Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

ProactiveVA: Proactive Visual Analytics with LLM-Based UI Agent

T0 review · 3 major / 6 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 ProactiveVA shows that an LLM-based UI agent can detect when visual analytics users are stuck and offer context-aware help without being asked, and its user study reports that assistance roughly doubled reported events (about 2 to about…

desk verdict Solid design study with a genuinely new pipeline, but the central quantitative benefit claim rests on a non-controlled cross-study comparison that needs a proper experiment. read the letter →

arxiv 2507.18165 v1 pith:B2X332QA submitted 2025-07-24 cs.HC

Yuheng Zhao , Xueli Shu , Liwen Fan , Lin Gao , Yu Zhang , Siming Chen This is my paper
classification cs.HC
keywords proactivevisualanalyticsLLM-basedUIagentmixed-initiativesystemsuserintentinferencehelp-neededeventdetectionhuman-AIcollaborationinteractionlogsintelligentassistance
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 visual analytics tools can be made proactive rather than merely reactive: an LLM-powered UI agent monitors a user's interaction log, detects moments when the user appears stuck or mistaken, infers what the user is trying to do, and offers non-intrusive, context-aware assistance. The authors ground the design in a formative study of 10 analysts, distilling when help is needed (onboarding, exploration, verification), what help should look like, and how it should be presented. They then build a three-stage perception–reasoning–acting agent and embed it in two visual analytics systems. In a user study, proactive assistance increased reported events from about 2 to about 4 per person, reported details from 5 to 11, and answer accuracy from 0.78 to 0.95. The claim matters because analysts often get lost in complex tools precisely when they do not know what to ask, and current LLM assistants only respond when explicitly invoked.

What carries the argument

The central mechanism is a three-stage LLM-based UI agent pipeline. A perception stage converts interaction logs into help-needed events using behavioral triggers (prolonged pauses, repetitive actions) and semantic triggers (errors or omissions in user notes), then infers the user's intent. A reasoning stage plans and iteratively refines exploration steps, and an acting stage executes those steps on the interface and feeds system responses back. Memory and knowledge stores carry interaction history, dataset state, task context, and available operations, while the interface gives users preview, frequency control, and toggles so the assistance stays non-intrusive.

What would settle it

Conduct a pre-registered between-subjects study on the same 15-minute event-analysis task: one group receives suggestions only when they explicitly ask, and another receives the same suggestions proactively; if reported events, details, and accuracy are statistically indistinguishable between the two groups, the claim that autonomous timing of assistance drives the improvement is falsified.

Watch

Extended reading notes

Core claim

The paper claims that help-needed moments in visual analytics can be detected automatically from interaction logs, and that acting on them improves analysis. The agent treats a pause longer than a ThinkTime threshold (default 3 seconds), repetitive actions such as toggling filters or re-clicking regions, and inconsistencies in user-authored notes as signals of difficulty. From those signals it infers intent (onboarding confusion, exploration goal, or verification error), proposes a suggestion, and, if the user accepts, executes a transparent step-by-step exploration of the interface. In the user study, the average number of reported events per person rose from about 2 to about 4, reported details from 5 to 11, and accuracy from 0.78 to 0.95; the paper concludes that proactive assistance improved breadth, depth, and accuracy of analysis.

Load-bearing premise

The framework assumes that a user who pauses for more than about three seconds or repeats the same action several times is genuinely stuck, and that this reading holds across different users and tasks; if those behaviors sometimes just reflect normal thinking or exploration, the agent will interrupt at the wrong moment.

Editorial extensions

If this is right

  • Proactive assistance can be delivered without an explicit user request, and in this study it roughly doubled reported events and details while raising accuracy from 0.78 to 0.95.
  • The same perception–reasoning–acting pipeline transferred to two different visual analytics systems, supporting the claim that the framework generalizes beyond a single bespoke tool.
  • Help-needed detection can be built from interaction-log features and LLM few-shot inference rather than from trained per-application models, which lowers the barrier to adding proactive assistance to existing visual analytics systems.
  • Interpretability and control, including preview of suggestions, adjustable frequency, toggles for assistance types, and step-by-step annotations, are part of what makes proactive intervention acceptable rather than intrusive.
  • Users adjust the intervention threshold over time, with several lowering the pause threshold from three seconds to one second and some later reverting, indicating that the level of proactivity is itself a parameter users want to control.

