acceptodds
Under review as a conference paper at ICLR 2027

Acquisition Function-Guided Generation for Bayesian Optimisation over Discrete Spaces

Abstract

Bayesian optimization (BO) over discrete spaces is challenging due to missing gradients and the sparsity of valid solutions. Latent-space BO sidesteps both issues by fitting a surrogate in the latent space of a generative model, which however substantially complicates the direct incorporation of domain knowledge that is naturally expressed in the original discrete space. In this work, we instead combine a pretrained generative model with a surrogate defined directly in data space, where inductive biases such as domain-specific kernels can be readily incorporated. Rather than maximizing the acquisition function over the discrete design space, our method samples candidates from the generative model under a distribution tilted by the acquisition function. To recover useful gradients for sampling, we embed this tilted distribution into a diffusion bridge that yields a smooth, posterior-averaged acquisition function. This leads to *acquisition function guided generation* (AFGG), which we instantiate with a categorical diffusion model and an autoregressive normalizing flow. Across a broad suite of molecular benchmarks, the diffusion model variant performs on par with leading BO methods, while the flow model variant outperforms state-of-the-art methods in most cases and remains highly competitive otherwise. Our implementation is available at https://anonymous.4open.science/r/afgg.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian Optimization, Difffusion Models, Molecular Optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_djvAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper reformulates the inner loop of discrete-space Bayesian optimization. Instead of maximizing the acquisition function, the proposed method (AFGG) samples from P⋆(w)∝Pθ(w) ℓ(w)P^\star(w) \propto P_\theta(w)\,\ell(w), obtained by tilting a fixed pre-trained generative model with the acquisition function. This keeps the surrogate in data space, which allows domain-specific kernels such as Tanimoto and avoids retraining the latent space during BO. The authors make three technical contributions:

  1. Since ℓ(τ(x))\ell(\tau(x)) is piecewise constant over decoding cells, they smooth it with a Gaussian noise kernel to obtain a usable guidance signal (Prop. 1).
  2. They show that the guidance amounts to replacing the posterior mean with an acquisition-weighted mean z^tilt\hat{z}^{\text{tilt}}, estimated via SNIS (Prop. 2), so the acquisition function only needs to be evaluable pointwise.
  3. They apply the framework to both a categorical diffusion model (CDCD) and an autoregressive flow (SeqFlow), with trust-region initialization around the incumbent
Critical strengths

1. A general and clean formulation. This approach allows for the direct use of Expected Improvement (EI), as it only requires the acquisition function to be non-negative and point-wise evaluable. While the generative model remains fixed, the surrogate model operates within the data space. This separation enables the application of powerful, pre-trained generative models to Bayesian Optimization (BO) without the need for retraining.

2. Significant empirical performance improvements on standard benchmarks. Across 25 PMO objectives, the SeqFlow variants outperformed the baseline models on most tasks with margins that frequently exceeded the standard deviation by a wide margin.

Critical weaknesses

1. Lack of empirical evidence for the problems addressed. The authors motivate their research by identifying three limitations of Latent Space Bayesian Optimization (LSBO): (a) the inability to incorporate domain-specific knowledge into kernel functions within the latent space; (b) the risk of overfitting in small-data regimes, as the latent space is required to serve dual, simultaneous roles; and (c) the potential for the acquisition function optimization process to stray from the data manifold, rendering the resulting decodings unreliable. While point (a) appears intuitively sound, the paper fails to analyze the frequency or severity with which issues (b) and (c) actually occur in practice. Consequently, the authors provide insufficient justification for whether the problems they aim to solve are truly worth addressing.

2. Lack of analysis demonstrating that the problems have actually been resolved. While AFGG achieves significant performance improvements, the authors do not analyze whether these gains are, in fact, the result of addressing the three previously identified issues. It is acknowledged that the proposed algorithm is structurally capable of circumventing all three problems; however, whether the observed performance improvements are truly a consequence of resolving these specific issues is a separate matter. Further details regarding this weakness are addressed in the 'Question' section below.


Minor comments

  1. The notation used deviates from standard conventions, which hinders readability. Here, X\mathcal{X} denotes a discrete design space with elements represented as ww, while the lowercase xx refers to a continuous embedding, E(w)∈RL×dE(w) \in \mathbb{R}^{L \times d}.

  2. The caption for Table 10 states that setting λ=0\lambda=0 results in uniform SNIS weights. However, according to Equations (45)–(46), setting λ=0\lambda=0 merely removes the guidance term while leaving the weights unchanged; the weights become uniform only when γ→∞\gamma \to \infty.

  3. In Equation (45), the target to which λ\lambda is applied is denoted as z^tilt,γ\hat{z}^{\text{tilt}, \gamma}, where γ\gamma is the temperature parameter that appears in Equation (46).

Questions for authors

We request that the authors provide the following analyses. Each analysis should address the underlying cause of a claimed issue or improvement.

(a) Domain Knowledge. Compare the predictive accuracy (e.g., NLL or rank correlation) of Gaussian Processes (GPs) in the data space (using a Tanimoto kernel) versus the latent space for the PMO task. Alternatively, report the Bayesian Optimization (BO) performance of AFGG when the latent space kernel is replaced with a Tanimoto kernel.

