acceptodds
Under review as a conference paper at ICLR 2027

HapBO: Harnessed α–π Bayesian Optimization via Expert Natural Language Priors

20 views

Abstract

Bayesian Optimization (BO) can be substantially more sample-efficient when informed by prior knowledge about the problem. In practice, however, such knowledge is often available only in natural language, e.g., expert intuitions or documentation, rather than as the explicit probability distribution landscape that prior-guided BO methods may require. Leveraging it effectively demands both an expressive mapping from language to priors and robustness to information that may be misleading. We propose HapBO (Harnessed - Bayesian Optimization), which elicits beliefs once from a human expert, a large language model (LLM), or both, and deterministically compiles a structured specification into a reusable belief function (), with a rich representation not restricted to a Gaussian family. Building on -BO, HapBO applies this belief function through a bounded positive multiplier on any nonnegative acquisition, preserving positive acquisition support and requiring no LLM calls during optimization. We prove that, for the Expected Improvement acquisition function (EI), HapBO preserves the standard simple-regret rate up to a bounded factor, even under adversarial beliefs. We evaluate HapBO on 40 HPOBench tasks, a synthetic suite in 6–15 dimensions and two real-world applications: cortical neurostimulation and constrained solar design. HapBO achieves the lowest aggregate HPOBench regret and strong real-world task performance, with zero in-loop token cost. In real-world applications it improves the optimization outcome even at budgets of 8-30 queries, and at a small fraction of the wall-clock and token costs of other LLM-based methods.

Primary area:
optimization
Keywords:
Bayesian optimization, domain knowledge integration, prior-weighted acquisition function, large language models

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.75 over 4 reviews, above 99% of the others.
Reviewer_Kg2COpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes HAPBO (Harnessed α\alpha-π\pi Bayesian Optimization), a framework for incorporating expert prior knowledge into Bayesian optimization (BO). Unlike conventional approaches that typically encode prior knowledge as probability distributions or preference information, HAPBO starts from natural-language descriptions of expert knowledge. It uses a large language model (LLM) to convert such knowledge into a structured Prior Specification Language (PSL), which is then deterministically compiled into a numerical belief function. The resulting prior information is incorporated into the acquisition function through a bounded multiplicative factor, in a manner related to π\piBO.

Critical strengths
  1. The proposed framework is clearly motivated and well presented. The overall methodology is coherent, and the paper provides a structured approach to translating natural-language expert knowledge into a form that can be used by BO. The topic is meaningful, and the proposed framework offers an interesting and potentially useful contribution.
  2. The experimental evaluation is extensive, covering multiple benchmark tasks and real-world applications, with additional ablation studies and robustness analyses. The paper also provides theoretical analysis to support the proposed approach, making the overall study comprehensive.
Critical weaknesses

I did not identify any major flaws in the overall framework. However, several aspects of the experimental evaluation and methodological design require further clarification. My main concerns are detailed in the questions below.

Questions for authors
  1. Motivation for the bounded multiplicative factor. The proposed method selects the next query according to $ x_t \in \arg\max_{x\in\mathcal{X}} A_t(x)M(x), \quad M(x)=(1-\alpha)+\alpha\pi(x). $ In contrast, π\piBO directly uses M(x)=π(x)M(x)=\pi(x). What is the motivation for introducing the parameter α\alpha and the additive term 1−α1-\alpha? What advantages does this bounded formulation offer over the original π\piBO formulation, particularly in terms of robustness to inaccurate or misleading prior information? Could the authors provide further empirical or theoretical evidence to clarify the benefits of this design choice?

  2. Unexpected performance of random search and the missing π\piBO baseline. In Figure 2, random search appears to outperform several BO-based baselines, ranking approximately third among the compared methods. This result seems somewhat unexpected, given the potential sample-efficiency advantages of BO. Could the authors explain why random search performs so well in this setting? In addition, why is the original π\piBO method not included as a baseline in Figure 2? Since π\piBO is directly related to the proposed approach, a comparison would help clarify the contribution of HAPBO beyond the use of prior information itself.

