acceptodds
Under review as a conference paper at ICLR 2027

Budget-First Conformal Robustness Control under Hard Deployment Budgets

Abstract

Decision-making usually relies on uncertainty sets for robust decisions, and existing methods like conformal robustness control (CRC) typically return an uncertainty set at a fixed coverage or robustness level. However, the uncertainty set may violate pointwise deployment requirements such as per-decision cost limits or set-size budgets. In this paper, we propose *Budget-First Conformal Robustness Control* (BCRC) to satisfy deployment budgets under both continuous and discrete regimes. Our main contribution is a budget-first conformal robustness framework that learns the deployment rule induced by the hard budget. In contrast to this risk-first order, BCRC works backward from the hard budget to construct the maximal feasible candidate set for an observed input. For the resulting budget-feasible deployment, we provide a finite-sample upper bound on the failure probability of the exact deployed rule using an independent certification set. Across various experiments, BCRC consistently satisfies the deployment budget and achieves lower failure risk than several baseline methods under the same budget.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal Prediction, Robust Decision-Making, Hard Deployment Budgets

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_xtWFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose Budget-First Conformal Robustness Control for decision-making under pointwise deployment budgets. The method learns budget-feasible decision rule and certifies failure risk control under independent data. Experiments include continuous robust decision-making, image classification, and finite-candidate answer shortlisting.

Critical strengths

The paper considers a practically relevant budget-first setting, where every output should satisfy a prescribed budget.

Critical weaknesses

However, I'm not sure whether the upper-bound objective in (8) is a suitable surrogate for actual failure risk, particularly when model probabilities are estimated inaccurately. Moreover, the advantage of this formulation over directly minimizing failure risk is insufficiently explained. These two concerns lead me to recommend weak rejection. I would consider raising my score if the authors clearly address both points.

Questions for authors
  1. In the continuous setting, the expected model-assigned probability outside the prediction set, E[αˉθ(X)]\mathbb E[\bar\alpha_\theta(X)], may differ substantially from the actual failure risk. Even if two models, θ1\theta_1 and θ2\theta_2, yield the same value of this quantity, their actual failure risks may differ substantially. In (8), is it reasonable to require this quantity to be an upper bound on the actual failure risk and then minimize it? Are there experimental results showing that this quantity is close to the actual failure risk for the trained model θ^\hat{\theta}?

  2. In the continuous experiment, the budget requires (r(x)\leq T). Why not set (r(x)\equiv T) and directly minimize the sample version of (\mathbb P{\phi(Y,z(X))>T})? What statistical or computational advantage does the proposed formulation offer over this direct objective? A comparison would help clarify its benefit.

5 comments
AuthorsOpenReview ↗

Response to Reviewer xtWF (Part 1/3)

We sincerely thank the reviewer for the thoughtful questions. In response, we (a) added a new Proposition R1 with empirical verification and (b) compared BCRC with the reviewer’s proposed direct method under matched settings. We respond to each question in detail below.


Q1. Is the objective in Eq. (8) a suitable surrogate for the failure risk, particularly when model probabilities are inaccurate?

Setup. The goal is the budget-first problem (4), which minimizes the failure probability P{ϕ(Y,z(X))>r(X)}\mathbb P\lbrace\phi(Y,z(X))>r(X)\rbrace subject to the pointwise budget C(z(x),r(x),x)≤T\mathcal C(z(x),r(x),x)\le T. BCRC restricts the rule to a nested family of sets Uθ(x;s)\mathcal U_\theta(x;s). For each input it deploys the largest set that meets the budget and takes the robust decision zθ(x)z_\theta(x) and its worst-case loss rθ(x)r_\theta(x) over that set. The target thus becomes minimizing the failure risk P(Eθ(X,Y)=1)\mathbb P(E_\theta(X,Y)=1) over θ\theta, where EθE_\theta indicates ϕ(Y,zθ(X))>rθ(X)\phi(Y,z_\theta(X))>r_\theta(X). BCRC solves it through Eq. (8), which minimizes E[αˉθ(X)]\mathbb E[\bar\alpha_\theta(X)] subject to P(Eθ(X,Y)=1)≤E[αˉθ(X)]\mathbb P(E_\theta(X,Y)=1)\le\mathbb E[\bar\alpha_\theta(X)], where αˉθ(x)=Pθ(Y∉Uθ(x)∣x)\bar\alpha_\theta(x)=P_\theta(Y\notin\mathcal U_\theta(x)\mid x) is the probability that the fitted model assigns outside the deployed set.

