acceptodds
Under review as a conference paper at ICLR 2027

Decision-Aware Safe Causal Bayesian Optimization via Structural Value of Information

Abstract

Causal Bayesian optimization seeks interventions that improve the decision objective under limited information. When the causal graph is unknown, the interplay among structural uncertainty, intervention effects and execution risk further complicates intervention selection. However, reducing structural uncertainty does not necessarily benefit downstream decision-making; the value of structural information depends on its impact on intervention decisions. In this paper, we propose a decision-aware and safe sequential causal Bayesian optimization framework that explicitly evaluates the value of structural information through its impact on downstream intervention decisions. We define Structural Value of Information (SVI), which quantifies the expected improvement in downstream decision value after executing an intervention and observing its outcome, and combine it with intervention cost to form a decision-oriented acquisition criterion. To improve computational efficiency, we perform decision-aware clustering of graph hypotheses based on their induced objective and risk maps, and aggregate each cluster's posterior mass onto a representative decision map. We independently evaluate candidate interventions using the full posterior predictive mixture and filter them based on the probability of violating the joint constraints to ensure decision safety and feasibility. Through synthetic experiments and real-world case studies, we demonstrate the effectiveness of the proposed method in balancing optimization performance, structural exploration and safe execution.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Causal Bayesian Optimization, Causal Reasoning, Value of Information, Safe Optimization, Decision-Making under Uncertainty

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

DASCBO is a framework selecting effective interventions under uncertain causal structure and safety constrains. Its main contributions is the Structure Value of Information, which learned the expected improvements in the intervention sets. The proposed approach leverages this feature with grouping similar decision consequences to reduce the computation costs and increase efficiency. Across the four benchmarks, DASCBO claims to achieve the best final optimal object value and perfect empirical feasibility.

Critical strengths
  1. Decision-relevant structural exploration. SVI provides a principled criterion for valuing causal information through its expected effect on intervention decisions. The result that SVI vanishes when all graph hypotheses share an optimal action makes this distinction precise by showing that informative experiments need not be useful experiments.

  2. Principled integration of efficiency and safety. The framework compresses decision maps while retaining full likelihood evidence, and independently certifies risk using fresh samples from the full posterior predictive model. Explicit approximation bounds and model-discrepancy conditions clarify what compression preserves and when safety guarantees transfer to the physical system.

Critical weaknesses
  1. Substantive discrepancies between the stated SCM and safety formulation and the released implementation.

    Section 2.1 defines interventions as replacing only the targeted structural equations. However, the Health implementation introduces intervention-family-specific offsets that also change nonintervened variables. Within the same fitted hypothesis, the expected panel mean of root node B changes from 1502.852 under do(C) to 1501.404 under do(C,G), although B has no parents and is not intervened upon. This change arises from the offsets, rather than sampling noise or posterior updating. Unless an additional intervention-family context is explicitly modeled, this behavior conflicts with the stated fixed-SCM semantics and undermines the causal interpretation of the graph hypotheses.

    Furthermore, the implementation certifies violations of a ten-sample mean, whereas the paper formulates risk over SCM constraint variables without specifying this aggregation. The released defaults also did not reproduce the claimed 100% Health feasibility: one complete run achieved 98/99 feasible interventions, and two of ten configurations violated the constraint on their first intervention. These findings question whether the experiments instantiate the stated formulation and reproduce the reported empirical result; they do not, by themselves, disprove the posterior-predictive certification theorem.

  2. Incomplete experimental reproducibility and insufficient evidence for the proposed components’ added value.

    The paper does not fully specify the SCMs underlying all four benchmarks, including their structural equations, noise distributions, and fitted mechanisms. The supplied code covers only Healthcare and omits the baseline implementations. It therefore remains difficult to verify how shared initialization data and evaluation protocols were enforced across separate implementations or to reproduce the full experimental comparison.

    Additionally, comparisons between complete algorithms do not isolate the contributions of SVI and decision-map compression. Controlled ablations are needed to establish whether these components account for the reported improvements; cumulative posterior movement alone does not demonstrate decision-relevant learning. Numerical summaries at matched budgets, with uncertainty estimates—including final feasible objectives and complementary measures such as GAP [1] and PA-GAP [2]—would strengthen the assessment. Together, the missing experimental artifacts and component-level evidence materially weaken confidence in the claimed advantages.

[1] Virginia Aglietti, Neil Dhir, Javier González, and Theodoros Damoulas. Dynamic Causal Bayesian Optimization. NeurIPS, 2021.

[2] Chenfeng Huang, Thuy T. Le, Zixuan Ma, and Hien Tran. Causal Bayesian Optimization: Foundations, Methods, and Applications. Transactions on Machine Learning Research, 2026.

Questions for authors
  1. Benchmark specification, especially Protein-Signaling: Can the authors provide the complete SCM for each benchmark, including its graph, structural equations, noise distributions, intervention domains, constraint thresholds, and fitting procedures? The fitted SCM used in the original cCBO Protein-Signaling experiments is reported as unreleased. Did the authors obtain that model, reconstruct an SCM from the Sachs data, or use a different model?

  2. Consistency with fixed-SCM intervention semantics: In the released Health implementation, intervention-family-specific offsets change nonintervened variables. Within the same fitted hypothesis, root B’s expected panel mean changes from 1502.852 under do(C) to 1501.404 under do(C,G), although B has no parents and is not intervened upon. How is this consistent with Section 2.1’s intervention semantics? Is an additional context variable intended, and if so, how is it represented in the graph and theoretical formulation?

  3. Safety definition and reported feasibility: Does the method certify violations of individual SCM constraint variables or violations of a ten-sample mean, as implemented in the Health code? Please clarify how the implemented event corresponds to Section 2.1. Furthermore, a released-default run with initial design 1 and algorithm seed 1001 achieved 98/99 feasible interventions, and two of ten configurations violated the constraint on their first intervention. Which configurations produced the reported 100% Health feasibility, and what explains this discrepancy?

  4. Contribution of SVI and compression: Can the authors provide controlled ablations with SVI disabled, SVI replaced by an information-gain criterion, and compression disabled, while keeping the remaining experimental settings fixed? How much does each component contribute to feasible-objective improvement, empirical feasibility, and runtime?

  5. Quantitative performance comparison: Can the authors report numerical summaries at matched budgets, with uncertainty estimates, alongside the existing curves—for example, final best-feasible objectives and complementary GAP/PA-GAP scores? If normalized metrics are used, how are reference values, infeasible evaluations, abstentions, and runs without an initially feasible point handled?

13 comments

We sincerely thank the reviewer for recognizing the novelty and readability of our work and for carefully examining both its theoretical development and implementation. Your comments are valuable for strengthening the theory and making the method more readily reproducible.

1. Q1 / W2: Complete Benchmark Specifications and the Protein-Signaling Setup

We thank the reviewer for emphasizing the importance of providing complete SCMs, causal graphs, and benchmark settings for accuracy and reproducibility. In response, we have compiled the specifications of all four benchmarks, including their causal graphs, structural equations, noise distributions, intervention domains, constraint thresholds, and experimental settings in the revised manuscript.

For Protein-Signaling, we constructed the SCM using the publicly available protein-signaling data of Sachs et al. (2005), with reference to Huang et al., (2026), to reflect the underlying pathway structure.

