acceptodds
Under review as a conference paper at ICLR 2027

CaliBO: Calibrating Incremental LLM Advice for Bayesian Optimization

Abstract

Large language models (LLMs) have shown promise for augmenting Bayesian optimization (BO), particularly when objective evaluations are scarce. When both an LLM and a Gaussian process (GP) provide objective predictions, a central question is how their means should influence acquisition. We introduce CaliBO, which calibrates LLM guidance by estimating its incremental predictive value beyond the current GP. Using predictions recorded before each evaluation, CaliBO learns whether the disagreement between the LLM and GP means predicts the GP’s subsequent residual error. A signed, shrunk coefficient produces a bounded mean correction, while GP posterior uncertainty remains unchanged for exploration. This allows CaliBO to follow, downweight, or reverse LLM advice according to its estimated incremental value, without requiring the LLM to outperform the GP as a standalone predictor. We bound acquisition perturbation and, under a zero-usefulness null and stated GP and noise-estimation assumptions, establish a sublinear advice-dependent term in the regret bound of an idealized GP-LCB variant. Across 49 hyperparameter-optimization tasks, CaliBO is competitive with ten baselines and significantly outperforms several of them. Ablation studies support the effectiveness of incremental calibration, while controlled experiments with synthetic advisors demonstrate benefits from signed correction under systematically inverted advice and reveal a failure case when uninformative advice still tracks the GP's error.

Primary area:
optimization
Keywords:
Large Language Model, Optimization Problem

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7RV7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The proposed method uses LLM guidance to improve bayesian optimization. The method learns how much to trust the LLM based on its ability to predict the GP's errors, and then uses this information to adjust the GP predictions. The authors provide theoretical analysis and experiments on comprehensive benchmarks to show the effectiveness of their method.

Critical strengths
  • The intuition and motivation is clearly described.
  • The framework is simple and lightweight.
  • The experimental result is comprehensive.
Critical weaknesses
  • The proposed method's performance improvement over GP-Mean-Only control is mainly on HPObench. And GP-Mean-Only has similar performance in the 180-dimensional experiment, which weakens the evidence for the effectiveness of incremental calibration.

  • The proposed method measures how well LLM advice predicts GP errors, but this doesn't show that the advice can always help optimization. The Pool Mean advisor experiments show that unhelpful advice can still receive high weight and hurt performance.

  • The presentation of Figure 1 can be improved, as it's overly crowded and hard to follow.

Questions for authors
  • Could the authors provide additional experiments to further demonstrate the independent contribution of incremental calibration beyond LLM-based candidate generation?
  • Could the authors provide a more detailed analysis of the trade-off between computational cost and optimization performance?
Reviewer_n8jwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces CaliBO for incorporating LLM predictions into Bayesian optimization. Instead of directly trusting the LLM, it estimates the incremental value of LLM advice relative to a GP. The historical GP prediction residuals on the disagreement between LLM and GP predictions are used to obtain a shrinking coefficient that determines the direction and strength of a bounded correction to the acquisition mean. Theoretical results were provided to bound the acquisition values and to establish a regret bound. The empirical validation was conducted on 25 Bayesmark tasks and 24 HPOBench tasks against 10 baselines, including LLM-augmented ones and classical ones. The authors also conduct ablations and synthetic-advisor experiments to investigate when calibration is beneficial and when it fails.

Critical strengths
  1. The research question is relevant and the design is intuitive. The paper addresses an important challenge in LLM-augmented BO regarding the accuracy of LLM predictions. Its choice of calibrating the disagreement rather than relying on standalone LLM accuracy is a meaningful design choice.

  2. The signed coefficient introduces controlled influence and interpretability in following, ignoring, or reversing LLM advice. Clipping the correction bounds its effect on the acquisition value, and the untouched GP uncertainty avoids relying on the LLM for uncertainty estimation.

  3. The evaluation is broad, and the limitations are transparent. The comparison with 10 baselines across 49 HPO tasks followed by statistical tests are sufficient to draw conclusions about the performance of CaliBO.

Critical weaknesses
  1. CaliBO uses one scalar coefficient to control the LLM's influence across the candidate pool, which assumes a stable relationship between LLM-GP disagreement. In practice, the LLM may be useful in some regions and misleading in others. In addition, the distribution of observed residuals is determined by the BO sampling policy and may not necessarily be adequate enough to represent the candidate regions where advice will subsequently be applied. For example, it is possible to design a method that locally adjusts LLM prediction in the form of a prior mean using observed discrepancies, which is adaptive across the whole search space, giving the advantage of adaptive shrinkage instead of a fixed coefficient.

