos et al. 2025) are used to convert CSV to DocTags metadata, and Granite Vision Team (2025) is one of the evaluated models. None of these carry the central scale, diversity, or performance conclusions. The paper also explicitly flags the main validity risk in Sec. 8: 'reliance on synthetic data carries risks of propagating inaccuracies if not adequately validated,' and Sec. 3.1's filter checks only executability and near-duplicate images, not semantic correctness or legibility. That is a data-quality caveat, not a circular step. Overall, no circular step can be exhibited with a specific reduction; score 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on assumed reliability of the two generation models, the external judge, and the quality filter. No new theoretical entities are introduced; DocTags is inherited from prior work (Nassar et al. 2025).

free parameters (1)
  • augmentation iterations = ~20
    Number of sequential LLM augmentation passes per seed script; chosen by hand and directly controls the final dataset scale (222.5K pairs).
assumptions (4)
  • domain assumption phi-3.5-vision-instruct can reconstruct sufficient chart content from real ChartQA images into executable Python code
    Stage 1 of the pipeline depends on this for the initial chart-to-code conversion.
  • domain assumption Codestral-22B augmentations preserve chart semantics while diversifying style and chart type
    Stage 2 depends on this to produce valid chart-code pairs; the paper does not semantically validate the augmented charts.
  • domain assumption GPT-4o judgment aligns with human judgment for data fidelity, code similarity, and image similarity
    The evaluation protocol in Section 4.2 uses GPT-4o as the sole judge without human agreement analysis.
  • ad hoc to paper Executability and image-deduplication are sufficient quality filters
    Section 3.1 uses these two filters to guarantee dataset quality, but neither verifies chart semantic correctness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ChartGen: Scaling Chart Understanding Via Code-Guided Synthetic Chart Generation." pith.science (2026). https://pith.science/paper/CZM3QNOJ

@misc{pith2026250719492,
  author       = {Pith},
  title        = {Pith review of: ChartGen: Scaling Chart Understanding Via Code-Guided Synthetic Chart Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CZM3QNOJ}},
  note         = {Machine review of arXiv:2507.19492}
}
read the original abstract

Chart-to-code reconstruction -- the task of recovering executable plotting scripts from chart images -- provides important insights into a model's ability to ground data visualizations in precise, machine-readable form. Yet many existing multimodal benchmarks largely focus primarily on answering questions about charts or summarizing them. To bridge this gap, we present ChartGen, a fully-automated pipeline for code-guided synthetic chart generation. Starting from seed chart images, ChartGen (i) prompts a vision-language model (VLM) to reconstruct each image into a python script, and (ii) iteratively augments that script with a code-oriented large language model (LLM). Using ChartGen, we create 222.5K unique chart-image code pairs from 13K seed chart images, and present an open-source synthetic chart dataset covering 27 chart types, 11 plotting libraries, and multiple data modalities (image, code, text, CSV, DocTags). From this corpus, we curate a held-out chart-to-code evaluation subset of 4.3K chart image-code pairs, and evaluate six open-weight VLMs (3B - 26B parameters), highlighting substantial room for progress. We release the pipeline, prompts, and the dataset to help accelerate efforts towards robust chart understanding and vision-conditioned code generation: https://github.com/SD122025/ChartGen/

Figures

Figures reproduced from arXiv: 2507.19492 by the authors.

Figure 1
Figure 1. An overview of ChartGen: an automated pipeline for code-guided synthetic chart generation. First, a seed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An illustration of synthetic chart images generated from a single seed chart using the ChartGen pipeline. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An overview of the synthetic chart dataset obtained with ChartGen pipeline. The dataset includes 200K chart [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Infinity-Parser2 Technical Report

    cs.AI 2026-07 unverdicted novelty 6.0 of 10

    Infinity-Parser2 pairs a 5M synthetic bilingual document corpus with multi-task RL to claim SOTA document parsing on olmOCR-Bench and ParseBench.