Here, we present the detailed SCM definitions and experimental settings for the Health and Protein-Signaling benchmarks. The corresponding specifications for the Synthetic and Honda Energy benchmarks will be provided in the appendix of the revised manuscript. We write σ(z)=(1+e−z)−1\sigma(z)=(1+e^{-z})^{-1}, [z]+=max⁡(z,0)[z]_+=\max(z,0), and TN(a,b)\mathrm{TN}(a,b) for a standard normal distribution truncated to [a,b][a,b].

A. Health

The ten nodes are Age, BMR, Cal, Height, Weight, BMI, Statin, Aspirin, Cancer, and PSA. Their non-root parent sets are

\mathrm{Pa}(\mathrm{Weight})&=\{\mathrm{Age,BMR,Cal,Height}\},\\ \mathrm{Pa}(\mathrm{BMI})&=\{\mathrm{Weight,Height}\},\\ \mathrm{Pa}(\mathrm{Aspirin})&=\{\mathrm{Age,BMI}\},\\ \mathrm{Pa}(\mathrm{Cancer})&=\{\mathrm{Age,BMI,Statin,Aspirin}\},\\ \mathrm{Pa}(\mathrm{PSA})&=\{\mathrm{Age,BMI,Statin,Aspirin,Cancer}\}. \end{aligned}$$ The graph contains ten nodes and seventeen edges. The root nodes are generated independently: $\mathrm{Age}\sim U(55,75)$, $\mathrm{BMR}=1500+10\mathrm{TN}(-1,2)$, $\mathrm{Cal}\sim U(-100,100)$, $\mathrm{Height}=175+10\mathrm{TN}(-0.5,0.5)$, and $\mathrm{Statin}\sim U(0,1)$. The non-root mechanisms are $$\begin{aligned} \mathrm{Weight}&=(\mathrm{BMR}+6.8\mathrm{Age}-5\mathrm{Height})/13.7+150\mathrm{Cal}/7716,\\ \mathrm{BMI}&=\mathrm{Weight}/(\mathrm{Height}/100)^2,\\ \mathrm{Aspirin}&=\sigma(-8+0.10\mathrm{Age}+0.03\mathrm{BMI}),\\ \mathrm{Cancer}&=\sigma(2.2-0.05\mathrm{Age}+0.01\mathrm{BMI}-0.04\mathrm{Statin}+0.02\mathrm{Aspirin}),\\ \mathrm{PSA}&=6.8+0.04\mathrm{Age}-0.15\mathrm{BMI}-0.60\mathrm{Statin} +0.55\mathrm{Aspirin}+\mathrm{Cancer}+\varepsilon_{\mathrm{PSA}}, \end{aligned}$$ Here $\varepsilon_{\mathrm{PSA}}\sim\mathcal N(0,0.4^2)$ and is independent of the root disturbances. The objective is to minimize PSA subject to $\mathrm{BMI}<25$. The intervention domains are $\mathrm{Cal}\in[-400,400]$ and $\mathrm{Statin},\mathrm{Aspirin}\in[0,1]$. The permitted intervention sets are $\{\mathrm{Cal}\}$, $\{\mathrm{Cal,Statin}\}$, $\{\mathrm{Cal,Aspirin}\}$, and $\{\mathrm{Cal,Statin,Aspirin}\}$. **B. Protein-Signaling** The eight nodes are PKC, PKA, Jnk, P38, Akt, Raf, Mek, and Erk. Their parent sets are $$\begin{aligned} \mathrm{Pa}(\mathrm{PKC})&=\varnothing,&\mathrm{Pa}(\mathrm{PKA})&=\{\mathrm{PKC}\},\\ \mathrm{Pa}(\mathrm{Jnk})&=\mathrm{Pa}(\mathrm{P38})=\mathrm{Pa}(\mathrm{Raf})=\{\mathrm{PKC,PKA}\},\\ \mathrm{Pa}(\mathrm{Akt})&=\{\mathrm{PKA}\},&\mathrm{Pa}(\mathrm{Mek})&=\{\mathrm{PKC,PKA,Raf}\},\\ \mathrm{Pa}(\mathrm{Erk})&=\{\mathrm{PKA,Mek,Akt,Raf}\}. \end{aligned}$$ The graph has fifteen edges. The root variable is $\mathrm{PKC}\sim U(1,106)$, and the intermediate structural mechanisms are $$\begin{aligned} \mathrm{PKA}&=554.390731+0.841153\mathrm{PKC},\\ \mathrm{Jnk}&=52.953603-0.764801\mathrm{PKC}-0.005306\mathrm{PKA},\\ \mathrm{P38}&=15.144328+1.234783\mathrm{PKC}+0.000591\mathrm{PKA},\\ \mathrm{Akt}&=-31.110747+0.128905\mathrm{PKA},\\ \mathrm{Raf}&=62.199046-0.177745\mathrm{PKC}-0.000379\mathrm{PKA},\\ \mathrm{Mek}&=-1.090275+0.039662\mathrm{PKC}-0.000652\mathrm{PKA}+0.520845\mathrm{Raf}. \end{aligned}$$ The objective mechanism is $$\begin{aligned} \mathrm{Erk}={}&15.4025+3.0\times10^{-5}(\mathrm{PKA}-700)^2 +3.0\times10^{-4}(\mathrm{Mek}-44.098)^2\\ &+1.0\times10^{-5}(\mathrm{Akt}-100)^2 +5.0\times10^{-2}(\mathrm{Raf}-44)^2\\ &+3.0\times10^{-5}(\mathrm{PKA}-700)(\mathrm{Mek}-44.098) +\varepsilon_{\mathrm{Erk}}, \end{aligned}$$ where $\varepsilon_{\mathrm{Erk}}\sim\mathcal N(0,0.4^2)$. The objective is to minimize Erk subject to $\mathrm{Raf}\ge45$ and $\mathrm{PKA}>255.4$. The intervention domains are $\mathrm{PKC}\in[1,105]$, $\mathrm{PKA}\in[439.2449,4491.5]$, $\mathrm{Mek}\in[0.5,389.5]$, and $\mathrm{Akt}\in[1.2,3555.5]$. The permitted intervention sets are $\{\mathrm{PKC}\}$, $\{\mathrm{PKA}\}$, $\{\mathrm{Mek}\}$, $\{\mathrm{PKC,PKA}\}$, $\{\mathrm{PKC,Mek}\}$, and $\{\mathrm{PKA,Mek}\}$.

2. Q2 / W1: Consistency with Fixed-SCM Intervention Semantics

We thank the reviewer for carefully inspecting the Healthcare implementation and identifying a potential inconsistency with the fixed structural causal model (SCM) intervention semantics in Section 2.1. We agree that, under a fixed SCM hypothesis hh, a standard surgical intervention replaces only the structural equations of the intervened variables; neither the mechanisms nor the exogenous noise distributions of non-intervened nodes should depend on the chosen intervention set.