2 comments

We very much appreciate reviewer Kg2C's engagement with the methodological and experimental aspects of HapBO, and we thank them for the opportunity to clarify our contributions and further strengthen the paper.

To address the questions by reviewer Kg2C:

  1. Bounded shaping and comparison with π\piBO.

    (a) Motivation for bounded shaping. The additive term (1−α)(1-\alpha) ensures that the prior cannot eliminate candidates with positive acquisition. Thus, unlike π\piBO, α\alpha-π\piBO bounds the influence of the prior, and α\alpha controls the strength of that influence. Both methods have one scalar parameter (α\alpha versus β\beta). (Guay-Hottin et al., Knowledge-Based Systems, 2025)[DOI 10.1016/j.knosys.2025.113039] demonstrated substantially greater robustness of α\alpha-π\piBO to misleading priors across benchmarks. We have highlighted this evidence in Section 3.3 and expanded Fig. 8, where the advantage of bounded shaping is particularly apparent under adversarial priors (HPO adversarial) and medium quality (synthetic ρ\rho=0.5).

    (b) π\piBO as an ablation rather than a baseline. HapBO is deliberately modular and π\piBO and α\alpha-π\piBO are interchangeable prior-injection components, both compatible with our PSL elicitation and deterministic compilation pipeline. Neither, by itself, provides the language-to-prior machinery that constitutes HapBO's central contribution. Therefore, π\piBO is not an independent end-to-end LLM-guided BO baseline, but an alternative module within our framework. Fig. 8 provides the more informative controlled comparison. This isolates precisely the contribution of choosing α\alpha-π\piBO over π\piBO, without confounding it with differences in prior construction.

  2. Unexpected performance of random search. We thank the reviewer for catching this anomaly, which prompted us to identify and correct an implementation bug. We refactored the affected code and reran the experiments. The revised Figs. 2 and 5 now show the expected relative performance of random search and BO-based methods on HPO. We found that the correction did not alter our main findings, as HapBO's performance advantage remains, and the qualitative conclusions of the figures throughout the paper are unchanged. We are particularly grateful for this observation, which helped us improve the reliability of our work.

In response to this and all reviewers, we expanded our experimental evaluation from 15 to 21 figures, added 4 additional SOTA baselines, and substantially broadened our ablations. We sincerely thank reviewer Kg2C and all reviewers for their constructive feedback.

Reviewer_Kg2COpenReview ↗

Author responses are satisfactory and no further concerns remain. The revised manuscript is clearly improved; I recommend acceptance.

Thank you for the authors’ responses to my questions and for the corrections made to the related issues. I have no further concerns at this point. I have reviewed the newly uploaded PDF, and the current version appears much improved over the previous one; the constructive points raised by the other reviewers also seem to have been addressed accordingly. Best wishes.

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

Baysian Optimisation is an application agnostic (black box) optimisation algorithm that encodes prior knowledge in the distribution over functions that are used as a surrogate model the the expensive black box. Another way to incorporate prior knowledge is to weight regions of the search space and apply the weight to the acquisition function by multiplication.

This weighting function is non-negative and high for regions of interest or high chance to contain optima, and low elsewhere capturing an expert's knowledge. However if such knowledge is written as text or can be intuitively derived from a problem specification or is well known general knowledge, converting such information into a weighting function is a tricky process.

This paper proposes a Prior Specification Language, or a "PSL" (a play on domain specific languages or DSLs), one may prompt an LLM with handwritten expert knowledge and the PSL and prompt it to reply in the PSL using the given information, or its own knowledge, or both. The response can then be converted into a prior weight function directly and used downstream in a BO task.

This paper constructs such a pipeline and

Critical strengths

Many potential issues may arise from introducing an LLM, this paper nicely tries to answer all such issues

  • a good prior: speeds things up, as expected
  • a bad prior: slow things down but not too much and matches baselines
  • sensitivity to choice of LLM: the authors use a variety of models
  • data leakage: does the LLM training data include public benchmark answers
  • PSL design: the authors compare the full PSL with heavily truncated PSL and show improvement.
Critical weaknesses

I have no critical weaknesses that I can think of.

Questions for authors

Can the choice of kernel be included in the main body?

