Pith. sign in

REVIEW 4 major objections 6 minor 76 references

Efficient Visual Appearance Optimization by Learning from Prior Preferences

T0 review · 4 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 This paper claims that Meta-PO, a meta-learning extension of Preferential Bayesian Optimization that reuses prior users' preference models, lets people reach a satisfactory image appearance in 5.86 iterations instead of 9.54, and still…

desk verdict Solid meta-learning integration for preferential BO with a clean same-theme result, but the cross-theme headline is confounded and the decay parameters need a closer look. read the letter →

arxiv 2507.15355 v2 pith:DB4NOBYL submitted 2025-07-21 cs.HC cs.LG

Zhipeng Li , Yi-Chi Liao , Christian Holz This is my paper · ORCID
classification cs.HCcs.LG
keywords human-in-the-loopoptimizationpreferentialBayesianmeta-learningvisualappearancedesignsequentialgallerytransferlearningGaussianprocessimageenhancement
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

Meta-PO is a method for making preference-based visual design optimization usable by ordinary end-users. Instead of requiring every user to explore a 12-parameter appearance space from scratch, it stores Gaussian-process models of previous users' preference sessions and aggregates their acquisition values with the current user's model, weighting each prior model by how consistently its predictions rank options the same way the current user does. The paper reports that with this meta-learning layer, participants reached a satisfactory image appearance in 5.86 iterations for same-theme transfer and 7.41 for cross-theme transfer, versus 9.54 without transfer. A second study on virtual-reality lighting showed the same pattern, with satisfaction at 4.05 iterations instead of 6.85. The significance is that iterative preference optimization, previously practical mostly for designers, becomes fast enough for casual everyday tuning.

What carries the argument

The load-bearing machinery is TAF-R, a Transfer Acquisition Function variant that weights each stored population Gaussian process by the rank correlation between its preference predictions and the current user's evolving model, so prior models that have been agreeing with the user dominate early suggestions and disagreeing ones are downweighted. Around this, Meta-PO constructs the Sequential-Gallery 2D search plane from three points: the current best design $x^+_k$, the TAF-R maximizer $x^{AF}_{k,1}$, and a third point $x^{AF}_{k,2}$ chosen by a two-step lookahead acquisition that simulates evaluating $x^{AF}_{k,1}$ before committing to the next corner. A time-dependent decay factor with parameters $d_1=3$, $d_2=8$ then reduces population-model influence as the current user's own observations accumulate. These three components together carry the claimed transfer: guided early exploration, diverse plane geometry, and a smooth handoff to personalization.

What would settle it

Run the image-enhancement study again with the same four themes in both the transfer and no-transfer conditions, counterbalanced across participants, and measure iterations to satisfaction; if the cross-theme advantage vanishes once theme sets are matched, the headline generalization result is an artifact of theme difficulty rather than transfer.

Watch

Extended reading notes

Core claim

The paper claims that Preferential Bayesian Optimization for visual appearance can be made substantially more sample-efficient by meta-learning across users and themes. Its central finding is that when a gallery of Gaussian-process preference models built from earlier users' optimization sessions is aggregated through a similarity-weighted acquisition function, new users reach a self-declared satisfactory appearance in significantly fewer iterations: 5.86 (SD = 1.20) for same-theme transfer versus 9.54 (SD = 2.19) with no transfer, and 7.41 (SD = 1.28) even when the target theme was unseen during population modeling. The authors further report that in a VR lighting task, Meta-PO reached satisfaction in 4.05 iterations versus 6.85 without transfer. The claim is that this speedup comes from selectively transferring prior preference structure while letting the current user's own observations gradually take over.

Load-bearing premise

The cross-theme result assumes the four target themes used in the transfer condition (winter, summer, pastel, vintage) are no harder to achieve than the four used in the no-transfer baseline (warm, cold, golden hour, blue hour), so any measured difference could come from theme difficulty or participant assignment instead of meta-transfer.

Editorial extensions

If this is right

  • Casual users can meaningfully tune image appearance in about half a dozen gallery selections instead of ten, which is the difference between a design-tool workflow and an everyday photo-editing interaction.
  • A population-model gallery built from one set of themes remains useful for unseen themes that share stylistic structure, so the system does not need to be retrained for every new aesthetic goal.
  • The same optimizer applies to continuous 3D interaction spaces such as VR lighting, not only to discrete 2D galleries, because the acquisition machinery is independent of the display format.
  • The simulated ablations imply that similarity-weighted transfer and two-step lookahead are complementary: each improves regret on its own, and their combination is the configuration carried into user studies.