Our inspection found that the original Healthcare surrogate model introduced intervention-family-specific residual anchoring offsets when fitting finite observation batches. Some offsets were applied to the predictive mechanisms of non-intervened nodes, causing their predictive distributions to depend on the intervention set even within the same fitted hypothesis. We thank the reviewer for identifying this issue. We have removed these offsets from the revised code, reconstructed the Healthcare SCM hypothesis library with fixed structural mechanisms, and re-evaluated performance under the revised setup. We emphasize that this was an implementation-specific deviation in the Healthcare surrogate, not an issue with the mathematical definitions or conditional theoretical results for SVI, decision-map compression, or posterior-predictive certification. We have also conducted preliminary validation across all four benchmark tasks, including Healthcare; the results indicate that this correction does not alter the effectiveness of the proposed method or the paper's central conclusions.

Specifically, in the revised Healthcare generative environment, the non-intervened root node BB follows the fixed structural equation

\varepsilon_B\sim\mathcal N(0,1) \text{ truncated to }[-1,2].$$ Neither this mechanism nor its exogenous noise distribution depends on $C$, $G$, or the intervention set. Thus, in the true generative environment, $$P(B\mid \mathrm{do}(C=-80)) = P(B\mid \mathrm{do}(C=-80,G=1))$$ and both interventions have the same theoretical expectation: $$\mathbb E[B]\approx1502.296372.$$ For any fixed fitted hypothesis $h$ in the revised model library, the non-intervened variable $B$ must likewise satisfy $$P(B\mid \mathrm{do}(C=-80)) = P(B\mid \mathrm{do}(C=-80,G=1))$$ Importantly, 1502.296372 is the theoretical expectation under the true Healthcare generative environment, not a parameter estimate shared by every fitted hypothesis. We inspected the SCM hypotheses constructed under the revision: in each, $B$ has no parents, and its simulated mechanism no longer includes offsets depending on the intervention family. We also examined the values 1502.852 and 1501.404 reported by the reviewer. In the revised fixed-SCM generative environment, using the original sequential random-number interface and generating two separate ten-sample batches with the same integer seed (`seed=1`) indeed reproduces - $\operatorname{do}(C=-80)$: sample mean of $B$ = 1502.851866; - $\operatorname{do}(C=-80,G=1)$: sample mean of $B$ = 1501.403855. However, these are empirical means from separately generated finite batches, not analytical expectations under a fixed SCM. Under sequential random-number sampling, intervening on $G$ removes the need to draw its random value, shifting subsequent consumption of the random stream. Consequently, even with the same integer seed, the two interventions need not share identical exogenous noise values node by node. The finite-batch difference in the revised environment can therefore be explained by unpaired random-number streams. To verify structural invariance more directly, we performed a counterfactual paired test using shared node-specific exogenous noise. Holding the same ten draws of the exogenous noise for $B$ fixed under both interventions yielded exactly identical values for all ten observations of $B$, with a sample mean of 1502.548502 in both cases. This directly verifies that, in the examined generative mechanism, the non-intervened root node $B$ is invariant to the intervention set. We have also rerun performance comparisons on all four benchmarks under the revised experimental settings. DASCBO achieves favorable empirical feasibility and nPA-GAP results across these tasks. The numerical results, uncertainty estimates, and comparisons are presented in the quantitative-performance response below. We again thank the reviewer for this careful examination and constructive recommendation. We believe these revisions provide a stronger basis for reproducing the empirical findings and clarify the conditions under which the method and its safety certificates should be interpreted.

3. Q3 / W1: Safety Definitions and Feasibility Reporting

We thank the reviewer for highlighting ambiguities concerning safety events and batched observations in the original code and manuscript. We agree that the safety definition must be unambiguous in both theory and implementation. DASCBO certifies the joint-constraint violation event for a single SCM outcome, and we have revised the manuscript and relevant code accordingly to align certification with the chance-constraint formulation. We will further distinguish single-outcome safety events, batched feedback, and batch-mean evaluation and specify a unified observation protocol in the revision.

Each intervention returns a batch of predetermined size mm; for example, m=10m=10 in Healthcare:

V_t^{(k)}\mid(h,x_t)\overset{\mathrm{i.i.d.}}{\sim}P_h(\cdot\mid\operatorname{do}(x_t)).$$ The objective feedback is $\bar Y_t=m^{-1}\sum_kY_t^{(k)}$, which estimates the expected objective. For a fixed hypothesis $h$, the full-batch likelihood is $L_h(O_t^{(m)}\mid x_t)=\prod_kp_h(V_t^{(k)}\mid x_t)$, and the posterior-predictive batch likelihood is $$\sum_h w_{t,h}\prod_{k=1}^m p_h(V_t^{(k)}\mid x_t).$$ Thus, the entire batch shares the same hypothesis. The current Health posterior update accumulates log likelihoods over the complete batch, and hypothetical observations for SVI use the same batch unit. The safety event is defined separately as $$Z(V)=\mathbf1\\{\exists j:C_j(V)\notin\mathcal A_j\\},\qquad r_t(x)=\mathbb E_{P_t(\cdot\mid x)}[Z(V)].$$ For Health, the admissible set is $\mathrm{BMI}<25$, so $Z(V)=\mathbf1\\{\mathrm{BMI}\ge25\\}$. The certifier generates independent, single-outcome SCM rollouts from the full posterior-predictive mixture over causal graphs and mechanisms, estimates the violation rate, and uses a Hoeffding upper confidence bound to assess the risk budget. The certification sample count $N_t^{\mathrm{cert}}$ and feedback batch size $m$ are independent parameters. Decision-map compression does not replace the certification distribution; the risk pilot likewise uses the violation indicator for a single outcome. We acknowledge that the original description of "100% empirical feasibility" on Health could conflate feasibility of observed batch means with joint feasibility of individual SCM outcomes, potentially suggesting that no individual violations occurred or that all interventions satisfied the true chance constraint. In the revision, we no longer use that insufficiently qualified statement and instead consistently report the sample-level joint feasibility rate $FR_{ind}$, with corresponding manuscript and code updates. Let $\mathcal T_{\mathrm{exec}}\subseteq\\{1,\ldots,T\\}$ denote the rounds in which an intervention was actually executed, and let $m_s$ be the number of SCM outcome rows received in round $s$. We define $$FR_{ind} = \frac{\sum_{s\in T_{exec}} \sum_{k=1}^{m_s} 1\\{ \text{for all } j\in\{1,\ldots,q\}:C_j(V_s^{(k)})\in A_j\\}} {\sum_{s\in T_{exec}}m_s} = 1- \frac{\sum_{s\in T_{exec}} \sum_{k=1}^{m_s}Z(V_s^{(k)})} {\sum_{s\in T_{exec}}m_s}.$$ If every executed online batch has the same size $m$, and $N_{exec}=|T_{exec}|$, this simplifies to $$FR_{ind} = \frac{1}{mN_{exec}} \sum_{s\in T_{exec}} \sum_{k=1}^{m}\big[1-Z(V_s^{(k)})\big].$$
Reviewer_sDZ7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DASCBO, a decision-aware and safe sequential causal Bayesian optimization framework for constrained intervention selection under structural uncertainty. Its main contributions are: (i) Structural Value of Information (SVI), which values graph-posterior updates by their expected effect on the best risk-penalized intervention decision rather than by information gain alone; (ii) decision-map-based clustering that compresses graph hypotheses while preserving full likelihood evidence, with approximation-error bounds; and (iii) full-posterior-predictive safety certification with simultaneous model-based risk guarantees and a conditional physical-risk transfer argument. The method is validated on a synthetic SCM and three application-derived SCMs (healthcare, protein signaling, and building energy), comparing against CBO, CCBO, CBOU, MCBO, and CEO on best feasible objective, empirical feasibility, runtime, structural exploration, and sensitivity to library size, constraint threshold, and risk tolerance.