Reading between the lines

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

  • If pause- and repetition-based detection is the active ingredient, the same triggers could be ported to other analytic interfaces such as notebooks or business-intelligence dashboards; a testable prediction is that transfer works whenever interaction logs contain timestamps and repeated actions.
  • The user-study comparison is against the formative study's unaided baseline rather than a randomized within-subject control, so a head-to-head comparison with a reactive assistant that gives the same suggestions only on request would separate the value of timing from the value of LLM-generated content.
  • Because participants manually adjusted the pause threshold as their expertise changed, an agent that profiles the user's pace over time could reduce both mistimed interruptions and missed help.
  • The note-validation mechanism suggests a broader design: treat user-authored artifacts such as notes, labels, and summaries as first-class signals for detecting errors and omissions, which could generalize to any analysis tool where users record conclusions.
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 / 6 minor

Summary. The paper proposes ProactiveVA, an LLM-powered UI agent framework for proactive visual analytics. The authors conduct a formative study with 10 participants to identify when, what, and how to assist, derive three design requirements, and implement a three-stage agent pipeline (perception, reasoning, acting) in two VA systems: a social-media event analysis system and a Tableau profitability dashboard. The evaluation comprises a 100-task algorithm benchmark scored by an LLM and two human experts, case and expert studies, and a 10-participant user study. The main claimed result is that proactive assistance improves analysis breadth, depth, and accuracy, with reported increases from about 2 to about 4 reported events, from 5 to 11 reported details, and from 0.78 to 0.95 accuracy (Section 5.3.1, Fig. 8).

Significance. If the effectiveness claims are supported, ProactiveVA would be a meaningful contribution to mixed-initiative visual analytics, addressing the underexplored problem of proactive, context-aware assistance. The design requirements from the formative study are useful, and the two implementations demonstrate architectural generalizability. The qualitative feedback from the expert studies and the user study is encouraging, and the inclusion of an algorithm evaluation with human expert scoring is a strength. However, the central quantitative claim of improved breadth, depth, and accuracy is not yet supported by the evidence as presented, because the main comparison is uncontrolled and partially self-referential. A controlled user study or a carefully reframed descriptive analysis is necessary before the effectiveness claims can be accepted.

major comments (3)
  1. [Section 5.3.1, Fig. 8(b)] The claim that proactive assistance improves breadth, depth, and accuracy relies on comparing the ProactiveVA user study with the formative study described in Section 3. These are not equivalent conditions: the two studies used different participant pools, different onboarding procedures (user-study participants were introduced to and guided through the system, while formative-study participants freely explored before a 15-minute think-aloud), and likely different task exposure. No significance tests, effect sizes, confidence intervals, or formal definitions of 'event', 'detail', and 'accuracy' are provided. As it stands, the observed differences cannot be attributed to the proactive agent. I recommend either a controlled within- or between-subject experiment with identical protocols and independent outcome coding, or a clear recharacterization of the comparison as descriptive rather than causal.
  2. [Section 5.1, Table 1] The algorithm evaluation is partly self-referential: DeepSeek R1 generates the 100 test tasks and scores the agent, while the agent itself is built on DeepSeek-V3, a model from the same family. This raises a risk of correlated bias between the task designer/judge and the system being evaluated. The claim that 'the LLM, as an automated evaluation mechanism, is a feasible option' is supported only by the closeness of average scores between LLM and human evaluation, with no per-task agreement, inter-rater reliability statistic, or significance test. Reporting exact-match rates or Cohen's kappa, and ideally using tasks from human analysts and a judge from a different model family, would substantially strengthen this evaluation.
  3. [Section 4.1, Fig. 3 and Section 5.3.2] The help-needed detector relies on a hand-set ThinkTime threshold (default 3 seconds) inferred from the 10-participant formative study. The user study reports that 6 of 10 participants adjusted this threshold during the session, which suggests the fixed threshold is not universally appropriate. No sensitivity analysis is provided, and the threshold is a free parameter in the system. Without evidence that the chosen value generalizes across users, tasks, and the two implemented systems, the claim that the detected events reliably indicate user confusion rather than normal exploration pauses remains under-supported.
minor comments (6)
  1. [Section 5.3.1] There is a typo in 'The agent’ recommended suggestion'; it should read 'The agent’s recommended suggestion'.
  2. [Section 5.2.1] The case study introduces the expert with female pronouns ('her') but later refers to 'Bob accepted the suggestion'; the pronoun/name inconsistency should be corrected.
  3. [Section 3.2] The text says the average thinkTime is 'around 2–3 seconds,' but the system threshold is set to 3 seconds; please clarify how the threshold was derived from the observed distribution (e.g., mean plus a multiple of standard deviation) and specify the units.
  4. [Section 5.1] The term 'Correlation & Impact Analysis' appears in the prose while Table 1 uses 'Correlation'; please unify the terminology.
  5. [Abstract and Section 4] The abstract uses 'LLM-powered UI agent' while Section 4 uses 'LLM-based UI agent'; please choose one phrasing and apply it consistently.
  6. [Section 5.3] The text states there are 'eleven effectiveness-related questions' but Fig. 8 shows only Q1 through Q9; please reconcile the number.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims rest on empirical studies and independent human evaluation, not on definitions that reduce to their inputs.