Reading between the lines

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

  • An untested implication is that the preference-model gallery is modality-agnostic: sound design, haptic personalization, or layout tuning could reuse the same TAF-R weighting if pairwise preference feedback is available.
  • A deployed system could estimate theme similarity beforehand from image features or text labels and pre-filter the population gallery, which would address the negative-transfer risk the authors acknowledge rather than relying on within-session downweighting alone.
  • The 1.55-iteration gap between same-theme and cross-theme transfer is a natural benchmark for future transfer methods: closing it would indicate that cross-theme generalization is approaching same-theme quality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces Meta-PO, a meta-learning extension of Preferential Bayesian Optimization (PBO) for visual appearance tuning. The method builds on the Sequential-Gallery interaction, using TAF-R (a ranking-similarity-based weight assignment) to combine multiple population GP models with the current user model, and a two-step acquisition function to select a third point when constructing each 2D search plane. A decay factor is introduced to reduce population-model influence over iterations. The authors report simulated ablation tests on four benchmark functions and two user studies (image enhancement and VR lighting design). The headline results are that Meta-PO reduces the mean number of iterations to satisfaction from 9.54 (No-Transfer) to 5.86 (Cross-Users, same theme) and to 7.41 (Cross-Themes) in the image-enhancement study, with supporting subjective improvements in exploration and satisfaction.

Significance. If the reported effects are valid, Meta-PO would be a practical contribution to HCI, making preferential optimization usable for everyday end-users by cutting interaction rounds roughly in half. The same-theme cross-user comparison is well designed and its result is plausible; the paper also offers a clear algorithmic recipe, open-source code, and simulated ablations that compare weighting schemes and plane-construction strategies. However, the central cross-theme generalization claim currently rests on a confounded between-subjects comparison, so the headline contribution cannot be accepted as demonstrated until that comparison is repaired. The same-theme result and the method's structure are valuable regardless of the cross-theme outcome.

major comments (4)
  1. [§5.1.2 and §5.1.5] The cross-theme generalization claim is confounded: the No-Transfer and Cross-Users conditions used the themes warm, cold, golden hour, and blue hour, while the Cross-Themes condition used winter, summer, pastel, and vintage. The comparison of Cross-Themes (M = 7.41, SD = 1.28) against No-Transfer (M = 9.54, SD = 2.19) therefore mixes the effect of meta-transfer with any differences in theme difficulty, participant familiarity, or task interpretability. Because condition and theme set are perfectly confounded, the claim that prior optimization experience transfers across divergent goals is not supported by this comparison. A No-Transfer control run on the same winter/summer/pastel/vintage themes, or a within-subject/counterbalanced design across a common theme set, is required to validate the cross-theme claim.
  2. [§4.3.3, Eq. (9)] With the reported pilot-based parameters d1 = 3 and d2 = 8, the decay factor d(k) equals 1 for k ≤ 3 and 0 for every integer k ≥ 4, because the second piece of Eq. (9) covers only the open interval (3, 3.125]. Thus all population-model weights are zero from the fourth iteration onward. Since the mean satisfaction iterations in the user study are 7.41–9.54, most of the measured optimization runs proceed without any population-model influence. The paper should report the effective integer-valued decay schedule, justify this abrupt cutoff, and provide a sensitivity analysis; otherwise it is unclear how much of the observed improvement is actually attributable to the meta-learned prior models versus the two-step acquisition mechanism present in both conditions.
  3. [§5.2.2 and §5.2.4] The virtual-scene lighting study does not provide a valid control for its cross-theme results. The Meta-PO group designed for golden hour and pastel in addition to warm and cold, while the No-Transfer group only designed for warm and cold; there is no baseline for the two new themes, and no inferential statistics are reported. The 'preliminary' qualification is appropriate, but the abstract and conclusion state cross-theme generalization as a general result that includes this 3D application. The claims should be restricted to the image-enhancement study, or a matched No-Transfer condition for the added themes should be collected.
  4. [§5.1.5 statistical analysis] The reported ANOVA has F(2,33), which implies a participant-level analysis with 12 participants per group, but the text describes trial-level outcomes ('Across all trials', 'excluding trials in which participants failed'). It should be stated explicitly whether each participant contributes one mean iteration count over 16 trials or whether trial-level data were used; if trial-level, the degrees of freedom and the treatment of repeated measures need to be described. Effect sizes or confidence intervals for the reported mean differences would also help assess the practical magnitude of the improvements.
minor comments (6)
  1. [Abstract and §1] The abstract says 'in 8 iterations' for cross-theme generalization, while the Introduction (§1) and Results (§5.1.5) report M = 7.41; these numbers should be made consistent.
  2. [§4.2, 'Core Elements'] The description of how population models are constructed from No-Transfer data is incomplete: it is not stated how many iterations of user data are used per model, whether models are trained after the full 15-iteration session, or how missing satisfaction responses are handled in the GP fitting.
  3. [§4.3.3, Eq. (9)] The same equation is used in the simulated tests with d1 = 5 and d2 = 0.1, but the manuscript does not discuss why the user-study parameters differ from the simulation parameters or how transferable the pilot-based choice is to the VR study.
  4. [§5.2.3, Participants & Apparatus] The demographic text reports identical means and standard deviations for the two VR groups (M = 27.17, SD = 1.59 for both No-Transfer and Meta-PO), despite the stated age ranges 25–29 and 24–30. This appears to be a copy-paste error and should be corrected.
  5. [§5.1.5, Figures 4–5] The gallery images are small and hard to interpret in the printed text; adding zoomed crops or perceptual difference annotations would make the qualitative convergence evidence more inspectable.
  6. [Throughout] Several equation references in the text use section-style labels ('Equation 3.1', 'Equation 3.4') instead of the manuscript's equation numbers; these should be renumbered to avoid confusion.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: Meta-PO's efficiency claims rest on new human-participant studies and external baselines; self-citations to prior Liao et al. work are contextual and not load-bearing.