(b) Latent Adaptation. Demonstrate whether the surrogate model maintains its accuracy throughout the BO process even without latent adaptation, and quantify the savings—in terms of wall-clock time and the number of oracle calls—achieved by omitting this adaptation step.

(c) Off-Manifold Drift. Compare the candidates proposed by AFGG with those proposed by the LSBO baseline in terms of validity, likelihood under the generative model, and FCD (Fréchet ChemNet Distance).

Sources of Improvement. Explain the results obtained when guidance (enabled vs. disabled) and trust regions (enabled vs. disabled) are removed from the SeqFlow variant that yields the primary results. Currently, all ablation studies are conducted using only CDCD.

4 comments
AuthorsOpenReview ↗

Weaknesses and Questions: (a) Domain Knowledge

We thank the reviewer for the time dedicated to our work, and are glad that our performance on the standard molecular benchmarks and the theoretical formulation, which lets a pretrained generative model be used for BO without retraining, are appreciated.

Weaknesses & Questions

  1. Lack of empirical evidence for the problems addressed. The authors motivate their research by identifying three limitations of Latent Space Bayesian Optimization (LSBO): (a) the inability to incorporate domain-specific knowledge into kernel functions within the latent space; (b) the risk of overfitting in small-data regimes, as the latent space is required to serve dual, simultaneous roles; and (c) the potential for the acquisition function optimization process to stray from the data manifold, rendering the resulting decodings unreliable. While point (a) appears intuitively sound, the paper fails to analyze the frequency or severity with which issues (b) and (c) actually occur in practice. Consequently, the authors provide insufficient justification for whether the problems they aim to solve are truly worth addressing.
  2. Lack of analysis demonstrating that the problems have actually been resolved. While AFGG achieves significant performance improvements, the authors do not analyze whether these gains are, in fact, the result of addressing the three previously identified issues. It is acknowledged that the proposed algorithm is structurally capable of circumventing all three problems; however, whether the observed performance improvements are truly a consequence of resolving these specific issues is a separate matter. Further details regarding this weakness are addressed in the 'Question' section below.

We thank the reviewer for their careful assessment and address these points one by one.

(a) Domain Knowledge. Compare the predictive accuracy (e.g., NLL or rank correlation) of Gaussian Processes (GPs) in the data space (using a Tanimoto kernel) versus the latent space for the PMO task. Alternatively, report the Bayesian Optimization (BO) performance of AFGG when the latent space kernel is replaced with a Tanimoto kernel.

The revision adds this comparison in Appendix F.7. We fit each surrogate on training sets of 100, 500, 1000 and 5000 molecules, evaluate on a held-out set of 500, and repeat over 5 seeds and 12 objectives, reporting NLL (calibration of the full predictive distribution), RMSE (accuracy of the mean prediction) and Spearman ρ\rho (how well the surrogate ranks candidates). The LSBO surrogates use the deep kernels of their released implementations, run without modification; ours uses a Tanimoto kernel in data space and no deep kernel at all. Crucially, at every size of the training data, a GP defined in data space with a domain-informed kernel outperforms the deep latent-space kernels on all three metrics. The appendix also shows a representative single-objective comparison of the three kernels at the smallest and largest training sets.

AuthorsOpenReview ↗

Weaknesses and Questions: (b) Latent Adaptation & (c) Off-Manifold Drift

(b) Latent Adaptation. Demonstrate whether the surrogate model maintains its accuracy throughout the BO process even without latent adaptation, and quantify the savings—in terms of wall-clock time and the number of oracle calls—achieved by omitting this adaptation step.

The revision tracks surrogate accuracy across the whole optimization run (Appendix F.7). We record test NLL on a fixed held-out set of 500 molecules at every BO iteration, over 12 objectives and 300 sequential iterations. For the LSBO baselines we run two variants: one updating the GP and the generative model jointly, as their released code does, and one updating only the GP while the generator stays frozen (i.e., no latent adaptation).

Jointly retraining the latent-space GP with its generative model does not consistently improve predictive accuracy in the small-data regime, and in several configurations it produces behaviour consistent with overfitting; in some cases updating the GP alone gives the better NLL. Our data-space GP paired with SeqFlow has the better test NLL across essentially all objectives and iterations. Paired with CDCD it does not: three of the twelve objectives show the same pathological behaviour as the latent-space arms, while the remaining nine match the SeqFlow pairing. Since the GP and its hyperparameter settings are identical across the two pairings and only the generative model changes, we attribute those three cases to the weaker prior rather than to the surrogate, consistent with the FCD gap of Appendix F.3.

Concerning the two kinds of savings, in oracle calls, AFGG spends none on adaptation, because the generator is frozen. CoBO and LOL-BO do spend them, realigning a latent with the molecule it decodes to, and their released code does not count those calls against the oracle budget; following [1] we modified both so that they do. This saving is therefore already reflected in the main results, where every method is compared at the same number of oracle evaluations. In wall-clock time we claim no saving. AFGG avoids generator retraining but spends more on guided sampling: The revised paper now reports wall-clock time for the entire BO run for all methods (Table 25, Appendix F.8). While AFGG is more expensive that the comparison methods, it still only takes at most 1-2 minutes for a single BO epoch, which is negligible in essentially all settings where BO is relevant, i.e., where evaluating the objective function dominates the cost, such as wet-lab experiments or expensive simulations.

