acceptodds
Under review as a conference paper at ICLR 2027

Latent-Time Steering for Bayesian Optimal Experimental Design

Abstract

Bayesian optimal experimental design (BOED) selects experiments to maximally reduce uncertainty about unknown physical parameters, but the expected information gain (EIG) objective commonly used in BOED is often computationally expensive and difficult to estimate under complex priors. In this work, we propose a steering-cost surrogate for EIG that leverages pretrained diffusion-based generative priors over parameters and estimates EIG as a latent-time steering cost induced by likelihood guidance. Specifically, we combine the MINDE score–KL identity with a diffusion posterior sampling (DPS) plug-in approximation, yielding a tractable integral of squared likelihood guidance. The same pretrained prior supports posterior inference and design scoring without training a separate design policy. Additionally, we characterize the approximation error introduced by the proposed estimator. We validate our method on a range of design problems, including source localization, constant elasticity of substitution, and Darcy flow parameter estimation. For Darcy source-position design on a grid, our method reduces mean relative reconstruction error by and after the first and second measurements, respectively, compared with ALINE at the same resolution over 20 test cases.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian Optimal Experimental Design, Diffusion model, Inverse Problem.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5gwqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method for sequential Bayesian optimal experimental design using a fixed diffusion prior. The main idea is to adapt the score-KL identity to estimate information gain, using DPS guidance in place of the likelihood score required at each noise level. The method evaluates candidate experiments by integrating the expected squared guidance across noise levels, using approximate posterior samples and, in the Darcy flow example, perturbed saved diffusion states. A key feature is that the same pretrained diffusion prior is reused as new observations are collected, without retraining the model. The paper provides a conditional error analysis and evaluates the approach on source localization, constant elasticity of substitution, and Darcy flow.

Critical strengths
  • The paper presents a useful approach to experimental design when a pretrained prior and a differentiable likelihood are available. Reusing the same prior throughout the experiment sequence avoids the need to train a separate acquisition policy, which is particularly appealing for scientific inverse problems involving spatially distributed parameters, where both inference and experimental design can be computationally expensive. The information identity is mathematically sound under the stated assumptions, and the analysis appropriately accounts for the residual error introduced by using a finite diffusion horizon.

  • The experimental evaluation includes several useful tests of the proposed method. In the source-localization example, the authors compare candidate rankings and selected designs against an sPCE reference, while the benchmark experiments use sPCE to evaluate entire acquisition sequences. The Darcy flow experiments further demonstrate the method on parameter-field reconstruction. The paper also discusses important limitations, including guidance bias and sensitivity to the sampling procedure, and clarifies that increasing the number of samples reduces Monte Carlo error but does not necessarily eliminate these other sources of error.

Critical weaknesses
  • My main concern is whether the proposed steering-cost surrogate can reliably identify informative experiments. In Equations (7) and (11), the DPS approximation replaces the likelihood averaged over the diffusion smoothing distribution with the likelihood evaluated at a denoised point estimate. This simplification can introduce systematic errors and potentially change the ranking of candidate experiments. Even in a linear Gaussian setting, where the prior denoiser and posterior samples are exact, the estimated scores may not agree with the true expected information gain. This issue is particularly important in sequential experimental design. As observations accumulate, the likelihood score should reflect the current posterior, whereas the proposed method continues to use the unconditional prior denoiser. Although Theorem 3 provides a conditional error bound, the error introduced by the guidance approximation remains unquantified. Moreover, the analysis considers an idealized estimator rather than the full method used in the experiments. The sensitivity results in Table 4 also show that the estimated information gain depends on the guidance strength and the number of diffusion steps, raising further questions about the stability of the resulting acquisition decisions. A controlled comparison with exact information gain, along with a clearer discussion of how sequential conditioning affects the approximation, would strengthen the paper's central claims.

  • The results on high-dimensional Darcy flow are promising, but it is difficult to determine how much of the reported improvement comes specifically from the proposed acquisition criterion. The comparison with ALINE involves differences in both the inference procedure and the acquisition strategy, making it challenging to attribute the gains in reconstruction quality to the information-gain estimate itself. A more direct comparison, in which different acquisition strategies use the same diffusion prior and posterior reconstruction method, would help isolate the contribution of the proposed approach. The random-patch comparison in Appendix G.2 is also limited to a few representative examples. Evaluating the method across more independent parameter fields and acquisition sequences would provide stronger evidence of its robustness. In addition, the paper does not provide enough information about computational cost to establish the practical benefits of repeatedly evaluating candidate experiments through guided diffusion sampling. Finally, the benefit of integrating likelihood guidance across diffusion time, rather than using a simpler score at a single noise level, remains unclear without a controlled comparison. These gaps matter because the main methodological contribution is the proposed acquisition criterion, and the experiments should more clearly demonstrate its advantages over simpler alternatives.

