acceptodds
Under review as a conference paper at ICLR 2027

Boundary Unlocking Value for Safe Bayesian Optimization under Monotone Safety Constraints

Abstract

Many safety-constrained optimization problems involve both a system configuration and an ordered intensity. When safety risk increases monotonically with intensity, each configuration is associated with an unknown maximum safe intensity. Under a limited evaluation budget, efficient optimization requires identifying which safe experiments are most valuable for subsequent optimization. Existing safe Bayesian optimization methods often guide query selection using criteria based on uncertainty, optimism, or information gain, but these criteria do not directly quantify the optimization value that may result from resolving different safety boundaries. The monotone configuration–intensity structure makes such value tractable to quantify: the feasibility uncertainty of each configuration can be represented by a latent maximum safe intensity, and each possible boundary location determines an accessible intensity range and the best objective value attainable within it. Building on this observation, we introduce Boundary Unlocking Value (BUV), which combines posterior boundary uncertainty with optimistic objective gains to quantify the exploration potential of each configuration. We further develop BUV-SafeBO, which evaluates candidate points within the certified safe set based on how informative they are about the safety frontiers of high-BUV configurations. It also incorporates expected improvement at each candidate to balance future safe expansion with immediate objective optimization. Experiments on multiple structured safe optimization benchmarks show that BUV-SafeBO achieves competitive optimization performance while maintaining high safe-query rates.

Primary area:
optimization
Keywords:
Safe Bayesian Optimization, Gaussian processes, Monotone Safety Constraints, Boundary Unlocking Value

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UMgFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies safe Bayesian optimization over configuration-intensity pairs, assuming that safety deteriorates monotonically with intensity while the objective can be nonmonotone. It introduces Boundary Unlocking Value (BUV), which averages optimistic objective gains beyond each configuration’s certified safe boundary under a sampled boundary distribution. BUV-SafeBO weights information about selected safety frontiers by these values and combines the resulting exploration score with expected improvement. Experiments on spinal cord stimulation, quadrotor control, and drug-dose optimization claim competitive optimization performance with high empirical safe-query rates.

Critical strengths

S1. A specific contribution to prioritizing safe expansion. BUV distinguishes boundary uncertainty that exposes little objective improvement from uncertainty associated with valuable extensions. The closest structural predecessor, M-SafeOpt, already uses objective information to eliminate suboptimal configurations and restrict expansion, through a binary test against the best safe LCB. BUV differs in both reference and role: it scores each configuration by a probability-weighted magnitude of prospective extension gains relative to its own certified prefix, and uses that score to select and weight frontier information targets rather than to gate which boundaries may be expanded.

Critical weaknesses

W1. Comparisons with baselines confound acquisition design with safety certification and query protocol. Methods use different safety multipliers βg\beta_g (Table 2), ranging from 2 to 5 on SCS. The paper’s own sensitivity analysis demonstrates that this choice matters: increasing BUV-SafeBO’s βg\beta_g from 2 to 4 raises its safe-query rate from 95.75% to 99.30% (Table 6). Query eligibility is also insufficiently specified: the no-repeat rule is stated only for BUV-SafeBO, leaving unclear whether baselines may spend their budgets re-querying evaluated actions, which yield identical responses on these deterministic tables. Consequently, the reported progress–safety trade-offs characterize the complete implementations, but cannot be attributed solely to acquisition design. A controlled comparison using shared certification and query-eligibility rules, or per-method confidence-parameter sweeps showing progress versus safety violations, would help resolve this issue. The repeat policies should also be reported.

W2. BUV’s incremental benefit needs clearer characterization against simpler alternatives. Under matched pipeline settings, EI-only, or the “BUV-EI” variant, Eq. (44), which computes no BUV, already achieves AUC₈₀₀ of 0.656 on SCS and 0.852 on quadrotor, exceeding the strongest reported baseline in each benchmark (ITL: approximately 0.624† and 0.841†). The full method provides additional gains of 0.032 and 0.034 (Table 1). The quadrotor increment is supported by the reported summaries, while the reliability of the SCS increment remains unresolved with five seeds and no paired uncertainty estimates. The entropy-weighting ablation supports objective-sensitive weighting, but does not isolate the benefit of probability-weighted extension gains over simpler objective-aware targeting, such as M-SafeOpt-style binary expander filtering. The evidence therefore supports an incremental benefit, while leaving its consistency and advantage over simpler alternatives incompletely characterized. Paired differences with confidence intervals, EI-only in the main comparison plots, and a controlled comparison with binary-filtered information targeting would strengthen the contribution assessment.