(c) Off-Manifold Drift. Compare the candidates proposed by AFGG with those proposed by the LSBO baseline in terms of validity, likelihood under the generative model, and FCD (Fréchet ChemNet Distance).

We address this question in the revised Appendix F.7 (Fig. 10) by analyzing the likelihood of proposed candidates under their respective pretrained generative models. While the two other metrics suggested by the reviewer (validity and FCD) are informative, they have some limitations in our setting. Since all methods generate SELFIES sequences, molecular validity is largely ensured by construction and is therefore unlikely to distinguish the methods meaningfully. FCD, on the other hand, measures distributional similarity, but its estimation can be noisy when based on the relatively small number of candidates proposed during a single BO run.

We therefore focus on likelihood under the frozen pretrained generative model as a more direct measure of departure from the learned prior. Since the methods employ different generative backbones and likelihood estimation procedures (exact likelihoods for normalizing flows, Hutchinson-based estimates for CDCD, and IWAE bounds with K=64K=64 for VAEs), raw log-likelihoods are not necessarily directly comparable. Instead, we standardize each candidate's log-likelihood relative to unconditional samples from the same model, using their mean and standard deviation. Values near zero indicate likelihoods typical of the pretrained distribution, while increasingly negative values indicate greater deviation toward lower-likelihood regions.

We do observe the drift, and it separates the methods. Compared against five different objective functions, LSBO methods based on a VAE tend to drift away substantially from their reference distribution. On the contrary, NFBO and AFGG, SeqFlow and CDCD, are less exposed toward this behaviour.

AuthorsOpenReview ↗

Sources of improvement and minor comments

Sources of Improvement. Explain the results obtained when guidance (enabled vs. disabled) and trust regions (enabled vs. disabled) are removed from the SeqFlow variant that yields the primary results. Currently, all ablation studies are conducted using only CDCD.

We agree with the reviwer and provide a more comprehensive assessment extending to AFGG with SeqFlow. The revision repeats the ablation on the SeqFlow backbone (Appendix F.6), under a protocol identical to the CDCD one: the same seven objectives, the same sequential setting, the same five seeds, each arm differing from the base configuration in exactly one setting.

Guidance is the component that matters most, and by a clear margin. Switching it off (using uniform SNIS weights, so the tilted posterior mean reduces to the untilted one) drops performance from 0.541 to 0.502, with losses on six of seven objectives (Table 21). This mirrors the CDCD results (0.480 to 0.446, five of seven objectives; Table 15). The same picture emerges on the two new Ehrlich protein instances, where guidance improves the average score from 0.442 to 0.572. For c=2c=2, two of five guided runs reach the optimum, while none of the unguided runs do (Table 11, Appendix F.2). A stronger generative backbone therefore does not make guidance redundant. Other design choices have a much smaller and less consistent impact: varying the acquisition function yields scores between 0.520 and 0.560 (Table 22), while changing where the temperature acts gives 0.524–0.545 (Table 24), with differences largely within the seed-to-seed variability.

For the trust region, the revised version studies the effect of its size on CDCD (Table 28). Widening the region consistently hurts performance: ρ=0.90\rho=0.90, 0.800.80, and 0.600.60 all perform progressively worse than the default ρ=0.95\rho=0.95. Tightening it further to ρ=0.99\rho=0.99 also comes at a cost. The default ρ=0.95\rho=0.95 thus strikes a good balance, suggesting that AFGG benefits from keeping the search relatively close to the incumbent without restricting it too strongly.

Minor comments

  1. The notation used deviates from standard conventions, which hinders readability. [...]

We thank the reviewer for spotting these three points, and we agree that the readability of the notation can be improved. We will adopt the more standard convention in the camera-ready version.

  1. The caption for Table 10 states that setting λ=0\lambda=0 results in uniform SNIS weights. However, according to Equations (45)–(46), setting λ=0\lambda=0 merely removes the guidance term while leaving the weights unchanged [...]

The reviewer is correct. Setting λ=0\lambda = 0 removes the guidance term so that the drift targets the untilted posterior mean, and it leaves the draws and the importance weights untouched, so the caption's description of that arm is not accurate. We will uniformise the description of the guidance-off arm across the captions and the main text in the camera-ready version.

  1. In Equation (45), the target to which [the gradient] is applied is denoted [...], where [...] is the temperature parameter that appears in Equation (46).

The two knobs are distinguished in the current revision: λ\lambda scales the drift after the tilted mean has been formed (Eq. 45), while γ\gamma is the temperature on the importance weights (Eq. 46), and the paragraph preceding Eq. (45) states that the two are different operations.

[1] Jaewon Chu, Jinyoung Park, Seunghun Lee, and Hyunwoo J. Kim. Inversion-based latent Bayesian optimization. In Advances in Neural Information Processing Systems, volume 37, 2024.

Dear Reviewer djvA,

As the discussion period ends soon, we kindly ask whether our rebuttal and revised manuscript have addressed your main concerns.

