acceptodds
Under review as a conference paper at ICLR 2027

Calibrating LLM Guidance for Acquisition Selection in Bayesian Optimization

Abstract

Bayesian optimization (BO) is a sample-efficient approach to black-box optimization, but its performance often depends on task-specific design choices, including acquisition functions and trust region methods. Identifying effective configurations often requires human expertise or substantial manual experimentation. Large language models (LLMs) offer a promising approach to automatically recommend these choices using task descriptions and optimization histories. However, it is reportedly difficult for LLMs to track long interaction histories efficiently, which can undermine the reliability of their recommendations for complex tasks. We propose LLM-Prior-Informed Gaussian Upper Confidence Bound (LPG-UCB), a framework that combines LLM guidance with empirical feedback to adaptively select acquisition functions. LPG-UCB encodes LLM recommendations as priors over acquisition function preferences and updates these preferences through Bayesian inference as the BO loop proceeds. This Bayesian update mechanism leverages summary statistics from the BO trajectory to refine and calibrate LLM guidance, reducing reliance on the accuracy of individual LLM recommendations. We model LPG-UCB as a non-stationary rotting-bandit framework, and derive an expected cumulative regret bound, which grows sublinearly under appropriate conditions. Extensive experiments on diverse synthetic benchmarks demonstrate faster convergence than competing LLM-based selection methods. LPG-UCB also achieves competitive performance on a high-dimensional black-box optimization problem.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian Optimization, Large Language Models, Adaptive Strategy Selection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_q67yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for using LLM to select acquisition functions during Bayesian optimization. Instead of following the recommendation from LLM directly, the author proposes to ask the LLM to assign preference scores to each acquisition function and treats these scores as prior over their expected reward. Then those priors are updated using observed reward feedback. The paper models acquisition-function selection as a non-stationary multi-armed bandit problem. The author provides a cumulative regret analysis and evaluates the method on several benchmarks and a high-dimensional black-box optimization problem.

Critical strengths

The idea of using LLM to choose a proper acquisition function and treat the selection process as a multi-armed bandit problem is quite novel and interesting. The author also provides a theoretical analysis of the cumulative regret and conducts several numerical analyses to validate the algorithm.

Critical weaknesses

The paper already formulates AF selection as a multi-armed bandit problem, for which many adaptive selection algorithms already exist. Since the number of candidate AFs is also relatively small, it is not very clear how much additional benefit is provided by introducing the LLM prior, especially since the final AF selection is still made by the Gaussian-UCB bandit algorithm. Although the paper compares against several existing adaptive AF-selection methods and includes a random-prior baseline, the contribution of the LLM prior is still not fully isolated. Furthermore, the authors show that the numerical selector frequently overrides the LLM, and 61% of improvement events are associated with override decisions. That is interesting evidence that the calibration mechanism is useful, but it also raises a question: how much of LPG-UCB’s advantage actually requires an LLM?

Questions for authors

1, Since the number of candidate AFs is relatively small, why is an LLM needed to guide the selection rather than relying only on an existing adaptive bandit algorithm? Could the author further explain why the LLM-based method has superior performance based on the numerical results?

2, As the final AF selection is still determined by the Gaussian-UCB selector, how much of the performance improvement actually comes from the LLM prior?

3, Could the authors provide a direct ablation using the same Gaussian-UCB selector but without LLM guidance, for example, with a uniform prior shared by all AFs? This would help isolate the contribution of the LLM from that of the bandit-based adaptive selection itself.

5 comments
AuthorsOpenReview ↗

Thank you for your comments!

Thank you for your detailed and constructive feedback. We have addressed your specific comments below and incorporated the corresponding revisions into the manuscript.

1. Benefits of LLM Guidance over Conventional Adaptive Bandit Algorithms

We thank the reviewer for this important question. We agree that the relatively small number of candidate AFs raises a reasonable question about the need for LLM guidance. Our motivation for incorporating an LLM is to provide additional context-aware guidance for AF selection, particularly when the available optimization feedback is limited. Conventional adaptive bandit methods, such as GP-Hedge, primarily adapt their AF selection based on accumulated reward feedback. While effective, these methods do not explicitly incorporate additional optimization-state descriptors, such as GP length scales, distances between previously evaluated points, and recent improvement history, into their AF selection policy. In contrast, LLMs offer a flexible mechanism for jointly interpreting these contextual features and translating them into informed prior preferences over candidate AFs. This capability is particularly valuable during the early stages of BO, when limited observations are available, as LLM-generated preferences can provide a knowledge-based warm start for AF selection before sufficient numerical feedback has accumulated. Furthermore, scientific optimization problems may provide additional domain-specific knowledge, such as physical constraints or prior experimental insights, which can also be incorporated into LLM-based guidance.

