Pith. sign in

REVIEW 3 major objections 4 minor 54 references

IDEA: Augmenting Design Intelligence through Design Space Exploration

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A framework that treats design as a search over discrete choices produces higher-quality outputs than direct LLM generation in two evaluated domains.

desk verdict A well-specified LLM-plus-search pipeline with a real construct-validity hole in the headline user study: the baseline drops both the ASP constraint semantics and the reasoning LLM, so the "superior design outcomes" claim is overstated, though the core idea is worth peer review. read the letter →

arxiv 2506.10587 v1 pith:YDP77V43 submitted 2025-06-12 cs.HC

Chuer Chen , Xiaoke Yan , Xiaoyu Qi , Nan Cao This is my paper · ORCID
classification cs.HC
keywords designspaceexplorationMonteCarloTreeSearchlargelanguagemodelsAnswerSetProgrammingconstraint-guideddata-drivenarticlespictorialvisualizationintelligent
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 sets out to show that design tasks can be reframed as search problems over a formally defined design space and that, within such a space, combining LLM-generated constraints with constraint-guided Monte Carlo Tree Search yields better design outcomes than asking an LLM to generate a design directly. The claim is supported in two domains: data-driven article composition and pictorial visualization generation, where the framework's decisions are explicit selections from named dimensions and elements. A sympathetic reader would take the contribution to be a reusable pipeline—define dimensions, elements, and action functions, then let LLM-written constraints steer search—that produces inspectable design decisions rather than opaque end-to-end generation.

What carries the argument

The central object is the structured design space $S=\{D_1,\dots,D_n\}$ with each $D_i$ an orthogonal dimension over discrete elements, together with the constraint-guided reward function $r_c(P)=-\alpha\cdot V_h/|C_h| + \beta\cdot V_p/|C_p| - \gamma\cdot V_n/|C_n|$ combined with the quantity penalty $r_q(P)$ into $\mathrm{reward}(P)=r_c(P)+\delta\cdot r_q(P)$. This reward, evaluated by an ASP solver against the LLM-generated constraints, drives the UCT selection in Eq. (9), so the search quality depends on how faithfully the hand-tuned weights ($\alpha=20$, $\beta=10$, $\gamma=1$, $\delta=0.5$, $c=5$) encode design quality.

What would settle it

Apply IDEA to a third domain, such as interior design or poster layout, using the same weights and constraints pipeline, and compare against the direct-LLM baseline: a failure to outperform would falsify the cross-domain adaptability claim. Alternatively, re-run the pictorial-visualization user study with the reward weights changed substantially (for instance, setting all weights equal) or with a random reward function; if IDEA still wins, the reward model is not what carries the result, and if it stops winning, the result hinges on the specific tuned weights.

Watch

Extended reading notes

Core claim

IDEA formalizes a design space as a set of orthogonal dimensions, each containing discrete selectable elements, with a design solution being a choice of elements and a set of action functions turning that choice into a concrete artifact. An LLM translates the user requirement and contextual information into constraints written in Answer Set Programming: hard prohibitions, soft positive recommendations, and soft negative warnings. These constraints feed a reward function, and Monte Carlo Tree Search selects elements by maximizing a constraint-aware UCT score; the best path is then materialized by domain-specific actions. The paper reports that this pipeline reaches near-optimal average rewards (9.692 out of 10) on synthetic constraint sets with the highest validity ratio among the compared search algorithms, and that in a user study IDEA's pictorial visualizations were preferred over a direct-LLM baseline in 64.8% of comparisons versus 20.8% for the baseline.

Load-bearing premise

The load-bearing premise is that the hand-tuned reward weights and the LLM-generated constraints make the numeric reward a faithful proxy for design quality in any new domain; the paper gives no sensitivity analysis showing the results hold for other weight values.

Editorial extensions

If this is right

  • IDEA can be ported to a new design domain by defining dimensions, elements, and action functions, with the same LLM-constraint-plus-MCTS pipeline producing decisions without retraining.
  • Because every design choice is an explicit element selection, the framework yields an inspectable and editable decision trace, unlike end-to-end generative models.
  • Search-based design generation can outperform a strong direct-LLM baseline on accuracy, requirement consistency, and overall preference, as measured in the pictorial-visualization user study.
  • Constraint-guided MCTS achieves higher reward and better constraint compliance than genetic algorithms, simulated annealing, and beam search on the evaluated constraint sets.