† Baseline AUCs were estimated from the figures’ vector paths because numerical values are not reported. The same extraction procedure reproduces all eight Table 1 entries to within 0.0001.

Questions for authors

Q1. How were the per-method βg\beta_g values in Table 2 selected? Do the progress–safety rankings persist under a common βg\beta_g per benchmark, or across per-method sweeps plotted as progress versus violation?

Q2. Were baselines permitted to re-query evaluated actions? Please report the number of unique actions actually queried by each method on each benchmark.

Q3. For SCS and quadrotor, could you report the mean AUC₈₀₀ differences between BUV-SafeBO and each of BUV-EI and BUV-Uncertainty, paired by seed and accompanied by confidence intervals? Please also include BUV-EI in the LINCS and Sim4Life comparisons.

Q4. Under the same GP models, safety rules, and repeat-query policy, does BUV outperform an M-SafeOpt-style binary filter that retains configurations only when their optimistic extension value exceeds the highest objective LCB in the current safe set? For this comparison, the binary-filter variant would assign equal weights to the selected configurations’ first uncertified points.

Q5. What fraction of configurations violates the monotonicity assumption in Eq. (1)? In particular, how often is a higher intensity safe while a lower intensity is unsafe for the same configuration? Was any preprocessing used to enforce monotonicity? Among unsafe queries, how many were certified by the current LCB rule, and how many were admitted only through historical certificates (Eq. 4)?

6 comments
AuthorsOpenReview ↗

Response to Reviewer UMgF

We thank the reviewer for the detailed assessment and constructive suggestions. We have revised the comparison protocol and further examined the incremental contribution of BUV.

W1 / Q1) Fair comparison under matched settings

We thank the reviewer for identifying the confounding effects of safety certification and query eligibility in our original comparisons. We agree that different βg\beta_g values can produce different certified-safe sets, affecting both optimization progress and safety performance.

In the original experiments, we used method-specific safety confidence multipliers to obtain reasonable performance–safety trade-offs for the respective implementations. However, we acknowledge that this did not fully isolate the contribution of acquisition design, and that the original manuscript did not adequately specify the repeat-query policies of all baselines.

To address these issues, we reran the comparisons under a unified protocol. Within each benchmark, all methods now share the same GP specification, initial observations, safety confidence multiplier, safety certification rule, repeat-query policy, query budget, and random seeds. In particular, all methods restrict query selection to previously unevaluated actions in the certified-safe set, excluding both initial observations and previous online queries. Repeated queries are not permitted.

We regenerated the optimization-performance and safety comparisons in Figures 2, 3, and 4 using these matched settings. The shared safety confidence multipliers are listed below.

MethodSCSQuadrotorLINCS
BUV-SafeBO4163
ISE4163
M-SafeOpt4163
M-SafeUCB4163
SafeOpt4163
ITL4163
HdSafeBO4163

The revised results broadly preserve the qualitative trends of the original experiments, with BUV-SafeBO maintaining competitive optimization performance and safety under the unified protocol.

Q2) Repeat-query policy and unique actions

The original presentation did not adequately specify repeat-query policies for all methods. In our audit of the earlier SCS comparison that unified only the safety multiplier, some baselines did re-query evaluated actions. Their mean numbers of unique online actions per 800 queries were 744.71 for M-SafeOpt, 747.83 for M-SafeUCB, 588.25 for HdSafeBO, 742.71 for SafeOpt, 193.48 for ITL, and 785.29 for ISE. The four BUV variants each queried 800 unique online actions.

In the revised unified protocol, all methods exclude previously evaluated actions, including both initial observations and earlier online queries.

W2 / Q3) Paired AUC differences and the contribution beyond EI-only

We thank the reviewer for this constructive suggestion. We agree that EI-only is a strong comparator and that BUV’s contribution should be characterized as an additional benefit beyond immediate objective optimization, rather than a replacement for EI.

The two components of our acquisition serve complementary purposes. EI evaluates the expected immediate objective improvement within the current certified-safe set, accounting for both the posterior mean and uncertainty of the objective. In contrast, the BUV-weighted information score Qt(x)Q_t(x) measures how informative a certified-safe query is about safety margins at the first uncertified intensities of selected configurations. BUV assigns greater weight to configurations whose potential safe expansions may unlock larger optimistic objective gains. Thus, EI supports immediate improvement, while BUV-weighted information acquisition guides learning about potentially valuable future safe expansions. We have added an explicit discussion of EI’s role in Section 3.3 and expanded the ablation discussion in Section 4.4 to acknowledge its contribution and explain its complementarity with BUV-weighted information acquisition.

