Pith. sign in

REVIEW 4 major objections 6 minor 44 references

This paper claims that chart-to-code generation fails from the wrong supervision signal, not from lack of data, and that a structured intermediate representation called Chart Specification makes a 7-billion-parameter vision-language model s

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A 7B VLM trained with a structured chart-specification reward beats larger and commercial models on chart-to-code benchmarks using only 3K-4K training samples.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A plausible and novel structural-spec RLVR recipe for chart-to-code, but the missing dedup analysis could fully explain the headline numbers, so treat the empirical claims as provisional. the 4 major comments →

arxiv 2602.10880 v2 pith:V2APBCVE submitted 2026-02-11 cs.CV

Chart Specification: Structural Representations for Incentivizing VLM Reasoning in Chart-to-Code Generation

Minggui He , Mingchen Dai , Jian Zhang , Yilun Liu , Shimin Tao , Pufan Zeng , Osamu Yoshie , Yuya Ieiri This is my paper
classification cs.CV
keywords chart-to-code generationvision-language modelsstructural representationverifiable rewardsreinforcement learningdata efficiencychart specificationstructural fidelity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 argues that the standard way of teaching vision-language models to turn chart images into plotting code—directly matching generated code tokens against reference code—teaches surface imitation rather than visual structure. To fix this, the authors introduce Chart Specification, a declarative intermediate representation that records the chart's topology, coordinate system, data domains, and analytic formulas, plus a runtime-captured version of the exact numbers fed into plotting calls. Using this specification, they build a structurally balanced training set and a fine-grained, verifiable reward that reinforcement learning can optimize. The central claim is that structural supervision is far more data-efficient than scaling supervised data: with roughly 3,000 training samples the model surpasses much larger and more heavily trained systems on three public chart-to-code benchmarks, and with 4,000 samples it reaches state-of-the-art results. If true, this means precise structural feedback, not raw data volume, is what unlocks faithful chart reconstruction.

Core claim

The paper's central claim is that chart-to-code generation fails not from a lack of training data but from the wrong supervision signal. Direct token-level training on plotting code makes a model imitate syntax while remaining blind to which visual structures matter. The authors propose Chart Specification as a canonical description of a chart's structure—chart type and panel layout, coordinate system, axis ranges and series labels, and analytic data transformations—together with a code-level specification obtained by intercepting the plotting library at runtime to capture exact data primitives. They then use this representation two ways: to curate ChartStruct, a training corpus balanced acr

What carries the argument

The load-bearing object is the Chart Specification, written as a pair of a semantic specification and a code specification. The semantic part abstracts a chart into four dimensions—global topology (chart type, panel count, layout), coordinate system, data domains (axis ranges and series labels), and analytic representations (explicit functional forms)—so that functionally equivalent plotting scripts collapse to the same structural description. The code part is produced by runtime interception of plotting primitives, recovering exact numeric values, wedge ratios, node-edge relations, and other data that are computed implicitly during execution. This hybrid representation carries the argument:

Load-bearing premise

The whole performance story depends on the training corpus, which was taken largely from an existing chart-to-code dataset, being disjoint from the test images of the benchmarks used for evaluation; the paper reports no deduplication check, so if overlap exists the reported gains are inflated.

