acceptodds
Under review as a conference paper at ICLR 2027

CooL-BO: Constraint-aligned Latent Bayesian Optimization for High-dimensional Problems

Abstract

Recent methods employing variational autoencoders (VAEs) to map discrete, high-dimensional input spaces to continuous latent representations, have achieved remarkable success in high-dimensional Bayesian optimization (BO) in unconstrained settings. However, real-world problems (e.g., molecule design, robotic control tasks) are often subject to practical constraints (e.g., drug likeliness of proposed molecules, safety constraints) that are not explicitly captured by standard VAE training. Consequently, the trained VAE may map feasible and infeasible inputs to overlapping latent regions. As a result, nearby latent points can decode to candidates with different feasibility, producing an irregular feasible region that complicates acquisition-function optimization and reduces sampling efficiency. In this paper, we propose CooL-BO, which explicitly shapes the learned latent manifold with respect to input feasibility. By introducing a discriminator and a dynamic mechanism for learning priors, we enforce separability in the latent domain during pre-training of the VAE, thereby facilitating the selection of promising feasible candidates for latent space BO. Consequently, CooL-BO achieves a higher BO objective than state-of-the-art latent space BO methods, while reducing computation time by up to times across various high-dimensional benchmark tasks. Our results highlight the importance of leveraging both feasible and infeasible inputs in representation learning for latent-space BO, motivating the need for learning representations tailored for latent BO tasks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Latent space BO, VAE

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

This paper addresses the problem of Bayesian Optimization (BO) in high-dimensional latent spaces where the objective function is computationally inexpensive, but pre-defined constraints exist. Standard Variational Autoencoders (VAEs) are trained without regard for whether inputs satisfy these constraints, leading to an intermingling of valid and invalid inputs within the latent space. The authors point out that this mixing complicates the optimization of the acquisition function and degrades sampling efficiency. The proposed method, CooL-BO, incorporates information regarding constraint satisfaction directly into the geometric structure of the latent space during the VAE pre-training phase. CooL-BO consists of the following three components:

  1. A linear discriminant term, Ω(T∣Z)\Omega(T|Z), is added to the Evidence Lower Bound (ELBO) of the joint distribution p(X,T)p(X, T).
  2. Class-specific Gaussian priors, N(μt,I)N(\mu_t, I), are employed for inputs that satisfy the constraints and those that do not. In each iteration, the center of each prior distribution is recalculated using the mean of the posterior distributions of the data points belonging to that respective class.
  3. The discriminator and the VAE are updated in an alternating fashion.
Critical strengths

1. A simple design that directly reduces the cost of constrained acquisition optimization. CooL-BO shapes the latent space so that feasibility becomes a single linear inequality θ⊤z≥Δ\theta^\top z \ge \Delta. This integrates directly with TuRBO’s box trust region and removes the need for separate constraint surrogates such as those in HDSafeBO. Since it adds almost no learnable parameters, it can be plugged into existing latent BO pipelines with minimal overhead.

2. The learned representations are evaluated directly, beyond final BO performance. The paper reports latent separability (Fisher LDA ratio, linear probe) and decoded feasibility of posterior samples (Table 5), compared against conditional VAE, VampPrior, and contrastive VAE. Table 5 is the most relevant of these, since decoded feasibility matters more for BO than latent separability.

Critical weaknesses

1. The paper does not show that the addressed problem incurs a real cost. Since the constraints are known and cheap to evaluate, infeasible proposals could simply be filtered by checking gi(x)g_i(x) after decoding, before calling the oracle. The paper does not compare against this alternative.

2. There is a logical gap between the method and sample efficiency. The linear constraint reduces optimization cost. Sample efficiency, however, depends on whether arbitrary zz in the half-space decodes to feasible inputs, while the discriminator only separates encoder outputs.

Questions for authors

1. Comparison with post-decoding filtering alternatives. Does the count of oracle evaluations include infeasible candidates? Could you provide a comparison of the following three approaches, all using the same TuRBO settings? (a) Standard VAE + filtering via known constraints after decoding (b) Standard VAE + constraints imposed by a linear classifier trained on the latent space (c) CooL-BO VAE + GP-based constraints This comparison would help disentangle whether the performance gains stem from the learned representations, the use of linear-constrained Bayesian optimization (BO), or the avoidance of wasted oracle evaluations on infeasible candidates.

2. Performance improvements in tasks with pre-existing linear separability. CooL-BO demonstrates performance gains even in synthetic tasks where the latent space of a standard VAE is already nearly linearly separable (as evidenced by accuracy ≥ 0.99 for Ackley20h, Keane50h, and Powell200h in Table 6). To what do you attribute these gains? If it is not due to the resolution of entanglement, could you explain what other mechanism might be at play?

Reviewer_D4SEOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

CooL-BO – constraint aligned latent Bayesian optimization

In this work, the authors propose a framework to enforce separability in the latent space during the pretraining of the VAE to enable a well-separated latent space for selection of feasible versus infeasible candidates with BO. They do so by introducing class specific priors explicitly learning disjoint the feasible and infeasible spaces. They test their model on a subset of tasks from the molecular generation benchmark Guacamole, on the mathematical expression task and other synthetic tasks.