Critical strengths

A critical strength is the decision-aware formulation of structural information: SVI values a graph-posterior update by its expected effect on the best risk-penalized intervention decision, and the paper establishes nonnegativity, a mutual-information upper bound, and the property that SVI vanishes when all graph hypotheses share an optimal action. This is consequential because it gives a principled alternative to information-gain-driven causal exploration and directly addresses the paper’s central claim that reducing structural uncertainty need not improve decisions. A second strength is the safety architecture: full-posterior-predictive certification yields simultaneous model-based risk guarantees independent of decision-map compression, while the compression scheme preserves full likelihood evidence and comes with explicit approximation-error bounds. This matters because it lets the method combine aggressive decision-aware exploration with a separately certified execution rule, making the framework more than a heuristic combination and providing a credible basis for the empirical feasibility results.

Critical weaknesses

A critical weakness is the absence of direct ablations isolating the two core mechanisms: there is no reported “DASCBO without SVI,” no sweep over the SVI weight, and no compression ablation reporting representative counts, realized covering radii, or the actual SVI distortion. Without these, the empirical gains cannot be confidently attributed to SVI or decision-map compression rather than to safety certification, cost terms, or hyperparameter choices, which directly undermines the paper’s central methodological claim. A second critical weakness is that the “safe” guarantee is only posterior-predictive within the finite model library; physical-risk transfer requires an explicit model-discrepancy bound that is neither calibrated nor empirically stress-tested, and all validation is on simulated or application-derived SCMs. This is consequential because the paper’s title and framing promise safe execution, yet the evidence supports only model-based certification, leaving the practical safety claim substantially unsubstantiated.

Questions for authors

Clarification requests: (1) Could you provide ablations that isolate SVI and decision-map compression—e.g., DASCBO without SVI, and reporting representative counts, realized covering radii, and actual SVI distortion—since without them the empirical gains cannot be confidently attributed to the paper’s core mechanisms? (2) Could you report per-benchmark certification budgets, abstain rates, and whether baselines use the same safety certification, to assess fairness and the practical meaning of the feasibility results?

6 comments

We sincerely thank the reviewer for recognizing the value of our decision-aware structural exploration and independent risk-certification mechanism, and for providing specific comments on module attribution, the interpretation of physical risk, and certification costs.

1. W1 / Q1: Ablations and the Individual Contributions of SVI and Decision-Map Compression

We thank the reviewer for the constructive suggestion to disentangle the contributions of SVI and decision-map compression. To examine the roles of these components, we conducted four controlled ablation variants on the Healthcare SCM: full DASCBO (Full), removal of the SVI reward (w/o SVI), replacement of SVI with graph information gain (Graph IG), and removal of decision-map compression (w/o Compression).

Specifically, w/o SVI sets the SVI reward weight in the acquisition function to zero, isolating the effect of the decision-relevant information reward. Graph IG replaces SVI with graph-structure information gain, I(G;Ox)I(G;O_x), using an information-reward weight of 0.20. The w/o Compression variant retains the SVI reward but increases the representative-graph cap to 100 and sets the compression tolerance to zero, thus retaining all active graphs with distinct decision maps. Full DASCBO uses λSVI=5\lambda_{\mathrm{SVI}}=5, a representative cap of 3, and a nominal compression tolerance of 0.1. Apart from the information reward and decision-map representation being ablated, all four variants use the same Healthcare SCM, initial design, training data, model-construction settings, candidate-generation strategy, full-posterior updating procedure, and independent safety-certification protocol. The model-risk budget is α=0.20\alpha=0.20 and the number of certification samples per candidate is Ncert=1500N^{\mathrm{cert}}=1500 in all variants. The results are summarized below.

Table 1: Controlled ablation results on Healthcare (mean ± SD).

VariantBest observed feasible objectiveFRindFR_{\mathrm{ind}}nPA-GAPAlgorithm time (s)
Full DASCBO5.3913 ± 0.0422(90.31 ± 0.30)%0.4553 ± 0.062799.5 ± 2.4
w/o SVI5.4839 ± 0.1147(90.08 ± 0.28)%0.2361 ± 0.0694108.0 ± 6.4
Graph IG5.5430 ± 0.1899(89.66 ± 1.13)%0.2749 ± 0.0869154.8 ± 45.0
w/o Compression5.5308 ± 0.1234(90.14 ± 0.36)%0.2934 ± 0.0768154.2 ± 53.6

Table 2: Mean paired improvements of Full relative to the ablation variants.

Paired comparisonImprovement in best objectiveChange in FRindFR_{\mathrm{ind}} (percentage points)Improvement in nPA-GAPMean time reduction (s)
Full vs. w/o SVI+0.0926+0.227+0.21918.5
Full vs. Graph IG+0.1518+0.648+0.180455.3
Full vs. w/o Compression+0.1395+0.162+0.161854.7

The mean paired nPA-GAP improvement of Full over each of the three variants is positive, indicating that the SVI reward and decision-map compression can materially affect sequential optimization under the parameter settings examined.

To further isolate the effect of decision-map compression itself, without conflating approximation differences with posterior differences arising from divergent online trajectories, we performed a same-state diagnostic along one 99-round posterior trajectory of Full DASCBO. At each checkpoint, we fixed the posterior, candidate set, and Pilot settings and reconstructed the compressed representation using different representative-graph caps. The trajectory contains an average of 8.38 active decision graphs, while Full uses an average of 3.00 representatives, reducing the representation size by approximately 64.2%. The observed mean maximum covering radii for different caps are shown below.

Table 3: Mean maximum covering radius.

Representative-graph cap13 (Full)58No compression
Mean observed maximum covering radius0.65890.35620.17220.08910

Although Full uses a nominal tolerance of 0.1, the representative cap is reached first, giving a mean observed maximum covering radius of 0.3562 and a maximum observed radius of 0.9520 over the trajectory. To quantify the effect of compression on numerical SVI estimates, we selected eight fixed posterior checkpoints and performed three paired estimates per checkpoint using shared hypothetical-observation randomness, with the same candidate set, Pilot settings, and random streams. Under Full's representative cap of 3, the mean absolute candidate-level difference between compressed and uncompressed SVI estimates was 0.01039, with a maximum observed difference of 0.05618. These measurements directly assess compression-related approximation differences under the same posterior and random-sampling conditions.

These two types of experiments are complementary and help clarify the roles of the core components and the scope of the available evidence.

2. W2: The Boundary Between Model-Risk Certification and True Physical Safety

We thank the reviewer for highlighting the important distinction between posterior-predictive safety and safety in a real physical system. We agree that posterior-predictive certification under a finite model library does not automatically guarantee safety in the presence of model misspecification. We therefore clarify the scope of our safety guarantee, the conditions needed to transfer it to the true system, and the boundaries of the current theoretical and empirical evidence.