full rationale

ProactiveVA is a system-building paper whose derivation chain is: formative study -> design requirements -> three-stage agent implementation -> evaluation. None of the headline claims is equivalent to an input by construction. The help-needed detector uses a hand-set ThinkTime threshold (default 3 seconds) informed by the Fig. 3 observation that average thinkTime is around 2-3 seconds, but this is an implementation trigger, not a fitted parameter disguised as a prediction; the user study then separately measures behavior with the full system. The quantitative benefit claim in Sec. 5.3.1 compares the ProactiveVA user study against the earlier formative study (Fig. 8b). That comparison is methodologically weak (different participant pools, different onboarding, no inferential statistics), but it is not circular: the 'without assistance' numbers are not derived from the 'with assistance' results, nor vice versa; the report is an uncontrolled empirical contrast, which is a validity concern, not a self-referential reduction. The algorithm-evaluation section uses DeepSeek R1 to generate 100 tasks and to score the agent, while the agent itself runs on DeepSeek-V3; this same-family overlap is a benchmark-validity limitation and is acknowledged in Sec. 6.2 ('LLM-based evaluations, while efficient, may misinterpret user intent...'). It does not make the claimed agent capability reduce to its input, and the paper also reports independent human expert scoring (Sec. 5.1: 'have two data analysis experts evaluate with the same criteria'), which closely tracks the LLM scores. Citations to the authors' prior work, e.g., alignment of assistance types with LEVA [78] and use of the VAST-challenge system [44], are ordinary related-work and system-reuse references, not load-bearing uniqueness theorems that force the design. No equation or fitted value is renamed as a prediction, and no central result is imported solely from a self-citation. The main empirical concerns (confounded comparison, same-family LLM judge) belong to correctness risk, not circularity, under the stated definition.

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

The framework rests on a small number of design choices: a hand-set pause threshold, the choice of DeepSeek-V3 as the LLM, and several generalization assumptions from a 10-person formative study. No parameters are fit to optimize evaluation outcomes, and no novel physical or conceptual entities are introduced.

free parameters (1)
  • ThinkTime threshold = 3 seconds (default), user-adjustable 1-3 s
    Pauses beyond this threshold trigger help-needed event detection (Section 4.1, Section 5.3.2). The value is hand-set based on the observed average thinkTime of 2-3 s in the formative study, not fit to outcomes.
assumptions (4)
  • domain assumption Prolonged pauses and repetitive interactions reliably indicate user confusion or help-needed moments.
    Used as the core trigger for proactive detection in Section 4.1; derived from the 10-participant formative study (Fig. 3) and prior bursty interaction research, but not validated as a general signal.
  • domain assumption DeepSeek-V3 can reliably perceive UI states, reason about analytical tasks, and execute interface operations through the defined tool set.
    The whole acting pipeline (Section 4.2) depends on the LLM's tool-use reliability; errors are handled by the action module but no quantitative reliability analysis is provided.
  • domain assumption The formative study findings generalize to other VA systems and user populations.
    Design requirements DR1-DR3 are distilled from one system and 10 experienced postgraduate users (Section 3), then applied to a different Tableau dashboard in Case 2.
  • ad hoc to paper DeepSeek R1-generated tasks and scores are a valid proxy for real user tasks and expert judgment in the algorithm evaluation.
    Section 5.1 uses LLM generation and LLM scoring to assess the agent, which is a pragmatic but unvalidated method; the authors acknowledge LLM evaluation may misinterpret intent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ProactiveVA: Proactive Visual Analytics with LLM-Based UI Agent." pith.science (2026). https://pith.science/paper/B2X332QA

@misc{pith2026250718165,
  author       = {Pith},
  title        = {Pith review of: ProactiveVA: Proactive Visual Analytics with LLM-Based UI Agent},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2X332QA}},
  note         = {Machine review of arXiv:2507.18165}
}
read the original abstract

Visual analytics (VA) is typically applied to complex data, thus requiring complex tools. While visual analytics empowers analysts in data analysis, analysts may get lost in the complexity occasionally. This highlights the need for intelligent assistance mechanisms. However, even the latest LLM-assisted VA systems only provide help when explicitly requested by the user, making them insufficiently intelligent to offer suggestions when analysts need them the most. We propose a ProactiveVA framework in which LLM-powered UI agent monitors user interactions and delivers context-aware assistance proactively. To design effective proactive assistance, we first conducted a formative study analyzing help-seeking behaviors in user interaction logs, identifying when users need proactive help, what assistance they require, and how the agent should intervene. Based on this analysis, we distilled key design requirements in terms of intent recognition, solution generation, interpretability and controllability. Guided by these requirements, we develop a three-stage UI agent pipeline including perception, reasoning, and acting. The agent autonomously perceives users' needs from VA interaction logs, providing tailored suggestions and intuitive guidance through interactive exploration of the system. We implemented the framework in two representative types of VA systems, demonstrating its generalizability, and evaluated the effectiveness through an algorithm evaluation, case and expert study and a user study. We also discuss current design trade-offs of proactive VA and areas for further exploration.