Importantly, LPG-UCB does not rely solely on LLM recommendations. Instead, it uses LLM-generated preferences as prior information within a Gaussian-UCB framework, allowing the initial guidance to be continuously updated based on optimization feedback. This integration combines context-aware LLM guidance with feedback-driven strategy selection, rather than replacing conventional bandit learning. Moreover, the proposed framework is not inherently restricted to AF selection and can be extended to the selection of other BO strategies, such as global and local search regimes.

To further clarify the benefits of using an LLM prior over conventional adaptive bandit algorithms, we have expanded the discussion in Section 7 to emphasize the warm-start capability of LLM-generated preferences, their integration with numerical feedback, and the potential extension to other BO strategy-selection problems. The corresponding revisions are highlighted in red in the manuscript.

AuthorsOpenReview ↗

Thank you for your comments!

2. Contribution of LLM prior on performance improvement

We thank the reviewer for this valuable comment. We agree that a more direct evaluation of the contribution of LLM-generated priors would strengthen our study. In the original manuscript, we included comparisons with conventional non-LLM adaptive AF selection methods, including GP-Hedge, No-PASt-BO, and SETUP-BO, as reported in Table 1. Across the 31 synthetic benchmarks, LPG-UCB achieved a mean RAUC of 1.51, compared with 2.48, 2.49, and 2.53 for GP-Hedge, No-PASt-BO, and SETUP-BO, respectively. However, we recognize that these comparisons alone do not fully isolate the contribution of LLM-generated priors from that of the Gaussian-UCB selection mechanism.

To address this limitation and better understand the specific contribution of LLM-generated priors, we conducted additional ablation experiments following the reviewer's suggestion. First, we considered a uniform prior, assigning a fixed prior score of 0.5 to all AFs. This experiment was designed to isolate the contribution of the LLM-generated prior from that of the Gaussian-UCB selector itself, allowing us to evaluate whether LLM guidance provides additional benefits beyond reward-based adaptive AF selection. Second, we considered a linearly scheduled prior to investigate whether the LLM-generated prior provides more effective guidance than a simple heuristic-based strategy. Specifically, the prior scores of exploration-oriented AFs were linearly decreased from 1 to 0, while those of exploitation-oriented AFs were linearly increased from 0 to 1 as the optimization budget progressed. This comparison allows us to assess whether LLM-generated preferences offer benefits beyond a predefined exploration-to-exploitation schedule. We additionally included a random prior, assigning randomly generated prior scores to AFs, to examine whether the observed improvements could be attributed merely to arbitrary variations in prior initialization.

Across the benchmark functions listed in Table 5, LPG-UCB achieved a mean RAUC of 1.38, compared with 2.56, 2.57, and 2.32 for the uniform, random, and linearly scheduled priors, respectively. These results correspond to reductions in mean RAUC of approximately 46.1%, 46.3%, and 40.5%, respectively. These findings demonstrate that the LLM-generated prior provides a meaningful contribution to the performance of LPG-UCB beyond the Gaussian-UCB selection mechanism alone. Furthermore, the superior performance over the linearly scheduled prior suggests that the LLM-generated prior can provide more effective guidance for AF selection than a simple heuristic by leveraging the optimization context.

We have incorporated these additional ablation results and the corresponding discussion into Section 5.1.3 and Appendix G.2 of the revised manuscript, with the changes highlighted in red.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Revised Manuscript

We would like to respectfully follow up on our rebuttal and revised manuscript. In response to your valuable comments and suggestions, we have further clarified the motivation for incorporating LLM guidance into adaptive AF selection and conducted additional ablation experiments using uniform, random, and linearly scheduled priors while keeping the Gaussian-UCB selector unchanged. These experiments were intended to directly address your central concern regarding the contribution of LLM-generated priors beyond the numerical selection mechanism. All corresponding revisions have been highlighted in red in the revised manuscript for your convenience.

