pith. sign in

arxiv: 2605.29692 · v1 · pith:2AVHFB4Jnew · submitted 2026-05-28 · 💻 cs.DB

Towards Reliable Agentic Progressive Text-to-Visualization with Verification Rules

Wenxin Xu , Chen Jason Zhang , Xiaoyong Wei , Haoyang Li , Hwanhee Kim , Yuanfeng Song , Raymond Chi-Wing Wong This is my paper

Pith reviewed 2026-06-29 00:20 UTC · model grok-4.3

classification 💻 cs.DB
keywords text-to-visualizationprogressive multi-turnagent-based frameworkverification rulesPMVisBenchexecution accuracymulti-agent systemReAct-style loop
0
0 comments X

The pith

Progressive multi-turn interactions with a validation agent improve text-to-visualization accuracy over one-shot methods.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes shifting text-to-visualization from single-round queries to a progressive multi-turn process that lets users refine their intents step by step. It introduces PMVisBench, a dataset built by simplifying visualization query languages and reconstructing natural language queries under explicit rules so each intermediate step stays valid. PMVisAgent then deploys three specialized agents in a ReAct-style loop where a Validation Agent checks outputs and repairs errors to limit accumulation across turns. Experiments show this yields higher execution accuracy than existing baselines on both single-table and multi-table cases. The approach directly targets the cognitive overload that occurs when users must specify every visualization detail at once.

Core claim

PMVisAgent, built on the progressive PMVis paradigm and the PMVisBench dataset, uses a User agent, a System agent, and a Validation Agent that applies verification and repair through a ReAct-style tool-use loop with explicit interaction and verification rules; this setup produces up to 17.57 percent and 23.21 percent gains in execution accuracy for single-table and multi-table settings respectively.

What carries the argument

The Validation Agent that performs verification and repair via a ReAct-style tool-use loop with explicit interaction and verification rules.

If this is right

  • Users can refine visualization requirements across multiple turns instead of specifying everything upfront.
  • Verification rules keep every intermediate visualization query valid and meaningful.
  • Error accumulation is reduced by the validation agent's repair loop across dialogue rounds.
  • Both single-table and multi-table scenarios benefit from the combined progressive interaction and clarification steps.

Where Pith is reading between the lines

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

  • The multi-agent verification pattern could transfer to other iterative generation tasks such as text-to-SQL or code synthesis.
  • The dataset construction method of rule-constrained simplification and reconstruction offers a template for building progressive benchmarks in additional domains.
  • Deployment in interactive data tools might allow non-experts to explore datasets through natural back-and-forth dialogue rather than precise single queries.

Load-bearing premise

The PMVisBench dataset, built through VQL simplification and NLQ reconstruction with explicit rule constraints, accurately captures the progressive and iterative nature of real-world user queries.

What would settle it

Running the same evaluation protocol on logs of actual non-expert users issuing iterative visualization requests and finding no accuracy gain or a drop relative to one-shot baselines.

Figures

Figures reproduced from arXiv: 2605.29692 by Chen Jason Zhang, Haoyang Li, Hwanhee Kim, Raymond Chi-Wing Wong, Wenxin Xu, Xiaoyong Wei, Yuanfeng Song.

Figure 1
Figure 1. Figure 1: Pipeline of PMVisBench dataset construction. In the VQL Simplification step, original VQLs are parsed into clauses [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dataset statistics of PMVisBench. (a) Distribution of samples across different chart types under four difficulty levels. (b) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The overall architecture of PMVisAgent. The pipeline simulates a progressive multi-turn text-to-vis workflow with [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: An example of PMVisAgent workflow. This case shows how Validation Agent detects and repairs both schema errors [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
read the original abstract