DASCBO's pre-execution certificate concerns rtmodel(u)r_t^{\mathrm{model}}(u), the probability, under the current full posterior-predictive mixture over graphs and mechanisms, that a single SCM outcome under candidate intervention uu violates at least one constraint. Theorem 3.3 states that, under independent certification samples, the corresponding Hoeffding upper confidence bounds, and confidence-budget allocation across rounds, with probability at least 1−δcert1-\delta_{\mathrm{cert}}, every certified candidate across all rounds satisfies

rtmodel(u)≤αt.r_t^{\mathrm{model}}(u)\leq\alpha_t.

This probability guarantee controls finite Monte Carlo certification error, not misspecification between the true SCM and the model library. Since certification always uses the full posterior-predictive mixture, each graph–mechanism hypothesis contributes to predicted risk according to its posterior weight. Certification therefore depends neither on a single maximum-a-posteriori graph nor on compressed representative graphs. Decision-map compression can thus reduce the computational cost of estimating information value without directly changing the pre-execution model-risk certification threshold.

We further emphasize that the original optimization problem in Section 2.1 concerns the true SCM chance constraint r⋆(u)≤αr^\star(u)\leq\alpha. Because the true causal structure, mechanisms, and noise distribution are unknown, the algorithm cannot directly compute this risk and instead constructs a computable certificate at the posterior-predictive level. Model-predictive risk certification and true physical-risk guarantees are distinct claims. Theorem 3.4 establishes their conditional connection: if, with probability at least 1−δmodel1-\delta_{\mathrm{model}}, a valid model-risk discrepancy bound ϵtphy\epsilon_t^{\mathrm{phy}} holds uniformly over all relevant rounds and candidate interventions, such that

r⋆(u)≤rtmodel(u)+ϵtphy,r^\star(u)\leq r_t^{\mathrm{model}}(u)+\epsilon_t^{\mathrm{phy}},

then all certified and executed interventions simultaneously satisfy

r⋆(ut)≤αt+ϵtphy,r^\star(u_t)\leq\alpha_t+\epsilon_t^{\mathrm{phy}},

with confidence at least 1−δcert−δmodel1-\delta_{\mathrm{cert}}-\delta_{\mathrm{model}}. Appendix D further identifies a uniform upper bound on the total-variation distance between the true and posterior-predictive interventional distributions as one sufficient condition for such risk transfer.

We believe that making this scope explicit does not diminish the theoretical significance or methodological value of the work. We do not claim unconditional physical safety in the absence of model-accuracy conditions. Rather, our contribution is a unified framework that accounts for posterior uncertainty over causal graphs and mechanisms, provides simultaneous finite-sample certification of posterior-predictive risk across rounds, and separates safety certification from decision-map compression. We also acknowledge that experiments based on synthetic and application-derived SCMs are insufficient to establish safety in real physical deployments; reliable estimation and calibration of model-discrepancy bounds remain subjects for further research.

We sincerely thank the reviewer for identifying this critical issue. Model-discrepancy calibration and the validation of true physical safety are important questions. We will discuss this limitation and potential approaches in the Discussion section of the revised manuscript and continue to investigate model-misspecification assessment, safety-risk calibration, and real-system validation in future work to improve the method's practical applicability and safety.

3. Q2: Per-Benchmark Certification Budgets, Abstention Rates, and Baseline Fairness

We thank the reviewer for this specific and important suggestion. We agree that reporting empirical feasibility only for executed interventions is insufficient to assess the benefits and costs of safe optimization. Pre-execution risk certification incurs additional model-simulation costs and may lead to abstention when no candidate passes certification. To clarify the practical meaning of feasibility comparisons, we provide per-benchmark risk thresholds, certification sample budgets, confidence budgets, candidate-set sizes, execution and abstention counts, and sample-level joint feasibility rates, and clarify whether baselines employ the same pre-execution safety certificate as DASCBO.

DASCBO's certification budget is determined by four quantities: the posterior-predictive joint-violation probability threshold αt\alpha_t, the number of certification samples per candidate NtcertN_t^{\mathrm{cert}}, the per-round certification failure-probability budget δtcert\delta_t^{\mathrm{cert}}, and the number of candidates nt=∣Ut∣n_t=|\mathcal U_t|. Once the candidate set is fixed, the algorithm draws fresh, independent single-outcome SCM rollouts for each candidate uu from the full, uncompressed posterior-predictive mixture over graphs and mechanisms, and records whether at least one constraint is violated. Define

\sum_{i=1}^{N_t^{\mathrm{cert}}} \mathbf 1\\!\left\\{\text{at least one constraint is violated in single SCM rollout }i\right\\},$$ $$U_t^{\mathrm{risk}}(u) =\min\\!\left\\{1,\widehat r_t(u)+ \sqrt{\frac{\log\\!\left(n_t/\delta_t^{\mathrm{cert}}\right)}{2N_t^{\mathrm{cert}}}}\right\\}, \qquad \mathcal S_t=\left\\{u\in\mathcal U_t:U_t^{\mathrm{risk}}(u)\leq\alpha_t\right\\}.$$ Indexing online rounds from $t=1$, we use $$\delta_t^{\mathrm{cert}}=\frac{6\delta_{\mathrm{cert}}}{\pi^2t^2}, \qquad \sum_{t=1}^{\infty}\delta_t^{\mathrm{cert}}=\delta_{\mathrm{cert}}.$$ Here, $\alpha_t$ is the permissible model-based joint-violation probability, and $\delta_{\mathrm{cert}}=0.05$ is the total confidence-failure budget for the finite Monte Carlo certification bounds. The algorithm executes an action only if $\mathcal S_t$ is nonempty; if it is empty and there is no independently validated fallback action, the algorithm abstains. Abstention produces no new environmental feedback but still consumes one online decision opportunity. Decision-map compression affects information-value approximation and candidate ranking, not full-posterior-predictive certification. To make certification costs auditable, Tables 4 and 5 report one complete run for each of the four benchmarks under its specified revised configuration. Each run includes $T=99$ online decision opportunities, excluding the initial design (trial 0). **Table 4: Certification budgets and model-simulation costs for one complete run on each benchmark.** | Benchmark | Model-risk threshold $\alpha_t$ | Certification samples per candidate $N_t^{\mathrm{cert}}$ | Total confidence budget $\delta_{\mathrm{cert}}$ | Candidates per round (range) | Total certified candidate instances $\sum_t n_t$ | Total certification rollouts $\sum_t n_tN_t^{\mathrm{cert}}$ | |---|---:|---:|---:|---:|---:|---:| | Synthetic | 0.20 | 2,000 | 0.05 | 58–113 | 8,628 | 17,256,000 | | Health | 0.20 | 8,000 | 0.05 | 44–182 | 8,976 | 71,808,000 | | Honda Energy | 0.20 | 8,000 | 0.05 | 152–220 | 18,302 | 146,416,000 | | Protein-Signaling | 0.10 | 2,000 | 0.05 | 2–102 | 5,942 | 11,884,000 | **Table 5: Online executions, abstentions, and sample-level feasibility for one complete run on each benchmark.** | Benchmark | Planned online opportunities | Executions | Abstentions | Abstention rate | Jointly feasible samples / generated samples | $\mathrm{FR}_{\mathrm{ind}}$ | |---|---:|---:|---:|---:|---:|---:| | Synthetic | 99 | 99 | 0 | 0.00% | 909 / 990 | 91.82% | | Health | 99 | 90 | 9 | 9.09% | 785 / 900 | 87.22% | | Honda Energy | 99 | 99 | 0 | 0.00% | 967 / 990 | 97.68% | | Protein-Signaling | 99 | 99 | 0 | 0.00% | 990 / 990 | 100.00% | Abstentions count toward the 99 planned opportunities but not toward the sample denominator of $\mathrm{FR}_{\mathrm{ind}}$; each raw SCM outcome in an executed feedback batch is counted separately.
Reviewer_mUBXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies the problem of Bayesian optimisation with causal information in the case where there are some probabilistic constraints on the variables. A new algorithm is proposed that produces certificates for constraint validation and performs graph clustering at the intermediate step. Some simple mathematical statements about the proposed method are made and experimental results on various datasets are reported. The mathematical results are correct; they are essentially direct consequences of the Pinsker and Hoeffding inequalities.