full rationale

The paper's central derivation chain combines published components — Sequential-Gallery [26], Transfer Acquisition Function [67], and two-step lookahead BO [68] — and then evaluates the integrated Meta-PO method against no-transfer baselines with new participants in both simulated and user studies. The reported efficiency gains (5.86 vs. 9.54 iterations same-theme; 7.41 vs. 9.54 cross-theme) are measured outcomes of those studies, not quantities derived from the method's own equations, so the claim is not self-definitional and no fitted parameter is renamed as a prediction. The only fitted choices are the decay hyperparameters (d1=3, d2=8) selected from a pilot study, which are implementation-level tuning rather than the target result. Self-citations to Liao et al. [36] for TAF-M, EI usage, and the decay-function form are real external prior work with independent experiments, and they are not the load-bearing evidence for Meta-PO's contribution; the paper explicitly contrasts with and extends that work. The cross-theme comparison does confound condition with theme set (winter/summer/pastel/vintage vs. warm/cold/golden hour/blue hour), which is a legitimate internal-validity concern, but it is not circularity because the method does not encode the measured outcome. Overall, no step in the derivation reduces by construction to its own inputs.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard GP/BTL assumptions, on the TAF-R similarity heuristic adopted from prior work, on the two-step lookahead approximation, and on synthetic user models in the simulations. The pilot-derived decay parameters are free hyperparameters not independently justified.

free parameters (3)
  • Decay parameters d1, d2 (user study) = d1=3, d2=8
    Set from a pilot study (Sec. 5.1.4); controls how quickly population model weights decay. With the stated formula, d2=8 gives negative weights at k=4.
  • Decay parameters d1, d2 (simulated tests) = d1=5, d2=0.1
    Used in Appendix A simulations for comparing methods; chosen by the authors, not derived from a theory.
  • Shift range and scale range for synthetic preference variation = shift_range=0.05, scale_range=0.1
    Appendix A.3; controls how varied synthetic users are in simulations and is not tied to real user data.
assumptions (6)
  • domain assumption User selections follow the Bradley-Terry-Luce model (Eq. 2)
    Assumes choice probabilities are softmax over latent goodness values; standard in PBO but not derived from psychological data.
  • domain assumption Gaussian process with ARD Matérn 5/2 kernel approximates users' implicit preference functions
    Invoked in Sec. 5.1.4; smoothness and stationarity assumptions that are not validated for human preference data here.
  • domain assumption Ranking alignment between current GP and population GPs is a valid similarity measure for transfer weighting (TAF-R)
    Core to Meta-PO; adopted from Wistuba et al. without independent validation in preferential settings.
  • standard math The two-step lookahead with Monte Carlo sampling approximates the true value of information
    Borrowed from Wu and Frazier; accepted approximation under GP assumptions.
  • ad hoc to paper Simulated users generated by shifting and scaling benchmark functions represent real user preference variation
    Appendix A.3; no evidence that such shifts capture human preference diversity.
  • ad hoc to paper The synthetic user chooses the highest-valued design on each search plane
    Appendix A.2; removes human noise and BTL stochasticity from the simulation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Visual Appearance Optimization by Learning from Prior Preferences." pith.science (2026). https://pith.science/paper/DB4NOBYL

