acceptodds
Under review as a conference paper at ICLR 2027

Gaussian Processes for Query-Efficient Boltzmann Sampling

Abstract

Neural samplers have achieved remarkable success in recent years, as efficient learning-based substitutes for traditional MCMC methods that significantly reduce the sampling cost after learning to sample. However, both neural samplers and MCMC methods remain sample-inefficient and therefore unsuitable for sampling problems where energy or likelihood evaluations are expensive. We develop a model-based sampling approach using a Gaussian process (GP) framework that trades computation for fewer evaluations by learning an energy surrogate and using its uncertainty to guide exploration. We distinguish two objectives: (1) producing accurate samples throughout exploration and (2) learning an accurate distribution for subsequent sampling. For (1), we analyze posterior-mean, optimistic and posterior-sampling policies using a KL-based notion of sampling regret and establish temperature-dependent bounds. For (2), we derive a Bayes-optimal one-step acquisition that selects evaluations to maximize the expected reduction in reverse KL divergence, together with practical approximations. Our experiments on low-dimensional sampling problems show the significant advantage of our model-based sampling method over neural samplers, MCMC methods, and existing GP-based sampling techniques.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Gaussian Processes, Boltzmann Sampling

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_qBAwOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a GP-based framework for query-efficient Boltzmann sampling, focusing on settings where energy evaluations are expensive. The authors consider two objectives: obtaining accurate samples during exploration and learning an accurate sampling distribution for subsequent use. For the first objective, they analyze three GP-based sampling policies and establish temperature-dependent sampling regret bounds. For the second, they derive a one-step acquisition function that minimizes the expected reverse KL divergence after an additional observation, together with several practical approximations.

The proposed methods are evaluated on synthetic distributions and a real-data inference task. The experiments demonstrate promising query efficiency compared with MCMC, neural samplers, and existing GP-based methods, particularly in low-dimensional settings.

Critical strengths

1. Principled methodology with theoretical analysis. The paper derives a one-step acquisition function directly from the expected reduction in reverse KL divergence, rather than relying on heuristic exploration criteria. It also establishes finite-budget sampling regret bounds for several GP-based policies, providing theoretical insights into how exploration depends on temperature. These results make the proposed framework more than a straightforward application of GP-based active learning.

2. Promising empirical performance: It demonstrates improved query efficiency on several tasks, both synthetic and realistic. In particular, the alanine experiment shows that OneStep can recover a minor mode missed by other GP sampling policies under the same query budget. This provides concrete evidence that the proposed method can improve distribution learning when energy evaluations are expensive.

Critical weaknesses
  1. The proposed framework is closely related to existing GP-based active posterior estimation methods, particularly BAPE [1] and GPry [2]. Although both are included in the experiments, the paper does not directly compare against BAPE's expected-divergence acquisition, and some GP baselines use different prior mean functions. As a result, it is difficult to determine whether the reported improvements arise from the proposed acquisition function or other implementation choices. This weakens the evidence for the paper's incremental novelty.

  2. A central motivation of the paper is to reduce sampling costs when energy evaluations are expensive. However, the experimental comparisons primarily focus on the number of energy evaluations, while the proposed acquisition functions introduce additional computation. Moreover, the reported sampling costs are not consistently adjusted for sample correlations across methods. A more comprehensive comparison using end-to-end runtime and estimation error at matched computational budgets would better substantiate the practical advantages.

[1] Kandasamy et al. (2015). Bayesian Active Learning for Posterior Estimation. IJCAI.

[2] El Gammal et al. (2023). Fast and Robust Bayesian Inference Using Gaussian Processes with GPry. Journal of Cosmology and Astroparticle Physics.

Questions for authors
  1. The paper acknowledges BAPE's expected-divergence acquisition as a closely related approach. Could the authors clarify the main methodological differences between this criterion and OneStep, beyond the choice of divergence direction and the tractable approximations? Which aspects of OneStep do the authors consider the most substantial improvements over prior work?

  2. The proposed method trades additional computation for fewer energy evaluations. Could the authors clarify how the computational overhead of OneStep should be considered when interpreting the reported efficiency improvements, particularly in comparison with existing GP-based approaches? Under what conditions do the authors expect this trade-off to be most beneficial?

2 comments

We thank the reviewer for their feedback. Below we address their weaknesses and questions.

W1 and Q1: Limited evidence of improvement [...] the paper does not directly compare against BAPE's expected-divergence acquisition, and some GP baselines use different prior mean functions. [...]