We sincerely hope that these additional experiments and explanations have helped clarify the distinct contribution of LLM guidance and address your concerns. We would be grateful for any further feedback you might have and would greatly appreciate knowing whether our rebuttal and revisions have affected your overall evaluation of the paper.

Thank you very much for your careful review and valuable suggestions, which have helped us strengthen our work.

Reviewer_wCsgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper improves acquisition function selection. It uses an LLM to track long interaction histories to construct a Gaussian UCB that helps propose acquisition functions. Essentially through this process of constructing a UCB, they tackle the problem of erroneos suggestions from the LLM. They demonstrate the results using 31 synthetic benchmark problems and one high dimensional ROVER benchmark data.

Critical strengths

The idea of dealing with the uncertainty of LLM priors through an arm specific gaussian update is interesting. The problem they tackle is a genuine one.

Critical weaknesses

• While the results show performance vs key baselines like Lmabo and labo in synthetic baselines, these are missing in the evaluation in real experimental settings. Why? I would consider this important. • Also, the extension to a new space which is high dimensional BO is not motivated well, as the fundamental problem in high dimension BO is search, not necessarily the acquisition function. A better fitting real world experiment would have been more convincing. • This is my greatest concern: Section 3.3: It appears BO has to be run again, not reusing old trajectories. This is not mentioned or accounted for in extra function evaluation. It appears that the comparison is not fair because of this additional budget.

Questions for authors

It also appears that POSMEAN is selected most frequently as the acquisition function - a bit surprising as in previous works it is more likely to be EI. POSMEAN is likely to be very exploitative, perhaps the authors can explain this.

4 comments
AuthorsOpenReview ↗

Thank you for your comments!

We sincerely appreciate your thoughtful and constructive comments. Below, we provide point-by-point responses to your feedback, and the corresponding changes have been reflected in the revised manuscript.

1. Additional result of LLM-based method on high-dimensional problem

We thank the reviewer for highlighting the importance of evaluating LLM-based baselines on real-world optimization problems. In response, we have extended our experiments on the Rover benchmark to include LMABO and LABO. For LMABO, we use the same eight-arm portfolio as LPG-UCB, consisting of four acquisition functions under both global and TuRBO search regimes, ensuring a consistent comparison. The additional results demonstrate that our TuRBO-enhanced LPG-UCB outperforms both LMABO and LABO on the Rover benchmark. These findings suggest that adaptively selecting both the search regime and acquisition function by combining LLM-generated preferences with numerical feedback through Gaussian-UCB can provide additional benefits in high-dimensional optimization.

We have added the experimental results to Table 2, the corresponding discussion to Section 5.2.2, and the detailed experimental settings to Appendix D.3, with the revisions highlighted in red.

2. Motivation of extension to HDBO

We thank the reviewer for this valuable comment. We agree that search is a fundamental challenge in high-dimensional BO and that the motivation for our extension should be more clearly explained. Various approaches have been proposed to address the challenges of high-dimensional BO, including local trust-region search (Eriksson et al., 2019), improved GP prior assumptions (Hvarfner et al., 2024), and alternative surrogate models (Doumont et al., 2026). Recent studies have also highlighted the importance of GP initialization and local search behavior (Papenmeier et al., 2025). These findings suggest that no single modeling or search strategy can be assumed to perform best across diverse high-dimensional optimization problems, highlighting the importance of selecting appropriate strategies based on problem characteristics and the evolving optimization state. This motivates adaptive strategy selection, particularly in high-dimensional settings where objective evaluations are expensive and evaluation budgets are limited. Importantly, our extension is not limited to AF selection. LPG-UCB jointly selects an acquisition function and a search regime (global search or TuRBO-based local trust-region search), enabling adaptive control over search strategy based on the evolving optimization state.

We have revised Section 5.2 of the manuscript to clarify this motivation and emphasize the role of adaptive search strategy selection in our high-dimensional extension. The corresponding changes are highlighted in red.

3. Discussion on budget efficiency of LPG-UCB

We thank the reviewer for raising this important concern. We fully agree that evaluation budget efficiency and fairness are critical considerations in BO, particularly when objective function evaluations are expensive. We appreciate the opportunity to clarify how reward feedback is obtained in LPG-UCB. In our framework, the selected AF receives feedback from the true objective evaluation, while unselected AFs receive GP-based pseudo-feedback computed from the GP posterior mean at their proposed candidate points. This allows the Gaussian-UCB selector to incorporate feedback from multiple AF arms with relatively low computational overhead.