@misc{pith2026250715355,
  author       = {Pith},
  title        = {Pith review of: Efficient Visual Appearance Optimization by Learning from Prior Preferences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DB4NOBYL}},
  note         = {Machine review of arXiv:2507.15355}
}
read the original abstract

Adjusting visual parameters such as brightness and contrast is common in our everyday experiences. Finding the optimal parameter setting is challenging due to the large search space and the lack of an explicit objective function, leaving users to rely solely on their implicit preferences. Prior work has explored Preferential Bayesian Optimization (PBO) to address this challenge, involving users to iteratively select preferred designs from candidate sets. However, PBO often requires many rounds of preference comparisons, making it more suitable for designers than everyday end-users. We propose Meta-PO, a novel method that integrates PBO with meta-learning to improve sample efficiency. Specifically, Meta-PO infers prior users' preferences and stores them as models, which are leveraged to intelligently suggest design candidates for the new users, enabling faster convergence and more personalized results. An experimental evaluation of our method for appearance design tasks on 2D and 3D content showed that participants achieved satisfactory appearance in 5.86 iterations using Meta-PO when participants shared similar goals with a population (e.g., tuning for a ``warm'' look) and in 8 iterations even generalizes across divergent goals (e.g., from ``vintage'', ``warm'', to ``holiday''). Meta-PO makes personalized visual optimization more applicable to end-users through a generalizable, more efficient optimization conditioned on preferences, with the potential to scale interface personalization more broadly.

Figures

Figures reproduced from arXiv: 2507.15355 by the authors.

Figure 1
Figure 1. We introduce Meta-PO, a computational method that models users’ implicit preferences of visual appearances given a theme (here "cold") as they iteratively select from a sampled set of candidates presented in a 2D gallery. Meta-PO integrates Preferential Bayesian Optimization with meta-learning to increase optimization efficiency by leveraging population models derived from prior optimization experiences — across use… view at source ↗
Figure 2
Figure 2. Meta-PO’s working mechanism during deployment. Meta-PO first stores a set of population models (shown in the gray [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Image enhancement interface. Users are presented with 25 candidate images and select the one that best aligns with [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Comparison of the intermediate results generated by Meta-PO and by Sequential Gallery (baseline) across four themes. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Intermediate results after the 1st, 3rd, and 5th iterations using Meta-PO across all eight themes. Leveraging previous [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: CSI questionnaire results for image enhancement [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Meta-PO-based optimization of immersive lighting appearances: Intermediate results for four themes. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: CSI questionnaire results for virtual lighting ap [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Ablation study results on 3D Hartmann. The left plot compares the optimization performance of [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Ablation study results on 6D Hartmann. The main conclusions remain consistent with the previous results. However, [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Ablation study results on 15D Isotropic Gaussian. The main conclusions remain consistent with the previous results. [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]
Figure 12
Figure 12. Figure 12: Ablation study results on 20D Rosenbrock. The main conclusions remain consistent with the previous results. The [PITH_FULL_IMAGE:figures/full_fig_p024_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 64 canonical work pages

  1. [1]

    Moritz Bächer, Emily Whiting, Bernd Bickel, and Olga Sorkine-Hornung. 2014. Spin-it: Optimizing moment of inertia for spinnable objects. ACM Transactions on Graphics (TOG) 33, 4 (2014), 1–10

    work page 2014

  2. [2]

    Tianyi Bai, Yang Li, Yu Shen, Xinyi Zhang, Wentao Zhang, and Bin Cui

    work page

  3. [3]

    Rémi Bardenet, Mátyás Brendel, Balázs Kégl, and Michele Sebag. 2013. Collabo- rative hyperparameter tuning. In International conference on machine learning . PMLR, 199–207

    work page 2013

  4. [4]