Questions for authors
  • How well does the guidance in Equation (11), which relies on a fixed, unconditional denoiser, approximate the history-conditioned likelihood in Equation (19)? I would appreciate clarification on how the reverse-drift and initialization assumptions in Appendix E relate to the posterior sampler used in the experiments. Does the error analysis also apply to the hybrid estimator used in the Darcy experiments?

  • Could the Darcy experiments be repeated using the same posterior reconstruction procedure for designs selected by MINDE-DPS, ALINE, and random acquisition? This would help isolate the contribution of the proposed acquisition criterion to the reported improvements. I would also appreciate clarification on the "static (t=0)" baseline in Table 3. In addition, how does the integrated steering-cost criterion compare with a score evaluated at a single noise level when both are given comparable computational budgets?

  • How are prospective observations generated when evaluating candidate designs? Appendix G.2 uses a fixed true parameter field to produce Figure 5, but it is not entirely clear how this analysis relates to the acquisition procedure used in the experiments. Could the authors clarify this connection and explain why the first-stage information-gain maps differ across the examples in Figure 6? For the source-position design problem, it would also be helpful to specify which solution values are observed after the forcing is applied.

  • What is the motivation for running a separate DPS sampler targeting the updated posterior for each prospective observation? An alternative would be to draw a parameter sample from the current posterior, generate a prospective observation conditioned on that sample, and then apply forward noising to the same parameter sample. If the current posterior can be sampled exactly, this approach appears to provide the joint distribution needed to evaluate the guidance expectation without an additional posterior-sampling loop. Is this alternative compatible with the proposed estimator, and if so, how would its performance and computational cost compare with the current implementation? A comparison of runtime and the number of forward-model evaluations would help clarify the practical efficiency of the proposed approach.

Reviewer_kTSeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a method for Bayesian optimal experimental design using diffusion models to represent posteriors over high-dimensional parameter spaces. The proposed algorithm combines recent techniques from the literature on diffusion generative models to derive a tractable approach to compute expected information gain with a pre-trained diffusion prior without needing to retrain the model. The algorithm is demonstrated on two standard BOED benchmarks and a high-dimensional problem involving the Darcy flow PDE.

Critical strengths

The proposed use of MINDE (Franzese et al., 2024) seems novel and quite interesting in the sense that it does not require refitting the diffusion model or any other model to compute the expected information gain. To complement that, the paper provides theoretical results confirming the soundness of the approach. Later applying DPS to derive approximate posterior updates keeps the method practical without requiring additional retraining to estimate posteriors.

The experiment with Darcy flow is also interesting as a high-dimensional benchmark using a function-space diffusion prior. The high-dimensional setting demonstrated where the proposed method, MINDE-DPS, seems to thrive.

The paper is generally well written and not too hard to follow.

Critical weaknesses

My main concern is with the experimental results. The proposed method does not seem to be doing well in comparison with the presented baselines in the first experiments. The high-dimensional setting with Darcy flow seems to be novel, and only ALINE is compared against in that setting. So there's not much precedent to check how existing approaches might perform to establish a proper baseline.

In terms of organisation, Sec. 2 seems to be conflating problem formulation, background and parts of the method's derivation. I'd suggest separating things a bit better. The paragraphs in lines 169 to 177 describe a critical component of the proposed approach, which would be better place under the method section. In general, Sec. 2.2 should be focusing on introducing MINDE and DPS more neutrally. Their explicit connection to the proposed method should be explained with the methodology. Otherwise it's easy to skip those details for readers who might be familiar with that background.

There are a few other minor issues. Eq. 2 might better define the EIG with an expectation over yk\mathbf{y}_k, not a generic y\mathbf{y}. Some acronyms are introduced without prior definition (e.g., VE-SDE and EDM). The notation pt=p0K0,tp_t = p_0 K_{0,t} can be confused with pointwise multiplication, whereas it makes more sense to mean integration w.r.t. p0p_0 in this setting, which could be clarified with a brief definition just after that equation in Proposition 1. I'd also suggest renaming Sec. 3 with something less generic.

Questions for authors

Have the authors considered other problems involving high-dimensional parameter spaces where the algorithm might demonstrate a stronger impact?

Reviewer_Rr96OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors present a novel reformulation of diffusion guidance for sequential BOED without training directly with simulated parameter-observation relationships. Given a pre-trained diffusion-prior, they combine a MINDE identity as a surrogate for the EIG with training-free diffusion posterior steering (DPS), evaluated on a fixed number of designs sampled from a prior together with a differentiable explicit likelihood. They derive a bound for their surrogate's error rate. They evaluate their method on two low-dimensional standard design problem benchmarks and a benchmark with high posterior dimensionality. Despite not training explicitly on EIG, the performance compared to existing state-of-the-art is mixed; there are cases where the method approaches or beats state-of-the-art competitor methods that use direct training.