2 comments

We are ecstatic about reviewer PJtg encouraging assessment and for recognizing a central motivation of our work, which is that LLMs offer valuable knowledge for optimization, but for maximizing this potential we need to carefully harness their generative production (as opposed to providing them unrestricted control over the BO loop).

Regarding the question on kernel choice: we understood the suggestion as an invitation to examine whether HapBO's performance depends on the choice of GP kernel, complementing our existing comparison of acquisition functions. We have therefore expanded Fig. 15 to include three kernel variants (Matérn-1/2, Matérn-5/2, and RBF), showing that the benefits of HapBO shaping persist across kernel choices.

The review process has been a genuinely valuable opportunity to strengthen this paper. In response to all reviewers, we expanded our experimental evaluation from 15 to 21 figures, added 4 additional SOTA baselines, and substantially broadened our ablations. We sincerely thank reviewer PJtg and all reviewers for their constructive feedback, which has helped us make the paper more comprehensive, and clear.

Reviewer_PJtgOpenReview ↗

thank you for the reply

I leave my score as accept.

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

This paper introduces HapBO, a framework for incorporating natural-language knowledge into Bayesian optimization while limiting the influence of misleading priors. Before optimization, an LLM translates expert or LLM-generated beliefs into a structured Prior Specification Language (PSL), which is deterministically compiled into a reusable belief function and trust level. Building on α\alpha-π\pi BO, HapBO uses a bounded positive multiplier to reweight the acquisition function without further LLM calls. For noiseless EI with a fixed kernel, the authors show that the classical simple-regret rate is preserved up to a constant factor. Experiments on HPOBench, synthetic functions, cortical neurostimulation, and constrained solar design report improved sample efficiency under small evaluation budgets and examine robustness to misleading priors.

Critical strengths

The paper studies a practical gap between natural-language domain knowledge and the numerical beliefs required by prior-guided BO. This matters when evaluations are expensive and budgets are limited.

The language-to-prior design is appealing. PSL provides a structured way to express beliefs, while deterministic compilation produces a numerical prior that weights the acquisition function throughout optimization, without further LLM calls.

The theory offers an interesting characterization of prior influence. For noiseless EI with a fixed kernel, it relates the simple-regret bound to the prior’s value at an optimum, providing an interpretable connection between prior quality and optimization guarantees.

The evaluation is broad, covering HPOBench, synthetic problems, and two application domains. Misleading-prior tests and ablations across knowledge sources, representations, and injection sites provide useful evidence about the framework’s benefits and limitations.

Critical weaknesses

The need to remove acquisition normalization for the EI guarantee is not fully justified. The normalized EI rule appears to satisfy the key inequality in Theorem 2, suggesting that the same regret bound may still hold.

The synthetic fidelity sweep controls the correlation between the true objective and the model supplied to the LLM, but does not quantify the quality of the final prior used by BO. This makes the robustness results harder to interpret.

The continuous optimizer adds prior peaks to acquisition-search candidates, but their use in Figs. 4 and 7E is unclear. If used, a vanilla BO control with the same prior peaks is needed to isolate the effect of acquisition weighting.

PSL’s added value over existing language-guided BO methods remains unclear. The representation ablations do not clearly isolate its benefits under matched input knowledge and elicitation budgets.

Questions for authors
  1. Section 3.3 attributes the EI guarantee to removing min–max normalization. However, for nonconstant EI, the normalized rule maximizes (EIn(x)−an)M(x)(\mathrm{EI}_n(x)-a_n)M(x), where an:=min⁡xEIn(x)≥0a_n:=\min_x\mathrm{EI}_n(x)\ge 0. The same optimality argument yields EIn(x⋆)≤Λ(M)EIn(xn+1)\mathrm{EI}_n(x^\star)\le\Lambda(M)\mathrm{EI}_n(x_{n+1}), apparently sufficient for Theorem 2 under its remaining assumptions. Is removing normalization necessary for the regret guarantee, or does it instead serve to preserve positive acquisition support?

