acceptodds
Under review as a conference paper at ICLR 2027

CLUE-BO: Adaptive Language Priors for High-Dimensional Bayesian Optimization

Abstract

Bayesian optimization (BO) becomes challenging in high dimensions, where limited evaluations make it difficult for Gaussian process surrogates to reliably infer dimension-wise relevance. However, many optimization problems provide semantic descriptions of their decision variables, offering structural information that standard BO does not directly exploit. We introduce CLUE-BO, a language-guided BO framework that uses an LLM to translate task semantics into dimension-specific priors over ARD lengthscales. Rather than treating these semantic priors as fixed or inherently reliable, CLUE-BO continually evaluates them through predictive comparisons between semantic and neutral GPs. Predictive evidence determines when to trigger reflection, whether a revised prior should be activated, and whether optimization should follow semantic or neutral guidance. Across six high-dimensional benchmarks ranging from 30 to 1,003 dimensions, CLUE-BO achieves the best mean final objective on five benchmarks and remains competitive on the sixth. Ablation and trajectory analyses further show that semantic priors complement sparse numerical observations, while evidence-driven control allows unreliable semantic hypotheses to be revised during optimization. These results demonstrate that task semantics can provide useful structural information for high-dimensional BO when observations alone are insufficient.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian Optimization, Language-Guided Optimization, ARD Gaussian Processes

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_HHxdOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CLUE-BO, a framework for high-dimensional Bayesian optimization that uses LLMs to construct priors over Gaussian process ARD lengthscales based on provided task descriptions and variable semantics. The method employs sequential predictive evidence to monitor the reliability of these priors, trigger LLM-based revisions, and validate updated priors before activation. Experiments on six benchmarks show that CLUE-BO achieves improved performance against a conventional high-dimensional BO methods. Ablation studies examine individual contributions of semantic guidance and the adaptive switching mechanism.

Critical strengths

The paper presents a well-motivated approach to incorporating semantic information into high-dimensional Bayesian optimization. The proposed framework is interesting, and the adaptation mechanism provides a principled way to assess and revise LLM priors, which is lacking in many existing LLM+BO approaches. The experimental results are convincing, demonstrated across a diverse set of high-dimensional benchmarks, and the ablation studies provide useful support for the individual components of the method.

Critical weaknesses

Limited evaluation of LLM generalization. The experiments rely on a single LLM and a set of established optimization benchmarks. It would be interesting to evaluate the method using:

  • a) additional LLMs
  • b) newly constructed synthetic benchmark that is unlikely to have been encountered during pretraining.

Such a benchmark could have a controlled ground-truth sensitivity structure and semantically meaningful variable descriptions. It could be constructed from scratch or by modifying the semantics of existing benchmarks. This would allow to more robustly assess whether the LLM can infer useful dimension-wise priors from the provided task information rather than relying on familiarity with existing benchmarks.

Limited comparison with related LLM-assisted BO methods. While the paper includes a comprehensive set of conventional high-dimensional BO baselines and ablations of the method, it lacks experimental comparisons with more closely related LLM-assisted approaches, particularly CAKE ([1], NeurIPS 2025) and LGBO ([2], ICLR 2026).

The kernel-selection mechanism proposed by CAKE incorporates EI alongside model fit, which could potentially provide an interesting alternative to CLUE-BO's predictive-likelihood-based switching criterion. LGBO, on the other hand, incorporates LLM preferences into the GP surrogate and is primarily motivated by improved suitability for high-dimensional problems (relative to existing LLM+BO approaches such as LLAMBO).

[1] https://arxiv.org/abs/2509.17998

[2] https://arxiv.org/abs/2605.17976

Questions for authors

The proposed controller relies on relative predictive likelihood to decide when semantic priors should be revised, accepted, or reactivated. However, superior predictive performance does not necessarily translate into better optimization performance. Could the authors elaborate on the motivation for using predictive likelihood as the switching criterion? Could a measure more directly related to optimization utility be used instead, e.g. like in CAKE? Is there empirical or theoretical evidence supporting the relationship between the chosen switching criterion and subsequent optimization performance?

8 comments
AuthorsOpenReview ↗

Author Rebuttal1

W1: Limited evaluation of LLM generalization. The experiments rely on a single LLM and a set of established optimization benchmarks. It would be interesting to evaluate the method using:

  • a) additional LLMs
  • b) newly constructed synthetic benchmark that is unlikely to have been encountered during pretraining.