    Eric Brochu, Tyson Brochu, and Nando De Freitas. 2010. A Bayesian interactive optimization approach to procedural animation design. In Proceedings of the 2010 ACM SIGGRAPH/Eurographics Symposium on Computer Animation . 103–112

    work page 2010

  5. [5]

    Bilal Catkin and Volkan Patoglu. 2023. Preference-based human-in-the-loop optimization for perceived realism of haptic rendering. IEEE Transactions on Haptics 16, 4 (2023), 470–476

    work page 2023

  6. [6]

    Liwei Chan, Yi-Chi Liao, George B Mo, John J Dudley, Chun-Lien Cheng, Per Ola Kristensson, and Antti Oulasvirta. 2022. Investigating positive and negative qualities of human-in-the-loop optimization for designing interaction techniques. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems . 1–14

    work page 2022

  7. [7]

    Jonlin Chen, Masaru Ishii, Kristin L Bater, Halley Darrach, David Liao, Pauline P Huynh, Isabel P Reh, Jason C Nellis, Anisha R Kumar, and Lisa E Ishii. 2019. Association between the use of social media and photograph editing applications, self-esteem, and cosmetic surgery acceptance. JAMA facial plastic surgery 21, 5 (2019), 361–367

    work page 2019

  8. [8]

    Yi Fei Cheng, Christoph Gebhardt, and Christian Holz. 2023. Interactionadapt: Interaction-driven workspace adaptation for situated virtual reality environments. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. 1–14

    work page 2023

Show all 76 references
  1. [9]

    Erin Cherry and Celine Latulipe. 2014. Quantifying the creativity support of digital tools through the creativity support index.ACM Transactions on Computer- Human Interaction (TOCHI) 21, 4 (2014), 1–25

    2014

  2. [10]

    Mark Colley, Pascal Jansen, Mugdha Keskar, and Enrico Rukzio. 2025. Improving External Communication of Automated Vehicles Using Bayesian Optimization. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machiner...

    2025

  3. [11]

    Nigel Cross. 2004. Expertise in design: an overview. Design studies 25, 5 (2004), 427–441. UIST ’25, September 28-October 1, 2025, Busan, Republic of Korea Zhipeng Li, Yi-Chi Liao, and Christian Holz

    2004

  4. [12]

    Ritendra Datta, Dhiraj Joshi, Jia Li, and James Z Wang. 2006. Studying aesthetics in photographic images using a computational approach. In Computer Vision– ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part III 9 . Springer...

    2006

  5. [13]

    James R Deneault, Woojae Kim, Jiseob Kim, Yuzhe Gu, Jorge Chang, Benji Maruyama, Jay I Myung, and Mark A Pitt. 2025. Preferential Bayesian opti- mization improves the efficiency of printing objects with subjective qualities. Digital Discovery 4, 3 (2025), 723–737

    2025

  6. [14]

    Yubin Deng, Chen Change Loy, and Xiaoou Tang. 2017. Image aesthetic assess- ment: An experimental survey. IEEE Signal Processing Magazine 34, 4 (2017), 80–106

    2017

  7. [15]

    Bartlett, Ilya Sutskever, and Pieter Abbeel

    Yan Duan, John Schulman, Xi Chen, Peter L. Bartlett, Ilya Sutskever, and Pieter Abbeel. 2016. RL$ˆ2$: Fast Reinforcement Learning via Slow Reinforcement Learning. CoRR abs/1611.02779 (2016). arXiv:1611.02779 http://arxiv.org/abs/ 1611.02779

    2016 arXiv

  8. [16]

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta- learning for fast adaptation of deep networks. In International conference on machine learning. PMLR, 1126–1135

    2017

  9. [17]

    Peter I Frazier. 2018. A tutorial on Bayesian optimization. arXiv preprint arXiv:1807.02811 (2018)

    2018 arXiv

  10. [18]

    David Ginsbourger and Rodolphe Le Riche. 2010. Towards Gaussian process- based optimization with finite time horizon. In mODa 9–Advances in Model- Oriented Design and Analysis: Proceedings of the 9th International Workshop in Model-Oriented Design and Analysis held in Bertinor...