For each seed, we first average AUC800 over the fixed tasks or targets and then subtract the comparator’s AUC from that of BUV-SafeBO. The resulting paired differences are summarized below as mean ± sample SD across seeds. Positive values favor BUV-SafeBO.

BenchmarkBUV − EI-onlyBUV − Uncertainty
SCS+0.0476 ± 0.0203+0.1179 ± 0.0558
Quadrotor+0.0383 ± 0.0202+0.0946 ± 0.0386
Sim4Life+0.0297 ± 0.0176+0.0484 ± 0.0310

We have not yet included BUV-EI in the LINCS comparison. LINCS permits only 272 additional queries and was therefore excluded from our common-budget AUC800 ablation study. These results show positive mean increments over both alternatives in the reported settings.

AuthorsOpenReview ↗

Response to Reviewer UMgF

Q4) Controlled comparison with binary-filtered information targeting

We implemented an M-SafeOpt-style binary-filter variant under the same GP models, initial observations, safety certification rule, repeat-query policy, and budget as BUV-SafeBO. Both methods use βg=4\beta_g=4, βf=2\beta_f=2, and the same EI term and normalization procedure.

For each configuration with a nonempty, unresolved certified-safe prefix, the variant retains the configuration only if its maximum objective UCB beyond that prefix exceeds the maximum objective LCB over the current certified-safe set. It assigns equal weights to the first uncertified points of all retained configurations in the same mutual-information score.

The binary variant uses all qualifying configurations, whereas BUV-SafeBO retains its original maximum of M=3M=3 information targets. Thus, the comparison changes target selection and weighting, while preserving the information-routing mechanism and EI component.

We evaluated this comparison on SCS using two targets and ten seeds, and 800 online queries following 200 initial observations. Results are mean ± SEM across seeds after averaging targets within each seed.

MethodAUC800Final progressSafe queries (%)Cumulative violation
BUV-SafeBO0.8080 ± 0.02290.9077 ± 0.033799.0417 ± 0.18070.0641 ± 0.0171
Binary filter + equal-weight MI + EI0.7634 ± 0.02970.8702 ± 0.014999.2396 ± 0.12010.0467 ± 0.0088

BUV-SafeBO achieves higher mean AUC and final progress, with a mean paired AUC difference of +0.0447. The binary variant has a slightly higher safe-query rate and lower cumulative violation.

Q5) Monotonicity and the source of unsafe admissions

We audited every configuration and intensity level in SCS and Sim4Life, distinguishing numerical violations of the nonincreasing safety margin from unsafe-lower/safe-higher reversals.

DatasetConfigurations × intensitiesConfigurations with nonmonotone safety marginsUnsafe-lower/safe-higher pairs
SCS21,800 × 102/21,800 (0.0092%)0/981,000
Sim4Life171 × 1000/171 (0%)0/846,450

The pair counts include all lower–higher intensity pairs within each configuration, not only adjacent levels.

The two SCS violations are small adjacent increases in the safety margin, approximately 7.09×10−67.09\times10^{-6} and 5.97×10−55.97\times10^{-5}. Neither changes the safety classification. Thus, numerical monotonicity is not exact everywhere in SCS, but no configuration exhibits a lower intensity that is unsafe while a higher intensity is safe.

Our local preprocessing does not enforce monotonicity through smoothing, monotone projection, or removal of violating rows. For SCS, the safety margin is computed directly from the maximum activation across response channels. For Sim4Life, preprocessing checks monotonicity and raises an error if the check fails, rather than modifying the responses.

Our audit found that, on both SCS and Sim4Life, all unsafe online queries were admitted by the current LCB-prefix rule. The number of unsafe queries admitted solely through historical certificates was zero in both datasets.

Dear Reviewer UMgF,

Thank you very much for your help with reviewing for ICLR'27!

The authors have now responded to your reviews. As the discussion period is approaching the end, I'm wondering if you could have a look at the authors' responses and see whether they have addressed your concerns. And could you also leave a comment to let the authors know your thoughts? If any concerns remain, please explain which ones have not been addressed and why.

Warm regards, Your AC

Reviewer_35EhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper is on safe BO where actions consist of a configuration and an ordered intensity, and the safety function is assumed to be monotone in the intensity. The authors estimate a distribution over the induced latent boundary from GP samples and define the Boundary Unlocking Value (BUV) as the expected gain that would become accessible if the boundary lies beyond the currently certified one. Based on this, the design BUV-SafeBO which combines BUV multiplied with the mutual information gain with expected improvement. The authors show results on three different benchmarks.