What would settle it

Run a contamination scan between the training corpus and the test splits of the three benchmarks (e.g., exact or near-duplicate image hashing and code similarity), or re-evaluate the trained model on a newly collected set of charts from a different source; if performance collapses or overlap is found, the data-efficiency claim is not the explanation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, chart-to-code models can be trained to high fidelity with a few thousand samples instead of hundreds of thousands, dramatically lowering compute and data requirements.
  • Structural supervision should generalize to other structured generation tasks where text tokens under-specify visual or executable outcomes, such as UI code generation or diagram-to-code.
  • The Spec-Align Reward design shows that verifiable rewards can be dense and hierarchical, not just binary, extending reinforcement learning with verifiable rewards beyond math and coding to visually grounded generation.
  • The 'think-then-answer' reasoning stage contributes most on complex, long-tail chart types; standard charts gain little, suggesting reasoning scaffolding should be allocated adaptively.
  • Supervised fine-tuning alone can regress performance relative to the base model, while spec-aligned reinforcement learning reliably improves it—a caution for recipe design in chart-to-code systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's comparison implies that contamination between its training corpus (derived largely from an existing chart-to-code dataset) and the evaluation benchmarks could materially inflate the headline numbers; an independent, freshly collected test set would be the cleanest way to confirm the gains.
  • If the structural-specification approach generalizes, chart editing and reuse could be reframed as specification editing: modify the layout, domain, or data mapping in the spec and regenerate code, enabling interactive, self-correcting chart tools.
  • The runtime-interception idea could be extended to other domains with implicit data, such as network diagrams, 3D scenes, or scientific visualizations, where the visible output is generated from computations not present in the source code.
  • The observed saturation between 3K and 4K samples suggests an inflection point; an inference worth testing is whether mixture-of-spec balancing within a fixed budget, rather than raw scale, is the dominant lever for further gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces Chart Specification, a structured intermediate representation for chart-to-code generation, together with a pipeline that uses this representation to curate a structurally balanced training corpus (ChartStruct) and to design a fine-grained Spec-Align reward for GRPO-based reinforcement learning. The representation has two parts: Phi_sem, a declarative semantic description of chart topology, coordinates, data domains, and analytic forms; and Phi_code, numerical facts obtained by runtime interception of plotting primitives. The authors evaluate their method on ChartMimic, Plot2Code, and ChartX, using Qwen2.5-VL-7B as the backbone. They claim that 3K training samples suffice to surpass strong baselines and that 4K samples establish state-of-the-art results across all evaluated metrics, with ablations supporting the importance of the structural reward, the data curation strategy, and the reasoning chain.

Significance. If the results hold, the paper makes a useful contribution: it identifies a concrete failure mode in token-level chart-to-code supervision and proposes a verifiable, structure-aware alternative that is both practical and well motivated by visualization grammar. The paper ships code and dataset, and the ablation design is thoughtful, including comparisons across model scales, data sizes, reward components, and reasoning variants. The reported gains are large (e.g., ChartX GPT-score 3.52 vs. 2.09 for ChartCoder), and the data-efficiency claim is falsifiable. However, the headline results rest on an unresolved external-validity threat concerning the provenance of ChartStruct, and the statistical support is thin because all evaluations are single runs without variance estimates. The contribution is therefore promising but not yet fully established.