  2. In Fig. 4, ρ\rho controls the correlation between the true objective and the perturbed model supplied to the LLM. Could the authors quantify the quality of the final prior actually used by BO, for example through π(x⋆)\pi(x^\star) at a true optimum, and report the actual α\alpha for each function and ρ\rho level? This would connect the experiment to the theoretical guarantee and clarify whether low-ρ\rho inputs remain misleading after LLM translation and compilation.

  3. In the code, prior peaks are added to HapBO’s acquisition-search candidates at every iteration, even when prior warm-starting is disabled. Was this used for Figs. 4 and 7E? If so, could the authors include a control that gives vanilla BO the same prior peaks while keeping its acquisition unweighted? This would separate gains from acquisition weighting from those due to the additional search candidates.

  4. Prior work has already incorporated LLM knowledge into reusable BO priors. BORA uses LLM-generated priors for ColaBO and HypBO in its baseline experiments [1], while LLM-EI distils LLM preferences into a utility function for BO [2]. What distinguishes PSL from these approaches? Do the representation ablations in Fig. 7D–F establish its value under matched input knowledge and BO settings?

[1] Cissé, A., et al. (2025). Language-Based Bayesian Optimization Research Assistant (BORA). IJCAI 2025, 4967–4975. (https://arxiv.org/abs/2501.16224).

[2] Patel, R. A., et al. (2026). Distilling and exploiting quantitative insights from large language models for enhanced Bayesian optimization of chemical reactions. Digital Discovery, 5, 2042–2051. https://doi.org/10.1039/D6DD00052E.

2 comments

We thank Reviewer oFWG for the careful and constructive assessment, particularly for recognizing the practical importance, the theoretical characterization, and the breadth of our experiments. The review raises important points that have helped us strengthen both the theory and empirical aspects.

  1. Acquisition normalization and the EI guarantee. We think the reviewer is correct: removing min–max normalization is not necessary for the regret guarantee. We have therefore removed all claims that our normalization choice is necessary for the theoretical guarantees and added a corollary extending the bounded-influence and EI regret results to the original α\alpha-π\piBO formulation. We retain raw acquisition shaping for three practical reasons: (i) it preserves positive acquisition support, whereas min–max normalization can assign zero to candidates with positive raw acquisition; (ii) it requires no special handling when the acquisition range is degenerate; and (iii) it applies directly to any nonnegative acquisition without computing its extrema. Importantly, the two formulations show essentially indistinguishable empirical performance in our tested settings, as demonstrated by an additional comparison available in our anonymous repository (https://anonymous.4open.science/r/HapBO-27E5/ in the folder: Reviewer_oFWG_response/acq_norm_curves.pdf).

  2. Quality of the final priors in synthetic. We thank the reviewer for this suggestion. Importantly, the elicitor is unaware of the fidelity level: it receives only the perturbed model, without knowing how well it correlates with the true objective. Consequently, the compiled trust parameter α remains the same across fidelity conditions. We have now added Table 3, directly quantifying the final compiled prior's correlation with both the supplied model and the true objective, as well as its value at their respective optima. This confirms that low-ρ\rho inputs remain misleading after LLM translation and compilation. For further insight into prior quality, we also introduce Fig. 16 (Downstream performance and elicitor quality), examining fidelity, precision, agreement, and representational diversity across different LLMs and elicitation strategies.

  3. Prior peaks in continuous acquisition optimization. We thank the reviewer for raising this point and have added the requested control (Fig. 11). We also wish to clarify that including prior peaks is a natural part of numerically obtaining the argmax of an acquisition function. In continuous domains, acquisition maximization requires candidate generation and local argmax optimization; searching only around peaks of the unshaped acquisition risks missing maxima of the product of acquisition and prior. Including peaks from both components therefore improves the numerical search for the actual acquisition argmax, rather than introducing a separate prior-guidance mechanism. Nevertheless, to isolate any possible advantage from candidate generation, in Fig. 11 vanilla BO has access to exactly the same prior-peak candidates. This is a particularly stringent control, since it gives vanilla BO access to HapBO's elicited prior without using it for acquisition shaping. Fig. 11 shows no substantial benefit from these additional candidates alone, confirming that the observed gains arise primarily from acquisition weighting.