Such a benchmark could have a controlled ground-truth sensitivity structure and semantically meaningful variable descriptions. It could be constructed from scratch or by modifying the semantics of existing benchmarks. This would allow to more robustly assess whether the LLM can infer useful dimension-wise priors from the provided task information rather than relying on familiarity with existing benchmarks.

Answer: Thank you so much for your suggestion.

  • a) additional LLMs

To investigate the effect of LLM choice, we conduct additional experiments using DeepSeek-V4-Pro (our default LLM), GPT-6-Sol, and Gemini 3.6 Flash, while keeping the remaining CLUE-BO components unchanged.

The results obtained so far are summarized below:

LLMNAS201Hopper-v4RoverDNASVMHumanoidStandup-v4
DeepSeek-V4-Pro (Ours)0.7322 (10)1751.13 (10)2.73 (10)−0.2882 (10)−0.1261 (10)138.33 (10)
GPT-6-Sol0.7314 (10)1781.90 (10)1.75 (10)−0.2891 (10)−0.1257 (10)141.30 (10)
Gemini 3.6 Flash0.7294 (10)1275.17 (10)1.33 (10)−0.2894 (10)-0.1287 (10)129.39 (10)

Numbers in parentheses indicate the number of completed random seeds. Higher values are better.

The results show that LLM choice has a noticeable but task-dependent impact on optimization performance. DeepSeek-V4-Pro outperforms Gemini 3.6 Flash on all benchmarks. In comparison, GPT-6-Sol performs better on Hopper-v4, SVM, and HumanoidStandup-v4, whereas DeepSeek-V4-Pro performs better on NAS201, Rover, and DNA. These findings suggest that sufficient reasoning capability may be important for extracting useful task semantics and constructing informative priors. However, stronger general reasoning capability does not necessarily translate into better optimization performance across all tasks. We have added a new subsection, Appendix C.5 (Effect of LLM Choice), to report these additional experiments and discuss the impact of LLM choice on CLUE-BO. The corresponding changes are highlighted in blue for easy reference.

  • b) newly constructed synthetic benchmark that is unlikely to have been encountered during pretraining.

This experiment is still under development. We are further refining its design and will provide an update before the rebuttal deadline.

AuthorsOpenReview ↗

Author Rebuttal2

W2: Limited comparison with related LLM-assisted BO methods. While the paper includes a comprehensive set of conventional high-dimensional BO baselines and ablations of the method, it lacks experimental comparisons with more closely related LLM-assisted approaches, particularly CAKE ([1], NeurIPS 2025) and LGBO ([2], ICLR 2026).

Answer: Thank you for raising this important question. Both CAKE and LGBO primarily evaluate their methods on relatively low-dimensional optimization tasks, with the highest-dimensional benchmarks in their reported experiments having fewer than 20 dimensions. Neither method has been extensively evaluated in the high-dimensional settings considered in our work. We attempted to reproduce both CAKE and LGBO using their publicly available implementations.

CAKE. We attempted to reproduce CAKE using its publicly available implementation. CAKE uses LLMs to generate candidate kernels through crossover and mutation based on accumulated input-output observations. According to the default configuration of its official repository, each BO iteration invokes the LLM once for crossover and once more for mutation with a probability of 0.7. This corresponds to an expected 1.7 LLM calls per iteration, or approximately 5,100 calls per benchmark under our experimental setting of 300 BO iterations and 10 random seeds, this corresponds to approximately 5,100 expected LLM calls per benchmark, incurring substantial time and token costs.

In addition, CAKE maintains a population of six GP kernels. At each iteration, it refits these kernels using the accumulated observations and evaluates their BIC scores. Newly generated crossover and mutation candidates also require separate GP fitting and BIC evaluation before population selection. These repeated GP fitting procedures can introduce substantial time costs in high-dimensional settings.

In contrast, CLUE-BO is designed with LLM inference cost in mind. Rather than invoking the LLM at every BO iteration, it uses an evidence-driven reflection mechanism that triggers LLM calls only when accumulated predictive evidence suggests that the current semantic prior may need revision. The numbers of reflection triggers across six benchmarks and ten random seeds are summarized below.

BenchmarkReflection Triggers
Rover92
Hopper-v485
HumanoidStandup-v455
DNA107
SVM156
NAS20176

Each benchmark contains 3,000 BO observations across ten random seeds.