    2010

  11. [19]

    Javier González, Zhenwen Dai, Andreas Damianou, and Neil D Lawrence. 2017. Preferential bayesian optimization. In International Conference on Machine Learn- ing. PMLR, 1282–1291

    2017

  12. [20]

    Bing-Jing Hsieh, Ping-Chun Hsieh, and Xi Liu. 2021. Reinforced few-shot acquisi- tion function learning for bayesian optimization. Advances in Neural Information Processing Systems 34 (2021), 7718–7731

    2021

  13. [21]

    Kuo-Ying Huang. 2009. Challenges in human-computer interaction design for mobile devices. In Proceedings of the World Congress on Engineering and Computer Science, Vol. 1. San Francisco, USA, 1–6

    2009

  14. [22]

    Koki Iwai, Yusuke Kumagae, Yuki Koyama, Masahiro Hamasaki, and Masataka Goto. 2025. Constrained Preferential Bayesian Optimization and Its Application in Banner Ad Design. arXiv preprint arXiv:2505.10954 (2025)

    2025 arXiv

  15. [23]

    Donald R Jones, Cary D Perttunen, and Bruce E Stuckman. 1993. Lipschitzian optimization without the Lipschitz constant. Journal of optimization Theory and Applications 79 (1993), 157–181

    1993

  16. [24]

    Andreas Karrenbauer and Antti Oulasvirta. 2014. Improvements to keyboard optimization with integer programming. In Proceedings of the 27th annual ACM symposium on User interface software and technology . 621–626

    2014

  17. [25]

    Khajah, Brett D

    Mohammad M. Khajah, Brett D. Roads, Robert V. Lindsey, Yun-En Liu, and Michael C. Mozer. 2016. Designing Engaging Games Using Bayesian Optimization. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems (San Jose, California, USA) (CHI ’16). Associati...

    2016

  18. [26]

    Yuki Koyama, Issei Sato, and Masataka Goto. 2020. Sequential gallery for interac- tive visual design optimization. ACM Transactions on Graphics (TOG) 39, 4 (2020), 88–1

    2020

  19. [27]

    Yuki Koyama, Issei Sato, Daisuke Sakamoto, and Takeo Igarashi. 2017. Sequential line search for efficient visual design optimization by crowds. ACM Transactions on Graphics (TOG) 36, 4 (2017), 1–11

    2017

  20. [28]

    Andreas Krause and Cheng Ong. 2011. Contextual gaussian process bandit optimization. Advances in neural information processing systems 24 (2011)

    2011

  21. [29]

    Remi Lam, Karen Willcox, and David H Wolpert. 2016. Bayesian optimization with a finite budget: An approximate dynamic programming approach. Advances in Neural Information Processing Systems 29 (2016)

    2016

  22. [30]

    Bryan Lawson. 2006. How designers think. Routledge

    2006

  23. [31]

    Dingzeyu Li, David IW Levin, Wojciech Matusik, and Changxi Zheng. 2016. Acoustic voxels: Computational optimization of modular acoustic filters. ACM Transactions on Graphics (TOG) 35, 4 (2016), 1–12

    2016

  24. [32]

    Wendi Li, Ting Wang, and Wing WY Ng. 2021. Population-based hyperparameter tuning with multitask collaboration. IEEE transactions on neural networks and learning systems 34, 9 (2021), 5719–5731

    2021

  25. [33]

    Yang Li, Yu Shen, Huaijun Jiang, Wentao Zhang, Zhi Yang, Ce Zhang, and Bin Cui

  26. [34]

    Zhipeng Li, Christoph Gebhardt, Yves Inglin, Nicolas Steck, Paul Streli, and Christian Holz. 2024. Situationadapt: Contextual ui optimization in mixed reality with situation awareness via llm reasoning. In Proceedings of the 37th Annual ACM Symposium on User Interface Software...

    2024

  27. [35]

    Yi-Chi Liao. 2023. Human-in-the-Loop Design Optimization. (2023)

    2023

  28. [36]

    Yi-Chi Liao, Ruta Desai, Alec M Pierce, Krista E Taylor, Hrvoje Benko, Tanya R Jonker, and Aakar Gupta. 2024. A Meta-Bayesian Approach for Rapid Online Parametric Optimization for Wrist-based Interactions. In Proceedings of the 2024 CHI Conference on Human Factors in Computing...