  4. Additional baselines. We clarify that our main-text ColaBO baseline already follows BORA's LLM-prior recipe from its supplementary material. We have now additionally evaluated HypBO with LLM-generated hypotheses and LLM-EI (Appendix L, Fig. 21). Importantly, LLM-EI has substantial scalability challenges because it requires pair-by-pair preference elicitation.

PSL's contribution goes beyond generating a prior. First, expressiveness: BORA's ColaBO prior uses a single isotropic Gaussian, while HypBO relies on axis-aligned hypothesis regions. PSL can represent these structures alongside multiple optima, trends, thresholds, interactions, penalized regions, and exact anchors. Second, separation of shape and trust: PSL independently represents where promising regions lie (π\pi) and how much to trust them (α\alpha). Third, portability: the same compiled prior can be used across multiple BO prior-injection mechanisms. Finally, the LLM produces an auditable specification, while a deterministic compiler (not the LLM) constructs the numerical prior.

In response to this review, we added one theory corollary, one quantitative table, three new figures, two additional baselines. We sincerely thank Reviewer oFWG for these insightful suggestions, which have strengthened the theoretical foundations and experimental breadth of our paper.

Thank you for the clarifications and additional experiments. The revised EI analysis, compiled-prior diagnostics, and matched search control address my main concerns. The new baselines help put PSL in context, though its advantage under the same elicitation budget remains less clear. I will maintain my rating.

Reviewer_56pGOpenReview ↗
Rating
4(+3)
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

The paper proposes to construct priors for use in Bayesian optimization with an LLM. The priors are constructed based on a proposed task card and a vocabulary card using a domain specific language and a compilation strategy. Experiments cover a wide range of HPO tasks, two real workd tasks, and synthetic functions.

Critical strengths

Constructing priors for use in Bayesian optimization with an LLM is a novel and interesting idea. I see this as potentially very relevant to practitioners.

The benchmarking covers 40 HPOBench tasks, two real-world tasks, and experiments on synthetic functions.

Critical weaknesses

The main contribution lies in the construction of the priors based on a task card and a vocabulary card. Methodologically it is a bit thin, but more crucial to my assessment, is that the presentation and experiments are misaligned with this being the main contribution.

  1. The quality of the constructed priors itself is not explicitly studied in the experiments. There is one ablation study that compares to different prior-building languages that goes in this direction, but this study shows that gaussian-only actually performs very strongly compared to PSL and having this in the same evaluation style as the main one would be more informative
  2. Some more ablation studies would help: How important is the incorporation of the LLM-stated confidence? How important are the different aspects of belief elicitation?
  3. The crucial step of "deterministic compilation of Pi" is only covered in an algorithm in the Appendix and one sentence in the main paper. As I understand it, the belief elicitation and compilation into the prior Pi are the main methodological contributions and the latter is barely covered in the main paper.
Questions for authors
  • LLM based baselines from the literature seem to perform worse than RS. Is this in line with the literature?
  • Why not compare against PiBO, Priorband or PriMO in the main evaluation (instead of only ColaBO)?
3 comments

We thank reviewer 56pG for recognizing the potential of LLM-based structured prior construction for BO as well as the breadth of our experimental evaluation, and we share the enthusiasm for the practical relevance of our approach for practitioners.

We kindly bring to the discussion table that HapBO's methodological contribution proposes a fundamentally different way to harness LLM knowledge in BO: eliciting a structured landscaped prior (compiled from a document in natural language – that LLM are excellent in building), and injecting it directly into the acquisition function throughout optimization. Unlike other methods, like the (highly-cited) LLAMBO (Liu et al., ICLR 2024), our approach preserves the standard very functional GP-BO machinery. Our contributions are also theoretical: we worked on PSL representation and completeness, and proved preservation of the classical EI regret rate. We also extend the π\piBO family theory to UCB, deriving a regret bound for noisy GP-UCB under prior-guided acquisition shaping, which is novel. We also provide extensive comparisons and ablations.

We seek to offer an answer to important points raised by reviewer 56pG.