Reference graph

Works this paper leans on

26 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [2]

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton

    URL https://arxiv.org/abs/2410.07073. Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton. Program synthesis with large language models,

    arXiv

  2. [4]

    For chart code augmentation we rely on Codestral-22B-v0.1, configured to output at most 1200 tokens with sampling disabled

    A.1 The ChartGen Pipeline For the first stage of the ChartGen pipeline, we employ the vision–language model phi-3.5-vision-instruct, with maximum generation length set to 2000 new tokens, and temperature set to τ = 0.0 to obtain deterministic outputs and increase the likelihood of faithful chart reconstruction. For chart code augmentation we rely on Codes...

    work page 2000

  3. [6]

    URL https://arxiv.org/abs/2107.03374. Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, Lixin Gu, Xuehui Wang, Qingyun Li, Yimin Ren, Zixuan Chen, Jiapeng Luo, Jiahao Wang, Tan Jiang, Bo Wang, Conghui He, Botian Shi, Xingcheng Zhang, Han Lv, Yi Wang, Wenqi Shao, Pei Chu, Zhongying Tu, ...

    arXiv

  4. [8]

    URL https://arxiv.org/abs/2409.17146. Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang,...

    arXiv

  5. [9]

    Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Akos Kadar, Adam Trischler, and Yoshua Bengio

    URL https://zenodo.org/records/10256836. Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Akos Kadar, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning,

    arXiv

  6. [12]

    URL https://arxiv.org/abs/2203.10244. Meta AI. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation, April

    arXiv

  7. [13]

    Accessed: 2025-04-23

    URL https://ai.meta.com/blog/llama-4-multimodal-intelligence/ . Accessed: 2025-04-23. Ahmed Nassar, Andres Marafioti, Matteo Omenetti, Maksym Lysak, Nikolaos Livathinos, Christoph Auer, Lucas Morin, Rafael Teixeira de Lima, Yusik Kim, A. Said Gurbuz, Michele Dolfi, Miquel Farré, and Peter W. J. Staar. Smoldocling: An ultra-compact vision-language model fo...

    work page 2025

  8. [14]

    URL https://arxiv.org/abs/2503.11576. 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. Learning transferable visual models from natural language supervision,

    arXiv

Show all 26 references
  1. [15]

    Jonathan Roberts, Kai Han, Neil Houlsby, and Samuel Albanie

    URL https://arxiv.org/abs/2103.00020. Jonathan Roberts, Kai Han, Neil Houlsby, and Samuel Albanie. Scifibench: Benchmarking large multimodal models for scientific figure interpretation,

    arXiv

  2. [16]

    Chenglei Si, Yanzhe Zhang, Ryan Li, Zhengyuan Yang, Ruibo Liu, and Diyi Yang

    URL https://arxiv.org/abs/2405.08807. Chenglei Si, Yanzhe Zhang, Ryan Li, Zhengyuan Yang, Ruibo Liu, and Diyi Yang. Design2code: Benchmarking multimodal code generation for automated front-end engineering,

    arXiv

  3. [17]

    URL https://arxiv.org/abs/2305.14637. 11 ChartGen: Scaling Chart Understanding Via Code-Guided Synthetic Chart Generation A PREPRINT Granite Vision Team, Leonid Karlinsky, Assaf Arbelle, Abraham Daniels, Ahmed Nassar, Amit Alfassi, Bo Wu, Eli Schwartz, Dhiraj Joshi, Jovana Kon...

    arXiv

  4. [18]

    Jingxuan Wei, Nan Xu, Guiyong Chang, Yin Luo, BiHui Yu, and Ruifeng Guo

    URL https://arxiv.org/abs/2502.09927. Jingxuan Wei, Nan Xu, Guiyong Chang, Yin Luo, BiHui Yu, and Ruifeng Guo. mchartqa: A universal benchmark for multimodal chart question answer based on vision-language alignment and reasoning,

    arXiv

  5. [19]

    URL https: //arxiv.org/abs/2404.01548. David Wood, Boris Lublinsky, Alexy Roytman, Shivdeep Singh, Constantin Adam, Abdulhamid Adebayo, Sungeun An, Yuan Chi Chang, Xuan-Hong Dang, Nirmit Desai, Michele Dolfi, Hajar Emami-Gohari, Revital Eres, Takuya Goto, Dhiraj Joshi, Yan Koy...