Reading between the lines

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

  • A natural extension the paper leaves implicit is to use the search tree itself as an explanation interface, showing which elements were pruned by hard constraints and why, to support human review and trust.
  • If the reward weights were calibrated per domain or learned from user feedback instead of hand-tuned, the same machinery could become a general optimization backbone for generative design tools.
  • A third-domain test, such as the interior-design scenario the paper sketches, would be the cleanest check of the adaptability claim; failure there would localize the result to the two evaluated spaces.
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 / 4 minor

Summary. The paper proposes a formalization of design spaces as orthogonal dimensions with discrete selectable elements, and introduces IDEA, a three-module pipeline: an LLM generates ASP-style constraints from user requirements, a constraint-guided Monte Carlo Tree Search explores the design space, and domain-specific action functions instantiate the selected solution. The framework is demonstrated in two domains: data-driven article composition and pictorial visualization generation. Evaluation includes example outputs, interviews with five domain experts, a quantitative search-algorithm comparison on a synthetic design space, and a user study (N=60) in which participants compared IDEA output against a single-LLM baseline. The paper's central claim is that IDEA reliably produces superior design outcomes and generalizes across domains.

Significance. If the central claim is upheld, IDEA would be a genuinely reusable and inspectable alternative to end-to-end generative design systems, and the structured design-space model could be a useful foundation for further work. The paper's strengths include a clean mathematical formulation (Eqs. 1-4), a concrete instantiation in two distinct domains, and a user study with a standard non-parametric test. The expert interviews add qualitative grounding for the system's usability and decision quality. However, the key evidence for 'superior design outcomes' is currently undermined by a construct-validity threat in the user-study baseline, and the search-algorithm comparison is circular with respect to design quality because it optimizes and evaluates with the same hand-designed reward. These are fixable with additional experiments and more cautious claims, but they are load-bearing for the abstract's central assertion.

major comments (3)
  1. [Section 6.3.2] The user study compares IDEA against a baseline that differs in at least three respects: (i) IDEA uses the ASP constraint-semantics prompting step while the baseline excludes it, (ii) IDEA generates constraints with DeepSeek-R1 and only invokes Claude 3.7 Sonnet for HTML synthesis, whereas the baseline uses Claude 3.7 Sonnet for both element selection and rendering, and (iii) IDEA performs MCTS search while the baseline selects elements directly from a single prompt. The reported preference for IDEA (64.8% vs. 20.8%) therefore conflates richer prompting, LLM choice, and search. Since the abstract credits IDEA with 'superior design outcomes,' this comparison does not support the attribution of that superiority to the framework's MCTS component. Please run an ablation that fixes the constraint-generation prompt and the reasoning LLM across conditions and varies only whether MCTS search is applied; if that experiment is infeasible, the causal claims should be softened to describe a holistic system comparison rather than the effect of search.
  2. [Section 4.5, Table 2] The quantitative comparison of MCTS against GA, SA, and BS is performed on a single synthetic design space and reports no significance tests, confidence intervals, or per-constraint-set breakdowns. Because the reward used to evaluate all algorithms is the same hand-designed reward from Eq. (8) that MCTS is designed to optimize, this experiment primarily measures optimization of a self-defined objective, not externally validated design quality. As it stands, the claim that 'MCTS is the optimal search algorithm' overstates the evidence. Please report statistical tests across the ten constraint sets (e.g., paired tests per metric), and clarify that this benchmark validates computational efficiency and constraint satisfaction under the assumed reward, not human-judged design quality.
  3. [Section 4.3, Eq. (8); Section 6.3.2] The reward function is the sole quantitative proxy for design quality, and all of its parameters (α=20, β=10, γ=1, δ=0.5, and UCT constant c=5) are set by the authors after empirical tuning. No sensitivity analysis is provided, and no evidence is given that reward scores correlate with human judgments of design quality. The user study cannot validate this proxy because the whole pipeline (constraint semantics, LLM, and search) differs from the baseline. To make the reward-based evaluation credible, please report the correlation between reward scores and human ratings on a common set of stimuli, or at least perform a sensitivity analysis over plausible weight ranges and show that the main conclusions are unchanged.
minor comments (4)
  1. [Equation (4)] The composition operator is typeset as a circled-plus with 'm' in the submitted text but the explanatory sentence refers to 'L'; please unify the notation and define the operator before its first use.
  2. [Section 4.3] In the sentence 'obtaining counts of satisfied hard Vh, soft positive Vp, and soft positive constraints Vn,' the third count should be 'soft negative constraints Vn'; as written, the definition of Vn is inconsistent with Eq. (6).
  3. [Table 2] The table reports mean rewards and times but no standard deviations or per-run variation. Please add measures of dispersion and state how convergence-window overhead is included in the reported times.
  4. [Section 5.3.2] The expert congruence analysis is based on only five experts and reports an aggregate '59% congruence' with modal choices. Please provide the full per-dimension contingency table and explicitly discuss the small-sample limitation in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reward-based search benchmark is self-contained and the central claim of superior design outcomes is grounded in external human evaluation.