Figures

Figures reproduced from arXiv: 2507.18165 by the authors.

Figure 1
Figure 1. Illustration of three different visual analytics paradigms. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. This table shows various user interaction patterns identified during the study, along with the interpretations, problem subcategories and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of user thinking time during analysis. (a) Box plot shows distribution of ThinkTime (in seconds) for different action types, with outliers indicating variability in decision-making. (b) A heatmap shows interactions across 30 steps, with each dark square indicating a longer thinking time than that type of action’s average thinking time. Inspired by this, we examined the distribution of thinkTime in our … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The workflow of ProactiveVA. The agent is responsible for detecting help-needed events (c), inferring user intent and providing suggestions (d), analyzing data to decide the next operation (e), performing actions on the interface (f), and storing history to facilitate …
Figure 5
Figure 5. Figure 5: The overview interface of our UI agent-assisted system: (A) Chat View displays proactive suggestions from the agent alongside its reasoning [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Three cases of proactive agent interventions: (a) onboarding guidance for data filtering, (b) error correction in event analysis, and (c) [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: The results of quantitative evaluation of the effectiveness. (a) [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 7
Figure 7. Figure 7: Case 2: Integrating ProactiveVA into Tableau Sales Data Analysis [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

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. ProEvent: An Event-centric Benchmark for Proactive Agents

    cs.AI 2026-07 conditional novelty 6.0 of 10

    ProEvent is a benchmark showing LLM agents keep a user's event timetable from chats poorly, with the best fully-correct score at 27.2%.

Reference graph

Works this paper leans on

80 extracted references · 65 canonical work pages · cited by 1 Pith paper

  1. [1]

    Amershi, M

    S. Amershi, M. Cakmak, W. B. Knox, and T. Kulesza. Power to the People: The role of humans in interactive machine learning. AI magazine, 35(4):105–120, 2014. 9

    work page 2014

  2. [2]

    Andrienko, G

    N. Andrienko, G. Andrienko, S. Chen, and B. Fisher. Seeking patterns of visual pattern discovery for knowledge building. Computer Graphics Forum, 41(6):124–148, 2022. 2

    work page 2022

  3. [3]

    Battle and J

    L. Battle and J. Heer. Characterizing exploratory visual analysis: A litera- ture review and evaluation of analytic provenance in tableau. Computer Graphics Forum, 38(3):145–159, June 2019. 2, 4

    work page 2019

  4. [4]

    Boukhelifa, E

    N. Boukhelifa, E. Lutton, and A. Bezerianos. A case study of using analytic provenance to reconstruct user trust in a guided visual analytics system. In 2021 IEEE Workshop on TRust and EXpertise in Visual Analytics , TREX’21, pp. 45–51. IEEE, Oct. 2021. 2

    work page 2021

  5. [5]

    Brehmer and T

    M. Brehmer and T. Munzner. A multi-level typology of abstract visualiza- tion tasks. IEEE Transactions on Visualization and Computer Graphics, 19(12):2376–2385, Dec. 2013. 2

    work page 2013

  6. [6]

    E. T. Brown, A. Ottley, H. Zhao, Q. Lin, R. Souvenir, A. Endert, and R. Chang. Finding waldo: Learning about users from their interactions. IEEE Transactions on Visualization and Computer Graphics, 20(12):1663– 1672, Dec. 2014. 2

    work page 2014

  7. [7]

    Carenini, C

    G. Carenini, C. Conati, E. Hoque, B. Steichen, D. Toker, and J. Enns. High- lighting interventions and user differences: Informing adaptive information visualization support. In In proceedings of the 2014 CHI Conference on Human Factors in Computing Systems, CHI ’14, pp. 1835–1844, 2014. 9

    work page 2014

  8. [8]

    Ceneda, T

    D. Ceneda, T. Gschwandtner, T. May, S. Miksch, H.-J. Schulz, M. Streit, and C. Tominski. Characterizing guidance in visual analytics. IEEE transactions on visualization and computer graphics, 23(1):111–120, 2016. 1, 9

    work page 2016

Show all 80 references
  1. [9]

    V . Chen, A. Zhu, S. Zhao, H. Mozannar, D. Sontag, and A. Tal- walkar. Need help? designing proactive ai assistants for programming. arXiv:2410.04596, Oct. 2024. 2, 3, 4

    2024 arXiv

  2. [10]

    Conati, S

    C. Conati, S. Lallé, M. A. Rahman, and D. Toker. Comparing and com- bining interaction data and eye-tracking data for the real-time prediction of user cognitive abilities in visualization tasks. ACM Transactions on Interactive Intelligent Systems (TiiS), 10(2):1–41, 2020. 9

    2020

  3. [11]

    K. Cook, N. Cramer, D. Israel, M. Wolverton, J. Bruce, R. Burtner, and A. Endert. Mixed-initiative visual analytics using task-driven recom- mendations. In 2015 IEEE Conference on Visual Analytics Science and Technology (VAST), V AST ’15, pp. 9–16. IEEE, Oct. 2015. 1, 2