Direct failure-risk minimization determines the deployed rule but leaves the model-assigned risk level unconstrained, since different parameterizations can induce the same rule while assigning different probabilities outside the set. Because the true miscoverage of the deployed set upper-bounds the failure risk, the model-assigned miscoverage is the natural counterpart of the level α\alpha in (2), and BCRC uses its average as the learning objective. This upper-bound relation may fail for estimated probabilities, so Eq. (8) requires the level to upper-bound the failure risk. Under the conditions of Proposition D.4, its solutions keep the minimum failure risk, and their level equals that risk.

Theoretical guarantees. Let θA\theta_A solve Eq. (8), and let θR\theta_R minimize the failure risk among the parameters that satisfy its constraint. Proposition D.2 gives

0≤P(EθA=1)−P(EθR=1)≤E[αˉθR(X)]−P(EθR=1).0\le\mathbb P(E_{\theta_A}=1)-\mathbb P(E_{\theta_R}=1)\le\mathbb E[\bar\alpha_{\theta_R}(X)]-\mathbb P(E_{\theta_R}=1).

We strengthen this result with a new Proposition R1 (Proposition D.4 in the revision): if a risk-minimizing rule in the family can be reparameterized so that its deployed decision and certificate stay the same while E[αˉθ(X)]\mathbb E[\bar\alpha_\theta(X)] equals its failure risk, then every minimizer of (8) satisfies

E[αˉθA(X)]=P(EθA(X,Y)=1)=min⁡θ∈ΘP(Eθ(X,Y)=1).\mathbb E[\bar\alpha_{\theta_A}(X)]=\mathbb P(E_{\theta_A}(X,Y)=1)=\min_{\theta\in\Theta}\mathbb P(E_\theta(X,Y)=1).

In the continuous Gaussian family, rescaling the covariance provides such a reparameterization where the budget is binding (Appendix D.1). This also addresses the reviewer’s example. If θ1\theta_1 and θ2\theta_2 satisfy the constraint with the same E[αˉθ(X)]\mathbb E[\bar\alpha_\theta(X)] and θ1\theta_1 has the lower failure risk, then θ1\theta_1 can be rescaled to a strictly smaller objective with the same decisions, so θ2\theta_2 is not a solution.

AuthorsOpenReview ↗

Response to Reviewer xtWF (Part 2/3)

Q1 (continued).

Empirical verification. Table R1 compares the level with the failure risk of the exact deployed rule on independent test data. Besides BCRC, it reports a Direct-risk variant that uses the same set family and deployment rule but minimizes the smoothed failure risk directly, without the level term and the constraint.

Table R1. Continuous experiment. Gap is E[αˉθ^(X)]\mathbb E[\bar\alpha_{\hat\theta}(X)] minus failure risk.

Budget TTBCRC levelBCRC failure riskGap (pp)Direct-risk levelDirect-risk failure riskGap (pp)
50.4860.489−0.300.4720.489−1.72
80.3740.375−0.050.3700.375−0.53
100.3180.308+0.990.2830.307−2.40
120.2380.249−1.110.2140.249−3.53

BCRC and Direct-risk have similar observed failure risks across all four budgets, with all paired 95% confidence intervals for their differences containing zero. Their levels differ. BCRC’s level is within 1.1 pp of its failure risk, while Direct-risk’s is lower at all four budgets, by up to 3.5 pp.

Robustness to inaccurate probabilities. We replaced only the true noise with heavy-tailed multivariate t3t_3 noise or skewed lognormal noise (σ=0.8\sigma=0.8, standardized to mean zero and unit variance). Table R2 also includes the best budget-adapted conformal baseline and the direct method of Q2 (Direct Decision Network). At every budget BCRC has the exceedance risk closest to the oracle, and its level stays within 1.2 pp of its failure risk.