  2. The appendix addresses the compliance of output format, but numerical accuracy and consistency were unaddressed. The paper did not adequately evaluate whether LLM predictions are stable across repeated queries or reasonable prompt perturbations.

  3. A thorough sensitivity analysis of design choices is not sufficiently provided, and adding multiple hyperparameters also weakens the applicability of the proposed method. CaliBO introduces several constants, selected on a separate design task set. The paper does not demonstrate robustness to these choices. As such, it is hard to tell if the reported results generalize to new problems.

  4. The evaluation scope is considerably limited. Firstly, doing experiments with only 5 seeds can lead to highly unstable standard deviation. Secondly, the empirical results rely on only two proprietary OpenAI LLMs. Therefore, it limits the evidence that CaliBO can work well with other LLM families, particularly open-weight models.

  5. Section 2 is not sufficiently comprehensive. Requeima et al. (2024) was a pioneering work in using LLMs as the surrogate model, but it was not mentioned. Many other works listed below about using LLMs in BO are also not mentioned, both in using LLMs as a numerical predictor or other aspects.

Requeima, J., Bronskill, J., Choi, D., Turner, R. E., & Duvenaud, D. (2024). Llm processes: Numerical predictive distributions conditioned on natural language. Advances in Neural Information Processing Systems, 37, 109609-109671.

Yuan, X., Chen, Z., Zhang, J., Xiong, H., Ye, N., Li, Y., & Gu, Q. (2026). Unleashing LLMs in Bayesian Optimization: Preference-Guided Framework for Scientific Discovery. The Fourteenth International Conference on Learning Representations. Retrieved from https://openreview.net/forum?id=LktUOZayG9

Aglietti, V., Ktena, I., Schrouff, J., Sgouritsa, E., Ruiz, F., Malek, A., … Chiappa, S. (2025). FunBO: Discovering Acquisition Functions for Bayesian Optimization with FunSearch. Forty-Second International Conference on Machine Learning. Retrieved from https://openreview.net/forum?id=XjbJR9374o

Ngo, G., Trong, D. P., Nguyen, D., Gupta, S., & Venkatesh, S. (2026). Adaptive Acquisition Selection for Bayesian Optimization with Large Language Models. The Fourteenth International Conference on Learning Representations. Retrieved from https://openreview.net/forum?id=EPKmSgXvRe

Zeng, Y., Maus, N., Jones, H. T., Tao, J., Wan, F., Torres, M. D. T., … Gardner, J. R. (2026). Scaling Multi-Task Bayesian Optimization with Large Language Models. The Fourteenth International Conference on Learning Representations. Retrieved from https://openreview.net/forum?id=ekmUkRYnkN

Questions for authors
  1. How sensitive are the results to the hyperparameters used? What tasks were used to select the fixed constants?

  2. How much does the estimated coefficient and final optimization performance vary across repeated LLM queries and modest prompt changes?

  3. The derived posterior mean was used directly in a lower confidence bound as the acquisition function as shown in Eq. (7). However, this choice is not well motivated/explained. It seems possible that other acquisition functions can be used in place of the lower confidence bound, because the proposed method only addressed the posterior mean. Did the author ever consider other acquisition functions?

Reviewer_cTJuOpenReview ↗
Rating
1
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose CaliBO, which use disagreement between LLM and BO to predict GP's error. a coefficient is fit to measured how well pask LLM-GP disagreement can predict the GPe error, and the GP prediction is corrected using the coefficient.

Critical strengths
  1. How to combine LLM and BO is an interesting problem.

  2. The paper seems to be well-prepared except for the fact that many sentence read like AI-generated.

Critical weaknesses
  1. Most of the sentence read like AI-generated. e.g. "The disagreement is a candidate correction,not evidence that the LLM is more reliable." Please make sure the readability.

  2. The main figure also shows some AI pattern (e.g. uneven letter spacing) Please fix.

  3. The theory part in the main paper makes it a bit hard to understand. Those theory does not seems to assess the proposed method directly. I would move most of them to the appendix.

  4. The experiments seem to be too simple

Questions for authors

What if both of the LLM and GP are mismatched to the true model?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.