    2015

  4. [12]

    Z. Cui, S. K. Badam, A. Yalçin, and N. Elmqvist. DataSite: Proactive vi- sual data exploration with computation of insight-based recommendations. arXiv:1802.08621, Sept. 2018. 1, 2

    2018 arXiv

  5. [13]

    Dabek and J

    F. Dabek and J. J. Caban. A grammar-based approach for modeling user interactions and generating suggestions during the data exploration process. IEEE Transactions on Visualization and Computer Graphics, 23(1):41–50, Jan. 2017. 1, 2

    2017

  6. [14]

    Y . Deng, W. Lei, W. Lam, and T.-S. Chua. A survey on proactive dialogue systems: Problems, methods, and prospects. arXiv:2305.02750, May 2023. 2

    2023 arXiv

  7. [15]

    Y . Deng, L. Liao, Z. Zheng, G. H. Yang, and T.-S. Chua. Towards human- centered proactive conversational agents. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR’21, pp. 807–818. ACM, July 2024. 4

    2024

  8. [16]

    V . Dibia. LIDA: A tool for automatic generation of grammar- agnostic visualizations and infographics using large language models. arXiv:2303.02927, Mar. 2023. 2

    2023 arXiv

  9. [17]

    L. Gao, J. Lu, Z. Shao, Z. Lin, S. Yue, C. Ieong, Y . Sun, R. J. Zauner, Z. Wei, and S. Chen. Fine-tuned large language model for visualization sys- tem: A study on self-regulated learning in education. arXiv:2407.20570, July 2024. 2

    2024 arXiv

  10. [18]

    Gotz and Z

    D. Gotz and Z. Wen. Behavior-driven visualization recommendation. In Proceedings of the 14th International Conference on Intelligent User Interfaces, pp. 315–324. ACM, Feb. 2009. 2, 4

    2009

  11. [19]

    Gotz and M

    D. Gotz and M. X. Zhou. Characterizing users’ visual analytic activity for insight provenance. Information Visualization, 8(1):42–55, Jan. 2009. 2, 9

    2009

  12. [20]

    A. M. Grant and S. J. Ashford. The dynamics of proactivity at work. Research in organizational behavior, 28:3–34, 2008. 2

    2008

  13. [21]

    H. Guo, S. R. Gomez, C. Ziemkiewicz, and D. H. Laidlaw. A case study using visualization interaction logs and insight metrics to understand how analysts arrive at insights. IEEE Transactions on Visualization and Computer Graphics, 22(1):51–60, Jan. 2016. 2

    2016

  14. [22]

    P. J. Guo, S. Kandel, J. M. Hellerstein, and J. Heer. Proactive Wrangling: Mixed-initiative end-user programming of data transformation scripts. In Proceedings of the 24th Annual ACM Symposium on User Interface Software and Technology, UIST ’24, pp. 65–74. ACM, Oct. 2011. 2

    2011

  15. [23]

    D. Hagen. Superstore dashboard - profitability overview. https://public.tableau.com/app/profile/dave.hagen/viz/ Superstore_embedded_800x800/Overview, Jan 2023. Accessed: Jul. 2, 2025. 6, 7

    2023

  16. [24]

    J. Han, W. Buntine, and E. Shareghi. Towards uncertainty-aware language agent. arXiv preprint arXiv:2401.14016, 2024. 9

    2024 arXiv

  17. [25]

    Harrison, K

    G. Harrison, K. Bryson, A. E. B. Bamba, L. Dovichi, A. H. Binion, A. Borem, and B. Ur. Jupyterlab in retrograde: Contextual notifications that highlight fairness and bias issues for data scientists. In Proceedings of the CHI Conference on Human Factors in Computing Systems, CH...

    2024

  18. [26]

    Heer and B

    J. Heer and B. Shneiderman. Interactive dynamics for visual analysis. Commun. ACM, 55(4):45–54, 10 pages, Apr. 2012. 1, 3

    2012

  19. [27]

    E. Horvitz. Principles of mixed-initiative user interfaces. InIn proceedings of the 1999 CHI Conference on Human Factors in Computing Systems , CHI ’99, pp. 159–166. ACM Press, 1999. 1, 2

    1999

  20. [28]

    Kadoma, M

    K. Kadoma, M. Aubin Le Quere, X. J. Fu, C. Munsch, D. Metaxa, and M. Naaman. The role of inclusion, control, and ownership in workplace ai-mediated communication. In Proceedings of the CHI Conference on Human Factors in Computing Systems, CHI ’24, pp. 1–10, 2024. 2

    2024

  21. [29]

    D. A. Keim, F. Mansmann, J. Schneidewind, J. Thomas, and H. Ziegler. Visual Analytics: Scope and challenges. In S. J. Simoff, M. H. Böhlen, and A. Mazeika, eds., Visual Data Mining, vol. 4404, pp. 76–90. Springer Berlin Heidelberg, 2008. 1, 2