Critical strengths

S1) I see the general direction of efficiently leveraging prior information such as monotonicity in SafeBO application as very interesting and relevant.

S2) Also the key idea of efficiently optimizing for information gain about the boundary values is interesting and well motivated.

Critical weaknesses

W1) My major concern is regarding the empirical comparison of the proposed method against other SafeBO-like baselines. Looking at table 2, the authors vary beta_g, which the the multiplier defining the safe set, not only between different experiments (which one could maybe argue for) but also they change it within one class of experiment and vary it between methods. I believe this is then a clearly unfair comparison as beta_g directly influences the number of violations. beta_g essentially is meant to guarantee P(E_t) >= 1 - delta through uniform error bounds (eg [R1] but also others) and using different beta_gs implies structurally different assumptions on the function class g actually comes from (usually though a bound on the RKHS norm, nice discussion on this in [R2])

W2) Most of the papers framing is around the contribution of the BUV but looking at the results in Table 1 and also the sweep over eta it seems that EI seems to do a lot of heavy lifting which is IMO not well reflected in the main body of the paper.


(more minor)

W3) The SCS experiments are only run with 5 seeds which I dont think is sufficient especially since the results in Figure 2 (left) may not be statistically significant.

(Minor IMO but maybe more critical for other in the safeBO community) W4) Most of SafeBO papers go deeper into the theory of preserving safety with high probability under specific assumptions on the class of functions (see the ref stated above). This is IMO not super critical as these assumptions can usually not be varified in practice.

(Minor) W5) I believe there is some related work on goal-directed safe exploration missing such as research directions coming out of algorithms like GoSafe (and GoSafeOpt). Also, constrained BO papers should be cited here as they also in form mirror to some extent the construction of the BUV, ie, here usually the probability of feasibly is multiplied by eg the expected improvement.


[R1] Berkenkamp, F., Krause, A. & Schoellig, A.P. Bayesian optimization with safety constraints: safe and automatic parameter tuning in robotics. Machine Learning 112, 3713–3747 (2023).

[R2] Fiedler, Christian, et al. "On safety in safe Bayesian optimization." TMLR (2024).

[R3] Gardner, Jacob R., et al. "Bayesian optimization with inequality constraints." Icml. Vol. 2014. 2014.

Questions for authors

Q1) How do the result look if one uses consistent beta_gs for all methods? Also interesting could be a sweep over different beta_gs and how the different methods behave in different regeimes of beta_g

Q2) In your acquisition function you normalize both contributing factors, ie Q and EI. Can you give some intuition on how sensitive this is to choosing the quantiles for the normalization? Also, dont you with this somehow loose the scale that can also be important? Essentially this can significantly increase the weight put on EI even if "actual" EI is very flat.

8 comments
AuthorsOpenReview ↗

Response to Reviewer 35Eh

We thank the reviewer for the insightful and constructive comments, which have been very helpful in improving our work.

W1) Fair comparison under matched safety settings

We thank the reviewer for raising this important concern. We agree that the safety confidence multiplier βg\beta_g directly determines the width of the safety confidence bounds and therefore affects the certified-safe set, admissible queries, and the observed performance–safety trade-off. Using different safety confidence multipliers for different methods within the same benchmark affects the fairness of the comparison.

In the original experiments, we used method-specific safety confidence multipliers to obtain reasonable performance–safety trade-offs for the respective implementations. However, we acknowledge that this did not provide a fully controlled comparison of acquisition strategies, because differences in safety-certification conservativeness could affect the results.

To address this concern, we conducted additional experiments under matched settings. Within each benchmark, all methods share the same GP model specification, initial observations, βg\beta_g, safety certification rule, repeat-query policy, query budget, and random seeds. We adapted the methods to these shared rules and regenerated the optimization-performance and safety comparisons in Figures 2, 3, and 4, as well as the ablation results in Table 1. The revised results broadly preserve the qualitative trends of the original experiments. The shared safety confidence multipliers are reported below.

MethodSCSQuadrotorLINCS
BUV-SafeBO4163
ISE4163
M-SafeOpt4163
M-SafeUCB4163
SafeOpt4163
ITL4163
HdSafeBO4163

We have incorporated these controlled experiments and their settings into the revised manuscript to provide a fairer assessment of the contributions of different acquisition strategies. We thank the reviewer for identifying this issue and helping us improve the rigor of our experimental comparisons.