full rationale

The paper's derivation chain does not reduce to its own inputs. The design-space model (Eqs. 1-4) is an explicit formalization credited to prior work by Shaw and Schulz, not a prediction generated from the model itself. The constraint-generation module (Eq. 5) and reward function (Eqs. 6-8) define the objective that MCTS optimizes; Section 4.5 compares MCTS against GA, SA, and BS on that same reward, which is a standard algorithmic benchmark rather than a claim that maximizing the reward is equivalent to external design quality. The external validity of the framework rests on the expert interviews (Sec. 5.3.2) and the preference user study (Sec. 6.3.2), where human participants judge requirement consistency, data accuracy, and overall quality independently of the reward weights and tuned constants. The user-study baseline does not perfectly isolate MCTS as the causal factor, because it also omits IDEA's ASP constraint semantics and uses a different reasoning model for selection; however, that is a construct-validity concern or confound, not a circular reduction, since IDEA's output is not defined as the winning output by construction. Self-citations to prior work from the same laboratory are contextual and not load-bearing; no uniqueness theorem is imported and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The framework's central claim rests on manually specified design spaces, an LLM's constraint generation, and a reward function with hand-tuned weights. These are modeling choices and domain assumptions rather than independently verified facts. The user studies provide external grounding for the final outputs, but the search objective itself is not externally calibrated.

free parameters (4)
  • reward weights α, β, γ = α=20, β=10, γ=1
    Weights in Eq. (6) that balance hard, soft positive, and soft negative constraint violations; set 'after empirically tuning' with no sensitivity analysis reported.
  • quantity penalty weight δ = 0.5
    Weights the count-mismatch penalty rq in Eq. (8); fixed in IDEA with no reported sensitivity analysis.
  • UCT exploration constant c = 5
    Exploration coefficient in Eq. (9); set 'to c=5 after experiments' with no sensitivity analysis reported.
  • convergence window W and threshold ε = W=100, ε=0.1 (deduced from Section 4.5)
    Termination conditions in Algorithm 1 are given as symbols W and ε; the evaluation text uses a 100-iteration window and a 0.1 fluctuation threshold, but the algorithm pseudocode does not specify values.
assumptions (6)
  • domain assumption Design dimensions are orthogonal and independent
    Eq. (1) declares each D_i an orthogonal design dimension. Interdependencies are handled only later through constraints, not within the base model.
  • domain assumption The manually specified design spaces are complete and correct
    Tables 1, 3, and 4 list dimensions and elements for narrative composition, data facts, and pictorial visualizations. If a relevant dimension or valid element is missing, the search cannot discover it.
  • domain assumption LLM-generated ASP constraints faithfully capture user requirements and are executable
    Section 4.2 relies on an LLM to produce constraints from R, I, and S. No validation of constraint correctness or parsing error rate is reported.
  • ad hoc to paper The reward function is a valid proxy for design quality
    Eqs. (6)-(9) use hand-tuned weights α, β, γ, δ, and c. No external benchmark or calibration study justifies these weights as a general measure of design effectiveness.
  • standard math MCTS with UCT balances exploration and exploitation to find near-optimal solutions
    The paper adopts a standard MCTS algorithm; no optimality guarantee for the finite but constrained design spaces is provided.
  • domain assumption LLM action functions faithfully instantiate abstract design decisions
    The generate_article and generate_pictorial_vis actions call DeepSeek-R1 and Claude 3.7 Sonnet. The authors acknowledge in Section 7.2 that LLM hallucinations can cause misaligned choices and artificial outputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IDEA: Augmenting Design Intelligence through Design Space Exploration." pith.science (2026). https://pith.science/paper/YDP77V43