Thank you for your time, The Authors

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

The authors consider black-box optimization of discrete sequences like molecules, a popular topic at the intersection between bayesian optimization (BO) and generative modelling. There, the historical approach has been to carry BO directly in the latent space of a generative model using a probabilistic surrogate mapping a latent code to the output. Recently, state-of-the-art performance was achieved using a decoupled approach, where the generative model only serves for generation, and the surrogate maps the structured object directly to the output (e.g., using a Tanimoto kernel over molecular fingerprints for molecules). The main idea is to introduce a relevant sampling scheme to steer the generative models toward regions of high interest with respect to the surrogate.

This work operates in this setting and proposes to tilt the generative model's sampling distribution by an acquisition function. More specifically, the authors embed the tilted distribution into a diffusion bridge, yielding a smooth, posterior-averaged acquisition function whose log-gradient provides the guidance signal for sampling. After introducing their method, the authors demonstrate its superiority on molecular design benchmarks against relevant baselines

Critical strengths
  • The method outperforms competitor baselines in all experiments.
Critical weaknesses
  • I believe the novelty aspect is more limited than what the author makes it sound like. The method is overall quite similar to COWBOYS [1]. Even if the authors mention L52, "To the best of our knowledge, COWBOYS (Moss et al., 2025) is the only prior BO method that follows this strategy, but it is tied to a VAE as the generative model and probability of improvement as the acquisition function.". I do not think that COWBOYS is "tied" to these choices, and I view these as implementation choices. Other generative models could have been used and other acquisition functions as well, typically SeqFlow and LogEI, as used by AFDD. Besides that, COWBOYS and AFDD share the same surrogate, and both freeze the generative model during the BO trial. The difference that remains is the guided sampling scheme: the diffusion bridge proposed by the authors, versus the pCN MCMC scheme of COWBOYS. In other words, an ablation/experimental result comparing AFDD-SeqFlow-EI with COWBOYS-SeqFlow-EI would be better suited to tell whether AFDD provides a clear improvement. Overall, reading the method section, I felt like this connection was a bit hidden, as it can damage the novelty aspect of the work.

  • While the related work mentions active generation as a concurrent line of work, none of these approaches are present in the experiments, only "classical LSBO baselines" are reported there. Also see question below.

  • The experiment section in the main text is limited to a description of the results that anyone could make from reading the figures. In particular, many additional ablation studies are relegated to Appendix E.2 and mentioned in a 2-line sentence in the main text. From the main text only, it is basically impossible to understand what makes this method better than the others, and how do its moving parts influence performance.

[1] Return of the Latent Space COWBOYS: Re-thinking the use of VAEs for Bayesian Optimisation of Structured Spaces, ICML 2025

Questions for authors
  • Coming back to the first weakness I mentioned, for the generative model in COWBOYS, did you use SeqFlow or a VAE? COWBOYS' original implementation uses a VAE, but this submission's text is misleading:

L52: "To the best of our knowledge, COWBOYS (Moss et al., 2025) is the only prior BO method that follows this strategy, but it is tied to a VAE as the generative model and probability of improvement as the acquisition function." L373: "[...] These include both the closest prior approach, COWBOYS, and NFBO, which uses the same SeqFlow generative backbone as one of our AFGG variants."

  • Would it be possible to compare against other generative approaches like Diff-BBO, VSD, GenBO? [1,2,3] At the moment, only GenBO is mentioned, in a quite brief manner, in the related work. If these approaches are not applicable, that could be mentioned.

  • Could you provide the runtime of all the baselines for a full BO trial?

[1] Diff-BBO: Diffusion-Based Inverse Modeling for Black-Box Optimization, NeurIPS 2024

[2] Variational Search Distributions, ICLR 2025

[3] Generative Bayesian Optimization: Generative Models as Acquisition Functions, ICLR 2026

4 comments
AuthorsOpenReview ↗

Author Rebuttal: Weaknesses (Part I)

We thank the reviewer for the insightful comments and are glad that our work is perceived as understandable, informative, and sound, and that our strong empirical results are appreciated.

Weaknesses

  1. I believe the novelty aspect is more limited than what the author makes it sound like. The method is overall quite similar to COWBOYS [1]. [...] I do not think that COWBOYS is "tied" to these choices, and I view these as implementation choices. [...] The difference that remains is the guided sampling scheme: the diffusion bridge proposed by the authors, versus the pCN MCMC scheme of COWBOYS. In other words, an ablation/experimental result comparing AFDD-SeqFlow-EI with COWBOYS-SeqFlow-EI would be better suited to tell whether AFDD provides a clear improvement.

We thank the reviewer for pointing this out and agree that "tied to" was too strong: COWBOYS could in principle be combined with other generative models and acquisition functions. We have revised this claim in the paper. Nonetheless, a VAE with probability of improvement is the only setting COWBOYS considers, and it approximates the VAE decoder as deterministic. AFGG makes no such approximation.