LGBO. LGBO incorporates LLM-generated point or region preferences into the GP surrogate at every BO iteration. According to its official implementation, each iteration requires an LLM call to generate a new preference based on the task information and accumulated observations, with an additional call if the output cannot be parsed. Under our experimental setting of 300 BO iterations and 10 random seeds, this corresponds to approximately 3,000 LLM calls per benchmark, excluding possible retries. This introduces considerable LLM inference cost. We are further examining the LGBO implementation and its applicability to our high-dimensional benchmarks.

We are currently working on reproducing both CAKE and LGBO under our high-dimensional experimental settings. Although preliminary experimental results are not yet available, we will make every effort to provide the results that can be obtained before the rebuttal deadline.

AuthorsOpenReview ↗

Author Rebuttal3

Q1: The proposed controller relies on relative predictive likelihood to decide when semantic priors should be revised, accepted, or reactivated. However, superior predictive performance does not necessarily translate into better optimization performance. Could the authors elaborate on the motivation for using predictive likelihood as the switching criterion? Could a measure more directly related to optimization utility be used instead, e.g. like in CAKE? Is there empirical or theoretical evidence supporting the relationship between the chosen switching criterion and subsequent optimization performance?

Answer: Thank you for your insightful question. Our choice of relative predictive likelihood is motivated by the need to continuously assess whether the semantic prior remains suitable for GP modeling as new observations become available. As discussed in Appendix A, the one-step log predictive ratio admits a relative KL-divergence interpretation, and its sequential accumulation has an e-process interpretation under the corresponding assumptions. This provides a principled basis for monitoring persistent predictive differences rather than switching priors based on individual predictions, while triggering LLM reflection and subsequent validation only when sufficient evidence accumulates.

We agree that approaches such as CAKE, which combine BIC and EI, offer an interesting alternative by jointly considering model fit and optimization potential. More generally, predictive quality and optimization utility play complementary roles in BO. GP posterior predictions characterize the objective function and provide the basis for acquisition functions, while acquisition functions such as EI directly guide the selection of promising evaluation points. Our relative predictive likelihood is intended to monitor the suitability of the GP prior, rather than directly measure the optimization utility of a candidate prior. Directly incorporating optimization-oriented criteria into our sequential evidence framework would require an additional statistical formulation to establish corresponding guarantees.

Predictive quality is an important consideration in Bayesian optimization, as GP posterior predictions directly influence subsequent acquisition decisions. To further examine the relationship between prior quality and optimization performance, we conduct independent function variation diagnostics on DNA and SVM, with the corresponding analyses added to Appendices C.1 and C.2. The corresponding changes are highlighted in blue for easy reference.

In particular, the SVM diagnostic shows that the initial semantic priors in eight of the ten runs are less consistent with the observed function variation, and these runs achieve lower final objective values than the other two runs. Nevertheless, seven of these eight runs produce at least one accepted reflection candidate that improves this consistency. Despite the less favorable initial allocations, the final objective values of these eight runs are comparable to those of the w/o LLM baseline. Moreover, the ablation results show that w/o Reflection performs worse than w/o LLM. These findings suggest that monitoring predictive evidence and validating reflected proposals can help mitigate the negative effects of potentially unreliable initial priors. Although these results do not establish a direct correspondence between predictive likelihood and optimization utility, they provide empirical support for using predictive evidence to monitor prior suitability during optimization. Exploring optimization-aware sequential evidence remains an interesting direction for future work. We have added this discussion to the Future Work section in Appendix G, with the corresponding changes highlighted in blue for easy reference.

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

The paper introduces CLUE-BO, a language-guided framework for high-dimensional Bayesian optimization that leverages task descriptions and variable-level semantics to construct dimension specific priors over GP ARD lengthscales. LLM generates relative lengthscale preferences, which are incorporated into the GP through LogNormal priors. Rather than treating these semantic priors as inherently reliable, the framework compares the predictive performance of semantic and neutral GPs using a sequential e-process. Based on the accumulated predictive evidence, CLUE-BO can do LLM reflection to revise the existing priors, validate revised priors using subsequent observations, or switch between semantic and neutral guidance during optimization. The method is evaluated on six high-dimensional benchmarks ranging from 30 to 1,003 dimensions, achieving the highest mean final objective on five tasks. The authors further provide ablation and sensitivity studies to investigate the contributions of semantic guidance, reflection, and evidence-driven prior adaptation.