Text-to-Visualization (Text-to-Vis) translates natural language queries into visualization query languages, enabling non-expert users to perform data analysis. However, most existing methods follow a one-shot paradigm that requires users to specify all visualization details in a single round, often leading to cognitive overload and incorrect visualizations. In this paper, we propose PMVis, a progressive multi-turn paradigm for text-to-vis, where users' intents are refined through multi-turn interactions. To support research in this paradigm, we construct PMVisBench, the first dataset designed to capture the progressive and iterative nature of real-world user queries. It is built through VQL simplification and NLQ reconstruction, with explicit rule constraints to ensure each intermediate VQL remains valid and meaningful. Building upon PMVis, we further introduce PMVisAgent, an agent-based framework that simulates realistic user-system dialogues. PMVisAgent consists of a User, a System, and a Validation Agent that performs verification and repair via a ReAct-style tool-use loop to mitigate error accumulation across rounds, with explicit interaction and verification rules to ensure reliability of the multi-agent system. Extensive experiments on PMVisBench demonstrate that PMVisAgent significantly outperforms state-of-the-art text-to-vis baselines. It achieves up to 17.57\% and 23.21\% improvements in execution accuracy in single-table and multi-table settings, respectively, while ablation studies confirm the importance of combining progressive interaction with clarification. The code is available at https://github.com/wxxv/PMVis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper proposes PMVis, a progressive multi-turn paradigm for text-to-visualization that refines user intents iteratively, constructs PMVisBench as the first dataset for this setting via VQL simplification and NLQ reconstruction under explicit validity rule constraints, and introduces PMVisAgent, a multi-agent framework (User, System, Validation Agent) that uses a ReAct-style tool-use loop with verification and repair rules to mitigate error accumulation. Experiments on PMVisBench report that PMVisAgent outperforms state-of-the-art one-shot text-to-vis baselines by up to 17.57% and 23.21% execution accuracy in single-table and multi-table settings, with ablations confirming the value of progressive interaction plus clarification; code is released.

Significance. If the results hold under fair evaluation, the work is significant for shifting text-to-vis research from one-shot to realistic progressive interactions, providing the first benchmark capturing iterative query refinement, and demonstrating a reliable agentic architecture with explicit verification rules. The open-sourced code is a clear strength for reproducibility.

major comments (1)
  1. [PMVisBench construction] PMVisBench construction (abstract and dataset section): the process of simplifying existing VQLs and reconstructing NLQs under explicit rule constraints that keep every intermediate VQL valid risks systematically producing query sequences whose error patterns match exactly those targeted by the Validation Agent's repair rules. This raises the possibility that the reported 17.57%/23.21% gains reflect benchmark engineering rather than intrinsic superiority of the progressive + ReAct design over one-shot baselines.
minor comments (1)
  1. [Abstract] Abstract: the headline percentage improvements are stated without naming the baselines, reporting statistical significance, error bars, or the number of runs; the full experimental section should make these details explicit.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thoughtful review and for raising this important concern about potential benchmark construction artifacts. We address the comment directly below.

read point-by-point responses

  1. Referee: [PMVisBench construction] PMVisBench construction (abstract and dataset section): the process of simplifying existing VQLs and reconstructing NLQs under explicit rule constraints that keep every intermediate VQL valid risks systematically producing query sequences whose error patterns match exactly those targeted by the Validation Agent's repair rules. This raises the possibility that the reported 17.57%/23.21% gains reflect benchmark engineering rather than intrinsic superiority of the progressive + ReAct design over one-shot baselines.

    Authors: The PMVisBench construction starts from real VQL queries drawn from existing datasets and applies simplification to derive multi-turn sequences. The explicit rule constraints are general VQL validity rules (schema compliance, valid column references, executable aggregations, etc.) whose sole purpose is to guarantee that every intermediate ground-truth VQL is meaningful and executable; they do not encode or inject particular error types. The Validation Agent’s ReAct-style verification and repair rules, by contrast, operate on the agent’s own reasoning trace and tool calls to detect and correct generation-time mistakes (e.g., incomplete intent, wrong tool parameters). These two rule sets address orthogonal concerns: one ensures benchmark ground truth is valid, the other mitigates error accumulation inside the agent loop. Because one-shot baselines are evaluated on the final NLQ alone, they receive no benefit from the progressive structure. Ablation results further isolate the contribution of progressive interaction and the validation component. We will add an explicit subsection clarifying the distinction between construction-time validity rules and agent-time repair rules. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's claims rest on empirical execution-accuracy gains of PMVisAgent versus baselines when evaluated on the newly constructed PMVisBench. Benchmark construction (VQL simplification + NLQ reconstruction under rule constraints) and the agent's ReAct-style verification rules are presented as separate design elements intended to model progressive queries; no equations, fitted parameters, or self-citation chains are shown that reduce the reported deltas to the inputs by construction. The evaluation is therefore self-contained against the stated baselines and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no explicit free parameters, axioms, or invented entities are stated.

pith-pipeline@v0.9.1-grok · 5822 in / 1119 out tokens · 35312 ms · 2026-06-29T00:20:56.294187+00:00 · methodology

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

41 extracted references · 3 canonical work pages

  1. [1]

    Synthesizing natural language to visualization (nl2vis) benchmarks from nl2sql benchmarks,

    Y . Luo, N. Tang, G. Li, C. Chai, W. Li, and X. Qin, “Synthesizing natural language to visualization (nl2vis) benchmarks from nl2sql benchmarks,” inProceedings of the 2021 ACM SIGMOD International Conference on Management of Data, 2021, pp. 1235–1247