    2008

  22. [30]

    J. Kil, C. H. Song, B. Zheng, X. Deng, Y . Su, and W.-L. Chao. Dual-view visual contextualization for web navigation. arXiv:2402.04476, Mar. 2024. 9

    2024 arXiv

  23. [31]

    J. Kim, S. Lee, H. Jeon, K.-J. Lee, H.-J. Bae, B. Kim, and J. Seo. Phe- noFlow: A human-llm driven visual analytics system for exploring large and complex stroke datasets. IEEE Transactions on Visualization and Computer Graphics, pp. 1–11, 2024. 1

    2024

  24. [32]

    H. Lai, X. Liu, I. L. Iong, S. Yao, Y . Chen, P. Shen, H. Yu, H. Zhang, X. Zhang, Y . Dong, and J. Tang. AutoWebGLM: A large language model- based web navigating agent. arXiv:2404.03648, Oct. 2024. 2

    2024 arXiv

  25. [33]

    Lallé, C

    S. Lallé, C. Conati, and G. Carenini. Predicting confusion in information visualization from eye tracking and interaction data. In IJCAI, pp. 2529– 2535, 2016. 9

    2016

  26. [34]

    Y . Li, Y . Qi, Y . Shi, Q. Chen, N. Cao, and S. Chen. Diverse interaction recommendation for public users exploring multi-view visualization us- ing deep learning. IEEE Transactions on Visualization and Computer Graphics, pp. 1–11, 2022. 1, 2

    2022

  27. [35]

    Y . Lin, H. Li, L. Yang, A. Wu, and H. Qu. InkSight: Leveraging sketch interaction for documenting chart findings in computational notebooks. IEEE Transactions on Visualization and Computer Graphics, 30(1):944– 954, 2023. 9

    2023

  28. [36]

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 4

    2024 arXiv

  29. [37]

    S. Liu, H. Miao, Z. Li, M. Olson, V . Pascucci, and P.-T. Bremer. A V A: Towards autonomous visualization agents through visual perception-driven decision-making. In Computer Graphics Forum, vol. 43, p. e15093. Wiley Online Library, 2024. 2

    2024

  30. [38]

    T. Liu, H. Zhao, Y . Liu, X. Wang, and Z. Peng. ComPeer: A generative conversational agent for proactive peer support. UIST ’24: Proceedings of the 37th Annual ACM Symposium on User Interface Software and Technology, pp. 1–22, 2024. 2

    2024

  31. [39]

    X. B. Liu, S. Fang, W. Shi, C.-S. Wu, T. Igarashi, and X. A. Chen. Proactive conversational agents with inner thoughts. arXiv:2501.00383, Dec. 2024. 2

    2024 arXiv

  32. [40]

    Y . Lu, S. Yang, C. Qian, G. Chen, Q. Luo, Y . Wu, H. Wang, X. Cong, Z. Zhang, Y . Lin, W. Liu, Y . Wang, Z. Liu, F. Liu, and M. Sun. Proactive Agent: Shifting llm agents from reactive responses to active assistance. arXiv:2410.12361, Dec. 2024. 2, 9

    arXiv

  33. [41]

    B. Min, A. Chen, Y . Cao, and H. Xia. Malleable overview-detail inter- faces. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, article no. 688, 25 pages. Association for Computing Machinery, 2025. 9

    2025

  34. [42]

    Musleh, R

    M. Musleh, R. G. Raidou, and D. Ceneda. TrustME: A context-aware explainability model to promote user trust in guidance. IEEE Transactions on Visualization and Computer Graphics, 2025. 4

    2025

  35. [43]

    J. Oh, W. Kim, S. Kim, H. Im, and S. Lee. Better to Ask Than As- sume: Proactive voice assistants’ communication strategies that respect user agency in a smart home environment. In Proceedings of the CHI Conference on Human Factors in Computing Systems, CHI ’24, pp. 1–17. ACM,...

    2024

  36. [44]

    L. Peng, Y . Zhao, Y . Hou, Q. Wang, S. Shen, X. Lai, J. Gao, J. Dong, Z. Lin, and S. Chen. Mixed-initiative visual exploration of social media text and events. In Proceedings of the IEEE Conference on Visualization and Visual Analytics, VIS ’21, 2021. 2, 6

    2021

  37. [45]

    Z. Peng, Y . Kwon, J. Lu, Z. Wu, and X. Ma. Design and evaluation of service robot’s proactivity in decision-making support process. In proceedings of the 2019 CHI conference on human factors in computing systems, CHI ’19, pp. 1–13, 2019. 2

    2019

  38. [46]

    Pérez-Messina, M

    I. Pérez-Messina, M. Angelini, D. Ceneda, C. Tominski, and S. Miksch. Coupling guidance and progressiveness in visual analytics. In Computer Graphics Forum, p. e70115. Wiley Online Library, 2025. 4, 9