Critical strengths
  1. Integrating of LLM semantic information into lengthscales prior: Rather taking the route of LLM as optimizers it tackles a narrower direction to express knowledge into the ARD length scale priors: proposing hypotheses about variable-wise sensitivity while leaving acquisition optimization to the GP.

  2. Mechanism for monitoring and revising potentially unreliable semantic priors: Instead of assuming that the initial LLM-generated priors remain reliable throughout optimization, the framework continuously compares semantic and neutral GP predictions and uses accumulated evidence to determine when the current prior should be reconsidered and also distinction between generating a revised prior and allowing that prior to influence acquisition: newly proposed priors must undergo prospective validation. The additional ability to temporarily fall back to neutral guidance and subsequently reactivate a previously discarded semantic prior provides flexibility that is absent from approaches relying on fixed semantic assumptions. Importantly, the authors also acknowledge the model-based limitations of the sequential evidence interpretation.

Critical weaknesses
  1. Potential mismatch between semantic variable importance and ARD lengthscale interpretation: The proposed framework assumes that LLM-inferred variable sensitivity can be translated into shorter or longer ARD lengthscale priors. However, ARD lengthscales characterize the scale of function variation and are not necessarily reliable measures of predictive variable relevance. Piironen and Vehtari (2016) demonstrate limitations of lengthscale-based relevance assessment, while Paananen et al. (2019) show that ARD can systematically favor nonlinear variables over equally relevant linear variables. This distinction is important because a variable can strongly influence the objective without necessarily requiring a shorter GP lengthscale. The paper would benefit from a clearer justification of how the LLM distinguishes semantic importance from the smoothness properties relevant to the GP kernel.

  2. Limited empirical evidence that the generated prior allocations capture meaningful dimension-wise sensitivity: Although the authors provide illustrative LLM outputs, the reported examples for Rover and HumanoidStandup-v4 only shorten selected lengthscales, without demonstrating how the method identifies dimensions that should receive longer lengthscales. Furthermore, several adjustments are modest relative to the LogNormal prior spread. For example, with σ_prior = sqrt(3), the HumanoidStandup ratios of 0.85 and 0.92 correspond to shifts of approximately 0.094 and 0.048 log-prior standard deviations, respectively. The paper also does not present the actual initial semantic allocations for DNA and SVM, making it difficult to evaluate whether the LLM identifies meaningful heterogeneous sensitivity in these tasks. Given prior work on statistical relevance modeling through ARD and sparsity-inducing priors (Eriksson & Jankowiak, 2021; Paananen et al., 2019), a comparison between the generated allocations and empirical sensitivity diagnostics would provide stronger support for the claimed semantic contribution.

  3. Incomplete comparison with alternative methods: Although the paper evaluates several competitive high-dimensional BO methods, the selected baselines do not fully cover approaches that explicitly identify or exploit influential input dimensions. Relevant examples include SAASBO (Eriksson & Jankowiak, 2021), MCTS-VS (Song et al., 2022), and MSR (Papenmeier et al., 2025)

More importantly, the experiments lack a simple metadata-informed baseline that uses the same variable descriptions without LLM reasoning. For example, the actuator identities and time-step structure in HumanoidStandup-v4 could be used to construct grouped or partially tied ARD lengthscales. Comparing against such a baseline would help determine whether CLUE-BO benefits from meaningful language-based reasoning or primarily from readily available structural information and reduced flexibility in GP hyperparameter estimation.

Questions for authors
  1. Alternative forms of prior incorporation: Prior-guided methods such as phi-BO (Hvarfner et al., 2022) and ColaBO (Hvarfner et al., 2024) incorporate beliefs about likely optimizer locations or values rather than dimension-wise smoothness. Could the authors discuss the circumstances under which semantic information is better represented as an ARD lengthscale prior versus another type of prior knowledge, and the practical advantages of their choice?

  2. Separating global kernel scaling from semantic structure: The proposed dimension-wise lengthscale ratios modify both the relative ARD allocation and the aggregate kernel distance scale. Could the authors compare CLUE-BO against a uniform lengthscale allocation matched to the semantic prior's aggregate kernel scale, as well as randomly permuted semantic allocations that preserve the original ratios? These comparisons would help determine whether the improvements arise from meaningful semantic correspondence between variables and their lengthscales, rather than favorable global scaling or heterogeneous initialization.