For the BAPE-NED, this method is added as a baseline; please see G2 for discussion of its limitations. We also added ablations on hyperparameters such as the prior mean, and we see that our method is more robust than GPry to these choices; see Table 19. Additionally, GPry's acquisition lacks a principled derivation, despite its good performance in practice.

W2 and Q2: Insufficient evidence for the claimed computational efficiency [...] while the proposed acquisition functions introduce additional computation. [...] A more comprehensive comparison using end-to-end runtime and estimation error at matched computational budgets would better substantiate the practical advantages.

We added figures comparing acquisition runtimes with all other settings fixed for Planck-lite and H202; see Figure 2 and 10. The time allocated to computation includes fitting the GP and calculating the acquisition function. This cost is dominated by the first step of fitting the GP, so different GP-based methods do not differ significantly. However, we would like to clarify that we separate three axes of (1) cost—query, (2) computation/learning, and (3) sampling. As shown in Figure 1, in the appropriate regime, GP methods can achieve orders of magnitude lower oracle cost while matching neural samplers' computation and sampling costs.

Thank you for the response. The additional comparison and analysis help address my main concerns and it remains unclear how query requirements and computational costs grow with dimension. However, I think the results within the low-dimensional, expensive-evaluation setting still make a worthwhile contribution. Overall, I will maintain my original rating.

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

This paper proposes a Gaussian-process-based approach to query-efficient Boltzmann sampling. It learns an energy surrogate and uses its uncertainty to select informative evaluations, aiming to reduce the number of expensive energy queries needed to learn a sampling distribution.

First of all, the reviewer admits that he/she is familiar with the sampling problem while not an expert in the Gaussian-process literature, therefore might not be able to provide correct judgement on the GP side; and the review is more from a sampling perspective. Regardless of the methodology itself, the most significant issue of the paper is the scalability. Therefore, I will give my assessment with confidence lowered by my unfamiliarity of GP.

Critical strengths

The experiments provide an initial proof of concept for using Gaussian-process energy surrogates and uncertainty-guided query selection in low-dimensional sampling problems.

Critical weaknesses
  1. Limited evidence for dimensional scalability. The evaluation comprises two-dimensional synthetic targets, alanine dipeptide represented by two torsional coordinates, a six-parameter cosmological posterior, and DW-4 represented by eight centered Cartesian coordinates. These experiments demonstrate the approach in small state spaces, but do not establish how sampling accuracy, query requirements, and computational cost change as dimensionality increases.

    The authors explicitly acknowledge this limitation and focus on low-dimensional problems with expensive energy evaluations. Nevertheless, a controlled dimensional-scaling study, even within a moderately low-dimensional regime, would help establish the method's practical range. Exact GP fitting has cubic computational cost in the number of observations, while the observations needed to learn an accurate energy surrogate may themselves increase substantially with dimension.

  2. Unresolved sensitivity to the kernel and its hyperparameters. For distance-based kernels, predictive performance depends on input scaling and length-scale selection. My concern is whether the surrogate and acquisition strategy remain accurate and query-efficient under these choices as dimensionality increases. The current evaluation does not sufficiently resolve this concern.

  3. Limited demonstrated scope relative to existing neural samplers. Neural sampling methods such as Adjoint Sampling [1] have been evaluated on substantially higher-dimensional targets. These methods address a different computational regime, so their scalability does not directly invalidate the proposed approach. However, this contrast highlights the limited breadth of the present empirical validation.


[1] Aaron J. Havens et al. “Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching.” Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:22204–22237, 2025.

Questions for authors
  1. Within the intended low-dimensional, expensive-oracle regime, how do sampling accuracy, required oracle evaluations, and total computation change as dimensionality increases?

  2. How sensitive are the reported results to the kernel family, length-scale settings, and input normalization? What evidence supports the robustness of these choices?

3 comments

We thank the reviewer for their feedback. Below we address their weaknesses and questions.

Q1: Within the intended low-dimensional, expensive-oracle regime, how do sampling accuracy, required oracle evaluations, and total computation change as dimensionality increases?

We added an eleven-dimensional experiment, posterior estimation for a black-hole binary observed by LISA. The computational cost of the GP is linear in the dimension of the data.

Q2: How sensitive are the reported results to the kernel family, length-scale settings, and input normalization? What evidence supports the robustness of these choices?

See general answer G3

CW1 and CW3: Limited evidence for dimensional scalability. Limited demonstrated scope relative to existing neural samplers.