Table R2. Misspecified noise with the Gaussian model unchanged. Gap is E[αˉθ^(X)]\mathbb E[\bar\alpha_{\hat\theta}(X)] minus failure risk.

NoiseTTOracleBCRCBest adapted baselineDirect Decision NetworkGap (pp)
t3t_350.48520.48580.48880.4875−0.34
80.37020.37120.37340.3732−1.17
100.30780.30870.31140.3186−0.76
120.24750.24800.25040.2590−0.67
Lognormal50.48400.48630.48760.4873−0.65
80.36940.37150.37230.3737−0.69
100.30670.30850.30960.3189−0.35
120.24610.24760.24830.2614−0.63
AuthorsOpenReview ↗

Response to Reviewer xtWF (Part 3/3)

Q2. Why not set r(x)≡Tr(x)\equiv T and directly minimize failure risk? What statistical or computational advantage does BCRC offer?

In the continuous experiment, the budget is C=r\mathcal C=r, so setting r≡Tr\equiv T and directly minimizing P{ϕ(Y,z(X))>T}\mathbb P\lbrace\phi(Y,z(X))>T\rbrace is a valid alternative, with the same oracle minimum α⋆(T)\alpha^\star(T) as Eq. (4). Compared with directly learning a decision policy, BCRC achieves better sample efficiency in our experiments and provides input-dependent certificates that are tighter than TT on many inputs. It also applies to budgets other than C=r\mathcal C=r.

Budgets other than C=r\mathcal C=r. Setting r≡Tr\equiv T is optimal only when the certificate is the only quantity under the budget. For the portfolio cost C(z,r,x)=r+κ∥z−z0(x)∥1\mathcal C(z,r,x)=r+\kappa\Vert z-z_0(x)\Vert_1 of Section 3.1, setting r≡Tr\equiv T violates the budget. BCRC needs no change, since the robust problem (5) includes this cost and the budget selects the set size.

Statistical advantage. Fixing r≡Tr\equiv T simplifies the optimization problem, but learning a low-risk decision policy z(x)z(x) from finite data remains challenging. The direct method parameterizes z(x)z(x) and optimizes a smoothed threshold-exceedance loss, whose gradient becomes small when the loss is far from TT. BCRC instead learns a conditional outcome model and derives decisions through budget-feasible uncertainty sets. This structured approach provides additional learning signals from the full outcomes, which we find improves sample efficiency. On inputs where the budget constraint is binding, the resulting decision minimizes the probability of exceeding TT under the fitted Gaussian model. To test this, we implemented the reviewer's proposal as a Direct Decision Network and compared the two methods across training-set sizes.

Table R3. BCRC minus Direct Decision Network in realized test exceedance rate (pp).

ntrainn_{\rm train}T=5T=5T=8T=8T=10T=10T=12T=12
200−0.59−1.03−2.43−3.47
500−0.48−0.59−1.17−2.09
1,000−0.20−0.33−1.27−1.87
1,500−0.21−0.21−1.11−1.76

All 16 paired bootstrap 95% CIs exclude zero, and the gap is largest for small training sets. At ntrain=1,500n_{\rm train}=1{,}500, BCRC is within 0.02–0.15 pp of the exact oracle α⋆(T)\alpha^\star(T), and the Direct Decision Network within 0.2–1.9 pp.

Input-dependent certificates. With r≡Tr\equiv T, every decision is reported with the same bound TT. BCRC reports the worst-case loss over the deployed set, which equals TT only where the budget is binding. On 4% / 26% / 25% / 37% of test inputs at T=5/8/10/12T=5/8/10/12, the largest set in the family already fits within the budget. The certificate on these inputs averages 2.6 / 4.5 / 6.0 / 7.8. Across all test inputs, measuring failure against BCRC's own certificates rather than TT increases empirical failure risk by at most 0.013 percentage points. Thus, the tighter certificates incur virtually no additional observed failure risk.


Thank you again for the careful and constructive review. We hope these responses address both of your concerns, and we are happy to provide any further clarifications to help your evaluation.

Reviewer_4v6xOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes an uncertainty set construction method that can satisfy pointwise deployment for robust decision-making. In problems with continuous output spaces, the set determines the robust decision and loss certificate, whereas in finite output spaces, it is a shortlist. Experiments on synthetic, image classification, and multiple-choice language-model tasks report input-level budget compliance and low failure risk.

