- 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 , 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:
- Since is piecewise constant over decoding cells, they smooth it with a Gaussian noise kernel to obtain a usable guidance signal (Prop. 1).
- They show that the guidance amounts to replacing the posterior mean with an acquisition-weighted mean , estimated via SNIS (Prop. 2), so the acquisition function only needs to be evaluable pointwise.
- 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
-
The notation used deviates from standard conventions, which hinders readability. Here, denotes a discrete design space with elements represented as , while the lowercase refers to a continuous embedding, .
-
The caption for Table 10 states that setting results in uniform SNIS weights. However, according to Equations (45)–(46), setting merely removes the guidance term while leaving the weights unchanged; the weights become uniform only when .
-
In Equation (45), the target to which is applied is denoted as , where 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
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 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.
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 , 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: , , and all perform progressively worse than the default . Tightening it further to also comes at a cost. The default 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
- 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.
- The caption for Table 10 states that setting results in uniform SNIS weights. However, according to Equations (45)–(46), setting merely removes the guidance term while leaving the weights unchanged [...]
The reviewer is correct. Setting 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.
- 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: scales the drift after the tilted mean has been formed (Eq. 45), while 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
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
We thank the reviewer for their careful assessment and address these points one by one.
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 (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.