  3. Robustness to misleading semantic priors: Motivation for CLUE-BO ability to reconsider unreliable semantic assumptions. Could the authors evaluate controlled perturbations of initially informative prior allocations, preferably on problems where relevant dimensions or sensitivity structure are known? Examining whether reflection identifies and corrects the perturbed dimensions, and how quickly optimization performance recovers, would provide more direct evidence for the claimed reliability of semantic adaptation.

  4. Empirical behavior of the evidence-driven controller: Could the authors report the frequency of neutral fallback, reflection, candidate acceptance or rejection, and semantic reactivation across benchmarks and random seeds? In particular, what is the distribution of candidate-validation durations, and how frequently are reflections triggered by validation timeouts rather than evidence-threshold crossings? This would help clarify how the proposed controller behaves beyond the representative Rover example, where a revised prior is accepted after only one subsequent observation.

References:

  1. Piironen, J., & Vehtari, A. (2016) -- https://arxiv.org/pdf/1510.04813v3

  2. Paananen et al. (2019) -- https://proceedings.mlr.press/v89/paananen19a/paananen19a.pdf

  3. Eriksson & Jankowiak, 2021 -- https://proceedings.mlr.press/v161/eriksson21a.html

  4. MCTS-VS ---- https://proceedings.neurips.cc/paper_files/paper/2022/file/b6a171867138c80de2a35a6125d6757c-Paper-Conference.pdf

  5. MSR --- https://proceedings.mlr.press/v267/papenmeier25a.html

  6. Phi-Bo --- https://arxiv.org/pdf/2204.11051

  7. ColaBO -- https://proceedings.iclr.cc/paper_files/paper/2024/file/b95f3f8b6cff9e7ac6553a56b895cf70-Paper-Conference.pdf

12 comments
AuthorsOpenReview ↗

Author Rebuttal1

W1: Potential mismatch between semantic variable importance and ARD lengthscale interpretation: The proposed framework assumes that LLM-inferred variable sensitivity can be translated into shorter or longer ARD lengthscale priors. However, ARD lengthscales characterize the scale of function variation and are not necessarily reliable measures of predictive variable relevance. Piironen and Vehtari (2016) demonstrate limitations of lengthscale-based relevance assessment, while Paananen et al. (2019) show that ARD can systematically favor nonlinear variables over equally relevant linear variables. This distinction is important because a variable can strongly influence the objective without necessarily requiring a shorter GP lengthscale. The paper would benefit from a clearer justification of how the LLM distinguishes semantic importance from the smoothness properties relevant to the GP kernel.

Answer: Thank you for mentioning this concern. In our framework, sensitivity refers to the modeling emphasis assigned to different dimensions, rather than a direct estimate of their predictive relevance. Specifically, we intend for the LLM to reason about the functional roles of different variables based on task descriptions and domain knowledge, and identify dimensions that may benefit from greater modeling flexibility. For these dimensions, the LLM may assign shorter lengthscale priors, allowing the GP to better capture potentially faster function variation along them.

As specified in the Initial Semantic Prior Prompt in Appendix D, "below one indicates a shorter prior lengthscale and greater sensitivity to changes along that dimension; one is neutral; and above one indicates a longer prior lengthscale and smoother variation." We do not assume that these initial preferences perfectly capture the true function variation structure. Instead, their predictive usefulness is assessed and adapted through subsequent observations.

To clarify this distinction, we have revised the relevant descriptions throughout the manuscript. Specifically, we replace "dimension-wise relevance" in the abstract with "dimension-wise patterns of function variation" and clarify "relative sensitivity of different dimensions" in the Introduction by adding "in terms of how the objective function responds to changes along each dimension." All corresponding revisions are highlighted in blue for easy reference.