Critical strengths

The most important strengths, in my opinion, are

  1. The paper is the first one that combines causal information with Bayesian optimization under structural graph uncertainty and constraints on the variables. These kinds of settings might appear in practice. Having a paper that provides a method for dealing with it is a valuable contribution.
  2. Numerical experiments presented in the paper show that the proposed method is feasible and achieves its goal.
  3. The paper provides some mathematical results (although simple) that quantify some properties of the method.
Critical weaknesses

The paper has no serious weakness that would justify a strong rejection. However,

  1. The mathematical results are almost trivial. Put differently, there is no non-trivial mathematical result quantifying the proposed method's performance. In particular, no result on the regret.
  2. The paper is not easy to read. In particular, it would be almost inaccessible to readers who are not familiar with Bayesian optimization and structural causal models. However, I should also stress that it would be very hard (almost impossible) to respect the page limitation and to present the material of the paper in a clear way.
  3. The proposed method contains a step of clustering of graphs. It is clear that this step leads to a speed-up, but the impact of it is not quantified.

Specific remarks:

  1. First paragraph of section 3.1: Repetition of "Each hypothesis specifies a causal graph, structural mechanisms, and a noise model;"
  2. Section 2.1: please clarify on which space PtrueP_{true} is acting.
  3. Section 2.1, last sentence: please clarify what the output oto_t is.
  4. Section 2.1: is there a reason for requiring X\mathcal X to be compact?
  5. Line 165: please clarify the meaning of "agree". Should they be equal?
  6. It is worth mentioning that lines 165-179 perform graph clustering within Gt+\mathcal G_t^+.
  7. Prop 3.1 is just trivial. I am not sure it deserves to be stated as a proposition.
  8. I wonder whether "Value of Structural Information" (VSI) is not more appropriate than "Structural Value of Information" (SVI).
  9. The content of lines 201-205 is almost impossible to understand at this stage. As I understand it, its goal is to explain the forthcoming definitions. In that case, it might be more suitable to postpone it to a later place in Section 3.2.
  10. Line 213: Could you explain why qt∘q_t^\circ does not depend on xx ?
  11. It is not clear what is c(x)c(x) in eq 7.
  12. I am not sure that stating eq 6 at this stage adds any value. The reader cannot really understand it, since ata_t and btb_t are not defined.
  13. Independent rollouts generated in Section 3.3 for computing R^t\hat R_t seem to be independent of those used for computing L^t\hat L_t. Is this independence really necessary? What is the problem with having the same rollout?
  14. It is not necessary to have the minimum with 1 in eq 8. The threshold αt\alpha_t will always be chosen smaller than 11, so if R^t(u)+ηt\hat R_t(u)+\eta_t is larger than 11, the corresponding uu will not be included in StS_t.
  15. In the discussion of the related work, it might be useful to mention that causal BO has also been studied in the contextual framework. Some relevant references are
  • Functional Causal Bayesian Optimization. Limor Gultchin, Virginia Aglietti, Alexis Bellot, Silvia Chiappa. arXiv:2306.06409
  • Causal Bayesian Optimization: Foundations, Methods, and Applications. Chenfeng Huang, Thuy T. Le, Zixuan Ma, Hien Tran arXiv:2609.24112
  • Contextual Causal Bayesian Optimisation. Vahan Arsenyan, Antoine Grosnit, Haitham Bou-Ammar, Arnak Dalalyan arXiv:2301.12412
Questions for authors

I put some questions in the box above. See items 4, 10, 13.

4 comments

We sincerely thank the reviewer for the careful evaluation of our manuscript and for recognizing the value of studying constrained causal optimization under structural uncertainty. We are encouraged by your positive assessment of the contribution that our work makes to this research area.

1. The Role of the Compactness Assumption (Comment 4)

We thank the reviewer for pointing out that the scope of this assumption was insufficiently clear. Compactness is primarily needed for the continuous-domain comparison in Appendix F, not for finite-candidate action selection, SVI computation, or the risk certification in Theorem 3.3. At each round, the algorithm evaluates decision losses and constructs the certified set St\mathcal S_t over a finite candidate set Ut\mathcal U_t; it abstains if St\mathcal S_t is empty.

Appendix F considers a comparator with a safety margin Δ>0\Delta>0:

x_\Delta^\star\in\arg\min_{x\in K_\Delta}J^\star(x).$$ Under the Lipschitz-continuity assumptions of that appendix, a nonempty $K_\Delta$ is a closed subset of the compact domain $\mathcal X$, so the continuous objective attains its minimum on $K_\Delta$. Corollary F.1 relates this comparator to the finite candidate set through its fill distance $h_t$, adding a term $L_J^\star h_t$ under the stated safety-margin conditions. This comparison concerns the tightened risk budget $\alpha-\Delta$; comparison with the optimum under the original budget $\alpha$ would additionally require controlling the objective loss induced by constraint tightening. We will move the compactness and intervention-family-specific distance assumptions to the continuous-domain extension rather than present them as requirements for all finite-candidate results. > **2. Action Dependence of the Hypothetical Posterior (Comment 10)** We thank the reviewer for identifying an ambiguity in our notation. The hypothetical posterior does depend on the candidate intervention $x$; the original notation $q_t^o$ merely suppressed this dependence. For observations with positive predictive likelihood, we will write $$q_t^{o,x}(G)=\frac{q_t(G)p_t(o\mid x,G)}{p_t^{\mathrm{full}}(o\mid x)},\qquad p_t^{\mathrm{full}}(o\mid x)=\sum_{G'}q_t(G')p_t(o\mid x,G'),$$ and define SVI explicitly as $$\text{SVI}_t^{\text{full}}(x) = L_t^{\text{full}}(q_t) - E_o\big[ L_t^{\text{full}}(q_t^{o,x}) \big], \quad o \sim p_t^{\text{full}}(\cdot | x)$$ The action $x$ affects both the distribution of hypothetical observations and the likelihood-induced posterior update. During SVI evaluation, the current candidate set and pilot loss map $\ell_t(u,G)=\widehat J_t(u,G)+\zeta_{\mathrm{info}}\widehat r_t(u,G)$ remain fixed. SVI measures the improvement in the optimal risk-penalized loss from reweighting the structural posterior under that fixed decision map; it does not refit mechanisms or regenerate candidates after each hypothetical observation. Actual feedback updates all graph–mechanism weights through Eq. (10). For the compressed representation, we will likewise use $Q_t^{o,x}(g)=\sum_{G\in C_t(g)}q_t^{o,x}(G)$, preserving both action dependence and the full within-cluster likelihood evidence. > **3. Is Independent Certification Sampling Necessary? (Comment 13)** We appreciate the reviewer's suggestion that samples might be reused. Fresh independent certification sampling is a sufficient design for our conditional proof, not a necessary feature of every valid certification procedure. Conditional on the pre-certification history, candidate set, pilot estimates, and SVI computations, we draw $N_t^{\mathrm{cert}}$ fresh single-SCM rollouts for each candidate from the full posterior predictive mixture, resampling the graph–mechanism hypothesis and noise on each draw. The joint-violation indicators are then conditionally i.i.d. within each candidate, allowing a direct Hoeffding bound and union bounds across candidates and rounds. Independence between different candidates' sampling batches is not required. Sample reuse is possible in principle if the candidate set and sample size are fixed before the reused draws and a simultaneous risk bound is established for the entire candidate set. Samples used to construct an acquisition decision cannot subsequently be treated as fresh independent certification samples conditional on that same decision information. Moreover, graph-conditional pilot samples cannot be treated directly as i.i.d. samples from the full posterior mixture. A hypothetical ten-row SVI batch shares a model hypothesis, so its rows are generally dependent after marginalizing over that hypothesis and cannot simply be counted as ten independent mixture samples. We use separate sampling streams to align the implementation's sampling units with the assumptions of Theorem 3.3, rather than to rule out all sample-reuse schemes.