    2024

  29. [37]

    Yi-Chi Liao, John J Dudley, George B Mo, Chun-Lien Cheng, Liwei Chan, Antti Oulasvirta, and Per Ola Kristensson. 2023. Interaction design with multi-objective bayesian optimization. IEEE Pervasive Computing 22, 1 (2023), 29–38

    2023

  30. [38]

    Yi-Chi Liao, Sunjun Kim, Byungjoo Lee, and Antti Oulasvirta. 2020. Button simulation and design via fdvv models. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems . 1–14

    2020

  31. [39]

    Yi-Chi Liao, George B Mo, John J Dudley, Chun-Lien Cheng, Liwei Chan, Per Ola Kristensson, and Antti Oulasvirta. 2024. Practical approaches to group-level multi-objective Bayesian optimization in interaction technique design. Collective Intelligence 3, 1 (2024), 26339137241241313

    2024

  32. [40]

    Yi-Chi Liao, Paul Streli, Zhipeng Li, Christoph Gebhardt, and Christian Holz

  33. [41]

    Dong C Liu and Jorge Nocedal. 1989. On the limited memory BFGS method for large scale optimization. Mathematical programming 45, 1 (1989), 503–528

    1989

  34. [42]

    Lydia Manikonda, Yuheng Hu, and Subbarao Kambhampati. 2014. Analyzing user activities, demographics, social network structure and user-generated content on Instagram. arXiv preprint arXiv:1410.8099 (2014)

    2014 arXiv

  35. [43]

    Luca-Maxim Meinhardt, Clara Schramm, Pascal Jansen, Mark Colley, and Enrico Rukzio. 2025. Fly Away: Evaluating the Impact of Motion Fidelity on Optimized User Interface Design via Bayesian Optimization in Automated Urban Air Mobil- ity Simulations. In Proceedings of the 2025 C...

    2025

  36. [44]

    George Mo, John Dudley, Liwei Chan, Yi-Chi Liao, Antti Oulasvirta, and Per Ola Kristensson. 2024. Cooperative multi-objective Bayesian design optimization. ACM Transactions on Interactive Intelligent Systems 14, 2 (2024), 1–28

    2024

  37. [45]

    Jonas Mockus. 1975. On the Bayes methods for seeking the extremal point. IFAC Proceedings Volumes 8, 1 (1975), 428–431

    1975

  38. [46]

    Alex Nichol and John Schulman. 2018. Reptile: a scalable metalearning algorithm. arXiv preprint arXiv:1803.02999 2, 3 (2018), 4

    2018 arXiv

  39. [47]

    Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. 2015. Designscape: Design with interactive layout suggestions. In Proceedings of the 33rd annual ACM conference on human factors in computing systems . 1221–1224

    2015

  40. [48]

    Michael A Osborne, Roman Garnett, and Stephen J Roberts. 2009. Gaussian processes for global optimization. (2009)

    2009

  41. [49]

    Romain Prévost, Emily Whiting, Sylvain Lefebvre, and Olga Sorkine-Hornung

  42. [50]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10684–10695

    2022

  43. [51]

    Antonio Sabbatella, Andrea Ponti, Ilaria Giordani, Antonio Candelieri, and Francesco Archetti. 2024. Prompt optimization in large language models. Mathe- matics 12, 6 (2024), 929

    2024

  44. [52]

    Nicolas Schweighofer and Kenji Doya. 2003. Meta-learning in reinforcement learning. Neural Networks 16, 1 (2003), 5–9

    2003

  45. [53]

    Bobak Shahriari, Kevin Swersky, Ziyu Wang, Ryan P Adams, and Nando De Fre- itas. 2015. Taking the human out of the loop: A review of Bayesian optimization. Proc. IEEE 104, 1 (2015), 148–175

    2015

  46. [54]

    Jasper Snoek, Hugo Larochelle, and Ryan P Adams. 2012. Practical bayesian optimization of machine learning algorithms. Advances in neural information processing systems 25 (2012)

    2012

  47. [55]

    Kevin Swersky, Jasper Snoek, and Ryan P Adams. 2013. Multi-task bayesian optimization. Advances in neural information processing systems 26 (2013)

    2013

  48. [56]

    Shion Takeno, Masahiro Nomura, and M Karasuma. 2022. Preferential Bayesian optimization with hallucination believer. In Proc. NeurIPS Workshop Gaussian Processes, Spatiotemporal Model., Decision-Making Syst. 1–11

    2022

  49. [57]

    Yuki Tatsukawa, I-Chao Shen, Mustafa Doga Dogan, Anran Qi, Yuki Koyama, Ariel Shamir, and Takeo Igarashi. 2025. FontCraft: Multimodal Font Design Using Interactive Bayesian Optimization. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems . 1–14