    2025

  39. [47]

    Pérez-Messina, D

    I. Pérez-Messina, D. Ceneda, and S. Miksch. Guided visual analytics for image selection in time and space. IEEE Transactions on Visualization and Computer Graphics, 30(1):66–75, 2023. 2

    2023

  40. [48]

    Pister, P

    A. Pister, P. Buono, J.-D. Fekete, C. Plaisant, and P. Valdivia. Integrating prior knowledge in mixed-initiative social network clustering. IEEE Transactions on Visualization and Computer Graphics, 27(2):1775–1785, Feb. 2021. 2

    2021

  41. [49]

    C. O. Retzlaff, S. Das, C. Wayllace, P. Mousavi, M. Afshari, T. Yang, A. Saranti, A. Angerschmid, M. E. Taylor, and A. Holzinger. Human-in- the-loop reinforcement learning: A survey and position on requirements, challenges, and opportunities. Journal of Artificial Intelligence...

    2024

  42. [50]

    K. T. Rosenberg, R. H. Kazi, L.-Y . Wei, H. Xia, and K. Perlin. DrawTalking: Building interactive worlds by sketching and speaking. arXiv:2401.05631, Aug. 2024. 9

    2024 arXiv

  43. [51]

    S. I. Ross, F. Martinez, S. Houde, M. Muller, and J. D. Weisz. The programmer’s assistant: Conversational interaction with a large language model for software development. In Proceedings of the 28th International Conference on Intelligent User Interfaces, pp. 491–514, 2023. 2

    2023

  44. [52]

    S. Shin, I. Na, and N. Elmqvist. Drillboards: Adaptive visualization dashboards for dynamic personalization of visualization experiences. arXiv:2410.12744, Oct. 2024. 4

    2024 arXiv

  45. [53]

    Shneiderman

    B. Shneiderman. Human-centered AI. Oxford University Press, 2022. 9

    2022

  46. [54]

    Sperrle, D

    F. Sperrle, D. Ceneda, and M. El-Assady. Lotse: A practical framework for guidance in visual analytics. IEEE Transactions on Visualization and Computer Graphics, 29(1):1124–1134, 2023. 9

    2023

  47. [55]

    Sperrle, H

    F. Sperrle, H. Schäfer, D. Keim, and M. El-Assady. Learning contextual- ized user preferences for co-adaptive guidance in mixed-initiative topic model refinement. Computer Graphics Forum, 40(3):215–226, 2021. 1, 2

    2021

  48. [56]

    Srinivasan, J

    A. Srinivasan, J. Purich, M. Correll, L. Battle, V . Setlur, and A. Crisan. From dashboard zoo to census: A case study with tableau public. IEEE Transactions on Visualization and Computer Graphics, pp. 1–15, 2024. 5, 7

    2024

  49. [57]

    Steichen, G

    B. Steichen, G. Carenini, and C. Conati. User-adaptive information visual- ization: using eye gaze data to infer visualization tasks and user cognitive abilities. In Proceedings of the 2013 International Conference on Intelli- gent User Interfaces, IUI ’23, pp. 317–328, 2013. 9

    2013

  50. [58]

    Stoiber, D

    C. Stoiber, D. Ceneda, M. Wagner, V . Schetinger, T. Gschwandtner, M. Streit, S. Miksch, and W. Aigner. Perspectives of visualization on- boarding and guidance in va. Visual Informatics, 6(1):68–83, Mar. 2022. 4

    2022

  51. [59]

    C. D. Stolper, A. Perer, and D. Gotz. Progressive visual analytics: User- driven visual exploration of in-progress analytics. IEEE Transactions on Visualization and Computer Graphics, 20(12):1653–1662, Dec. 2014. 4

    2014

  52. [60]

    Subramonyam, C

    H. Subramonyam, C. Seifert, P. Shah, and E. Adar. texSketch: Active diagramming through pen-and-ink annotations. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, CHI ’20, pp. 1–13. ACM, Apr. 2020. 9

    2020

  53. [61]

    Tabalba, C

    R. Tabalba, C. J. Lee, G. Tran, N. Kirshenbaum, and J. Leigh. Articu- latePro: A comparative study on a proactive and non-proactive assistant in a climate data exploration task. arXiv:2409.10797, Sept. 2024. 2, 3

    2024 arXiv

  54. [62]

    Tableau embedding api

    Tableau. Tableau embedding api. https://help.tableau.com/ current/api/embedding_api/en-us/index.html, 2003. Accessed: Jul. 2, 2025. 7

    2003

  55. [63]

    Tankelevitch, V

    L. Tankelevitch, V . Kewenig, A. Simkute, A. E. Scott, A. Sarkar, A. Sellen, and S. Rintel. The metacognitive demands and opportunities of generative ai. In Proceedings of the CHI Conference on Human Factors in Computing Systems, CHI ’24, pp. 1–24. ACM. 9