@misc{pith2026250610587,
  author       = {Pith},
  title        = {Pith review of: IDEA: Augmenting Design Intelligence through Design Space Exploration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YDP77V43}},
  note         = {Machine review of arXiv:2506.10587}
}
read the original abstract

Design spaces serve as a conceptual framework that enables designers to explore feasible solutions through the selection and combination of design elements. However, effective decision-making remains heavily dependent on the designer's experience, and the absence of mathematical formalization prevents computational support for automated design processes. To bridge this gap, we introduce a structured representation that models design spaces with orthogonal dimensions and discrete selectable elements. Building on this model, we present IDEA, a decision-making framework for augmenting design intelligence through design space exploration to generate effective outcomes. Specifically, IDEA leverages large language models (LLMs) for constraint generation, incorporates a Monte Carlo Tree Search (MCTS) algorithm guided by these constraints to explore the design space efficiently, and instantiates abstract decisions into domain-specific implementations. We validate IDEA in two design scenarios: data-driven article composition and pictorial visualization generation, supported by example results, expert interviews, and a user study. The evaluation demonstrates the IDEA's adaptability across domains and its capability to produce superior design outcomes.

Figures

Figures reproduced from arXiv: 2506.10587 by the authors.

Figure 1
Figure 1. The overview of IDEA, including three core modules: constraint generation, design solution search, and action execution. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The user interface of the automated data-driven article composition system consists of five main components: (1) the input view, (2) the [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Data-driven articles generated by IDEA. (a) is a compelling buying guide about purchasing new energy vehicles; (b) is an in-depth industry [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The ratings of system, article, and visualization from different [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Examples generated by IDEA based on the following requirements: (a) ”A clear, visually focused design using a cheerful and playful color [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Participant selection proportions for IDEA, the baseline, and [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 39 canonical work pages

  1. [1]

    What makes a data-gif understandable?

    X. Shu, A. Wu, J. Tang, B. Bach, Y. Wu, and H. Qu, “What makes a data-gif understandable?”IEEE Transactions on Visualization and Computer Graphics, vol. 27, no. 2, pp. 1492–1502, 2020

    work page 2020

  2. [2]

    Filtering and informing the de- sign space: Towards design-space thinking,

    K. Halskov and C. Lundqvist, “Filtering and informing the de- sign space: Towards design-space thinking,”ACM Transactions on Computer-Human Interaction (TOCHI), vol. 28, no. 1, pp. 1–28, 2021

    work page 2021

  3. [3]

    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

    work page 2024

  4. [4]

    Vizbelle: A design space of embellishments for data visualiza- tion,

    Q. Chen, Z. Liu, C. Wang, X. Lan, Y. Chen, S. Chen, and N. Cao, “Vizbelle: A design space of embellishments for data visualiza- tion,”arXiv preprint arXiv:2209.03642, 2022

    arXiv 2022

  5. [5]

    Kineticharts: Aug- menting affective expressiveness of charts in data stories with animation design,

    X. Lan, Y. Shi, Y. Wu, X. Jiao, and N. Cao, “Kineticharts: Aug- menting affective expressiveness of charts in data stories with animation design,”IEEE Transactions on Visualization and Computer Graphics, vol. 28, no. 1, pp. 933–943, 2021

    work page 2021

  6. [6]

    Text-to-viz: Automatic generation of in- fographics from proportion-related natural language statements,

    W. Cui, X. Zhang, Y. Wang, H. Huang, B. Chen, L. Fang, H. Zhang, J.-G. Lou, and D. Zhang, “Text-to-viz: Automatic generation of in- fographics from proportion-related natural language statements,” IEEE transactions on visualization and computer graphics, vol. 26, no. 1, pp. 906–916, 2019

    work page 2019

  7. [7]

    Augmenting sports videos with viscommentator,

    C. Zhu-Tian, S. Ye, X. Chu, H. Xia, H. Zhang, H. Qu, and Y. Wu, “Augmenting sports videos with viscommentator,”IEEE Trans- actions on Visualization and Computer Graphics, vol. 28, no. 1, pp. 824–834, 2021

    work page 2021

  8. [8]

    Automating the design of graphical presentations of relational information,

    J. Mackinlay, “Automating the design of graphical presentations of relational information,”Acm Transactions On Graphics (Tog), vol. 5, no. 2, pp. 110–141, 1986

    work page 1986

Show all 54 references
  1. [9]