W2) The contribution of EI and its role in BUV-SafeBO

We thank the reviewer for this insightful comment. We agree that EI makes a substantial contribution to the optimization performance of BUV-SafeBO, and that this contribution was not sufficiently emphasized in the original manuscript. We have clarified its role in the revised manuscript.

Our acquisition combines two complementary sources of value. EI evaluates the expected immediate objective improvement at a candidate within the current certified-safe set, accounting for both the posterior mean and uncertainty of the objective. In contrast, the BUV-weighted information score Qt(x)Q_t(x) measures how informative a certified-safe query is about the safety margins at the first uncertified intensities of selected high-BUV configurations. BUV weights this information according to the optimistic objective gains that the corresponding boundary extensions may unlock.

Thus, EI promotes immediate objective improvement within the current certified-safe region, while Qt(x)Q_t(x) prioritizes information about potentially valuable future safe expansions. EI alone does not explicitly value information about unresolved safety boundaries, whereas the information term alone does not explicitly account for the immediate objective improvement available within the certified-safe set. Their combination is intended to balance these complementary objectives.

We have made these complementary roles more explicit in Section 3.3 and expanded the discussion of the ablation results in Section 4.4. The η\eta sensitivity analysis in Appendix C.3.3 further demonstrates the substantial contribution of EI.

W3) Number of random seeds

We have expanded the revised SCS experiments to 10 random seeds for each of two targets, as shown in Figure 2.

W4) Scope of the safety analysis

We thank the reviewer for this comment. We agree that high-probability safety guarantees are an important aspect of Safe Bayesian Optimization, even though their underlying assumptions are often difficult to verify in practice.

Our contribution focuses on value-directed query selection under monotone safety constraints. Appendix A.3 provides a conditional safety analysis: assuming that the true safety function is nonincreasing in intensity and that the initial safety certificates are valid, every query selected from the certified-safe set is safe on the joint lower-confidence-bound event over all candidates and iterations. If a separate confidence analysis establishes that this event holds with probability at least 1−δ1-\delta, and the initial certificates are valid almost surely, the same probability bound applies to the safety of all executed queries.

We have clarified this scope in Section 3.1, immediately after defining the certified-safe set, and added an explicit reference to Appendix A.3.

AuthorsOpenReview ↗

Response to Reviewer 35Eh

W5) Additional related work

We have added GoSafe (Baumann et al., 2021), GoSafeOpt (Sukhija et al., 2023), and representative constrained BO studies (Gardner et al., 2014; Hernández-Lobato et al., 2016) to the Related Work section. We also discuss their connections to BUV and clarify the distinction between candidate-level feasibility-weighted improvement and BUV’s configuration-level valuation of potential safety-boundary extensions.

Q1) Shared safety confidence multipliers and sensitivity to βg\beta_g

The revised Figures 2, 3, and 4 report comparisons in which all methods use the same βg\beta_g and safety certification rule within each benchmark. BUV-SafeBO achieves competitive optimization performance across the benchmarks, while maintaining a high safe-query rate.

To further examine the effect of the safety confidence multiplier, we conducted a sensitivity analysis of BUV-SafeBO on SCS with βg∈{1,2,3,4,5,6}\beta_g \in \{1,2,3,4,5,6\}, holding the remaining parameters, initial observations, targets, and seeds fixed. The results are reported in Appendix Table 6 and reproduced below as mean ± SEM.

βg\beta_gProgress AUCFinal progressSafe queries (%)Cumulative violation
10.7267 ± 0.02490.8388 ± 0.010792.0312 ± 0.16540.9214 ± 0.0503
20.7358 ± 0.03160.8517 ± 0.028796.3750 ± 0.11270.3159 ± 0.0190
30.7755 ± 0.03500.8423 ± 0.038798.4271 ± 0.16760.1204 ± 0.0140
40.8080 ± 0.02290.9077 ± 0.033799.0417 ± 0.18070.0641 ± 0.0171
50.7328 ± 0.01030.8268 ± 0.023899.7188 ± 0.06510.0206 ± 0.0073
60.7194 ± 0.02410.8238 ± 0.024699.8438 ± 0.04770.0067 ± 0.0004

As βg\beta_g increases, the mean safe-query rate increases and cumulative violation decreases, consistent with more conservative safety certification. Optimization performance, however, is non-monotonic: mean AUC increases up to βg=4\beta_g=4 and then decreases, while mean final progress also reaches its maximum at βg=4\beta_g=4. One possible explanation is that a small βg\beta_g may allocate queries to candidates that are actually unsafe, whereas a large βg\beta_g may delay access to valuable regions.