major comments (4)
  1. [Sec. 3.2, Table 1, Fig. 2] The ChartStruct corpus is stated to be curated 'primarily from ChartCoder [7]', but the paper does not report any deduplication or overlap analysis against the test splits of ChartMimic, Plot2Code, or ChartX. Figure 2 labels a 'Verification & De-duplication' step, yet the text gives no details about what is deduplicated or against what. Because ChartCoder is itself a chart-to-code model trained and evaluated in this benchmark ecosystem, exact or near-duplicate image/code pairs in the 3,008/3,996 training triplets would directly inflate the reported SOTA numbers and the data-efficiency claim. Please document the deduplication procedure (exact hashing plus near-duplicate detection) and report overlap counts against each of the three test splits; if overlaps are found, rerun the experiments on a cleaned corpus and report both sets of numbers.
  2. [Tables 3–5, Fig. 4] All experimental results are single runs with no error bars, confidence intervals, or significance tests. This matters particularly because several headline comparisons are small in absolute terms (e.g., ChartSpec 4k Pass Rate 88.7 vs. GPT-4o 88.6 in Table 4; ChartSpec 4k Overall 82.4 vs. GPT-4o 81.2 in Table 3) and because the High-Level, Rating, and GPT-score metrics are obtained from GPT-4o, which is stochastic. Without repeated evaluations or at least bootstrapped intervals, the claims of 'surpassing' and 'state-of-the-art' are not statistically supported. Please provide multiple seeds or a bootstrap analysis over the evaluation samples, and temper the conclusions accordingly.
  3. [Sec. 3.3, Table 8] The paper argues that 'binary execution feedback is overly sparse' and that the fine-grained Spec-Align reward is the key mechanism. However, Table 8 only compares Semantic Spec Only vs. Code Spec (Full); there is no control condition using only Format and Execution rewards (i.e., binary execution plus format checks). As presented, the gains over SFT could in principle come from applying RL at all rather than from the structural specification. Add an 'execution-only' (or format+execution) control to isolate the contribution of R_sem and R_code. This is essential for the central mechanistic claim.
  4. [Sec. 3.1, Sec. 3.3] The Chart Specification is extracted by Qwen3-32B with a 'carefully designed prompt', but manual verification is reported only on 200 instances from the reference training corpus (10 per chart type). During RL, the reward is computed from specifications extracted from model-generated code, which may be significantly more error-prone than clean reference scripts. Since the reward's verifiability is claimed as a key advantage, please report extraction accuracy on generated outputs (e.g., human-check a random sample of rollouts) and quantify the impact of extraction errors on the reward signal. Without this, the 'verifiable' property is not empirically established.
minor comments (6)
  1. [Ref. [11]] The reference cited as 'Chart2Code' appears to actually be a paper on iterative dual preference learning for chart-to-code; please verify the citation and correct it if needed.
  2. [Fig. 2] The 'Verification & De-duplication' stage shown in Figure 2 is not described in Section 3.2. If deduplication was performed, its scope and procedure should be described in the text; otherwise the label should be removed.
  3. [Table 1] The column alignment in Table 1 is hard to parse as rendered (the total row reads '55483423996 100%'), and the 'ChartCoder Count' values do not obviously sum to the stated total. Please reformat the table and re-verify the totals.
  4. [Fig. 5] The abbreviations in Figure 5 ('CB', 'HR', 'PIP', 'multidiff', etc.) are not expanded. Please include a legend or table mapping them to the chart families.
  5. [Abstract / Sec. 5] The abstract's 'up to 61.7%' improvement is not tied to a specific table or metric. Please state explicitly which baseline and benchmark this refers to.
  6. [Sec. 4.2 / Table 4] Table 4 says 'All results are taken directly from the official Plot2Code and ChartCoder evaluations.' Please clarify which baselines were re-evaluated by the authors and which were copied from prior papers, and report the evaluation protocol (temperature, number of samples, GPT-4o version) for reproducibility.

Circularity Check

0 steps flagged

No significant circularity — reported results are judged on external benchmarks; self-citations are background only.

full rationale

The paper's central claim is an empirical comparison against three external benchmarks (ChartMimic, Plot2Code, ChartX) using official metrics (Execution Rate, Low-Level, Text-Match, GPT-score). These metrics are computed from generated code and reference charts, not from the Chart Specification representation. The Spec-Align reward (Eq. 3) is only a training objective; it does not define any reported test quantity, and the 61.7% improvement is an arithmetic comparison of published GPT-scores (3.38 vs 2.09 on ChartX), not a fitted parameter. The ablations compare spec-driven curation against random curation under otherwise matched training, so the data-efficiency conclusion is not forced by construction. Self-citations to R1-T1 [26] and R-Log [27] occur only as related-work examples of RLVR and are not load-bearing. The main unresolved concern is possible benchmark contamination: ChartStruct is curated 'primarily from ChartCoder [7]' with no deduplication against test splits reported. That is an external-validity threat, not a circular derivation, so under the given rubric it does not raise the circularity score.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The central claim rests on several unverified or weakly verified premises: the extraction LLM is reliable, the curated corpus does not leak into evaluation benchmarks, GPT-4o scoring reflects structure, GRPO generalizes, and the Grammar-of-Graphics decomposition is sufficient. Reward coefficients and sampling densities are hand-set and not sensitivity-tested.