As the reviewer notes, the crucial difference lies in the sampler. COWBOYS samples the tilted latent distribution with preconditioned Crank–Nicolson (pCN), a random-walk MCMC method that leaves the Gaussian prior invariant but---importantly---makes no use of gradients of the acquisition function. AFGG instead obtains these gradients from the posterior-averaged acquisition function along the bridge and uses them to guide generation at every noise level. This matters especially in our setting where the latent space is high-dimensional: pCN's acceptance rate is robust to dimension [6], but its proposal is blind to the tilt, and gradient-free random walks need far more steps than gradient-informed ones to traverse a high-dimensional target [7].

To isolate exactly this difference, the revision includes the suggested comparison to COWBOYS-SeqFlow-EI (Table 5, Appendix E): AFGG-SeqFlow-EI with its bridge replaced by the pCN sampler of COWBOYS, targeting the same tilted distribution N(z;0,I)ℓ(z)γ\mathcal{N}(z; 0, I) \ell(z)^{\gamma} with the same number of decoder and acquisition evaluations per BO iteration and no more wall time. pCN performs significantly worse: it scores below the bridge on 20 of 25 objectives, significantly (more than two standard errors) on 14, and significantly above it on only one (QED, by 0.0020.002). The gaps are often large, e.g., 0.840.84 vs. 0.500.50 on Celecoxib rediscovery, 0.780.78 vs. 0.440.44 on JNK3, and 0.980.98 vs. 0.690.69 on Albuterol similarity. With backbone, surrogate, and acquisition function held fixed, the improvement is therefore due to the gradient-guided bridge.

  1. While the related work mentions active generation as a concurrent line of work, none of these approaches are present in the experiments, only "classical LSBO baselines" are reported there.

This is a fair point, and we have addressed it experimentally. Neither VSD nor GenBO reports any molecular results, so there is no published configuration of either method for our benchmark. Rather than port them to a setting their authors never tested, the revision evaluates AFGG on their own benchmark: two Ehrlich functions [4] (32 amino acids, c∈2,4c \in \\{2,4\\} motifs, optimum 1), reported by both, with 10 initial observations and 1,500 oracle evaluations over 5 seeds (Figure 3; setup in Appendix F.2). AFGG is applied without modification, and VSD and GenBO run with their authors' code and their own length-32 Ehrlich configurations.

Among the generative and BO methods compared, AFGG (SeqFlow) is the strongest on both instances and the only one to reach the optimum (2 of 5 seeds on c=2c=2):

c=2c=2c=4c=4
AFGG (SeqFlow)0.800 ± 0.2090.345 ± 0.150
AFGG (CDCD)0.525 ± 0.1570.352 ± 0.099
NFBO0.525 ± 0.0840.211 ± 0.041
VSD0.475 ± 0.0950.211 ± 0.000
GenBO0.275 ± 0.0560.122 ± 0.026
AuthorsOpenReview ↗

Author Rebuttal: Weaknesses (Part II)

  1. The experiment section in the main text is limited to a description of the results that anyone could make from reading the figures. In particular, many additional ablation studies are relegated to Appendix E.2 and mentioned in a 2-line sentence in the main text. [...]

We thank the reviewer for pointing out that results important for understanding the method were buried in the appendix. The revision extends the experiment section in the main text with a paragraph, "Which components matter," that states what drives performance and what does not.

In short: the generative prior matters. Our bridge needs only a deterministic bijection, so CDCD's probability-flow ODE drops in for the flow and lets us swap the backbone with the sampler held fixed (Appendix F.4). Keeping our sampler but replacing the SeqFlow backbone with CDCD's lowers the final score by 0.081±0.0160.081\pm0.016, five standard errors; keeping the backbone but replacing the sampler changes nothing (−0.007±0.007-0.007\pm0.007). The backbone carries the entire gap between our two variants. The same holds for NFBO, which drops from 0.4690.469 to 0.3730.373 when its SeqFlow backbone is replaced by CDCD (Table 13).

The guidance nevertheless matters just as much. At the same CDCD backbone, AFGG reaches 0.4580.458 against NFBO's 0.3730.373; turning the guidance off costs 0.480→0.4460.480\to0.446 (Table 15); REINFORCE in place of the closed-form SNIS estimator costs 0.480→0.4090.480\to0.409 (Table 16); and the gradient-free pCN sampler of COWBOYS is worse on 20 of 25 objectives (Appendix E, see W1). The acquisition function, the pool size beyond 10, and where the temperature acts all move the mean by less than a seed-to-seed standard deviation.

Since the previous ablations used CDCD only---as both this weakness and Reviewer smgw's Q4 note---the revision repeats the full ablation on SeqFlow (Appendix F.6). The conclusions carry over: switching the guidance off costs 0.541→0.5020.541\to0.502 (Table 21), so the stronger prior does not make the guidance redundant.

AuthorsOpenReview ↗

Author Rebuttal: Questions

Questions

  1. Coming back to the first weakness I mentioned, for the generative model in COWBOYS, did you use SeqFlow or a VAE? [...]

COWBOYS uses its original VAE, which we retrain on GuacaMol from the released code since no checkpoint is provided (Appendix F.1). SeqFlow, an autoregressive normalizing flow, is used only by AFGG-SeqFlow and NFBO. We agree that L373 was confusing and have revised it to state which backbone each baseline uses, alongside the "tied to" claim (see W1). The SeqFlow-EI comparison to COWBOYS' sampler is given in Table 5 (see W1).