    The role of design spaces,

    M. Shaw, “The role of design spaces,”IEEE software, vol. 29, no. 1, pp. 46–50, 2011

    2011

  2. [10]

    The design space of implicit hierarchy visualization: A survey,

    H.-J. Schulz, S. Hadlak, and H. Schumann, “The design space of implicit hierarchy visualization: A survey,”IEEE transactions on visualization and computer graphics, vol. 17, no. 4, pp. 393–411, 2010

    2010

  3. [11]

    Scatterplots: Tasks, data, and de- signs,

    A. Sarikaya and M. Gleicher, “Scatterplots: Tasks, data, and de- signs,”IEEE transactions on visualization and computer graphics, vol. 24, no. 1, pp. 402–412, 2017

    2017

  4. [12]

    The state of the art in map-like visualization,

    M. Hogr ¨afer, M. Heitzler, and H.-J. Schulz, “The state of the art in map-like visualization,” inComputer Graphics Forum, vol. 39, no. 3. Wiley Online Library, 2020, pp. 647–674

    2020

  5. [13]

    Unfolding edges: Adding context to edges in multivariate graph visualization,

    M.-J. Bludau, M. D ¨ork, and C. Tominski, “Unfolding edges: Adding context to edges in multivariate graph visualization,” in Computer Graphics Forum, vol. 42, no. 3. Wiley Online Library, 2023, pp. 297–309

    2023

  6. [14]

    The state of the art in visualizing multivariate networks,

    C. Nobre, M. Meyer, M. Streit, and A. Lex, “The state of the art in visualizing multivariate networks,” inComputer Graphics Forum, vol. 38, no. 3. Wiley Online Library, 2019, pp. 807–832

    2019

  7. [15]

    Treematrix: A hybrid visualization of compound graphs,

    S. Rufiange, M. J. McGuffin, and C. P . Fuhrman, “Treematrix: A hybrid visualization of compound graphs,” inComputer Graphics Forum, vol. 31, no. 1. Wiley Online Library, 2012, pp. 89–101

    2012

  8. [16]

    Dashboard design patterns,

    B. Bach, E. Freeman, A. Abdul-Rahman, C. Turkay, S. Khan, Y. Fan, and M. Chen, “Dashboard design patterns,”IEEE transactions on visualization and computer graphics, vol. 29, no. 1, pp. 342–352, 2022

    2022

  9. [17]

    A design space for applying the freytag’s pyramid structure to data stories,

    L. Yang, X. Xu, X. Lan, Z. Liu, S. Guo, Y. Shi, H. Qu, and N. Cao, “A design space for applying the freytag’s pyramid structure to data stories,”IEEE Transactions on Visualization and Computer Graphics, vol. 28, no. 1, pp. 922–932, 2021

    2021

  10. [18]

    Freytag,Technique of the drama: An exposition of dramatic compo- sition and art

    G. Freytag,Technique of the drama: An exposition of dramatic compo- sition and art. S. Griggs, 1895

  11. [19]

    What characterizes personalities of graphic designs?

    N. Zhao, Y. Cao, and R. W. Lau, “What characterizes personalities of graphic designs?”ACM Transactions on Graphics (TOG), vol. 37, no. 4, pp. 1–15, 2018

    2018

  12. [20]

    Screenai: A vision- language model for ui and infographics understanding,

    G. Baechler, S. Sunkara, M. Wang, F. Zubach, H. Mansoor, V . Etter, V . C˘arbune, J. Lin, J. Chen, and A. Sharma, “Screenai: A vision- language model for ui and infographics understanding,”arXiv preprint arXiv:2402.04615, 2024

    2024 arXiv

  13. [21]

    Automatic gener- ation of visual-textual presentation layout,

    X. Yang, T. Mei, Y.-Q. Xu, Y. Rui, and S. Li, “Automatic gener- ation of visual-textual presentation layout,”ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), vol. 12, no. 2, pp. 1–22, 2016

    2016

  14. [22]

    Posta: A go-to framework for customized artistic poster generation,

    H. Chen, X. Xu, W. Li, J. Ren, T. Ye, S. Liu, Y.-C. Chen, L. Zhu, and X. Wang, “Posta: A go-to framework for customized artistic poster generation,”arXiv preprint arXiv:2503.14908, 2025

    2025 arXiv

  15. [23]

    Designdiffu- sion: High-quality text-to-design image generation with diffusion models,