free parameters (3)
  • Reward coefficients and thresholds = beta=0.5; Rf=-2; Re=0.5/-1; topology gate=3
    Hand-set scalars in Eq. 3 and Table 2; no sensitivity analysis is reported, so results may depend on these values.
  • Complexity-adaptive sampling density per tier = rho = 90 / 72 / 54 for Tier 1/2/3
    Chosen to balance structural difficulty; this controls ChartStruct composition and all downstream results.
  • Training data scale = 3,008 and 3,996 samples
    The 4K instantiation is the default; the 3K variant is used for ablations. These budgets determine the headline efficiency claims.
axioms (5)
  • domain assumption Chart Specification extracted by Qwen3-32B accurately and completely captures the structural semantics needed for reward and data curation.
    Manual verification covers only 200 instances; extraction noise would propagate into dataset labels and reward signals.
  • ad hoc to paper ChartStruct samples curated from ChartCoder are disjoint from the test sets of ChartMimic, Plot2Code, and ChartX.
    No dedup/overlap analysis is given; if false, benchmark results are inflated.
  • domain assumption GPT-4o high-level scores and execution checks are valid proxies for structural fidelity.
    Headline comparisons on ChartMimic/Plot2Code/ChartX rely on GPT-4o evaluation and execution rate.
  • domain assumption GRPO with composite rewards converges to a policy that generalizes beyond the training distribution.
    Standard RLVR assumption; no formal guarantee is provided.
  • domain assumption Grammar of Graphics / Vega-Lite decomposition is a sufficient basis for chart structure representation.
    The spec design follows this factorization; the paper does not prove minimal sufficiency.
invented entities (1)
  • Chart Specification (Phi = <Phi_sem, Phi_code>) no independent evidence
    purpose: Intermediate structural representation used for dataset balancing, reward computation, and policy training.
    It is a new construct with no falsifiable handle outside this paper's pipeline; its validity is evidenced only by downstream benchmark results.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Chart Specification: Structural Representations for Incentivizing VLM Reasoning in Chart-to-Code Generation." pith.science (2026). https://pith.science/paper/V2APBCVE

@misc{pith2026260210880,
  author       = {Pith},
  title        = {Pith review of: Chart Specification: Structural Representations for Incentivizing VLM Reasoning in Chart-to-Code Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V2APBCVE}},
  note         = {Machine review of arXiv:2602.10880}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Vision-Language Models (VLMs) have shown promise in generating plotting code from chart images, yet achieving structural fidelity remains challenging. Existing approaches largely rely on supervised fine-tuning, encouraging surface-level token imitation rather than faithful modeling of underlying chart structure, which often leads to hallucinated or semantically inconsistent outputs. We propose Chart Specification, a structured intermediate representation that shifts training from text imitation to semantically grounded supervision. Chart Specification filters syntactic noise to construct a structurally balanced training set and supports a Spec-Align Reward that provides fine-grained, verifiable feedback on structural correctness, enabling reinforcement learning to enforce consistent plotting logic. Experiments on three public benchmarks show that our method consistently outperforms prior approaches. With only 3K training samples, we achieve strong data efficiency, surpassing leading baselines by up to 61.7% on complex benchmarks, and scaling to 4K samples establishes new state-of-the-art results across all evaluated metrics. Overall, our results demonstrate that precise structural supervision offers an efficient pathway to high-fidelity chart-to-code generation. Code and dataset are available at: https://github.com/Mighten/chart-specification-paper

Figures

Figures reproduced from arXiv: 2602.10880 by Jian Zhang, Mingchen Dai, Minggui He, Osamu Yoshie, Pufan Zeng, Shimin Tao, Yilun Liu, Yuya Ieiri.