AuthorsOpenReview ↗

Author Rebuttal2 (1/2)

W2: Limited empirical evidence that the generated prior allocations capture meaningful dimension-wise sensitivity: Although the authors provide illustrative LLM outputs, the reported examples for Rover and HumanoidStandup-v4 only shorten selected lengthscales, without demonstrating how the method identifies dimensions that should receive longer lengthscales. Furthermore, several adjustments are modest relative to the LogNormal prior spread. For example, with σ_prior = sqrt(3), the HumanoidStandup ratios of 0.85 and 0.92 correspond to shifts of approximately 0.094 and 0.048 log-prior standard deviations, respectively. The paper also does not present the actual initial semantic allocations for DNA and SVM, making it difficult to evaluate whether the LLM identifies meaningful heterogeneous sensitivity in these tasks. Given prior work on statistical relevance modeling through ARD and sparsity-inducing priors (Eriksson & Jankowiak, 2021; Paananen et al., 2019), a comparison between the generated allocations and empirical sensitivity diagnostics would provide stronger support for the claimed semantic contribution.

Answer: Thank you for your insightful comment.

  • The LLM also generates longer lengthscale priors in the DNA and SVM tasks. For example, the following shows the actual initial semantic allocation generated for SVM (seed 0):
{
  "action": "allocate",
  "default_mode_ratio": 1.0,
  "overrides": [
    {
      "dimensions": [0],
      "mode_ratio": 2.0,
      "hypothesis": "The patient identifier is a categorical variable without meaningful ordinal geometry. A longer lengthscale helps avoid fitting spurious variation along this dimension."
    },
    {
      "dimensions": [385, 386, 387],
      "mode_ratio": 0.5,
      "hypothesis": "The SVR hyperparameters (epsilon, gamma, and C) are known to have significant nonlinear effects on model performance. Shorter lengthscales allow the GP to capture rapid variations in the objective."
    }
  ],
  "rationale": "The patient identifier is a categorical label rather than a continuous variable, making a longer lengthscale appropriate. The SVR hyperparameters have substantial effects on model performance and warrant shorter lengthscales. The bone and air histogram features are retained at the neutral ratio, as individual features may have relatively weak and smooth effects."
}

This example illustrates that the LLM can assign both shorter and longer lengthscale priors based on different variable semantics. We have also added the actual initial semantic allocations for both DNA and SVM to Appendix E.1. The corresponding changes are highlighted in blue for easy reference.

  • Regarding the HumanoidStandup-v4 example, the ratios of 0.85 and 0.92 correspond to shifts of approximately 0.094 and 0.048 log-prior standard deviations, respectively. Although these shifts are modest relative to the prior spread, they represent multiplicative changes in the prior modes. For an ARD kernel, reducing a dimension's lengthscale by a factor of r increases its corresponding squared-distance contribution by a factor of 1/r². Specifically, ratios of 0.85 and 0.92 increase the respective contributions by approximately 38.4% and 18.1%. In our example, these adjustments cover 649 dimensions (11 trunk and lower-limb actuators across 59 time steps) and 150 dimensions (6 upper-limb actuators across the first 25 time steps), respectively. Under the illustrative assumption of equal coordinate differences across all 1,003 dimensions, the combined allocation increases the ARD squared distance by approximately 27.6% relative to the neutral configuration, evaluated at the prior modes. This illustrates how moderate per-dimension adjustments can accumulate to influence the kernel geometry, although their actual effects depend on subsequent GP fitting.
AuthorsOpenReview ↗

Author Rebuttal2 (2/2)

W2: A comparison between the generated allocations and empirical sensitivity diagnostics would provide stronger support for the claimed semantic contribution.

Answer: Thank you for this helpful suggestion. We conduct an independent diagnostic on DNA and SVM to examine whether LLM-generated lengthscale allocations reflect dimension-wise function variation.

Diagnostic metric. We use the relationship between lengthscales and function variation in our Matérn-5/2 GP. When two inputs differ only along dimension jj by a distance δ\delta, their prior correlation is