  56. [64]

    J. Thomas. Illuminating the Path: The Research and Development Agenda for Visual Analytics. IEEE, 2005. 1, 2

    2005

  57. [65]

    Y . Tian, W. Cui, D. Deng, X. Yi, Y . Yang, H. Zhang, and Y . Wu. Chart- GPT: Leveraging llms to generate charts from abstract natural language. arXiv:2311.01920, Nov. 2023. 2

    2023 arXiv

  58. [66]

    E. Wall, A. Arcalgud, K. Gupta, and A. Jo. A markov model of users’ interactive behavior in scatterplots. In 2019 IEEE visualization conference, pp. 81–85. IEEE, 2019. 1

    2019

  59. [67]

    L. Weng, X. Wang, J. Lu, Y . Feng, Y . Liu, H. Feng, D. Huang, and W. Chen. InsightLens: Augmenting llm-powered data analysis with interactive insight management and navigation. IEEE Transactions on Visualization and Computer Graphics, 2025. 2

    2025

  60. [68]

    T. Wu, E. Jiang, A. Donsbach, J. Gray, A. Molina, M. Terry, and C. J. Cai. PromptChainer: Chaining large language model prompts through visual programming. In Extended Abstracts of the 2022 CHI Conference on Human Factors in Computing Systems, CHI EA ’22, article no. 359, 10 p...

    2022

  61. [69]

    Y . Wu, J. Zhang, N. Hu, L. Tang, G. Qi, J. Shao, J. Ren, and W. Song. MLDT: Multi-level decomposition for complex long-horizon robotic task planning with open-source large language model. arXiv:2403.18760, Apr

    arXiv

  62. [70]

    K. Xu, A. Ottley, C. Walchshofer, M. Streit, R. Chang, and J. Wenskovitch. Survey on the analysis of user interactions and visualization provenance. Computer Graphics Forum, 39(3):757–783, June 2020. 2

    2020

  63. [71]

    Yanez, C

    F. Yanez, C. Conati, A. Ottley, and C. Nobre. The state of the art in user-adaptive visualizations. In Computer Graphics Forum, vol. 44, p. e15271. Wiley Online Library, 2025. 4, 9

    2025

  64. [72]

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao. ReAct: Synergizing reasoning and acting in language models. arXiv:2210.03629, Mar. 2023. 5

    2023 arXiv

  65. [73]

    Zargham, L

    N. Zargham, L. Reicherts, M. Bonfert, S. T. V oelkel, J. Schoening, R. Malaka, and Y . Rogers. Understanding circumstances for desirable proactive behaviour of voice assistants: The proactivity dilemma. In Pro- ceedings of the 4th Conference on Conversational User Interfaces ,...

    2022

  66. [74]

    Zhang, L

    C. Zhang, L. Li, S. He, X. Zhang, B. Qiao, S. Qin, M. Ma, Y . Kang, Q. Lin, S. Rajmohan, et al. UFO: A ui-focused agent for windows os interaction. arXiv preprint arXiv:2402.07939, 2024. 2

    2024 arXiv

  67. [75]

    Zhang, M

    G. Zhang, M. A. N. Ahmed, Z. Hu, and A. Bulling. SummAct: Uncov- ering user intentions through interactive behaviour summarisation. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, pp. 1–17. ACM, Apr. 2025. 9

    2025

  68. [76]

    Zhang, Y

    X. Zhang, Y . Deng, Z. Ren, S.-K. Ng, and T.-S. Chua. Ask-before-plan: Proactive language agents for real-world planning. arXiv:2406.12639, June 2024. 2

    2024 arXiv

  69. [77]

    Y . Zhao, J. Wang, L. Xiang, X. Zhang, Z. Guo, C. Turkay, Y . Zhang, and S. Chen. LightV A: Lightweight visual analytics with llm agent-based task planning and execution. IEEE Transactions on Visualization and Computer Graphics, pp. 1–13, 2024. 2

    2024

  70. [78]

    Y . Zhao, Y . Zhang, Y . Zhang, X. Zhao, J. Wang, Z. Shao, C. Turkay, and S. Chen. LEV A: Using large language models to enhance visual analytics. IEEE Transactions on Visualization and Computer Graphics, pp. 1–17,

  71. [79]

    Zhao and Z

    Z. Zhao and Z. Dou. Generating multi-turn clarification for web informa- tion seeking. In Proceedings of the ACM on Web Conference 2024, WWW ’24, pp. 1539–1548, 2024. 2

    2024

  72. [80]

    S. Zhou, F. F. Xu, H. Zhu, X. Zhou, R. Lo, A. Sridhar, X. Cheng, T. Ou, Y . Bisk, D. Fried, U. Alon, and G. Neubig. WebArena: A realistic web environment for building autonomous agents. arXiv:2307.13854, Apr

    arXiv

Pith tools

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