Figure 1
Figure 1. Figure 1: Motivation for structure-aware chart reasoning. (Top) Direct chart-to-code models rely on surface-level imitation and often hallucinate structural dependencies. (Bottom) By explicitly modeling chart structure via Chart Specification, our approach enforces constraint-consistent plotting logic and faithful visual reconstruction. Furthermore, layout-aware models like Pix2Struct [17] and MatCha [18] pre-train … view at source ↗
Figure 2
Figure 2. Figure 2: The Overview of Our Framework. (A) Specification-Driven Data Curation: Adopting Chart Specification () to extract semantic intent (𝑠𝑒𝑚) and physical execution data (𝑐𝑜𝑑𝑒) from raw scripts, and guiding the curation of the ChartStruct corpus. (B) Group Relative Policy Optimization: The VLM policy is optimized using group-based advantage estimation. (C) Hierarchical Reward Tree: A fine-grained reward mecha… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of our Chart Specification () across four distinct chart types. The grey panels represent the 𝑠𝑒𝑚, capturing declarative intents like topology and data domains. The orange panels (bottom-left and bottom-right) illustrate the 𝑐𝑜𝑑𝑒, which uses runtime interception to capture implicit data, such as calculated wedge ratios in Ring charts or node-edge relationships in Network graphs. address th… view at source ↗
Figure 4
Figure 4. Figure 4: Impact of training data scale on chart-to-code performance. We compare SFT and Spec-Align RL across varying data sizes on ChartMimic (a) and Plot2Code (b), evaluated using Pass Rate, Low-level accuracy, and Text-Match. SFT is prone to structural degradation in complex chart types. Although it maintains competence on standard for￾mats, performance drops below the baseline on intricate geometries like Contou… view at source ↗
Figure 5
Figure 5. Figure 5: Type-wise performance comparison at the 3K data scale on Chartmimic. Here, low level average accuracy across different chart types. Columns correspond to chart categories, and rows denote different training settings: Base (Qwen2.5VL-7B￾Instruct), SFT (Standard Fine-Tuning), RL (no-CoT) (Spec-Align RL without reasoning), and RL (CoT) (Spec-Align RL with reasoning). The Overall column reports weighted averag… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison on the ChartMimic validation set. Left to Right: Ground Truth, GPT-4o, InternVL3-8B, Qwen2.5- VL-7B, and Ours (ChartSpec). While base models (columns 3-4) struggle with complex layouts like Radar and Boxplots, our method (Column 5) achieves high fidelity comparable to GPT-4o. Notably, in the Boxplot (Row 3), our model successfully preserves statistical outliers (small circles) which … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 26 linked inside Pith

  1. [1]

    Siegel, S

    N. Siegel, S. Styles, S. K. Divvala, Figureseer: Parsing result-figures inresearchpapers, ACMTransactionsonGraphics(TOG)35(2016) 111

    2016

  2. [2]

    J. Poco, J. Heer, Reverse-engineering visualizations: Recovering visualencodingsfromchartimages, in:Proceedingsofthe2017CHI Conference on Human Factors in Computing Systems, ACM, 2017, pp. 5707–5717

    2017

  3. [3]

    Kafle, B

    K. Kafle, B. Price, S. Cohen, C. Kanan, Dvqa: Understanding data visualizations via question answering, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 5648–5656

    2018

  4. [4]

    Zhang, J

    J. Zhang, J. Huang, S. Jin, S. Lu, Vision-language models for vision tasks: A survey, IEEE transactions on pattern analysis and machine intelligence 46 (2024) 5625–5644

    2024

  5. [5]

    C. Yang, C. Shi, Y. Liu, B. Shui, J. Wang, M. Jing, L. Xu, X. Zhu, S. Li, Y. Zhang, et al., Chartmimic: Evaluating lmm’s cross-modal reasoning capability via chart-to-code generation, arXiv preprint arXiv:2406.09961 (2024)

    Pith/arXiv arXiv 2024

  6. [6]

    C. Wu, Y. Ge, Q. Guo, J. Wang, Z. Liang, Z. Lu, Y. Shan, P. Luo, Plot2code: A comprehensive benchmark for evaluating multi-modal largelanguagemodelsincodegenerationfromscientificplots, arXiv preprint arXiv:2405.07990 (2024)

    Pith/arXiv arXiv 2024

  7. [7]

    X. Zhao, X. Luo, Q. Shi, C. Chen, S. Wang, Z. Liu, M. Sun, Chartcoder: Advancing multimodal large language model for chart- to-code generation, arXiv preprint arXiv:2501.06598 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    Lightman, et al., Let’s verify step by step, ICLR (2024)