At each iteration, the Gaussian-UCB selector combines the LLM-generated prior preferences with the reward feedback to determine the AF arm with the highest UCB value. The true objective function is then evaluated only at the candidate point proposed by the selected AF, resulting in a single true objective evaluation per iteration. A similar GP-based surrogate feedback mechanism is also employed in GP-Hedge, which uses GP predictions to estimate rewards for candidate acquisition functions without requiring additional true objective function evaluations. Therefore, LPG-UCB does not require additional BO runs or extra true objective function evaluations to update the Gaussian-UCB selector. LPG-UCB operates under the same objective function evaluation budget as the compared baselines, ensuring a fair comparison.

We have also clarified this point in Section 3.3 of the revised manuscript, with the corresponding changes highlighted in red.

AuthorsOpenReview ↗

Thank you for your comments!

4. Discussion on preference over certain acquisition function

We thank the reviewer for this insightful observation. We agree that PosMean is a highly exploitative acquisition function, and its frequent selection deserves further explanation. Interestingly, our analysis of raw LLM preference scores indicates that PosMean is not among the most strongly preferred AFs, even in the later stages of optimization. This suggests that its high selection frequency is not primarily driven by the LLM-generated prior, but may instead be associated with the reward-feedback mechanism of the Gaussian-UCB selector. In LPG-UCB, unselected AFs receive GP-based pseudo-feedback, with their rewards calculated using the GP posterior mean at their proposed candidate points:

\frac{f_{\mathrm{best}}^{t-1} - \mu_{t-1}^{\mathrm{GP}}(x_t^{(i)})} {\sigma_y^t + \epsilon}, -c, c \right), \quad i \neq I_t.$$ Since our optimization tasks are formulated as minimization problems, PosMean selects the candidate point with the lowest GP posterior mean, $$x_t^{\mathrm{PosMean}} \in \arg\min_x \mu_{t-1}^{\mathrm{GP}}(x).$$ For unselected AFs, the GP-based pseudo-reward is calculated such that lower GP-predicted objective values correspond to higher rewards. Consequently, PosMean directly optimizes the same quantity used to determine these pseudo-rewards, giving it a structural advantage in receiving favorable feedback compared with other AFs when it is not selected. This structural advantage may become more influential as the Gaussian-UCB selector progressively places greater weight on accumulated reward feedback than on the initial LLM-generated prior. To further investigate whether the frequent selection of PosMean is driven by LLM preferences, we analyzed the AF selection frequencies under our random-prior ablation across all 31 benchmarks. In this setting, LLM-generated preference scores are replaced with random scores, while the Gaussian-UCB selection and GP-based reward-feedback mechanisms remain unchanged. As shown in Tables 10 and 11, PosMean remains the most frequently selected AF, accounting for 48.55% of all selections under the random-prior setting, compared with 55.02% in LPG-UCB. Furthermore, although PosMean is inherently exploitative, its increasing selection frequency in later optimization stages is consistent with a greater emphasis on exploitation as BO progresses. As observations accumulate, the GP surrogate may provide a more informative approximation of the objective landscape, making it increasingly beneficial to focus on promising regions. Together with the mathematical relationship between PosMean and GP-based pseudo-feedback, these results suggest that the frequent selection of PosMean is associated with the interaction between surrogate-based reward estimation and the Gaussian-UCB selection mechanism. We have added the corresponding discussion to **Section 6** of the revised manuscript and highlighted the changes in red.
AuthorsOpenReview ↗

Follow-up on Rebuttal and Revised Manuscript

We would like to respectfully follow up on our rebuttal and revised manuscript. In response to your thoughtful comments and concerns, we have conducted additional comparisons with LLM-based baselines on the Rover benchmark, clarified the motivation for our high-dimensional BO extension, and provided further explanations regarding objective evaluation budget fairness and the frequent selection of PosMean. All corresponding revisions have been highlighted in red in the revised manuscript for your convenience.

We hope that our additional experiments and explanations have provided satisfactory answers to your questions, particularly regarding the evaluation budget and the high-dimensional experiments. We would greatly appreciate any further feedback you may have and would be grateful to know whether our revisions have affected your assessment of the paper.