It appears the reviewer agrees that we address a different query-efficiency regime than neural samplers, as shown in our results (e.g., Figure 1). We have included the missing Adjoint Sampling baseline and new experiments, demonstrating the viability of our methods to up to 11 dimensions. We have also further emphasized limitations in the introduction and the limitations section. In the regime we target, where only 10² to 10⁴ energy evaluations are affordable, the neural samplers, including Adjoint Sampling, are less accurate than the best GP-based methods (Figure 1 and Table 15).

I appreciate the authors' efforts for answering my questions. I appreciate the new 11-dimensional experiments, yet it is still in a very low dimension setting, where a lot of neural samplers that can't scale could achieve nice results. Therefore, I prefer tol keep my original rating.

Thank you for your feedback. We understand your concerns about scaling to more dimensions.

In the regime we consider, there are already many interesting scientific tasks that would benefit from our approach. Examples can be found in cosmology, force field calibration, gravitational waves posterior estimation, and many others. Those are all important use cases where the oracle can take from minutes to hours of computation, which makes neural samplers impractical. We have included a table of standard neural samplers in the literature detailing why they are impractical for the specific black box expensive oracle setup (Table 20).

We believe that the scaling issue is not limited to GP methods. The limitation stems from the expensive-oracle setup, which makes it difficult for any method to learn in high dimension from a few evaluations. To support this, we included a neural sampler baseline to the 11 dimensional LISA experiment (Figure 4(b)). We see that this neural sampler is not as effective as our GP based methods.

Reviewer_zwfbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper use the Gaussian Process to construct approximated energy and to score the candidates points/samples. It does not need the gradient information of the energy E. However, the method may introduce big bias especially missing some local minimum of the energy. It also lacks scalability in higher dimensions.

Critical strengths
  1. the method does not need the gradients of E (though methods relying on \nabla E without knowing E is preferred in applications)

  2. OneStep directly targets expected reverse-KL reduction, with practical approximations.

  3. The method is relatively easy to implement

Critical weaknesses
  1. single-step optimality of OneStep does not establish how the final distribution error behaves
  2. Why the Gaussian process assumption could give good predictions needs some understanding, especially for energy landscape with several local minima
  3. The comparison includes BAPE-EV but omits BAPE-NED, leaving OneStep’s benefit over a closely related expected-divergence acquisition unclear.
  4. The sampling is often required in high dimensions. This approach may have limits in this aspect
Questions for authors
  1. Does the one-step analysis imply about final distribution accuracy?

  2. How do the samples correlated? How will affect the total sampling error? This method computes the approximated energy and it may casue big bias. Any estimate on the bias?

  3. How does OneStep compare with BAPE-NED under the same GP implementation and energy-evaluation budget?

1 comment

We thank the reviewer for their feedback. Below we address their weaknesses and questions.

Q1: Does the one-step analysis imply about final distribution accuracy?

The OneStep acquisition is not necessarily Bayes-optimal over multiple steps. This limitation is common to all existing acquisition strategies, which are myopic but nevertheless widely used in Bayesian optimization. Our Gibbs policies do come with cumulative and final policy regret.

Q2: How do the samples correlated? How will affect the total sampling error? This method computes the approximated energy and it may casue big bias. Any estimate on the bias?

Given that the the correlation of our MC samples can be made arbitrarily small. The error in the approximated energy is exactly what is being measured by our evaluation, we show our method achieves lower or equal error to other GP based methods.

Q3: How does OneStep compare with BAPE-NED under the same GP implementation and energy-evaluation budget?

See the general answer G2.

Reviewer_uo2GOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how, and proposes a novel method, to sample from expensive-to-compute energy or likelihood functions more efficiently. The paper combines findings from two literatures that study this task, these being model-based RL and Gaussian Processes in Bayesian Optimization. The authors study two settings, one which evaluates three different sampling policies realized by GPs and the other a OneStep acquisition function that the authors define also using a GP. The proposed methods show improved sample efficiency in learning 2D toy example densities as well as in slightly higher dimensional and more realistic physics based benchmarks. A key limitation highlighted are the inherent scalability issues for GPs, which could be addressed via existing GP scaling methods.