    H. Lightman, et al., Let’s verify step by step, ICLR (2024). Minggui He et al.:Preprint submitted to ElsevierPage 14 of 17 Chart Specification

    2024

  9. [9]

    H. Le, Y. Wang, A. D. Gotmare, S. Silvio, S. C. H. Hoi, Coderl: Mastering code generation through pretrained models and deep rein- forcement learning, in: NeurIPS, 2022

    2022

  10. [10]

    Beltramelli, pix2code: Generating code from a graphical user interface screenshot, arXiv preprint arXiv:1705.07962 (2017)

    T. Beltramelli, pix2code: Generating code from a graphical user interface screenshot, arXiv preprint arXiv:1705.07962 (2017)

    Pith/arXiv arXiv 2017

  11. [11]

    Zhang, Y

    Z. Zhang, Y. Cao, L. Liao, Enhancing chart-to-code generation in multimodal large language models via iterative dual preference learning, in: arXiv preprint arXiv:2504.02906, 2025

    Pith/arXiv arXiv 2025

  12. [14]

    Akhtar, O

    M. Akhtar, O. Cocarascu, E. Simperl, Reading and reasoning over chart images for evidence-based automated fact-checking, in: A. Vlachos, I. Augenstein (Eds.), Findings of the Association for Computational Linguistics: EACL 2023, Association for Computa- tional Linguistics, Dubrovnik, Croatia, 2023, pp. 399–414. URL: https://aclanthology.org/2023.findings-...

    doi:10.18653/v1/ 2023

  13. [15]

    Y. Wang, S. Wu, Y. Zhang, W. Wang, Z. Liu, J. Luo, H. Fei, Mul- timodal chain-of-thought reasoning: A comprehensive survey, arXiv preprint arXiv:2503.12605 (2025)

    Pith/arXiv arXiv 2025

  14. [16]

    Q. Zhao, Y. Lu, M. J. Kim, Z. Fu, Z. Zhang, Y. Wu, Z. Li, Q. Ma, S. Han, C. Finn, A. Handa, M.-Y. Liu, D. Xiang, G. Wetzstein, Cot- vla: Visual chain-of-thought reasoning for vision-language-action models, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. URL:https://arxiv. org/abs/2503.22020

    Pith/arXiv arXiv 2025

  15. [17]

    K.Lee,C.Gan,Y.Cheng,H.Li,W.Wang,X.Chen,J.Hu,L.Wang, D. Liu, Z. Chen, et al., Pix2Struct: Screenshot parsing as pretraining for visual language understanding, arXiv preprint arXiv:2210.03347 (2022). Accepted at ICML

    Pith/arXiv arXiv 2022

  16. [18]

    F. Liu, F. Piccinno, S. Krichene, C. Pang, K. Lee, M. Joshi, Y. Altun, N. Collier, J. M. Eisenschlos, Matcha: Enhancing visual language pretrainingwithmathreasoningandchartderendering,arXivpreprint arXiv:2212.09662 (2022)

    Pith/arXiv arXiv 2022

  17. [19]

    Z. Li, J. Fu, L. Song, J. Bian, J. Zhang, R. Wang, Chain of functions: Aprogrammaticpipelineforfine-grainedchartreasoningdata, arXiv preprint arXiv:2503.16260 (2025)

    Pith/arXiv arXiv 2025

  18. [20]

    Ouyang, et al., Training language models to follow instructions with human feedback, in: NeurIPS, 2022

    L. Ouyang, et al., Training language models to follow instructions with human feedback, in: NeurIPS, 2022

    2022

  19. [21]

    Rafailov, et al., Direct preference optimization: Your language modelissecretlyarewardmodel, in:AdvancesinNeuralInformation Processing Systems, 2023

    R. Rafailov, et al., Direct preference optimization: Your language modelissecretlyarewardmodel, in:AdvancesinNeuralInformation Processing Systems, 2023

    2023

  20. [22]