    2025

  50. [58]

    Kashyap Todi, Daryl Weir, and Antti Oulasvirta. 2016. Sketchplore: Sketch and explore with a layout optimiser. In Proceedings of the 2016 ACM conference on designing interactive systems. 543–555

    2016

  51. [59]

    Kristi Tsukida, Maya R Gupta, et al. 2011. How to analyze paired comparison data. Department of Electrical Engineering University of Washington, Tech. Rep. UWEETR-2011-0004 1 (2011)

    2011

  52. [60]

    Ricardo Vilalta and Youssef Drissi. 2002. A perspective view and survey of meta-learning. Artificial intelligence review 18 (2002), 77–95

    2002

  53. [61]

    Michael Volpp, Lukas P Fröhlich, Kirsten Fischer, Andreas Doerr, Stefan Falkner, Frank Hutter, and Christian Daniel. 2019. Meta-learning acquisition functions for transfer learning in bayesian optimization. arXiv preprint arXiv:1904.02642 (2019)

    2019 arXiv

  54. [62]

    Jane X Wang. 2021. Meta-learning in natural and artificial intelligence. Current Opinion in Behavioral Sciences 38 (2021), 90–95. Efficient Visual Appearance Optimization by Learning from Prior Preferences UIST ’25, September 28-October 1, 2025, Busan, Republic of Korea

    2021

  55. [63]

    Xilu Wang, Yaochu Jin, Sebastian Schmitt, and Markus Olhofer. 2023. Recent advances in Bayesian optimization. Comput. Surveys 55, 13s (2023), 1–36

    2023

  56. [64]

    Karl Weiss, Taghi M Khoshgoftaar, and DingDing Wang. 2016. A survey of transfer learning. Journal of Big data 3 (2016), 1–40

    2016

  57. [65]

    John H Williamson, Antti Oulasvirta, Per Ola Kristensson, and Nikola Banovic

  58. [66]

    Martin Wistuba and Josif Grabocka. 2021. Few-shot bayesian optimization with deep kernel surrogates. arXiv preprint arXiv:2101.07667 (2021)

    2021 arXiv

  59. [67]

    Martin Wistuba, Nicolas Schilling, and Lars Schmidt-Thieme. 2018. Scalable gaussian process-based transfer surrogates for hyperparameter optimization. Machine Learning 107, 1 (2018), 43–78

    2018

  60. [68]

    Jian Wu and Peter Frazier. 2019. Practical two-step lookahead Bayesian optimiza- tion. Advances in neural information processing systems 32 (2019)

    2019

  61. [69]

    Zhaozhuo Xu, Zirui Liu, Beidi Chen, Yuxin Tang, Jue Wang, Kaixiong Zhou, Xia Hu, and Anshumali Shrivastava. 2023. Compress, then prompt: Improving accuracy-efficiency trade-off of llm inference with transferable prompt. arXiv preprint arXiv:2305.11186 (2023)

    2023 arXiv

  62. [70]

    Cambridge University Press

    Bayesian Methods for Interaction and Design . Cambridge University Press

  63. [71]

    Dani Yogatama and Gideon Mann. 2014. Efficient transfer learning method for automatic hyperparameter tuning. In Artificial intelligence and statistics. PMLR, 1077–1085. UIST ’25, September 28-October 1, 2025, Busan, Republic of Korea Zhipeng Li, Yi-Chi Liao, and Christian Holz...

    2014

  64. [75]

    Kenta Yamamoto, Yuki Koyama, and Yoichi Ochiai. 2022. Photographic lighting design with photographer-in-the-loop bayesian optimization. In Proceedings of the 35th annual ACM symposium on user interface software and technology . 1–11

    2022

  65. [2013]

    ACM Transactions on Graphics (TOG) 32, 4 (2013), 1–10

    Make it stand: balancing shapes for 3D fabrication. ACM Transactions on Graphics (TOG) 32, 4 (2013), 1–10

    2013

  66. [2022]

    In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Transbo: Hyperparameter optimization via two-phase transfer learning. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 956–966

  67. [2023]

    arXiv preprint arXiv:2302.05927 (2023)

    Transfer learning for bayesian optimization: A survey. arXiv preprint arXiv:2302.05927 (2023)

    2023 arXiv

  68. [2025]

    In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems

    Continual Human-in-the-Loop Optimization. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems . 1–26

    2025

Pith tools

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