4. Theoretical Contributions and Scope of Performance Guarantees (Major Concern 1; Comment 7)

We thank the reviewer for encouraging us to clarify the performance implications of our theory. We do not establish a cumulative-regret bound or a no-regret convergence rate. Instead, our theoretical contribution is to combine decision-aware information value, compression-based approximation, and pre-execution model-risk certification while identifying their respective errors. For a fixed loss map, SVI vanishes when all active graphs share an optimal action, even if graph information gain remains positive. Thus, learning the graph and improving the current decision need not have the same value. Proposition 3.2 bounds the SVI distortion from compression by 2Δt2\Delta_t, with equality between the full and compressed values when Δt=0\Delta_t=0; full-posterior certification is not directly affected by this decision-map compression error.

Regarding objective performance, Theorem E.1 in Appendix E provides a conditional one-step objective-gap bound. Its assumptions include uniform model accuracy over the current candidate set:

\max_{u\in\mathcal U_t}|r_t(u)-r^\star(u)|\leq\epsilon_t^r,$$ together with the two-sided certification accuracy event $\max_{u\in\mathcal U_t}|\widehat R_t(u)-r_t(u)|\leq\eta_t^{\mathrm{cert}}$. For a nonempty comparator set with sufficient safety margin, $$\mathcal U_t^{\mathrm{cmp}}=\{u\in\mathcal U_t: r^\star(u)+\epsilon_t^r+2\eta_t^{\mathrm{cert}}\leq\alpha_t\},\qquad u_t^\dagger\in\arg\min_{u\in\mathcal U_t^{\mathrm{cmp}}}J^\star(u),$$ the selected intervention satisfies $$J^\star(x_t)-J^\star(u_t^\dagger) \leq 2(\epsilon_t^J+e_{t,J})+\lambda_t^{\mathrm{info}}d_t^\ell+\gamma\\,\operatorname{osc}_{\mathcal U_t}c$$ $$e_{t,J}=\rho_{t,J}+D_{t,J}.$$ This bound separates the roles of model accuracy, pilot and compression errors, the information reward, and intervention costs. Its model-accuracy assumption is not guaranteed by certification sampling, and its comparator is restricted to sufficiently safe candidates. We therefore do not present it as a general regret guarantee. We also agree that Proposition 3.1 should be framed as an auxiliary error decomposition supporting later results, rather than as a principal performance theorem. > **5. Graph-Clustering Approximation, Computational Cost, and Empirical Effects (Major Concern 3)** We thank the reviewer for requesting quantitative evidence of compression benefits and approximation errors. For a fixed pilot loss map, $$\Delta_t=\sum_Gq_t(G)\max_{u\in\mathcal U_t} |\ell_t(u,G)-\ell_t(u,T_t(G))| \leq D_{t,J}+\zeta_{\mathrm{info}}D_{t,r}.$$ Proposition 3.2 establishes $|\operatorname{SVI} _{t}^{\mathrm{full}}(x)-\operatorname{SVI} _{t}^{\mathrm{cl}}(x)|\leq 2\Delta _t$. Corollary B.2 further gives, for $\zeta _{\mathrm{comp}}>0$, $$\Delta _t \leq \max\\{1,\zeta _{\mathrm{info}}/\zeta _{\mathrm{comp}}\\} R _{t}^{\mathrm{cov}}.$$ The bound uses the realized covering radius and concerns exact expectations. Compression reduces decision enumeration from $O(B_tn_t^2g_t)$ to $O(B_tn_t^2m_t)$, where $n_t$, $g_t$, $m_t$, and $B_t$ denote candidates, active graphs, representatives, and hypothetical observations per candidate. Likelihood evaluations and pilot/certification sampling remain unchanged. On a fixed 99-round Healthcare trajectory, compression reduces the mean graph count from 8.38 to 3.00 (64.2%). For representative caps of 1, 3, 5, and 8, the mean realized maximum covering radii are 0.6589, 0.3562, 0.1722, and 0.0891 (zero without compression). Full reaches its representative cap before its nominal tolerance of 0.1, with a maximum observed radius of 0.9520. Across eight posterior states with three paired estimates each, using shared candidates, pilots, and random streams, the mean absolute SVI estimation difference is 0.01039 (maximum 0.05618). The online comparison fixes the initial design, candidate generation, SVI weight (5), pilot samples (64), hypothetical observations (8), risk budget (0.20), and certification samples (1,500). Both variants use full-posterior certification over 99 opportunities. The representative cap/tolerance is 3/0.1 for Full and 100/0 without compression. Results are mean $\pm$ sample standard deviation (time in seconds): | Variant | Best observed feasible objective | $FR_{\mathrm{ind}}$ | nPA-GAP | Algorithm time (s) | |---|---:|---:|---:|---:| | Full DASCBO | 5.3913 $\pm$ 0.0422 | 90.31 $\pm$ 0.30% | 0.4553 $\pm$ 0.0627 | 99.5 $\pm$ 2.4 | | w/o Compression | 5.5308 $\pm$ 0.1234 | 90.14 $\pm$ 0.36% | 0.2934 $\pm$ 0.0768 | 154.2 $\pm$ 53.6 | The best observed feasible objective uses ten-sample batch-mean feasibility; $FR_{\mathrm{ind}}$ uses individual SCM outputs. nPA-GAP accounts for early feasible improvements and abstentions.

6. Probability Spaces, Observation Records, and Safety Events (Comments 2–3)

We thank the reviewer for highlighting the need for clearer probability-space and feedback definitions. In Section 2.1, we will distinguish the exogenous-noise, endogenous-variable, and recorded-observation spaces. The endogenous space is V=∏i=1pVi\mathcal V=\prod_{i=1}^p\mathcal V_i, equipped with the product σ\sigma-algebra. The true SCM's structural map and joint noise distribution PEP_E induce a distribution on this space; intervention do⁡(x)\operatorname{do}(x) changes the structural equations and thereby induces the interventional distribution Px⋆P_x^\star. This interventional distribution should not be conflated with an ordinary observational conditional distribution.