We additionally reran the Quadrotor comparison with a shared βg=4\beta_g=4 for all methods, as reported in Appendix C.4. Compared with the main setting of βg=16\beta_g=16, optimization progress approaches the optimum in fewer evaluations, accompanied by lower safe-query rates.

Q2) Quantile normalization and amplification of nearly flat EI

We use quantile normalization to balance the numerical ranges of QQ and EI while reducing the influence of extreme scores on the scaling denominators. The 5th–95th percentile range is a practical design choice, rather than a theoretically optimal one.

We agree that independently normalizing the two terms suppresses absolute scale information and can amplify small but nonzero variations in EI. Specifically, let

Ds=max⁡(q0.95(s)−q0.05(s),ϵ).D_s=\max\left(q_{0.95}(s)-q_{0.05}(s),\epsilon\right). For candidates where neither normalized term is clipped, multiplying the acquisition by DQD_Q and dropping candidate-independent constants gives the equivalent ranking score

Q(x)+ηDQDEIEI(x).Q(x)+\eta\frac{D_Q}{D_{\mathrm{EI}}}\mathrm{EI}(x).

Thus, a small EI span can increase its effective coefficient. Exactly constant EI maps to zero, but nearly constant EI need not have negligible influence.

To assess practical sensitivity, we evaluated three percentile ranges, raw scores clipped to [0,1][0,1], and raw scores without normalization or clipping. Within each benchmark setting, all variants use η=1\eta=1, the same GP specification, initial observations, certification rule, evaluation budget, and 20 paired seeds.

Normalization sensitivity: AUC₈₀₀, mean ± sample SD across seeds.

ScalingQuadrotor, βg=16\beta_g=16Quadrotor, βg=4\beta_g=4LINCS, βg=3\beta_g=3
5%–95% (default)0.8919 ± 0.01420.9149 ± 0.02550.7581 ± 0.0186
1%–99%0.8849 ± 0.01670.9137 ± 0.02210.7809 ± 0.0196
10%–90%0.8855 ± 0.01540.9139 ± 0.02300.7490 ± 0.0188
Raw scores clipped to [0,1][0,1]0.8816 ± 0.02180.9641 ± 0.01030.7092 ± 0.0158
Raw, no clipping0.8816 ± 0.02180.9641 ± 0.01030.7098 ± 0.0160

Changing the percentile endpoints produces relatively small mean AUC differences on Quadrotor, but larger differences on LINCS. Removing normalization also has task-dependent effects: relative to the default, raw scores without clipping improve mean AUC on Quadrotor with βg=4\beta_g=4, but reduce it on LINCS. The default percentile range is therefore not uniformly optimal.

These results show that the practical effects of normalization depend on the task. We therefore interpret quantile normalization as a heuristic for balancing relative score variation: it intentionally reduces absolute-scale differences, but can also increase the influence of nearly flat EI, as the reviewer points out.

Dear Reviewer 35Eh,

Thank you very much for your help with reviewing for ICLR'27!

The authors have now responded to your reviews. As the discussion period is approaching the end, I'm wondering if you could have a look at the authors' responses and see whether they have addressed your concerns. And could you also leave a comment to let the authors know your thoughts? If any concerns remain, please explain which ones have not been addressed and why.

Warm regards, Your AC

Reviewer_1kPkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies safe Bayesian optimization (SafeBO) under a structured setting where each input consists of a system configuration and an ordered intensity variable, with safety risk increasing monotonically with the intensity. This additional structure makes the value of safe exploration more tractable to quantify. The authors introduce Boundary Unlocking Value (BUV), which combines posterior uncertainty about the safety boundary with optimistic objective improvements to quantify the exploration potential of each configuration. Building on this concept, they develop BUV-SafeBO, an acquisition strategy that integrates BUV with, MI and EI. Experiments across multiple structured safe optimization benchmarks demonstrate that BUV-SafeBO achieves competitive optimization performance while maintaining high safe-query rates.

Critical strengths
  1. The motivation of the paper is clear. It addresses the challenge of selecting the most beneficial configuration among multiple candidates after the M-SafeOpt trimming step.

  2. Clean use of the monotone structure. The argument in A.1.1 is concise: monotonicity reduces the 2^K possible safety patterns to K+1 prefixes, so a single scalar b fully characterizes each configuration’s feasible set.