Thank you again for taking the time to review our work and for your valuable suggestions.

AuthorsOpenReview ↗

Summary of Contributions and Rebuttal

We sincerely thank the reviewer for the constructive feedback and for recognizing the importance of addressing uncertainty in LLM-generated priors. We would like to briefly summarize our main contributions and emphasize how the additional analyses and revisions motivated by your feedback have further strengthened our paper.

Our main contribution is combining LLM-guided acquisition selection with adaptive Gaussian-UCB calibration, enabling the selection of acquisition functions based on both contextual guidance and numerical feedback.

To address your main concern, we clarified that LPG-UCB requires only one true objective evaluation per BO iteration, with GP-based pseudo-feedback used for unselected acquisition functions. Thus, all methods are compared under the same evaluation budget.

We also added LMABO and LABO comparisons on the Rover benchmark, where LPG-UCB achieved the best mean performance among the compared methods, and clarified the motivation for jointly selecting acquisition functions and global/local search strategies.

Finally, we explained the frequent selection of PosMean through its interaction with GP-based pseudo-feedback.

We hope these clarifications and additional experiments address your concerns and strengthen the contributions of our work.

Reviewer_hHysOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LPG-UCB, a framework for adaptive acquisition function selection in Bayesian optimization. The idea is to use an LLM to generate preferences over candidate acquisition functions based on the current optimization state, and to treat these preferences as priors that are subsequently calibrated via empirical reward feedback and a Gaussian-UCB selector. Experiments on synthetic benchmarks demonstrate improvements over fixed acquisition functions, classical adaptive AF selection methods, and existing LLM-based BO approaches such as LLAMBO. The method is also extended to jointly select acquisition functions and global/local search strategies on a high-dimensional optimization benchmark. The authors additionally provide a regret analysis under a non-stationary rotting-bandit formulation.

Critical strengths

S1. Interesting and novel approach. Combining LLM-generated preferences with a principled selection of AFs is a novel way to leverage LLM priors for BO.

S2. Strong empirical evaluation. The authors evaluate their method across a diverse set of synthetic benchmarks and compare against a comprehensive collection of baselines, including fixed AFs, established adaptive portfolio methods, and recent LLM-based approaches. The results show consistent improvements.

S3. Potential for broader insights. Beyond the proposed algorithm, the use of LLMs to reason about the suitability of acquisition functions at different stages of optimization is an interesting direction. Understanding the patterns underlying these recommendations could provide valuable insights into adaptive acquisition-function selection more generally.

Critical weaknesses

My main concern is the limited analysis of the LLM-generated acquisition-function preferences themselves. Given that the central motivation of the paper is to leverage an LLM's ability to interpret the optimization state and recommend suitable strategies, I would like to better understand what the LLM actually recommends, how these recommendations change depending on context and optimization stage, and what information drives them.

Figure 2a shows an interesting pattern in which the final selector initially favors exploratory functions, mostly JES, before quickly concentrating on more exploitative strategies such as PosMean, PI, and LogPI. However, Figure 2a displays the combined effects of the LLM preferences and the Gaussian-UCB selection mechanism. It is difficult to determine how much of this behavior originates from the LLM itself.

I would encourage the authors to expand the analysis of the LLM preferences in several directions:

  • Evolution of raw LLM preferences. It would be informative to provide a plot analogous to Figure 2a, but showing the acquisition functions preferred by the LLM before the Gaussian-UCB selection step. This would help identify whether the LLM itself follows an exploration-to-exploitation schedule or whether this behavior primarily emerges from the numerical selector.
  • Understanding the LLM's reasoning. Representative examples of LLM recommendations and their accompanying reasoning, together with an analysis of how preferences depend on different aspects of the optimization state, would be valuable. Does the LLM primarily rely on simple indicators such as optimization progress, or does it adapt its recommendations based on problem-specific information, GP hyperparameters, and recent optimization outcomes? It would be interesting to perform a post-hoc analysis of the reasoning traces to understand the heuristics used by the LLM in its decision-making process.
  • Comparison with simpler heuristics. An especially interesting question is whether the behavior learned or exhibited by the LLM could be approximated by much cheaper heuristic strategies. For instance, one could consider a predefined schedule that favors exploration early and exploitation later, or a single UCB acquisition function with a decaying exploration coefficient. Replacing the LLM-generated preferences with heuristic preference scores while retaining the same Gaussian-UCB selector would help isolate the added value of LLM guidance.