Critical strengths
  1. The paper is well-organized and well-written. They do a great job of conceptually placing their framework among the wider VAE landscape.
  2. The overall idea and strategy is very useful and novel and provides a way to explicitly enforce separability in the latent space without discarding data. This approach could have general applicability in guiding generative models.
  3. Their method outperforms other latent space BO methods at the chosen tasks.
Critical weaknesses

Weaknesses This is noted in the limitations, but I will mention it here for the sake of completeness. It seems that there is a strong dependence on the choice of threshold (feasible/infeasible) chosen for BO but it is not fully explored in the paper. This could limit the applicability of this method.

Questions for authors

None.

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

This manuscript studies latent space Bayesian optimization (BO) and focuses on improving alignment of latent embeddings in cases where feasibility constraints for designs are known and available. Latent alignment is improved by incorporating a discriminator into the generative model, which effectively corresponds to decoding binary feasibility value. VAE model and hence disentangling latent embeddings for feasibility even before BO loop starts. This can help disentangling the latent space wrt feasibility and can improve the selection of feasible designs during BO steps. Feasibility information is used already while training

%Authors assume that if feasibility is not directly available, as may be the case in real applications, feasibility can be checked easily. Authors mention e.g. drug-likeness scores as proxies for feasibility in molecule design and prior expert demonstrations in safe control. Whether such expert knowledge is easy and cheap to convert into explicit constraints may be a judgement call, and definitely depends on application; thus raising a question whether the proposed method is generally applicable.

Authors discuss earlier works, but in my opinion do not sufficiently cover methods that aim to do similar or exact same thing as this work. Other alternative methods for guiding or disentangling the latent embedding of designs include, e.g.:

  • REF1: Deep metric learning method: Antoine Grosnit et al. “High-dimensional Bayesian optimisation with variational autoencoders and deep metric learning”. In: arXiv preprint arXiv:2106.03609 (2021).
  • REF2: jointly learning the reconstruction and feasibility: Natalie Maus et al. “Local Latent Space Bayesian Optimization over Structured Inputs”. In: Advances in Neural Information Processing Systems. Ed. by S. Koyejo et al. Vol. 35. Curran Associates, Inc., 2022, pp. 34505–34518.
  • REF3: using conditional generative models: Siddharth Ramchandran, et al. High-dimensional bayesian optimisation with gaussian process prior variational autoencoders. In International Conference on Learning Representations, 2025. In particular, in Section 3.3, rows195-207: To my understanding, this corresponds to what pretraining of a VAE using the method from REF2 would be able to accomplish, as well as methods REF1 and REF3. REF1 and REF2 are already cited in this manuscript. While methods REF1-REF3 (and possibly other methods too) may not be introduced in the context of feasibility it should be straightforward to adapt them to this specific case. On the other hand, having said that, methods in REF1-3 use slightly different computational techniques to incorporate additional information into the generative model training, except REF2, which is very closely related.

Page 5, rows 261-262: authors note the conceptual similarity with REF2. With change of variable, i.e., by defining y to be t. the training in REF2 is very closely related to that used in this work, even more so if linear kernel is used in REF2.

Introducing trainable, different priors for feasible and infeasible classes changes the generative model to conditional VAE. Also, discussion in “Relation to prior work” is confusing. Your choice of using a mixture prior corresponds to explicitly conditioning your prior with T, i.e., p(X|T), although your prior is not neural network parameterised. Further, this conditioning conflicts generative or probabilistic modeling principles and causes circular reasoning by creating a cycle in your probabilistic graphical model: your prior depends on T, from which you sample you latent Z, which guides your feasibility indicator, i.e., your probabilistic graphical model looks like: T -> Z -> T. This creates a flaw in your generative model and representation learning, although the actual BO can still be applied no matter how the latent representations are obtained. Consequently, I do not agree with the probabilistic graphical model in Appendix A, Fig 7 e, since your graph (on right) does not have dependency from T to Z (or from c to x, to use notation from Fig. 7).

Evaluation: authors define tasks that have task-specific constraints. For molecule generation, the constraint is somewhat counterintuitive as drug-likeness is upper-bounded.

Evaluation: Authors assume that feasibility information is available for all data items, say molecules, during VAE pre-training. Presumably feasibility information should then be available for all molecules also during BO steps. Consequently, all baseline methods should have access to feasibility information as well. During BO steps, when a baseline model proposes a new design, then the feasibility information should be available, which could be quick and cheap to test: if the new design is infeasible, then a new design should be requested from Thompson sampling. It is unclear if baselines have access to feasibility information.

Critical strengths

A strength of this manuscript is the concept and idea of trying to align the latent embeddings wrt feasibility. However, the methodology does not seem to be entirely valid, and discussion of and comparison to relevant previous work is missing.

Critical weaknesses

Deep generative modeling methodology seems to have a fundamental flaw.

While empirical comparisons look promising they do not include all relevant baselines methods.

Questions for authors

Main comments are already in the main review summary above, here are minor comments and questions:

Equation 3: \Gamma(X|Z) should perhaps be rather called as decoder likelihood of the deep generative model, instead of posterior, and in latent BO applications \Gamma(\cdot) is commonly non-continuous valid if X represent e.g. molecules.

Equation 5: you have decided to use variational approximation that does not depend on T. Can you elaborate why this modeling assumption? Also, the last term seems to contain a typo: should read \mathbb{E}_{Psi}(Z|X) [\log \Omega(T|Z)].

Can you please clarify the following statement: “Since the input domain of these tasks is orthogonal, we adapt these tasks to our problem of constrained latent space BO by first projecting the problem to a higher-dimensional space, …”

Ablation: it would be useful to include an ablation experiment where you train a model only on feasible subset.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.