  3. The proposed BUV is explainable and easy to calculate.

Critical weaknesses

The methodological contribution is relatively incremental, as most of the components are based on existing techniques. BUV essentially computes the expected UCB-based gain over a sampled distribution of safety boundaries. Moreover, the paper lacks theoretical guarantees or analysis to justify the effectiveness of the proposed acquisition strategy.

My main concern comes from the experiment side. In SafeBO, the safe region is defined by the model prediction and beta_g. While in application the model prior is tuned from initial observation. That results in the safe-query rate is tunable and mostly depends on how you choose the hyper-parameter. The paper should test on robustness of hyperparameter, such as given an incorrect prior, and check how much would the safe rate drop.

In table 2, different methods are assigned with different beta_g, while the cumulative-violation ordering on SCS closely tracks beta_g. The the safe-rate and violation curves overlap in figure 3, where beta_g is the same for all methods.

Questions for authors

See weakness section.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1kPk

We sincerely thank the reviewer for the constructive comments, which have helped strengthen the methodological rigor of our work.

W1) Methodological contribution and theoretical analysis

BUV can be interpreted as the expected optimistic objective gain under an estimated safety-boundary distribution. Our contribution is to exploit the monotone configuration–intensity structure to quantify the objective opportunities that different boundary realizations may unlock, and to use this configuration-level value to guide safe information acquisition.

We have added a tail-probability decomposition and its proof in the new Appendix A.1.4. Fix the objective posterior and a nonempty certified prefix ending at bmb_m on an ordered grid b1<⋯<bKb_1<\cdots<b_K. Let Uj=Ut(c,bj)U_j=U_t(c,b_j) denote the maximum optimistic utility over the intensity prefix ending at bjb_j, and define δj=Uj−Uj−1≥0\delta_j=U_j-U_{j-1}\geq0. Then

BUV(c;π)=∑j=m+1KδjPr⁡π(B≥bj).\mathrm{BUV}(c;\pi)=\sum_{j=m+1}^{K}\delta_j\Pr_{\pi}(B\geq b_j).

Thus, BUV weights each incremental optimistic utility gain by the probability that the safety boundary reaches that intensity. This immediately implies monotonicity under first-order stochastic dominance.

Moreover, if two boundary distributions satisfy

max⁡j>m∣Pr⁡π1(B≥bj)−Pr⁡π2(B≥bj)∣≤ε,\max_{j>m}\left|\Pr_{\pi_1}(B\geq b_j)-\Pr_{\pi_2}(B\geq b_j)\right|\leq\varepsilon,

then

∣BUV(c;π1)−BUV(c;π2)∣≤ε(UK−Um).\left|\mathrm{BUV}(c;\pi_1)-\mathrm{BUV}(c;\pi_2)\right|\leq\varepsilon(U_K-U_m).

This bound characterizes the sensitivity of the BUV score to errors in boundary probabilities. Together, these results formalize the monotonicity and stability of BUV’s value assessment.

W2) Robustness to safety-model assumptions and confidence multipliers

Safety performance depends on both βg\beta_g and the safety GP assumptions. We conducted analyses on SCS benchmark.

First, we perturbed the lengthscale of the safety GP’s random Fourier feature component to 0.5ℓ00.5\ell_0, ℓ0\ell_0, and 2ℓ02\ell_0. We fixed βg\beta_g, the objective model, initial observations, and all remaining settings.

Safety lengthscaleMean safe-query rate (%)Mean AUC800Mean cumulative violation
0.5ℓ00.5\ell_099.35420.74100.0474
ℓ0\ell_099.04170.80800.0641
2ℓ02\ell_099.38540.73540.0438

Neither perturbation reduced the mean safe-query rate or increased mean cumulative violation relative to the nominal setting. However, AUC decreased, indicating sensitivity in optimization efficiency despite the high safe-query rates.

Second, we swept βg\beta_g from 1 to 6 while fixing the model specification and other settings. Results are mean ± SEM across seeds after averaging targets within each seed.

βg\beta_gProgress AUCFinal progressSafe queries (%)Cumulative violation
10.7267 ± 0.02490.8388 ± 0.010792.0312 ± 0.16540.9214 ± 0.0503
20.7358 ± 0.03160.8517 ± 0.028796.3750 ± 0.11270.3159 ± 0.0190
30.7755 ± 0.03500.8423 ± 0.038798.4271 ± 0.16760.1204 ± 0.0140
40.8080 ± 0.02290.9077 ± 0.033799.0417 ± 0.18070.0641 ± 0.0171
50.7328 ± 0.01030.8268 ± 0.023899.7188 ± 0.06510.0206 ± 0.0073
60.7194 ± 0.02410.8238 ± 0.024699.8438 ± 0.04770.0067 ± 0.0004

Larger βg\beta_g improves the observed safety metrics, while optimization performance is non-monotonic, with both mean AUC and final progress highest at βg=4\beta_g=4. Increasing certification conservativeness therefore does not necessarily improve optimization performance.