\left(1+u+\frac{u^2}{3}\right)e^{-u}, \qquad u=\frac{\sqrt{5}\delta}{\ell_j}.$$ The expected squared function increment is $$\mathbb{E}\left[ (F(\mathbf{x}+\delta\mathbf{e}_j)-F(\mathbf{x}))^2 \right] =2\sigma_f^2(1-\rho_j(\delta)).$$ For fixed $\delta$, this quantity increases as $\ell_j$ decreases, motivating controlled dimension-wise perturbations. We therefore vary one dimension at a time while fixing the others. Let $y_{bjq}$ denote the objective value at grid point $q$, for background point $b$ and dimension $j$. We measure the mean squared increment at distance $\delta_h=h/(K-1)$ as $$A_j(\delta_h)= \frac{1}{B(K-h)} \sum_{b=1}^{B} \sum_{q=0}^{K-h-1} (y_{bj,q+h}-y_{bjq})^2.$$ To account for output-scale differences, we normalize the increments by the corresponding slice variance: $$V_j= \frac{1}{B}\sum_{b=1}^{B} \frac{1}{K}\sum_{q=0}^{K-1} (y_{bjq}-\bar y_{bj})^2, \qquad \bar y_{bj}=\frac{1}{K}\sum_{q=0}^{K-1}y_{bjq},$$ and define the relative function variation as $$\boxed{ S_j(\delta)=\frac{A_j(\delta)}{2V_j}. }$$ Larger $S_j(\delta)$ indicates greater relative function variation along dimension $j$. We group dimensions by their initial lengthscale ratios into short ($r_j<1$), neutral ($r_j=1$), and long ($r_j>1$) groups, and examine whether their empirical variation follows the expected ordering. For SVM, we randomly sample eight remaining neutral dimensions for comparison to limit evaluation cost. **Empirical results.** The table reports the results, including the evaluated dimensions and their initial prior ratios. | Benchmark | Prior group | Dimensions | Prior ratio | Mean $S_j(0.125)$ | |:--|:--|:--|:--:|:--:| | DNA (seed 1) | Short | 90--92; 87--89, 93--95 | 0.5 / 0.7 | 0.2300 | | | Neutral | Remaining 165 dimensions | 1.0 | 0.1890 | | | Long | 0--2; 177--179 | 1.5 | 0.1640 | | SVM (seed 2) | Short | 0 | 0.1 | 0.0805 | | | Neutral | 8, 36, 82, 194, 260, 273, 297, 374 | 1.0 | 0.0839 | | | Long | 385--387 | 2.0 | 0.1533 | On DNA, the group means are 0.230, 0.189, and 0.164 for the short, neutral, and long groups, respectively, following the expected decreasing order. This pattern also holds in nine out of ten seeds. This is consistent with the ablation results, where w/o Reflection outperforms w/o LLM, suggesting that the initial semantic guidance provides useful information for this task. SVM exhibits a different pattern. The representative results show slightly lower variation in the short group than in the neutral group, and substantially higher variation in the long group. Across ten seeds, only seeds 0 and 1 follow the expected ordering. Interestingly, these two seeds also achieve better final objective values, around -0.103, whereas the remaining seeds converge to approximately -0.13. This is also reflected in the ablation results, where w/o LLM outperforms w/o Reflection, suggesting that retaining an unreliable initial semantic prior without subsequent revision can be less effective than using a neutral prior. This diagnostic provides an empirical assessment based on sampled inputs, rather than an exact estimate of the underlying function lengthscales or predictive relevance. Since $S_j(\delta)$ is normalized by dimension-wise slice variance, the expected ordering is an empirical hypothesis, not a strict consequence of the Matérn kernel. Nevertheless, the results show that the correspondence between semantic priors and empirical function variation varies across tasks. We have updated Appendix C.1 to present the corresponding experiments. The changes are highlighted in blue for easy reference. **The SVM results also reveal how CLUE-BO responds to potentially misleading initial allocations.** We further analyze the reflection behavior across ten seeds to examine whether the controller can revise initial allocations inconsistent with the empirical diagnostic. This analysis helps explain the SVM ablation results and is discussed in our response to Q3.
Reviewer_e83nOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes to use an LLM to levarge semantic clues to construct priors over ARD lengthscales for bayesian optimization. It also includes a strategy for the adaptation and using of these priors. The experiments consider six high-dimensional benchmarks, many baselines, and also include a comprehensive ablation study.

Critical strengths

Adapting lengthscale preferences based on textual descriptions of a task is an interesting way to leverage LLMs for Bayesian optimization and to make use semantic prior knowledge.

Methodologically, some of the ideas may transfer to other settings, e.g., the evidence-driven reflection and prior switching.

Critical weaknesses

The main paper does not consider optimization trajectories in the evaluation except for the ablation study. The Appendix includes this, and it does show CLUE-BO less favorable than just looking at the final performance. I do not see these results as a critical weakness, but I suggest this Figure should be featured in the main paper.

How semantic information presents itself in text form is barely discussed. The specific information of the benchmarks in the experiments are detailed in the Appendix, but giving some more explanation on this in the methodology part would help in understanding the approach.