Critical strengths

1 .Hard-budget adherence is a very well-motivated problem, and existing conformal methods provide average-case guarantees on budget.

  1. The empirical performance of the method is quite strong. BCRC beats the baselines across all three distinct experimental setups.
Critical weaknesses
  1. “Conformal” aspects of this method have not been clearly established.The main high-probability risk bound appears to follow from concentration of independent binary failures of a fixed rule. I could not identify a conformal calibration step that determines the deployed rule, or a main guarantee that specifically relies on conformal inference.

  2. The methodological contribution seems limited relative to the paper’s scope. The proposed training objective appears to be the principal new methodological component. I would find the contribution more substantial with a result connecting the learned rule’s failure risk to the oracle budget-first risk α∗(T)\alpha^*(T), or establishing when the training objective is a reliable proxy for failure risk.

  3. The notation makes the central method difficult to follow. For example, Uz,rU_{z,r} and Uθ,αU_{\theta,\alpha} appear to describe different classes of uncertainty sets, but their relationship is not made sufficiently clear in the main text. Likewise, the distinct roles of the oracle risk α∗(T)\alpha^*(T) and the input-dependent budget-feasible index αˉθ(x)\bar\alpha_\theta(x) are easy to confuse. I found Sections 3.1 and 3.2, which are essential to understanding the method, difficult to follow as a result.

Questions for authors

In Section 5.1, what are the Avg. Certificate, Failure Risk. and Risk upper bound values for the baseline methods?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 4v6x (Part 1/4)

We sincerely thank the reviewer for the constructive feedback. Following your suggestions, we have systematically reorganized Section 3.1 and streamlined the notation. Specifically, we (a) clarify what is conformal in BCRC, (b) analyze the gap between the learned rule’s failure risk and the oracle α⋆(T)\alpha^\star(T) and strengthen the theoretical relationship between the training objective and the failure risk, and (c) clarify the notation. We respond to each point in detail below.


W1. The "conformal" aspects of the method are not clearly established.

We thank the reviewer for raising this point. BCRC differs from standard conformal prediction in that the deployed set is determined by the hard budget rather than a conformal calibration threshold. Its connection to conformal inference lies in the backward construction and the subsequent risk assessment. Backward Conformal Prediction (Gauthier et al., 2025) similarly considers set-size constraints before determining the associated coverage level. In BCRC, αˉθ(x)\bar\alpha_\theta(x) represents the model probability excluded by the budget-feasible set; it is not itself a distribution-free coverage guarantee. For risk assessment, we follow the inverse conformal risk-control perspective of Zhou and Zhu (2026): rather than selecting a set to meet a prescribed risk level, we estimate the risk of the rule already selected for deployment. Specifically, for KmK_m failures among mm certification samples, Appendix G.1 gives the add-one risk estimate (Km+1)/(m+1)(K_m+1)/(m+1). Theorem 4.1 establishes its expected conservativeness under exchangeability, P(Eθ^=1)≤E[Km+1m+1]\mathbb P(E_{\hat\theta}=1)\le\mathbb E\left[\frac{K_m+1}{m+1}\right].

This guarantee is in expectation over certification sets. For a realized independent certification set, Proposition 4.2 instead provides the high-probability risk bound in Eq. (12), which is the bound reported in our experiments. We have clarified this distinction in the revised Section 4.

AuthorsOpenReview ↗

Response to Reviewer 4v6x (Part 2/4)

W2. Limited contribution beyond the training objective, and unclear link to the deployed failure risk and α⋆(T)\alpha^\star(T).

We appreciate the reviewer’s suggestion. Existing conformal methods fix the risk level first and do not enforce a pointwise budget. Problem (4) instead fixes the budget for every input and then minimizes the failure risk. This problem cannot be solved directly. Proposition D.1 shows that solving it over uncertainty sets gives the same optimal value α⋆(T)\alpha^\star(T) in the continuous setting. Since optimizing over all sets is still intractable, BCRC learns a parameterized nested family and deploys its largest member within the budget. Below, we clarify how the budget-induced training objective relates to actual deployment failure risk and the oracle α⋆(T)\alpha^\star(T).