Critical strengths
  • Using diffusion guidance for BOED is interesting and novel.
  • The paper is compact but informative and the quality of the presentation is high.
  • The paper has clear practical relevance for when a diffusion-based prior is already available.
  • The experiments show competitive performance in comparison with ALINE and other competitor methods.
Critical weaknesses
  • The scaling behavior of the method to high-dimensional settings is unclear and not sufficiently studied in the paper:
    • The method requires access to a tractable explicit likelihood, which may not be available for real high-dimensional applications.
    • The method chooses the optimal design from a set number of designs sampled from a prior, which is subject to the curse of dimensionality and may thus require exponentially more samples as the dimension of the design grows.
  • The paper motivates its method through high computational complexity of competitor methods but never mentions a concrete computational complexity and does not show any inference timing benchmarks.
Questions for authors
  • In line 98, one of the central contributions of the paper is presented as "high-dimensional sequential design", which reads as though the method is applied to design-problems where the design itself is high-dimensional. However, the highest-dimensional design that is studied is ∈R6\in \mathbb{R}^6. Could the authors adjust their wording to reflect they are referring to dimensionality of the parameters θ\theta?
  • Could the authors add a brief discussion on requiring access to an explicit and differentiable likelihood?
  • Could the authors present an ablation that analyses the trade-off between performance in EIG and number of evaluated designs from the design-prior? I would be particularly interested in its relationship to design dimensionality.
  • Could the authors mention what the computational complexity of the DPS is, especially when combined with their MINDE surrogate? I would also highly appreciate the addition of a comparison of inference timings for this method and the competitor methods shown in Table 1.
  • Could the authors add additional related work [1-3] in the introduction and related work sections? If the results are compatible, Table 1 could also be expanded with the numbers from these publications.

Nitpicks:

  • Please highlight the best values in Table 1 in bold.
  • The number of source locations and experimental steps for the source location and CES experiments should be shown in Table 1 or its caption.
  • Figure 1 contains text that is too small to read without significantly zooming into the PDF.
  • Figures 2, 3, 4, 6, and 8 have axis ticks and labels that are too small to read.
  • The source location labels in Figure 2 are low-contrast and a little small.
  • Formatting of tables is inconsistent: Table 2 is squeezed to the side and into the paragraph, making it hard to distinguish from the prose and the paragraph harder to read. Please format table 2 the same as table 1.
  • Table 1 does not mention where its uncertainty measures come from.

[1]: Bracher et al. (2026) - JADAI: Jointly Amortizing Adaptive Design and Bayesian Inference

[2]: Blau et al. (2022) - Optimizing Sequential Experimental Design with Deep Reinforcement Learning

[3]: Blau et al. (2023) - Statistically Efficient Bayesian Sequential Experiment Design via Reinforcement Learning with Cross-Entropy Estimators

Comments on the paper

3 comments

Thank you for the detailed responses. The common-DPS reconstruction comparisons and the additional Navier–Stokes evaluation have improved the empirical evidence. The explanations of the prior-only calibration baseline, the scope of Theorem 3, and the joint-sampling construction also address several of my questions. My main remaining concern is whether the implemented acquisition criterion reliably ranks candidate designs. The response acknowledges possible ranking errors, while the sequential calibration diagnostic uses a history-dependent Gaussian denoiser. Could you provide the requested controlled sequential linear-Gaussian comparison with exact EIG, including ranking agreement and the exact-EIG loss of the selected designs for the scoring implementation used in the experiments? Please reconcile the standard error versus standard deviation descriptions of the Navier–Stokes results and, if available, report uncertainty for paired performance differences. I acknowledge the improvements in the rebuttal, but currently retain my rating because the central acquisition validation concern remains open.

Thank you for the additional experiments and statistical clarification. The sequential linear Gaussian controls using fixed prior denoisers address my concern about the earlier diagnostic. Although ranking agreement is imperfect in the Source control, the reported exact-EIG loss is small. The paired reconstruction results give clearer support for the improvement after four observations. Taken together with the earlier common-reconstructor comparisons, these results address my main concerns. I increase my rating from 2 to 3. Please include the control specifications, observation histories, scoring settings, and corrected uncertainty labels in the manuscript. The conclusions should remain limited to the evaluated settings, with the conditional scope of the theoretical analysis and its limitations for the hybrid implementation mentioned clearly.

Thank you very much for your feedback and for reconsidering your assessment. We will include the requested control specifications, observation histories, scoring settings, and corrected uncertainty labels in the manuscript. We will also keep our conclusions limited to the evaluated settings and clearly state the conditional scope of the theoretical analysis and its limitations for the hybrid implementation.

It appears that the updated rating may not yet be reflected in the system. We would appreciate it if you could update it when convenient. Thank you again for your time and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.