W3) Fair comparison and overlapping safety curves

We acknowledge that using method-specific safety confidence multipliers in the original experiments introduced differences in certification conservativeness and confounded the comparison of acquisition strategies.

We have rerun comparisons under matched settings. Within each benchmark, all methods share the GP specification, initial observations, βg\beta_g, safety certification rule, repeat-query policy, query budget, and seeds. We adapted the methods to these shared rules and regenerated Figures 2, 3, and 4. The revised results broadly preserve the original qualitative trends.

MethodSCSQuadrotorLINCS
BUV-SafeBO4163
ISE4163
M-SafeOpt4163
M-SafeUCB4163
SafeOpt4163
ITL4163
HdSafeBO4163

The conservative βg=16\beta_g=16 setting in the original Quadrotor comparison led to near-saturated safe-query rates, making differences difficult to distinguish. We therefore added a controlled comparison with a shared βg=4\beta_g=4 in Appendix C.4. Safe-query rates are less saturated, and cumulative-violation curves show clearer differences. Optimization progress approaches the optimum in fewer evaluations, accompanied by lower safe-query rates.

These controlled comparisons and supplementary experiments have been incorporated into the revised manuscript.

Dear Reviewer 1kPk,

Thank you very much for your help with reviewing for ICLR'27!

The authors have now responded to your reviews. As the discussion period is approaching the end, I'm wondering if you could have a look at the authors' responses and see whether they have addressed your concerns. And could you also leave a comment to let the authors know your thoughts? If any concerns remain, please explain which ones have not been addressed and why.

Warm regards, Your AC

Thank you for the detailed response and additional experiments.

The additional results help clarify the properties of BUV and the sensitivity to the safety confidence multiplier. However, my main concerns remain. First, the theoretical results mainly characterize the BUV score itself, rather than providing guarantees for the proposed optimization strategy. Second, the sensitivity experiments confirm that the safe-query rate is strongly influenced by the choice of β_g, while the robustness evaluation under model misspecification remains limited.

I acknowledge the improvements made in the revision, but I do not find them sufficient to fully address my concerns. I will maintain my original score.

Thank you for your further feedback. However, there are still some aspects of the three concerns that we do not fully understand, and we would appreciate further clarification.

First, regarding the theoretical analysis, you noted that our theoretical results mainly characterize BUV itself rather than providing guarantees for the optimization strategy. However, we consider BUV to be the central component of our optimization strategy, rather than an auxiliary metric independent of it. BUV directly determines how the unresolved safety boundaries are evaluated according to their potential objective value, and further influences the information acquisition strategy. Therefore, the decomposition, monotonicity, and stability results we added are intended to characterize the mathematical properties of the core mechanism of our proposed optimization strategy. We are not entirely clear what kind of guarantees for the optimization strategy you specifically refer to. Do you primarily mean theoretical guarantees for the complete algorithm?

Second, regarding the sensitivity to βg\beta_g, we consider the dependence of the safe-query rate on βg\beta_g to be an expected consequence of its design, since this parameter directly controls the conservativeness of LCB-based safety certification. A smaller βg\beta_g leads to a larger certified-safe set, whereas a larger βg\beta_g results in more conservative safety certification. Therefore, it is natural for the safe-query rate to increase with βg\beta_g. Our sensitivity analysis is intended to quantify this performance–safety trade-off. Furthermore, we have conducted controlled comparisons using a shared βg\beta_g and matched evaluation protocols across methods to minimize the influence of different safety-certification parameters on the algorithmic comparisons.

Finally, regarding model misspecification, our additional lengthscale experiments directly examine the behavior of the method when the smoothness assumption of the safety GP is perturbed. Under both 0.5ℓ00.5\ell_0 and 2ℓ02\ell_0, the safe-query rate remains above 99%. Although optimization efficiency decreases, safety remains stable. We believe that some degradation in predictive accuracy or optimization performance under model misspecification is normal and expected. The purpose of robustness evaluation is not to require performance to remain completely unaffected, but rather to examine whether such misspecification leads to substantial degradation in safety.

Thank you again. We would appreciate any further clarification you could provide.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.