    2021

  2. [2]

    Aurora: Data-driven construction of visual graph query interfaces for graph databases,

    S. S. Bhowmick, K. Huang, H. E. Chua, Z. Yuan, B. Choi, and S. Zhou, “Aurora: Data-driven construction of visual graph query interfaces for graph databases,” inProceedings of the 2020 ACM SIGMOD Interna- tional Conference on Management of Data, 2020, pp. 2689–2692

    2020

  3. [3]

    Multivis-agent: A multi- agent framework with logic rules for reliable and comprehensive cross- modal data visualization,

    J. Lu, Y . Song, C. Zhang, and R. C.-W. Wong, “Multivis-agent: A multi- agent framework with logic rules for reliable and comprehensive cross- modal data visualization,” vol. 4, no. 1. ACM, feb 2026

    2026

  4. [4]

    Haichart: Human and ai paired visualization system,

    Y . Xie, Y . Luo, G. Li, and N. Tang, “Haichart: Human and ai paired visualization system,”Proceedings of the VLDB Endowment, vol. 17, no. 11, pp. 3178–3191, 2024

    2024

  5. [5]

    Navigating data repositories: Utilizing line charts to discover relevant datasets,

    D. Ji, H. Luo, Z. Bao, and S. Culpepper, “Navigating data repositories: Utilizing line charts to discover relevant datasets,”Proceedings of the VLDB Endowment, vol. 17, no. 12, pp. 4289–4292, 2024

    2024

  6. [6]

    Seedb: Efficient data-driven visualization recommendations to support visual analytics,

    M. Vartak, S. Rahman, S. Madden, A. Parameswaran, and N. Polyzotis, “Seedb: Efficient data-driven visualization recommendations to support visual analytics,” inProceedings of the VLDB Endowment International Conference on Very Large Data Bases, vol. 8, no. 13, 2015, p. 2182

    2015

  7. [7]

    Crowdchart: Crowdsourced data extraction from visualization charts,

    C. Chai, G. Li, J. Fan, and Y . Luo, “Crowdchart: Crowdsourced data extraction from visualization charts,”IEEE Transactions on Knowledge and Data Engineering, vol. 33, no. 11, pp. 3537–3549, 2020

    2020

  8. [8]

    Steerable self- driving data visualization,

    Y . Luo, X. Qin, C. Chai, N. Tang, G. Li, and W. Li, “Steerable self- driving data visualization,”IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 1, pp. 475–490, 2020

    2020

  9. [9]

    Natural language interfaces for tabular data querying and visualization: A survey,

    W. Zhang, Y . Wang, Y . Song, V . J. Wei, Y . Tian, Y . Qi, J. H. Chan, R. C.- W. Wong, and H. Yang, “Natural language interfaces for tabular data querying and visualization: A survey,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 11, pp. 6699–6718, 2024

    2024

  10. [10]

    Visualization recommendation through visual relation learning and visual preference learning,

    D. Ji, H. Luo, and Z. Bao, “Visualization recommendation through visual relation learning and visual preference learning,” in2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 2023, pp. 1860–1873

    2023

  11. [11]

    Towards robustness of text-to-visualization translation against lexical and phrasal variability,

    J. Lu, Y . Song, H. Zhang, C. J. Zhang, K. Wu, and R. C.-W. Wong, “Towards robustness of text-to-visualization translation against lexical and phrasal variability,” in2025 IEEE 41st International Conference on Data Engineering (ICDE). IEEE Computer Society, 2025, pp. 793–806

    2025

  12. [12]

    Datavist5: A pre-trained language model for jointly understanding text and data visualization,

    Z. Wan, Y . Song, S. Li, C. J. Zhang, and R. C.-W. Wong, “Datavist5: A pre-trained language model for jointly understanding text and data visualization,” in2025 IEEE 41st International Conference on Data Engineering (ICDE). IEEE, 2025, pp. 1704–1717

    2025

  13. [13]

    Vega- lite: A grammar of interactive graphics,

    A. Satyanarayan, D. Moritz, K. Wongsuphasawat, and J. Heer, “Vega- lite: A grammar of interactive graphics,”IEEE transactions on visual- ization and computer graphics, vol. 23, no. 1, pp. 341–350, 2016

    2016

  14. [14]

    ggplot2: elegant graphics for data analysis,

    R. A. M. Villanueva and Z. J. Chen, “ggplot2: elegant graphics for data analysis,” 2019

    2019

  15. [15]