    Z. Wang, J. Bao, S. Gu, D. Chen, W. Zhou, and H. Li, “Designdiffu- sion: High-quality text-to-design image generation with diffusion models,”arXiv preprint arXiv:2503.01645, 2025

    2025 arXiv

  16. [24]

    Decision trees,

    L. Rokach and O. Maimon, “Decision trees,”Data mining and knowledge discovery handbook, pp. 165–192, 2005

    2005

  17. [25]

    Genetic algorithm in search, optimization and machine learning, addison,

    D. E. Goldberg, “Genetic algorithm in search, optimization and machine learning, addison,”W esley Publishing Company, R eading, MA, vol. 1, no. 98, p. 9, 1989

    1989

  18. [26]

    R. S. Sutton, A. G. Bartoet al.,Reinforcement learning: An introduc- tion. MIT press Cambridge, 1998, vol. 1, no. 1

    1998

  19. [27]

    M. L. Puterman,Markov decision processes: discrete stochastic dynamic programming. John Wiley & Sons, 2014

    2014

  20. [28]

    Q-learning,

    C. J. Watkins and P . Dayan, “Q-learning,”Machine learning, vol. 8, pp. 279–292, 1992

    1992

  21. [29]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al., “Human-level control through deep reinforcement learning,” nature, vol. 518, no. 7540, pp. 529–533, 2015

    2015

  22. [30]

    Simple statistical gradient-following algorithms for connectionist reinforcement learning,

    R. J. Williams, “Simple statistical gradient-following algorithms for connectionist reinforcement learning,”Machine learning, vol. 8, pp. 229–256, 1992

    1992

  23. [31]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P . Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  24. [32]

    A survey of monte carlo tree search methods,

    C. B. Browne, E. Powley, D. Whitehouse, S. M. Lucas, P . I. Cowling, P . Rohlfshagen, S. Tavener, D. Perez, S. Samothrakis, and S. Colton, 13 “A survey of monte carlo tree search methods,”IEEE Transactions on Computational Intelligence and AI in games, vol. 4, no. 1, pp. 1–43, 2012

    2012

  25. [33]

    Survey on large language model- enhanced reinforcement learning: Concept, taxonomy, and meth- ods,

    Y. Cao, H. Zhao, Y. Cheng, T. Shu, Y. Chen, G. Liu, G. Liang, J. Zhao, J. Yan, and Y. Li, “Survey on large language model- enhanced reinforcement learning: Concept, taxonomy, and meth- ods,”IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  26. [34]

    Selfgoal: Your language agents already know how to achieve high-level goals,

    R. Yang, J. Chen, Y. Zhang, S. Yuan, A. Chen, K. Richardson, Y. Xiao, and D. Yang, “Selfgoal: Your language agents already know how to achieve high-level goals,” inNeurIPS 2024 Workshop on Open-World Agents, 2024

    2024

  27. [35]

    True knowledge comes from practice: Aligning llms with em- bodied environments via reinforcement learning,

    W. Tan, W. Zhang, S. Liu, L. Zheng, X. Wang, and B. An, “True knowledge comes from practice: Aligning llms with em- bodied environments via reinforcement learning,”arXiv preprint arXiv:2401.14151, 2024

    2024 arXiv

  28. [36]

    Do as i can, not as i say: Grounding language in robotic affordances,

    M. Ahn, A. Brohan, N. Brown, Y. Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Hausmanet al., “Do as i can, not as i say: Grounding language in robotic affordances,”arXiv preprint arXiv:2204.01691, 2022

    2022 arXiv

  29. [37]

    Ask more, know better: Reinforce-learned prompt questions for decision making with large language models,

    X. Yan, Y. Song, X. Cui, F. Christianos, H. Zhang, D. H. Mguni, and J. Wang, “Ask more, know better: Reinforce-learned prompt questions for decision making with large language models,”arXiv preprint arXiv:2310.18127, 2023

    2023 arXiv

  30. [38]

    Haichart: Human and ai paired visualization system,

    Y. Xie, Y. Luo, G. Li, and N. Tang, “Haichart: Human and ai paired visualization system,”arXiv preprint arXiv:2406.11033, 2024

    2024 arXiv

  31. [39]

    Calliope: Automatic visual data story generation from a spreadsheet,

    D. Shi, X. Xu, F. Sun, Y. Shi, and N. Cao, “Calliope: Automatic visual data story generation from a spreadsheet,”IEEE Transactions on Visualization and Computer Graphics, vol. 27, no. 2, pp. 453–463, 2020