    Liang, J

    X. Liang, J. Hu, D. Wang, Z. Ma, L. Zhao, R. Li, B. Wan, Q. Wang, Chexpo: Preference optimization for chest x-ray vlms with counter- factual rationale, in: Proceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 2606–2615

    2025

  21. [23]

    N.Lambert,J.Morrison,V.Pyatkin,S.Huang,H.Ivison,F.Brahman, L. J. V. Miranda, A. Liu, N. Dziri, S. Lyu, Y. Gu, S. Malik, V. Graf, J.D.Hwang,J.Yang,R.LeBras,O.Tafjord,C.Wilhelm,L.Soldaini, N. A. Smith, Y. Wang, P. Dasigi, H. Hajishirzi, TÜLU 3: Pushing frontiers in open language model post-training, in: Proceedings of the Conference on Language Modeling (CO...

    Pith/arXiv arXiv 2025

  22. [24]

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. Li, Y. Wu, et al., Deepseekmath: Pushing the limits of mathematical reasoning in open language models, arXiv preprint arXiv:2402.03300 (2024)

    Pith/arXiv arXiv 2024

  23. [25]

    URL:https://arxiv.org/abs/ 2501.12948.arXiv:2501.12948

    DeepSeek-AI, Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. URL:https://arxiv.org/abs/ 2501.12948.arXiv:2501.12948

    Pith/arXiv arXiv 2025

  24. [26]

    M. He, Y. Liu, S. Tao, Y. Luo, H. Zeng, C. Su, L. Zhang, H. Ma, D. Wei, W. Meng, et al., R1-t1: Fully incentivizing translation capa- bilityinllmsviareasoninglearning,arXivpreprintarXiv:2502.19735 (2025)

    Pith/arXiv arXiv 2025

  25. [27]

    Y. Liu, Z. Chen, S. Xu, M. He, S. Tao, W. Meng, Y. Xie, T. Han, C. Zhao, J. Du, D. Wei, S. Zhang, Y. Sun, R-log: Incen- tivizing log analysis capability in llms via reasoning-based rein- forcement learning, 2025. URL:https://arxiv.org/abs/2509.25987. arXiv:2509.25987

    arXiv 2025

  26. [28]

    Q. Yu, Z. Zhang, R. Zhu, et al., Dapo: An open-source llm rein- forcementlearningsystematscale, arXivpreprintarXiv:2503.14476 (2025)

    Pith/arXiv arXiv 2025

  27. [29]

    Zheng, S

    C. Zheng, S. Liu, M. Li, X.-H. Chen, et al., Group sequence policy optimization, arXiv preprint arXiv:2507.18071 (2025)

    Pith/arXiv arXiv 2025

  28. [30]

    S. Yang, C. Dou, P. Guo, K. Lu, Q. Ju, F. Deng, R. Xin, Dcpo: Dy- namicclippingpolicyoptimization, arXivpreprintarXiv:2509.02333 (2025)

    Pith/arXiv arXiv 2025

  29. [31]

    Wilkinson, The grammar of graphics, in: Handbook of computa- tionalstatistics:Conceptsandmethods,Springer,2011,pp.375–414

    L. Wilkinson, The grammar of graphics, in: Handbook of computa- tionalstatistics:Conceptsandmethods,Springer,2011,pp.375–414

    2011

  30. [32]

    Satyanarayan, D

    A. Satyanarayan, D. Moritz, K. Wongsuphasawat, J. Heer, Vega- lite: A grammar of interactive graphics, IEEE Transactions on Visualization and Computer Graphics 23 (2017) 341–350

    2017

  31. [33]

    Y. Han, C. Zhang, X. Chen, X. Yang, Z. Wang, G. Yu, B. Fu, H.Zhang, Chartllama:Amultimodalllmforchartunderstandingand generation, in:Unpublishedorworkshop,2023.Preprintavailableon HuggingFace / GitHub

    2023

  32. [34]

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, J. Zhou, Qwen2-vl: A versatile vision–language model for understanding, localization, text reading, and beyond, CoRR abs/2409.12191 (2024)