  1. Would it be possible to compare against other generative approaches like Diff-BBO, VSD, GenBO? If these approaches are not applicable, that could be mentioned.

VSD and GenBO are now compared against directly, on the Ehrlich protein benchmark that both of them report, where AFGG (SeqFlow) outperforms VSD and GenBO on both benchmarks (see W2 above for the setup). We do not run Diff-BBO [5], as it provides no publicly available implementation. The revision states this in the related work.

  1. Could you provide the runtime of all the baselines for a full BO trial?

The revision includes the wall-clock time of a full BO trial for the five main baselines and both AFGG variants (Table 25, Appendix F.8), in the sequential and batch settings, together with a per-iteration breakdown of AFGG's cost. In the sequential setting, a BO iteration takes 1–10 seconds for the baselines, whereas AFGG takes about 24 seconds (CDCD) and 71 seconds (SeqFlow) on a single RTX 8000 GPU. While AFGG is more expensive, this overhead is small in essentially all applications where BO is relevant, i.e., in settings where oracle evaluations can take hours or days to evaluate, such as wet-lab experiments or expensive simulations.


[1] Moss, H. B., et al. (2025). Return of the Latent Space COWBOYS: Rethinking the Use of VAEs for Bayesian Optimisation of Structured Spaces.

[2] Steinberg, D. M., Oliveira, R., Ong, C. S., & Bonilla, E. V. (2025). Variational Search Distributions. In International Conference on Learning Representations (ICLR).

[3] Oliveira, R., Steinberg, D. M., & Bonilla, E. V. (2026). Generative Bayesian Optimization: Generative Models as Acquisition Functions. In International Conference on Learning Representations (ICLR).

[4] Stanton, S., et al. (2024). Closed-Form Test Functions for Biophysical Sequence Optimization and Design.

[5] Wu, D., Kuang, N. L., Niu, R., Ma, Y.-A., & Yu, R. (2025). Diffusion-BBO: Diffusion-Based Inverse Modeling for Online Black-Box Optimization. arXiv:2407.00610.

[6] Cotter, S. L., Roberts, G. O., Stuart, A. M., & White, D. (2013). MCMC methods for functions: Modifying old algorithms to make them faster. Statistical Science, 28(3), 424–446.

[7] Roberts, G. O., & Rosenthal, J. S. (1998). Optimal scaling of discrete approximations to Langevin diffusions. Journal of the Royal Statistical Society: Series B, 60(1), 255–268.

Reviewer_VbGoOpenReview ↗

Happy with the rebuttal

I am happy with the rebuttal provided by the authors regarding my assessment and that of other reviewers, I will increase my score to 3!

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

The authors propose Acquisition Function-Guided Generation (AFGG), a framework for BO over discrete spaces. Instead of directly maximizing the acquisition function, AFGG guides pretrained generative models to sample promising candidates while maintaining a surrogate model in the original data space. The method is implemented using diffusion and normalizing flow models and evaluated on 25 molecular optimization tasks, demonstrating competitive or superior performance against existing baselines.

Critical strengths
  1. Methodological Novelty: The paper introduces a novel acquisition-guided generation framework that avoids directly optimizing acquisition functions over discrete spaces.

  2. Strong Theoretical Foundation: The proposed method is supported by mathematical derivations of the acquisition-tilted distribution and guidance score.

  3. Generality: The framework is demonstrated with two different generative models, including categorical diffusion and autoregressive normalizing flows.

  4. Comprehensive Experiments: The authors evaluate the method on 25 molecular optimization tasks in both sequential and batch settings, comparing against multiple strong baselines.

  5. Strong Empirical Performance: The proposed method achieves competitive or superior performance on many benchmarks, demonstrating its effectiveness.

Critical weaknesses
  1. Computational Overhead: The proposed method requires repeated posterior sampling and acquisition function evaluations during guided generation, which could potentially introduce significant computational costs compared to existing BO methods.

  2. Dependence on Pretrained Generative Models: The performance of AFGG heavily depends on the quality of the pretrained generative model. Promising candidates that are poorly represented by the generator may be difficult to discover.

  3. Limited Generalizability: Although the method is evaluated on 25 optimization tasks, all experiments focus on molecular optimization. Additional experiments on other discrete optimization domains would strengthen the generalizability claims.

  4. Approximation Reliability: The guidance term is estimated using self-normalized importance sampling, which may suffer from high variance or weight degeneracy in high-dimensional spaces. Further analysis of the approximation accuracy would be valuable.

  5. Backbone Sensitivity: The SeqFlow variant substantially outperforms the diffusion-based variant on several tasks. More analysis is needed to clarify how much of the improvement comes from the proposed guidance mechanism versus the underlying generative model.

Questions for authors
  1. The proposed method requires repeated posterior sampling and acquisition function evaluations during guided generation. Could the authors provide a detailed comparison of computational costs, including wall-clock time and computational overhead, against existing BO baselines? This is my deepest concern, which may help me assess my judgement in the submission if the authors provide more details.

  2. The experiments focus exclusively on molecular optimization tasks. Could the authors evaluate AFGG on other discrete optimization problems, such as protein sequence design or combinatorial optimization, to demonstrate its broader applicability?