I see the above as an opportunity to strengthen the contribution rather than a fundamental limitation of the proposed method. The experiments already provide evidence that the LLM-generated priors are beneficial compared with random priors. However, understanding why these priors are informative could be even more valuable.

If the LLM exhibits systematic, interpretable patterns in its recommendations, these could potentially be distilled into cheaper adaptive AF selection heuristics with comparable performance. Such an analysis could provide insights that extend beyond the proposed LPG-UCB framework and contribute more broadly to our understanding of effective acquisition-function selection in BO.

Questions for authors

Q1. Did you try ablating the prompt providing the taxonomy of AFs. Prompt in E3. provides strong priors on how different AFs behave. In particular it groups them into exploration-focused, balanced, and exploitation-focused. How does the method perform when no such grouping is provided?

Q2. How does the order in which the AFs are presented in the prompt influence the preferences of the LLM? It is well-known that LLMs have strong order biases.

Q3. The prompt in E.3. contains the following sentence: Empirically, this and other exploitation-focused AFs have performed strongly across a wide range of benchmarks, including several that are theoretically considered difficult -- do not assume exploitation-focused AFs are only appropriate in "easy" or late-stage settings. Was this part of the prompt added post-hoc? I.e. you observed that without this clause the LLM had a tendency to not choose these functions despite them performing well in other settings than "easy" or "late-stage".

7 comments
AuthorsOpenReview ↗

Thank you for your comments!

Thank you for your valuable comments and suggestions. We provide detailed responses to each of your points below and have updated the manuscript accordingly.

1. Evolution of raw LLM preferences

We thank the reviewer for this insightful comment. We have added an analysis of raw LLM preferences over the normalized optimization budget, showing both the mean preference scores and the frequency of top-ranked AFs before Gaussian-UCB selection. The results indicate that the LLM does not follow a clear exploration-to-exploitation schedule, as exploration-oriented AFs continue to receive relatively high preference scores even in later optimization stages. In particular, UCB and PES maintain relatively high LLM preference scores, while rarely being selected by Gaussian-UCB in the later stages. This contrast suggests that the observed shift toward exploitation-oriented AFs primarily emerges from the Gaussian-UCB selection mechanism rather than from a corresponding shift in the raw LLM preferences, as the selector adapts its decisions based on accumulated numerical reward feedback.

We have included the corresponding figures in Appendix H.2 and the discussion in Section 6 of the revised manuscript, with the changes highlighted in red.

2. Understanding the LLM's reasoning

We thank the reviewer for this valuable suggestion. To better understand the LLM's decision-making behavior, we conducted a post-hoc analysis of representative LLM-generated justifications under different optimization states. The examples illustrate how the LLM incorporates optimization context, including GP lengthscale variability, reduces preference scores following repeated failures, and considers the recent performance of multiple AFs when evaluating alternative strategies.

We have added these representative examples and their corresponding optimization states in Table 4 of Appendix F, along with a discussion of these findings in Section 6, with the changes highlighted in red.

3. Comparison with simpler heuristics

We thank the reviewer for this valuable comment. To investigate whether LLM-generated preferences provide advantages over simpler heuristic approaches, we conducted an additional ablation study using a linearly scheduled prior while keeping the Gaussian-UCB selector unchanged. Specifically, the prior means of exploration-oriented AFs were linearly decreased from 1 to 0, while those of exploitation-oriented AFs were linearly increased from 0 to 1 as the optimization budget progressed. For AFs categorized as balanced between exploration and exploitation in our prompt, the prior mean was fixed at 0.5 throughout the optimization. This configuration implements a simple exploration-to-exploitation heuristic without requiring LLM guidance. Across the benchmark functions listed in Table 5, LPG-UCB achieved a mean RAUC of 1.38, compared with 2.32 for the linearly scheduled prior, corresponding to a 40.5% reduction in mean RAUC. These results suggest that the LLM-generated prior can provide more effective guidance for AF selection than a simple heuristic by leveraging the optimization context, rather than relying solely on a predefined exploration-to-exploitation schedule.