The symbol oto_t denotes the actual observation record used in the likelihood update, not necessarily a scalar objective value. In Healthcare, each execution yields a complete ten-row batch ot=(vt(1),…,vt(10))∈V10o_t=(v_t^{(1)},\ldots,v_t^{(10)})\in\mathcal V^{10}. Under a fixed hypothesis hh, the likelihood is

=\prod_{k=1}^{10}p_h(v_t^{(k)}\mid\operatorname{do}(x_t)),\qquad p_t(o_t\mid x_t)=\sum_hw_{t,h}L_h(o_t\mid\operatorname{do}(x_t)).$$ The batch shares a model hypothesis: its full likelihood must be computed under each hypothesis before mixing over posterior weights. The objective may be summarized using the batch mean, whereas pre-execution certification concerns the joint violation event for a single SCM output. In Healthcare, feasibility requires $\mathrm{BMI}<25$, so a single-output violation is $\mathrm{BMI}\geq25$; the constraint-variable batch mean is used only for the reported best observed feasible objective. We will distinguish these three quantities explicitly in the revision. > **7. Decision-Map Equivalence, Intervention Cost, and Upper-Bound Clipping (Comments 5–6, 11, 14)** We thank the reviewer for the careful examination of these definitions. In Comments 5–6, “agree” means that two graphs have identical objective and risk maps for every candidate in the current set: $$J_t(u,G)=J_t(u,G'),\qquad r_t(u,G)=r_t(u,G'), \quad\forall u\in\mathcal U_t.$$ Clustering is restricted to $\mathcal G_t^+=\{G:q_t(G)>0\}$, and the implementation approximates these maps using pilot estimates. Equivalence does not require identical graph structures or likelihoods, nor identical predictions outside the candidate set; all within-cluster likelihood evidence is therefore retained. We will clarify this in the definition and traversal description. Regarding Comment 11, $c:\mathcal X\to[0,\infty)$ is a prespecified bounded intervention-cost function, weighted by $\gamma\geq0$ in Eq. (7). Healthcare uses $c(x)=|S_x|$; the default reproduction configuration sets $\gamma=0$, so the cost term does not affect action ranking in that setting. Intervention cost is distinct from computational runtime. Regarding Comment 14, the reviewer is correct that clipping at one is redundant for candidate selection when $\alpha_t<1$: $$\min\\{1,\widehat R_t(u)+\eta_t^{\mathrm{cert}}\\}\leq\alpha_t \iff\widehat R_t(u)+\eta_t^{\mathrm{cert}}\leq\alpha_t.$$ Clipping merely keeps the reported bound in the probability range and does not change the certified set. We will state this explicitly. > **8. Clarity, Presentation, and Terminology (Major Concern 2; Comments 1, 7–9, 12)** We thank the reviewer for balancing requests for clearer explanations with the manuscript's space constraints. We will remove the repeated sentence in Section 3.1 (Comment 1); present Proposition 3.1 as an auxiliary error decomposition rather than a main performance result (Comment 7); explain SVI at first use as the value of structural information for downstream intervention decisions (Comment 8); and move the discussion of fixed decision maps, risk penalties, and SVI nonnegativity after the corresponding posterior update and SVI definitions (Comment 9). Before Eq. (6), we will define $a_{t,J}$ and $a_{t,r}$ as pilot objective- and risk-estimation error radii, and $b_t$ as the SVI Monte Carlo error radius, before explaining how estimation, compression, and sampling errors are separated (Comment 12). We will also add an overview of the workflow before introducing detailed notation and bounds: the history determines the candidate set; pilot simulations estimate graph-specific decision consequences; compression and SVI determine action preferences; full-posterior risk certification selects admissible interventions; and actual feedback updates all model weights.

9. Contextual and Functional Causal Bayesian Optimization (Comment 15)

We thank the reviewer for recommending these relevant works. We will add and discuss Gultchin et al., Functional Causal Bayesian Optimization; Arsenyan et al., Contextual Causal Bayesian Optimisation; and Huang et al., Causal Bayesian Optimization: Foundations, Methods, and Applications, addressing functional interventions, contextual policies, and the broader causal BO framework, respectively. The former two lines of work study richer intervention representations or contextual policies, whereas our paper focuses on finite sets of constant-value interventions under structural uncertainty, combining decision-aware SVI with full-posterior model-risk certification.

We again thank the reviewer for the constructive feedback. We will use these suggestions to clarify the scope of our theoretical assumptions and performance guarantees, add compression diagnostics and implementation details, and improve the presentation of the equations and notation so that the paper's contributions and limitations are easier to assess.

We hope that the above revisions and clarifications adequately address the reviewer's concerns. We would be happy to provide further clarification or answer any additional questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal and the revised manuscript

We sincerely thank all three reviewers for their specific and constructive feedback. We are particularly encouraged that one reviewer recognized our work as the first to integrate causal information with Bayesian optimization under structural uncertainty and variable constraints. We also appreciate the reviewers' positive assessment of our methodological novelty, particularly the introduction of decision-aware structural information, which evaluates structural uncertainty through its impact on intervention decisions within a safe causal Bayesian optimization framework.

At the same time, the reviewers identified shortcomings in presentation clarity, the completeness of empirical evidence, and the interpretation of safety guarantees. We have carefully addressed these concerns through detailed responses and corresponding revisions. To facilitate the Area Chair's assessment, we briefly summarize the main concerns and our revisions below.

Concerns Raised by the Reviewers

∙\bullet Clarity and readability: Some notation and the order of presentation were insufficiently clear, making the paper less accessible to readers outside the immediate research area.

∙\bullet Component attribution and reproducibility: The paper lacked ablation studies isolating the contributions of its core components, complete SCM specifications for the benchmarks, and sufficiently detailed quantitative results with uncertainty estimates.

∙\bullet Safety guarantees and implementation consistency: The distinction between posterior-predictive model-risk certification and true physical safety needed clarification, and the implementation needed to be consistent with the intervention semantics of a fixed SCM and the theoretically defined safety events.

Main Revisions and Additional Evidence

∙\bullet Aligning theory and implementation: We reconstructed the model hypotheses to conform to the intervention semantics of a fixed SCM and aligned the safety-certification target with the joint constraint-violation event defined in Section 2.1.

∙\bullet Providing component-level evidence: We added controlled ablation studies and same-state compression diagnostics to examine the respective contributions of SVI and decision-map compression, including their approximation effects.

∙\bullet Clarifying safety guarantees and experimental reporting: We clarified the scope of posterior-predictive risk certification and the additional conditions required to transfer these guarantees to true physical safety. We also reported per-benchmark certification budgets, abstention statistics, and sample-level joint feasibility rates.

∙\bullet Improving clarity and reproducibility: We reorganized the notation and presentation, removed redundant explanations, and supplemented the four benchmarks with detailed causal graphs, structural equations, noise distributions, intervention domains, and constraint specifications.

We believe these revisions and additional evidence substantially strengthen the theoretical presentation, experimental evaluation, and reproducibility of our work. We hope that the revised manuscript adequately addresses the reviewers' concerns and would be happy to provide further explanations or answer any additional questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.