  3. The guidance term relies on importance sampling, and the performance varies substantially between SeqFlow and CDCD. Could the authors investigate how the number of posterior samples affects guidance accuracy and optimization performance, and clarify the extent to which the improvements depend on the underlying generative backbone?

4 comments
AuthorsOpenReview ↗

Author Rebuttal: Weaknesses (Part I)

We thank the reviewer for highlighting the novelty of the work, the theoretical foundations that let AFGG be instantiated with different generative models, and for appreciating our extensive set of experiments.

Weaknesses

  1. Computational Overhead: The proposed method requires repeated posterior sampling and acquisition function evaluations during guided generation, which could potentially introduce significant computational costs compared to existing BO methods.

Posterior sampling and repeated acquisition function evaluations indeed introduce some computational overhead. However, in our experiments both AFGG variants (SeqFlow and CDCD) require at most 1–2 minutes per BO iteration. This overhead is negligible in usual BO setting, often involving expensive wetlab experiments that require substantial financial resources and laboratory time, or computational chemistry simulations that can take hours to days even with extensive parallelization. Given the substantial performance improvements demonstrated by AFGG, we believe this modest additional cost is well justified.

Furthermore, our CDCD variant employs a mean-field posterior approximation, which significantly reduces computational overhead while maintaining competitive optimization performance. In particular, its computational cost is comparable to that of the baselines, making it an attractive alternative in settings where efficiency is a priority. Further details are provided in our answer to Q1 below.

  1. Dependence on Pretrained Generative Models: The performance of AFGG heavily depends on the quality of the pretrained generative model. Promising candidates that are poorly represented by the generator may be difficult to discover.

While it is true that AFGG depens on the quality of the generative model, we view the dependence on the pretrained generative model as a deliberate design choice and a key strength of AFGG, rather than a limitation. By decoupling the generative model from the surrogate model and acquisition function, AFGG can leverage powerful pretrained generators without requiring expensive retraining or fine-tuning throughout the BO process. Importantly, this modularity also allows AFGG to directly benefit from advances in generative modeling: improvements in the generative backbone can translate into improved BO performance without modifying the optimization framework.

Appendix F.3 quantifies the gap between our two backbones (CDCD and SeqFlow): validity, uniqueness and novelty do not separate them, while FCD differs by a factor of roughly 3.53.5 (6.026.02 for SeqFlow against 20.9920.99 for CDCD, Table 12).

  1. Limited Generalizability: Although the method is evaluated on 25 optimization tasks, all experiments focus on molecular optimization.

See Q2 below. In short, the revised version extends the experiments to include protein design benchmarks, where AFGG outperforms competing generative and BO approaches.

AuthorsOpenReview ↗

Author Rebuttal: Weaknesses (Part II)

  1. Approximation Reliability: The guidance term is estimated using self-normalized importance sampling, which may suffer from high variance or weight degeneracy in high-dimensional spaces.

While self-normalized importance sampling (SNIS) can suffer from weight degeneracy, we argue that this is not a significant concern in our setting, both for theoretical and empirical reasons.

First, the variance of the importance weights is not directly determined by the dimensionality of the search space, but rather by how strongly the acquisition function varies under the denoising posterior. In our case, the proposal distribution is precisely this posterior, and the target differs only through the acquisition-based reweighting ℓ1/γ\ell^{1/\gamma}, whose strength is controlled by the temperature parameter γ\gamma. Moreover, guidance is recomputed at every integration step using NN fresh posterior samples, reducing the dependence of the overall trajectory on any individual estimate.

Second, we provide empirical evidence supporting the reliability of our estimator in the newly added Appendix F.9 ("Effective Sample Size"). Across all 25 objectives, the normalized effective sample size (ESS) increases throughout the denoising trajectory, reaching 0.5–0.9 near the final steps (Figs. 11–12). This is consistent with the denoising posterior becoming increasingly concentrated as t→0t \to 0, resulting in more uniform importance weights. While ESS is lower at earlier steps, our ablations indicate that this does not meaningfully affect optimization performance. Specifically, halving or doubling the number of posterior samples NN yields essentially unchanged performance (Table 20). Furthermore, our closed-form SNIS estimator outperforms the REINFORCE-based alternative across all seven evaluated objectives (0.480 vs. 0.409; Table 16).

  1. Backbone Sensitivity: The SeqFlow variant substantially outperforms the diffusion-based variant on several tasks. More analysis is needed to clarify how much of the improvement comes from the proposed guidance mechanism versus the underlying generative model.

We disentangle the contributions of the generative backbone and our proposed guidance mechanism, by adding several ablation studies in the revised manuscript.

First, we investigate the influence of the generative backbone in Appendix F.4. Since our framework only requires a deterministic, invertible mapping, we can also employ the probability-flow ODE of CDCD within our flow-based sampling framework. This allows us to compare different generative backbones while keeping the sampling procedure fixed. We find that replacing SeqFlow with CDCD reduces optimization performance by 0.081±0.0160.081 \pm 0.016, whereas changing the sampling procedure while keeping the backbone fixed has essentially no effect (−0.007±0.007-0.007 \pm 0.007). This suggests that the performance gap between our SeqFlow and CDCD variants is primarily attributable to differences in the underlying generative models, rather than the sampling procedures.