    Natural language to visualization by neural machine translation,

    Y . Luo, N. Tang, G. Li, J. Tang, C. Chai, and X. Qin, “Natural language to visualization by neural machine translation,”IEEE Transactions on Visualization and Computer Graphics, vol. 28, no. 1, pp. 217–226, 2021

    2021

  16. [16]

    nvbench 2.0: Resolving ambiguity in text-to-visualization through stepwise reasoning,

    T. Luo, C. Huang, L. Shen, B. Li, S. Shen, W. Zeng, N. Tang, and Y . Luo, “nvbench 2.0: Resolving ambiguity in text-to-visualization through stepwise reasoning,”arXiv preprint arXiv:2503.12880, 2025

    work page arXiv 2025

  17. [17]

    Viseval: A benchmark for data visualization in the era of large language models,

    N. Chen, Y . Zhang, J. Xu, K. Ren, and Y . Yang, “Viseval: A benchmark for data visualization in the era of large language models,”IEEE Transactions on Visualization and Computer Graphics, 2024

    2024

  18. [18]

    Interactive text-to- visualization: Refining visualization outputs through natural language user feedback,

    X. Xiong, R. C.-W. Wong, and Y . Song, “Interactive text-to- visualization: Refining visualization outputs through natural language user feedback,” inProceedings of the 34th ACM International Confer- ence on Information and Knowledge Management, 2025, pp. 3571–3581

    2025

  19. [19]

    Marrying dialogue systems with data visualization: Interactive data visualization generation from natural language conversations,

    Y . Song, X. Zhao, and R. C.-W. Wong, “Marrying dialogue systems with data visualization: Interactive data visualization generation from natural language conversations,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 2733– 2744

    2024

  20. [20]

    Rgvisnet: A hybrid retrieval-generation neural framework towards automatic data visualiza- tion generation,

    Y . Song, X. Zhao, R. C.-W. Wong, and D. Jiang, “Rgvisnet: A hybrid retrieval-generation neural framework towards automatic data visualiza- tion generation,” inProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022, pp. 1646–1655

    2022

  21. [21]

    Automated data visualization from natural language via large language models: An exploratory study,

    Y . Wu, Y . Wan, H. Zhang, Y . Sui, W. Wei, W. Zhao, G. Xu, and H. Jin, “Automated data visualization from natural language via large language models: An exploratory study,”Proceedings of the ACM on Management of Data, vol. 2, no. 3, pp. 1–28, 2024

    2024

  22. [22]

    nvAgent: Automated data visualization from natural language via collaborative agent workflow,

    G. Ouyang, J. Chen, Z. Nie, Y . Gui, Y . Wan, H. Zhang, and D. Chen, “nvAgent: Automated data visualization from natural language via collaborative agent workflow,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, 2025, pp. 19 534–19 567

    2025

  23. [23]

    Automatic data visualization generation from chinese natural language questions,

    Y . Ge, V . J. Wei, Y . Song, J. C. Zhang, and R. C.-W. Wong, “Automatic data visualization generation from chinese natural language questions,” inProceedings of the 2024 Joint International Conference on Computa- tional Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 2024, pp. 1889–1898

    2024

  24. [24]

    Learning to recommend visualizations from data,

    X. Qian, R. A. Rossi, F. Du, S. Kim, E. Koh, S. Malik, T. Y . Lee, and J. Chan, “Learning to recommend visualizations from data,” inProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 2021, pp. 1359–1369

    2021

  25. [25]

    Sevi: Speech-to- visualization through neural machine translation,

    J. Tang, Y . Luo, M. Ouzzani, G. Li, and H. Chen, “Sevi: Speech-to- visualization through neural machine translation,” inProceedings of the 2022 ACM SIGMOD International Conference on Management of Data, 2022, pp. 2353–2356

    2022

  26. [26]

    Datatone: Managing ambiguity in natural language interfaces for data visual- ization,

    T. Gao, M. Dontcheva, E. Adar, Z. Liu, and K. G. Karahalios, “Datatone: Managing ambiguity in natural language interfaces for data visual- ization,” inProceedings of the 28th annual acm symposium on user interface software & technology, 2015, pp. 489–500

    2015

  27. [27]

    Applying pragmatics principles for interaction with visual analytics,

    E. Hoque, V . Setlur, M. Tory, and I. Dykeman, “Applying pragmatics principles for interaction with visual analytics,”IEEE transactions on visualization and computer graphics, vol. 24, no. 1, pp. 309–318, 2017. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

    2017

  28. [28]