Critical strengths
  • S1: The paper studies a well-motivated, interesting and novel method in the intersection of model-based RL sampling and Bayesian Optimization.
  • S2: The paper demonstrates convincing empirical evidence for the effectiveness of the method systematically, showing improved performance over baselines first in toy-examples and then more realistic tasks.
Critical weaknesses
  • W1: The presentation of the paper requires further work before acceptance. Mainly, the paper is currently missing a conclusion section and the main contributions of the paper could be made more clear in the contributions section. Furthermore, the novelty with respect to the prior related literature should be made more clear. Adding a reproducibility statement is also highly recommended. These weaknesses are critical in the sense that before getting fixed, the paper cannot be accepted, but not critical in the sense that they are fixable in the rebuttal period.
  • W2: The experimentation is done in low-dimensional benchmark settings where the energy is cheap to compute (which the authors are open about in the discussion and limitations sections), but given that the real benefits of the methods would show in an expensive to compute scenario, the paper would be substantially strengthened by adding at least one more larger scale experiment that demonstrates the methods performance in this setting. This is a critical weakness in the sense that it would be needed to demonstrate that the method works in the setting that the paper is concerned about the most (expensive energy evaluations).
Questions for authors
  • Q1 What are some applications with an expensive to compute energy function and where the scale of the problem allows for GPs to be utilized?
  • Q2 What are the main conclusions of the work, and what are the next promising steps to take for testing and advancing the method further? Could the authors please provide a complete conclusion section (and add this to their main text)?
  • Q3 Could the authors elaborate further on how the proposed OneStep acquisition function differs from the BAPE one-step criterion that is cited as a direct precedent? What is their relationship, and how does OneStep aim to improve on this?
  • Q4 What is the reason for not comparing the proposed method to the BAPE method, are the methods not comparable in some way?
  • Q5 Will the authors release the code for this method?
2 comments

We thank the reviewer for their feedback. Below we address their weaknesses and questions.

W1: The presentation of the paper requires further work ...

We have extensively revised the presentation, particularly that of the method section, for clarity and readability. We have also added a conclusion and reproducibility statement per the reviewer's suggestion.

W2 and Q1: The paper would be substantially strengthened by adding at least one more larger-scale experiment... What are some applications...

We added two longer experiments in addition to the Planck posterior: battery calibration and H2O2 conformation, each of which takes 10 to 15 seconds per oracle call. Those results demonstrate the advantages of query efficient methods and bring the total number of benchmarks to seven. We would also like to point out that the actual oracle cost is irrelevant in demonstrating an algorithm's superior performance, because one could replace cheap simulations for this kind of demonstration, noting that in real-world settings such simulations will be inferior in quality.

Q2: What are the main conclusions of the work...

We have added these under the Concluding Remarks section.

Q3 and Q4: Could the authors elaborate further on how the proposed OneStep acquisition function differs from the BAPE one-step ... comparing the proposed method to the BAPE method.

See the general answer G2.

Q5: Will the authors release the code for this method?

We added a reproducibility statement. If the reviewer would like access to any part of the code during the discussion period, please let us know.

Thank you for addressing my questions and concerns. The paper presentation has improved considerably, and the addition of the new experiments strengthens the paper also. I will raise my score from 2 to 3.

Comments on the paper

1 comment

General Questions:

Section, table, figure and line numbers below refer to the revised PDF. Several reviewers raised the same points; we answer them once here (G1--G3) and refer to these answers in the individual responses.

We thank all the reviewers for their valuable feedback, which has helped improve the paper. Below, we address general issues and address individual questions and concerns in comments under individual review.

G1. GP-based sampling is not new; what is the new contribution?

First, we identify the regime in which each family of classical, neural, and GP-based samplers is effective, considering the costs of sampling, learning, ann querying the energy as separate axes. Second, for the first time, we develop the GP sampling theory in analogy with the GP-based bandits and Bayesian optimization literature, with corresponding bounds for simple and cumulative regrets. Third, we show that our methods, derived from first principles, generally outperform or compare favourably against various baselines across a range of low-dimensional problems.

G2. Missing Baseline of BAPE-NED

We have added this comparison whenever BAPE-NED was found viable. The issue with BAPE-NED, according to that paper's authors, is that it is "computationally intractable" since it has to use a tensor grid to estimate the forward KL, and therefore it is only applied (also in the original paper) to up to three-dimensional problems. We briefly discuss this in Section 3 (and Appendix B). The authors proposed BAPE-EV as a simplification, but its approximation is not well justified, and we also find it not very competitive in our experiments.

G3. Sensitivity to choice of hyperparameters and matched comparison with other acquisitions

(qBAw W1; ApWf Q2)

Except for Gibbs policies, which have no acquisition, we compare all GP-based methods under the same GP scaffolding, varying only the acquisition. We have added ablations on the choice of kernel and prior mean (Table 5 and 19). We see that our method's performance is robust to several ablated choices. The length-scale is optimized as part of the inner loop using marginal likelihood for all GP methods. See Appendix C.4 and I.5

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.