Second, we explicitly evaluate the contribution of our proposed guidance mechanism. Disabling guidance (λ=0\lambda = 0), while retaining acquisition-based candidate selection, consistently reduces optimization performance for both backbones: from 0.4800.480 to 0.4460.446 for CDCD (Table 15) and from 0.5410.541 to 0.5020.502 for SeqFlow (Table 21). This demonstrates that the improvements are not solely due to the quality of the pretrained generators, but that our guidance mechanism provides substantial additional benefits regardless of the backbone.

Taken together, these results highlight two complementary findings: the choice of generative backbone affects the absolute optimization performance, while our proposed guidance mechanism consistently improves performance across both backbones. This also reinforces the modularity of AFGG, which can directly benefit from stronger generative models without requiring changes to the guidance framework.

AuthorsOpenReview ↗

Author Rebuttal: Questions

Questions

  1. Could the authors provide a detailed comparison of computational costs, including wall-clock time and computational overhead, against existing BO baselines?

We have added a detailed runtime comparison in Appendix F.8 (Table 25), covering five LSBO baselines and both AFGG variants in sequential and batch BO settings, averaged over 25 objectives and 5 random seeds on a single NVIDIA RTX 8000 GPU. While AFGG incurs higher computational costs than the baselines, its 1–2 minutes per BO iteration represent a modest overhead in expensive black-box optimization settings, where individual objective evaluations often take hours or days. Given the substantial performance improvements and state-of-the-art BO results achieved by AFGG, we believe this additional cost is well justified.

In sequential BO, the baselines require 1–2 seconds per iteration, compared to 23.6 seconds for AFGG (CDCD) and 71.4 seconds for AFGG (SeqFlow) (Appendix F.8, Table 25). Importantly, CDCD's main bottleneck is RDKit-based molecular fingerprint computation, which is not intrinsic to AFGG. This step is CPU-parallelizable, reducing runtime from 58.20 to 16.43 seconds with 2 versus 16 cores. For SeqFlow approximately 90% of runtime is spent on network and acquisition evaluations (Tables 26–27).

About the efficiency of posterior sampling. To limit overhead, CDCD uses a mean-field approximation that samples independently from token marginals obtained in a single network evaluation (Eq. (49), Appendix B). Table 3 shows that more expensive joint posterior sampling (K=32K=32), despite increasing runtime by ~24× (8.77 vs. 0.36 hours), yields no meaningful performance improvement. SeqFlow instead samples exactly from its Gaussian posterior without approximation.

  1. Could the authors evaluate AFGG on other discrete optimization problems, such as protein sequence design or combinatorial optimization, to demonstrate its broader applicability?

The revision adds protein sequence design (Fig. 3, Appendix F.2). We apply AFGG without modification to two Ehrlich functions (Stanton et al., 2024). We compare against NFBO on the same retrained SeqFlow backbone and against the active-generation methods VSD (Steinberg et al., 2025) and GenBO (Oliveira et al., 2026), run with their authors' code at the same budget. AFGG with SeqFlow outperforms all generative and BO baselines on both instances (0.800.80 against at most 0.520.52 on c=2c = 2, 0.350.35 against at most 0.210.21 on c=4c = 4) and is the only method to find the optimum.

  1. Could the authors investigate how the number of posterior samples affects guidance accuracy and optimization performance, and clarify the extent to which the improvements depend on the underlying generative backbone?

The revision adds this ablation (Table 20, Appendix F.5). Halving or doubling the number of SNIS posterior samples around our default of N=110N = 110 leaves performance unchanged: the means are 0.5490.549 at N=55N = 55, 0.5490.549 at N=110N = 110 and 0.5580.558 at N=220N = 220, all well inside the seed spread. The estimator matters more than its sample budget: REINFORCE targets the same quantity as the closed-form SNIS average but carries higher variance, and loses on all seven objectives (0.4090.409 against 0.4800.480, Table 16).

On the backbone, W5 above gives the decomposition: the backbone accounts for the gap between our two variants, the sampler does not. We see this as the point of the framework rather than a limitation of it --- AFGG is built so that a stronger pretrained generator yields a better search --- and the guidance contributes substantially on top of either prior.

Thank you for the detailed rebuttal and the additional experiments. Table 25 addresses my main question by quantifying the wall-clock trade-off. In sequential BO, CDCD and SeqFlow take 23.6 and 71.4 seconds per iteration, respectively, compared with roughly 1–2 seconds for the baselines. This overhead can be justified when objective evaluations dominate the total cost. I would encourage the authors to make this trade-off explicit in the main text, particularly for settings with inexpensive objective evaluations.

The ESS analysis and the posterior-sample ablation address my concern about the practical stability of SNIS in the tested settings. The backbone swap and guidance-off ablations also clarify where the gains come from. The two Ehrlich tasks extend the evaluation beyond molecules, and Table 11 adds useful nuance: guidance improves SeqFlow on both instances, while CDCD shows small decreases in mean performance. I suggest reflecting this dependence on the backbone and task when summarizing the benefits of guidance.

Overall, the revision addresses my main questions, and I am maintaining my rating of 3 (Weak Accept).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.