    arXiv

  6. [20]

    Chengyue Wu, Yixiao Ge, Qiushan Guo, Jiahao Wang, Zhixuan Liang, Zeyu Lu, Ying Shan, and Ping Luo

    URL https://arxiv.org/abs/2409.18164. Chengyue Wu, Yixiao Ge, Qiushan Guo, Jiahao Wang, Zhixuan Liang, Zeyu Lu, Ying Shan, and Ping Luo. Plot2code: A comprehensive benchmark for evaluating multi-modal large language models in code generation from scientific plots,

    arXiv

  7. [21]

    Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, and Yu Qiao

    URL https://arxiv.org/abs/2405.07990. Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, and Yu Qiao. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning,

    arXiv

  8. [22]

    Zhengzhuo Xu, Sinan Du, Yiyan Qi, Chengjin Xu, Chun Yuan, and Jian Guo

    URL https://arxiv.org/abs/2402.12185. Zhengzhuo Xu, Sinan Du, Yiyan Qi, Chengjin Xu, Chun Yuan, and Jian Guo. Chartbench: A benchmark for complex visual reasoning in charts,

    arXiv

  9. [23]

    Cheng Yang, Chufan Shi, Yaxin Liu, Bo Shui, Junjie Wang, Mohan Jing, Linran Xu, Xinyu Zhu, Siheng Li, Yuxiang Zhang, Gongye Liu, Xiaomei Nie, Deng Cai, and Yujiu Yang

    URL https://arxiv.org/abs/2312.15915. Cheng Yang, Chufan Shi, Yaxin Liu, Bo Shui, Junjie Wang, Mohan Jing, Linran Xu, Xinyu Zhu, Siheng Li, Yuxiang Zhang, Gongye Liu, Xiaomei Nie, Deng Cai, and Yujiu Yang. Chartmimic: Evaluating lmm’s cross-modal reasoning capability via chart...

    arXiv

  10. [24]

    Zifeng Zhu, Mengzhao Jia, Zhihan Zhang, Lang Li, and Meng Jiang

    URL https://arxiv.org/abs/ 2501.06598. Zifeng Zhu, Mengzhao Jia, Zhihan Zhang, Lang Li, and Meng Jiang. Multichartqa: Benchmarking vision-language models on multi-chart problems,

    arXiv

  11. [25]

    URL https://arxiv.org/abs/2410.14179. 12 ChartGen: Scaling Chart Understanding Via Code-Guided Synthetic Chart Generation A PREPRINT A Generation Configurations and Prompt Templates In this section, we specify the model and compute configurations, and the exact prompts used in...

    arXiv

  12. [2018]

    Shankar Kantharaj, Xuan Long Do, Rixie Tiffany Ko Leong, Jia Qing Tan, Enamul Hoque, and Shafiq Joty

    URL https://arxiv.org/abs/1710.07300. Shankar Kantharaj, Xuan Long Do, Rixie Tiffany Ko Leong, Jia Qing Tan, Enamul Hoque, and Shafiq Joty. Opencqa: Open-ended question answering with charts, 2022a. URL https://arxiv.org/abs/2210.06628. Shankar Kantharaj, Rixie Tiffany Ko Leon...

    arXiv

  13. [2020]

    URL https://arxiv.org/abs/2005.14165. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry...

    2005 arXiv

  14. [2021]

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou

    URL https://arxiv.org/abs/2108.07732. Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond,

    arXiv

  15. [2022]

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li

    URL https://arxiv.org/abs/2211.11501. Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024a. URL https://arxiv.org/abs/ 2408.03326. Jian Li, Weihe...

    arXiv

  16. [2023]

    URL https://arxiv.org/abs/2308.12966. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, ...

    arXiv

  17. [2024]

    URL https://arxiv.org/abs/2404.14219. Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliou, Paul Jacob, Albert Q. Jiang, Kartik Khan...

    arXiv

  18. [2025]

    URL https://arxiv.org/abs/2412.05271. Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, Jiasen Lu, Taira Anderson, Erin Bransom, Kiana Ehsani, Huong Ngo, YenSung Chen, Ajay Pate...

    arXiv

Pith tools

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