Questions for authors
  • What is the effect of the used LLM? How does performance vary with model type and capability. If you do not study this, mention it as a limitation among any other that may come up in the reviewing period (a discussion on limitations is missing)
  • Does your method also help in a lower dimensional setting?
2 comments
AuthorsOpenReview ↗

Author Rebuttal

W1: The main paper does not consider optimization trajectories in the evaluation except for the ablation study. The Appendix includes this, and it does show CLUE-BO less favorable than just looking at the final performance. I do not see these results as a critical weakness, but I suggest this Figure should be featured in the main paper.

Answer: Thank you for your valuable suggestion. We have moved the optimization trajectory figure from the Appendix to the main paper as Fig. 2 and added a discussion in Sec. 4.2. The corresponding changes are highlighted in blue for easy reference.

W2: How semantic information presents itself in text form is barely discussed. The specific information of the benchmarks in the experiments are detailed in the Appendix, but giving some more explanation on this in the methodology part would help in understanding the approach.

Answer: Thank you for your helpful suggestion. We have clarified in Sec. 3.2 how semantic information is provided to the LLM in natural language, including task-specific objectives and dimension-level variable semantics. The changes are highlighted in blue for easy reference.

Q1: What is the effect of the used LLM? How does performance vary with model type and capability. If you do not study this, mention it as a limitation among any other that may come up in the reviewing period (a discussion on limitations is missing)

Answer: Thank you for raising this important question. To investigate the effect of LLM choice, we conduct additional experiments using DeepSeek-V4-Pro (our default LLM), GPT-6-Sol, and Gemini 3.6 Flash, while keeping the remaining CLUE-BO components unchanged.

The results obtained so far are summarized below:

LLMNAS201Hopper-v4RoverDNASVMHumanoidStandup-v4
DeepSeek-V4-Pro (Ours)0.7322 (10)1751.13 (10)2.73 (10)−0.2882 (10)−0.1261 (10)138.33 (10)
GPT-6-Sol0.7314 (10)1781.90 (10)1.75 (10)−0.2891 (10)−0.1257 (10)141.30 (10)
Gemini 3.6 Flash0.7294 (10)1275.17 (10)1.33 (10)−0.2894 (10)-0.1287 (10)129.39 (10)

Numbers in parentheses indicate the number of completed random seeds. Higher values are better.

The results show that LLM choice has a noticeable but task-dependent impact on optimization performance. DeepSeek-V4-Pro outperforms Gemini 3.6 Flash on all benchmarks. In comparison, GPT-6-Sol performs better on Hopper-v4, SVM, and HumanoidStandup-v4, whereas DeepSeek-V4-Pro performs better on NAS201, Rover, and DNA. These findings suggest that sufficient reasoning capability may be important for extracting useful task semantics and constructing informative priors. However, stronger general reasoning capability does not necessarily translate into better optimization performance across all tasks. We have added a new subsection, Appendix C.5 (Effect of LLM Choice), to report these additional experiments and discuss the impact of LLM choice on CLUE-BO. The corresponding changes are highlighted in blue for easy reference.

We have added a new section, Appendix G (Limitations and Future Work), to discuss the limitations of CLUE-BO and potential directions for future research. The changes are highlighted in blue for easy reference.

Q2: Does your method also help in a lower dimensional setting?

Answer: Thank you for your insightful question. CLUE-BO is primarily designed for high-dimensional Bayesian optimization, where learning reliable ARD lengthscales from limited observations is challenging. Following the theoretical motivation of SBO[1] and VBO[2], we adopt d\sqrt{d} as the neutral lengthscale prior to mitigate the vanishing-gradient problem in GP hyperparameter learning. ARD kernels are also commonly used in lower-dimensional BO. However, in lower-dimensional settings, ARD lengthscales can generally be learned more reliably from data. Therefore, whether d\sqrt{d} remains suitable as the neutral prior requires further investigation.

Nevertheless, providing lengthscale priors that better reflect the actual variation patterns of the objective function is theoretically expected to benefit ARD modeling. However, extending CLUE-BO to lower-dimensional tasks requires further investigation into an appropriate neutral prior.

[1] https://arxiv.org/abs/2402.02746

[2] https://arxiv.org/abs/2402.02229

Dear Reviewer e83n,

As the discussion phase comes to an end, please acknowledge the author's response and consider leaving them a comment on whether it has resolved some or all of your concerns.

Thanks!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.