We have incorporated this additional ablation study and the corresponding discussion into Appendix G.2 and Section 5.1.3 of the revised manuscript, with the changes highlighted in red.

AuthorsOpenReview ↗

Thank you for your comments!

4. Ablation test on taxonomy of acquisition functions prompts

We thank the reviewer for this valuable suggestion. To assess the influence of the AF taxonomy, we conducted an additional ablation study in which the explicit grouping of AFs into exploration-focused, balanced, and exploitation-focused categories was removed from the LLM prompt. Importantly, we retained the individual descriptions of all AFs, thereby isolating the effect of the taxonomy while preserving the information about each acquisition function. The results show that removing the taxonomy has only a limited impact on the overall performance of LPG-UCB. Specifically, the mean RAUC changes slightly from 1.38 to 1.36, while the mean rank increases from 6.60 to 8.00. Additionally, the coefficient of variation of AUC decreases from 0.35 to 0.25 in the taxonomy-free setting. These findings suggest that the taxonomy provides useful structural guidance but is not essential to the effectiveness of LPG-UCB. In particular, explicit categorization may help a moderately sized LLM distinguish the characteristics of different AFs, while the comparable performance of the taxonomy-free setting indicates that the LLM can still generate effective preferences based on individual AF descriptions and the current optimization state.

We have added a discussion of this ablation study in Section 5.1.3, the corresponding prompt details in Appendix E.10, and the experimental settings and results in Appendix G.3, with all changes highlighted in red.

5. Sensitivity analysis on order of acquisition functions

We thank the reviewer for highlighting the potential influence of AF presentation order on LLM preferences. To investigate this issue, we constructed 10 prompt variants by randomly permuting the presentation order of the 12 AFs, while keeping their individual descriptions and all other prompt instructions unchanged. We evaluated these prompt variants across benchmarks listed in Table 5, using 10 random seeds. We focused on the first 5 BO iterations, where the influence of LLM-generated priors is expected to be strongest before substantial numerical feedback accumulates in the Gaussian-UCB selector. To avoid potential bias introduced by explicit AF categorization, we employed the taxonomy-free prompt described in Section E.10.

The results show that LLM preferences vary across AF presentation permutations, particularly in the frequency of top-ranked AFs. Exploration-oriented AFs, especially TS, JES, and PosSTD, frequently receive the highest preference scores, whereas exploitation-oriented AFs are rarely ranked first. However, this tendency is not uniform across exploration-oriented AFs, as MES, KG, and PES rarely receive the highest scores, even when presented near the beginning of the prompt. Furthermore, the analysis of mean preference scores by prompt position reveals no clear systematic positional bias: AFs presented earlier do not consistently receive higher scores, nor do those presented later. These findings suggest that although AF presentation order influences LLM preferences, the observed variations cannot be explained by a simple bias toward AFs appearing earlier or later in the prompt.

We have included the experimental settings, corresponding figures, and detailed discussion in the Appendix G.4 and Section 5.1.3 of the revised manuscript with the changes highlighted in red.

6. Post-hoc on prompt

We thank the reviewer for this insightful comment. We included this instruction in our prompt before conducting any BO experiments, rather than introducing it based on post-hoc observations of benchmark performance. The instruction was designed to mitigate potential biases in LLM-generated AF preferences by discouraging simplistic heuristics, such as assuming that exploitation-oriented AFs are appropriate only for easy problems or late-stage optimization, without adequately considering the current optimization context. Instead, we aimed to encourage the LLM to evaluate AFs based on the current optimization state and available BO information. Thus, this instruction was incorporated as a preventive measure against potential selection biases, rather than as a post-hoc adjustment to favor particular AFs.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Revised Manuscript

We would like to respectfully follow up on our rebuttal and revised manuscript. In response to your thoughtful comments and suggestions, we have provided additional analyses of raw LLM preferences and their reasoning, along with new ablation studies on heuristic priors, AF taxonomy, and presentation order. All corresponding revisions have been highlighted in red in the revised manuscript for your convenience.

We sincerely hope that these additional results and clarifications have helped address your concerns and better explain the contributions of our work. If you have had an opportunity to review our responses and revisions, we would be very grateful to hear whether they have addressed your concerns and whether they might have influenced your overall assessment.

Thank you very much for your time, careful consideration, and constructive feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.