    2020

  32. [40]

    From data to story: Towards automatic animated data video creation with llm-based multi- agent systems,

    L. Shen, H. Li, Y. Wang, and H. Qu, “From data to story: Towards automatic animated data video creation with llm-based multi- agent systems,” in2024 IEEE VIS Workshop on Data Storytelling in an Era of Generative AI (GEN4DS). IEEE, 2024, pp. 20–27

    2024

  33. [41]

    Design patterns for data-driven news articles,

    S. Hao, Z. Wang, B. Bach, and L. Pschetz, “Design patterns for data-driven news articles,” inProceedings of the 2024 CHI Conference on Human Factors in Computing Systems, 2024, pp. 1–16

    2024

  34. [42]

    Nar- rative design patterns for data-driven storytelling,

    B. Bach, M. Stefaner, J. Boy, S. Drucker, L. Bartram, J. Wood, P . Ciuccarelli, Y. Engelhardt, U. Koeppen, and B. Tversky, “Nar- rative design patterns for data-driven storytelling,” inData-driven storytelling. AK Peters/CRC Press, 2018, pp. 107–133

    2018

  35. [43]

    Patterns in award winning data sto- rytelling: Story types, enabling tools and competences,

    A. Ojo and B. Heravi, “Patterns in award winning data sto- rytelling: Story types, enabling tools and competences,”Digital journalism, vol. 6, no. 6, pp. 693–718, 2018

    2018

  36. [44]

    contributors

    W. contributors. (2025) Narration. [Online]. Available: https: //en.wikipedia.org/wiki/Narration

    2025

  37. [45]

    Lifschitz,Answer set programming

    V . Lifschitz,Answer set programming. Springer Cham, 2019, vol. 3

    2019

  38. [46]

    Clingo= asp+ control: Preliminary report,

    M. Gebser, R. Kaminski, B. Kaufmann, and T. Schaub, “Clingo= asp+ control: Preliminary report,”arXiv preprint arXiv:1405.3694, 2014

    2014 arXiv

  39. [47]

    Bandit based monte-carlo planning,

    L. Kocsis and C. Szepesv ´ari, “Bandit based monte-carlo planning,” inEuropean conference on machine learning. Springer, 2006, pp. 282– 293

    2006

  40. [48]

    Deepseek-r1: Incentivizing reason- ing capability in llms via reinforcement learning,

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P . Wang, X. Biet al., “Deepseek-r1: Incentivizing reason- ing capability in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  41. [49]

    Supporting expres- sive and faithful pictorial visualization design with visual style transfer,

    Y. Shi, P . Liu, S. Chen, M. Sun, and N. Cao, “Supporting expres- sive and faithful pictorial visualization design with visual style transfer,”IEEE Transactions on Visualization and Computer Graphics, vol. 29, no. 1, pp. 236–246, 2022

    2022

  42. [50]

    Claude 3.7 sonnet,

    Anthropic, “Claude 3.7 sonnet,” 2025, large language model developed by Anthropic. [Online]. Available: https://www. anthropic.com/news/claude-3-7-sonnet

    2025

  43. [51]

    F. Awesome. (2025) Font awesome. [Online]. Available: https: //fontawesome.com/

    2025

  44. [52]

    A new approach to interior design: Generating creative interior design videos of various design styles from indoor texture-free 3d models,

    Z. Shao, J. Chen, H. Zeng, W. Hu, Q. Xu, and Y. Zhang, “A new approach to interior design: Generating creative interior design videos of various design styles from indoor texture-free 3d models,”Buildings, vol. 14, no. 6, p. 1528, 2024

    2024

  45. [53]

    Generating interior design from text: A new diffusion model-based method for efficient creative design,

    J. Chen, Z. Shao, and B. Hu, “Generating interior design from text: A new diffusion model-based method for efficient creative design,”Buildings, vol. 13, no. 7, p. 1861, 2023. Chuer Chenreceived her MSc degree from the Department of Electrical and Computer En- gineering, Natio...

    2023

  46. [2021]

    degree as part of the Intelligent Big Data Visu- alization (iDVx) Lab, Tongji University

    She is currently working toward her Ph.D. degree as part of the Intelligent Big Data Visu- alization (iDVx) Lab, Tongji University. Her re- search interests include information visualization and intelligent design. Xiaoke Yanreceived her bachelor’s degree from the Department o...

    2024

Pith tools

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