    Flowsense: A natural language interface for visual data exploration within a dataflow system,

    B. Yu and C. T. Silva, “Flowsense: A natural language interface for visual data exploration within a dataflow system,”IEEE transactions on visualization and computer graphics, vol. 26, no. 1, pp. 1–11, 2019

    2019

  29. [29]

    Data2vis: Automatic generation of data visualizations using sequence-to-sequence recurrent neural networks,

    V . Dibia and C ¸ . Demiralp, “Data2vis: Automatic generation of data visualizations using sequence-to-sequence recurrent neural networks,” IEEE computer graphics and applications, vol. 39, no. 5, pp. 33–46, 2019

    2019

  30. [30]

    Chat2vis: Generating data visualizations via natural language using chatgpt, codex and gpt-3 large language models,

    P. Maddigan and T. Susnjak, “Chat2vis: Generating data visualizations via natural language using chatgpt, codex and gpt-3 large language models,”Ieee Access, vol. 11, pp. 45 181–45 193, 2023

    2023

  31. [31]

    prompt4vis: prompting large language models with example mining for tabular data visualization,

    S. Li, X. Chen, Y . Song, Y . Song, C. J. Zhang, F. Hao, and L. Chen, “prompt4vis: prompting large language models with example mining for tabular data visualization,”The VLDB Journal, vol. 34, no. 4, pp. 1–26, 2025

    2025

  32. [32]

    The design and implementation of xiaoice, an empathetic social chatbot,

    L. Zhou, J. Gao, D. Li, and H.-Y . Shum, “The design and implementation of xiaoice, an empathetic social chatbot,”Computational Linguistics, vol. 46, no. 1, pp. 53–93, 2020

    2020

  33. [33]

    2020.Towards a human-like open-domain chatbot

    D. Adiwardana, M.-T. Luong, D. R. So, J. Hall, N. Fiedel, R. Thoppilan, Z. Yang, A. Kulshreshtha, G. Nemade, Y . Luet al., “Towards a human- like open-domain chatbot,”arXiv preprint arXiv:2001.09977, 2020

    work page arXiv 2001

  34. [34]

    Blenderbot 3: a deployed conversational agent that continually learns to responsibly engage,

    K. Shuster, J. Xu, M. Komeili, D. Ju, E. M. Smith, S. Roller, M. Ung, M. Chen, K. Arora, J. Laneet al., “Blenderbot 3: a deployed con- versational agent that continually learns to responsibly engage,”arXiv preprint arXiv:2208.03188, 2022

    work page arXiv 2022

  35. [35]

    Dialogue act recognition via crf-attentive structured network,

    Z. Chen, R. Yang, Z. Zhao, D. Cai, and X. He, “Dialogue act recognition via crf-attentive structured network,” inThe 41st international acm sigir conference on research & development in information retrieval, 2018, pp. 225–234

    2018

  36. [36]

    Recent advances and challenges in task-oriented dialog systems,

    Z. Zhang, R. Takanobu, Q. Zhu, M. Huang, and X. Zhu, “Recent advances and challenges in task-oriented dialog systems,”Science China Technological Sciences, vol. 63, no. 10, pp. 2011–2027, 2020

    2011

  37. [37]

    Recent neural methods on dialogue state tracking for task-oriented dialogue systems: A survey,

    V . Balaraman, S. Sheikhalishahi, and B. Magnini, “Recent neural methods on dialogue state tracking for task-oriented dialogue systems: A survey,” inProceedings of the 22nd annual meeting of the special interest group on discourse and dialogue, 2021, pp. 239–251

    2021

  38. [38]

    End-to-end task-completion neural dialogue systems,

    X. Li, Y .-N. Chen, L. Li, J. Gao, and A. Celikyilmaz, “End-to-end task-completion neural dialogue systems,” inProceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2017, pp. 733–743

    2017

  39. [39]

    A network-based end-to-end trainable task-oriented dialogue system,

    T.-H. Wen, D. Vandyke, N. Mrk ˇsi´c, M. Gasic, L. M. R. Barahona, P.- H. Su, S. Ultes, and S. Young, “A network-based end-to-end trainable task-oriented dialogue system,” inProceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, 2017, pp. 438–449

    2017

  40. [40]

    Ubar: Towards fully end-to-end task- oriented dialog system with gpt-2,

    Y . Yang, Y . Li, and X. Quan, “Ubar: Towards fully end-to-end task- oriented dialog system with gpt-2,” inProceedings of the AAAI confer- ence on artificial intelligence, vol. 35, no. 16, 2021, pp. 14 230–14 238

    2021

  41. [41]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017