    Pith/arXiv arXiv 2024

  33. [35]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, et al., Qwen2. 5-vl technical report, arXiv preprint arXiv:2502.13923 (2025)

    Pith/arXiv arXiv 2025

  34. [36]

    Team, Google, Gemini: a family of highly capable multimodal models, arXiv preprint arXiv:2312.11805 (2023)

    G. Team, Google, Gemini: a family of highly capable multimodal models, arXiv preprint arXiv:2312.11805 (2023)

    Pith/arXiv arXiv 2023

  35. [37]

    Anthropic, The claude 3 model family: Opus, sonnet, haiku, An- thropic Technical Report (2024)

    2024

  36. [38]

    OpenAI, Gpt-4v(ision) system card, OpenAI Technical Report (2023)

    2023

  37. [39]

    A.Hurst,A.Lerer,A.P.Goucher,A.Perelman,A.Ramesh,A.Clark, A.Ostrow,A.Welihinda,A.Hayes,A.Radford,etal., Gpt-4osystem card, arXiv preprint arXiv:2410.21276 (2024)

    Pith/arXiv arXiv 2024

  38. [40]

    Zhang, X

    P. Zhang, X. Dong, Y. Zang, Y. Cao, R. Qian, L. Chen, Q. Guo, H. Duan, B. Wang, L. Ouyang, S. Zhang, W. Zhang, Y. Li, Y. Gao, P. Sun, X. Zhang, W. Li, J. Li, W. Wang, H. Yan, C. He, X. Zhang, K. Chen, J. Dai, Y. Qiao, D. Lin, J. Wang, Internlm-xcomposer-2.5: Aversatilevision–languagemodelsupportinglong-contextinputand output, CoRR abs/2407.03320 (2024)

    Pith/arXiv arXiv 2024

  39. [41]

    H. Lu, W. Liu, B. Zhang, B. Wang, K. Dong, B. Liu, J. Sun, T. Ren, Z.Li,Y.Sun,etal., Deepseek-vl:Towardsreal-worldvision-language understanding, arXiv preprint arXiv:2403.05525 (2024)

    Pith/arXiv arXiv 2024

  40. [42]

    H. Liu, C. Li, Y. Li, B. Li, Y. Zhang, S. Shen, Y. J. Lee, Llava- next: Improved reasoning, ocr, and world knowledge, 2024. URL: https://llava-vl.github.io/blog/2024-01-30-llava-next/

    2024

  41. [43]

    et al., Minicpm-llama3-v2.5: A gpt-4v-level multimodal llm with strong ocr and high-resolution capabilities, CoRR abs/2408.01800 (2024)

    Y. et al., Minicpm-llama3-v2.5: A gpt-4v-level multimodal llm with strong ocr and high-resolution capabilities, CoRR abs/2408.01800 (2024)

    Pith/arXiv arXiv 2024

  42. [44]

    F. Meng, W. Wang, Z. Zhou, H. Chen, Chartassisstant: A univer- sal chart multimodal language model via chart-to-table pre-training, arXiv preprint arXiv:2401.02384 (2024)

    Pith/arXiv arXiv 2024

  43. [45]

    Zhang, A

    L. Zhang, A. Hu, H. Xu, M. Yan, Y. Xu, Q. Jin, J. Zhang, F. Huang, Tinychart: Efficient chart understanding with program-of-thoughts learning and visual token merging, in: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2024, pp. 1882–1898. doi:10.18653/v1/2024.emnlp-main. Minggui He et al.:Preprint submitte...

    doi:10.18653/v1/2024.emnlp-main 2024

  44. [46]

    R. Xia, B. Zhang, H. Ye, X. Yan, Q. Liu, H. Zhou, Z. Chen, M. Dou, B. Shi, J. Yan, Y. Qiao, Chartx & chartvlm: A versatile benchmark andfoundationmodelforcomplicatedchartreasoning, arXivpreprint arXiv:2402.12185 (2024). MingguiHereceivedtheB.E.degreefromBeijing University of Chemical Technology, China and M.E. from Waseda University, Japan in 2016 and 201...

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.