  1. Quality of the constructed priors. We have substantially expanded this evaluation to match the request and revised Fig. 9 now presents the representation ablation in the same format as our main evaluation, including regret, rankings, and per-task breakdowns. Gaussian-only remains competitive on HPO adversarial priors, full PSL performs better in the other two suites. Moreover, we present a new Fig. 16 (Downstream performance and elicitor quality) which directly studies the constructed priors themselves, examining their fidelity, precision, agreement with the target landscape, and diversity and use of PSL components across different elicitors and elicitation strategies.
  2. α\alpha confidence and belief elicitation components. To match the request, we added a dedicated ablation (Fig. 10) that varies the trust weight α\alpha while holding the prior fixed, and we show that it modulates regret under misleading beliefs. Regarding elicitation components, PSL was made redundant by design, providing versatility across different problem structures while ensuring representational completeness. We expect the importance of individual components to be task-dependent: periodic landscapes may benefit from different primitives than sparse ones. We clarified this design rationale in the first contribution of the Introduction, and added to the limitations in the Discussion.
  3. Deterministic compilation. We agree that this contribution deserved more prominence and clarity. We have substantially expanded Section 3.2 of the main paper to explain the complete compilation pipeline, with explicit equations and a step-by-step explanation. We also added one figure in the appendix (Fig 7) allowing the reader to visualize the compilation process.

Q1. We agreed that there was an unexpected performance of random search compared to the SOTA baselines. We looked closely and found a mistake that is now corrected, and we re-ran all of the associated experiments. As expected, RS performs worse than all baselines in Fig 2. The main conclusion of the manuscript is preserved, but we are thankful for having helped us correct our experimental suite.

Q2. We expanded our evaluation to match this request. We have also added Fig. 20 (Appendix L), comparing HapBO against PriorBand and PriMO under both accurate and adversarial priors. Importantly, these methods have an additional advantage: they exploit cheaper, low-fidelity evaluations, whereas all other methods in the paper are full-fidelity only. We account for this conservatively using equivalent evaluation costs - but this means that these methods can perform more queries for the same total budget as the other methods, which is not like-for-like . Nevertheless, HapBO outperforms both overall. As discussed further in our response to reviewer Kg2C, π\piBO is an interchangeable prior-injection module within HapBO (see fig 8).

Overall, in response to this review, we have added four new figures, extensively revised one, and evaluated two additional SOTA baselines. We have also strengthened the presentation of our contributions, expanded the methodology, and clarified the discussion and limitations. We sincerely thank Reviewer 56pG for the constructive feedback, which has been very instrumental in strengthening our work.

Reviewer_56pGOpenReview ↗

Extensive Improvements and a Bug Fix

A. The authors fixed a crucial bug in their empirical evaluation in response to my review, related to RS outperforming most approaches. RS now performs as expected and the main conclusions stay unnaffected.

B. Every single one of my points has been thoroughly addressed, leading to substantial changes in the presentation of the manuscript, extended analysis, an adjustment to their stated limitations, and an added baseline.

C. I agree with the authors' points around their general methodological novelty and contributions. I also recognize the theoretical contributions as interesting.

I remain convinced of the strong practical relevance and have revised my scores. I recommend acceptance.

Additional points:

  • I strongly suggest the authors move Fig. 9a (and/or b) to the main paper/evaluation as the comparison to GP-only is a crucial result that scopes the contribution of the PSL.

  • Is it feasible for you to evaluate additional seeds for the Cortical ICMS experiment? The standard errors are so high, the performance of most methods is essentially indistinguishable.

We would like to thank reviewer 56pG for their assessment, their engagement and the suggestions to improve our manuscript.

  1. We accept the recommendation and the old Fig. 9 is now split between the new Fig. 6 (main text) and Fig. 10 (appendix).

  2. The SEM in the cortical experiment were driven by biological replicates, and not by seeds. Thus, we matched reviewer 56pG’s request by adding technical replicates. We did so by expanding the analysis to the whole rat cortical dataset (not just limited to leg flexor muscles as reported before). The result is substantially aligned to what was previously reported, but the SEM size now scales with n=40 technical replicates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.