1. Population level. Recall that αˉθ(X)\bar\alpha_\theta(X) is the model probability outside the deployed set, not the failure probability, and that Eq. (8) minimizes E[αˉθ(X)]\mathbb{E}[\bar\alpha_\theta(X)] subject to P(Eθ=1)≤E[αˉθ(X)]\mathbb{P}(E_\theta=1)\le\mathbb{E}[\bar\alpha_\theta(X)]. Let θA\theta_A solve (8) and θR\theta_R minimize the failure risk over the same feasible set, assuming both minima exist. Proposition D.2 (Appendix D.1) gives

0≤P(EθA=1)−P(EθR=1)≤E[αˉθR(X)]−P(EθR=1).0\le\mathbb{P}(E_{\theta_A}=1)-\mathbb{P}(E_{\theta_R}=1)\le\mathbb{E}[\bar\alpha_{\theta_R}(X)]-\mathbb{P}(E_{\theta_R}=1).

We strengthen this result with a new Proposition R1 (Proposition D.4 in the revision): if a risk-minimizing rule in the family can be reparameterized so that its deployed decision and certificate stay the same while E[αˉθ(X)]\mathbb E[\bar\alpha_\theta(X)] equals its failure risk, then every minimizer of (8) satisfies

E[αˉθA(X)]=P(EθA=1)=min⁡θ∈ΘP(Eθ=1).\mathbb{E}[\bar\alpha_{\theta_A}(X)]=\mathbb{P}(E_{\theta_A}=1)=\min_{\theta\in\Theta}\mathbb{P}(E_\theta=1).

In the continuous setting, rescaling the Gaussian dispersion provides such a reparameterization on inputs where the budget constraint is binding. To relate the solution of (8) to the unrestricted oracle, Appendix D.1 decomposes the gap in the continuous setting as P(EθA=1)−α⋆(T)≤[E[αˉθR(X)]−P(EθR=1)]+[P(EθR=1)−inf⁡θ∈ΘP(Eθ=1)]+[inf⁡θ∈ΘP(Eθ=1)−α⋆(T)].\mathbb{P}(E_{\theta_A}=1)-\alpha^\star(T)\le\big[\mathbb{E}[\bar\alpha_{\theta_R}(X)]-\mathbb{P}(E_{\theta_R}=1)\big]+\big[\mathbb{P}(E_{\theta_R}=1)-\inf_{\theta\in\Theta}\mathbb{P}(E_\theta=1)\big]+\big[\inf_{\theta\in\Theta}\mathbb{P}(E_\theta=1)-\alpha^\star(T)\big]. The first term is the right-hand side of Proposition D.2. The second is the loss from imposing the constraint in (8), and the third is the loss from restricting decisions to the parameterized family.

2. Finite-sample level. The results above concern population optimizers. For the parameter θ^\hat\theta obtained from training with smoothed surrogates, Theorem 3.1 shows that, under the conditions in Appendix F, with probability at least 1−δ1-\delta, P(Eθ^(X,Y)=1)≤E[αˉθ^(X)]+Δn+ξn(θ^).\mathbb{P}(E_{\hat\theta}(X,Y)=1)\le\mathbb{E}[\bar\alpha_{\hat\theta}(X)]+\Delta_n+\xi_n(\hat\theta). Here ξn(θ^)\xi_n(\hat\theta) is the amount by which the deployed rule violates the empirical constraint.

3. Empirical level. In the continuous experiment, the data-generating distribution is known. Since setting the certificate to TT can only reduce failure, Eq. (4) decouples across inputs as α⋆(T)=EX[inf⁡z∈ZP(ϕ(Y,z)>T∣X)]\alpha^\star(T)=\mathbb E_X\big[\inf_{z\in\mathcal Z}\mathbb P(\phi(Y,z)>T\mid X)\big]. We compute this oracle and the exact exceedance risk of each deployed rule on the same 7,500 test inputs. CRC, CRO and E2E-CRO keep their learned geometries but use the same budget-induced deployment as BCRC, instead of truncation.

Table R1. Exact conditional exceedance risk for the oracle and budget-induced deployed rules (lower is better).

TTOracleBCRCCRC geometryCRO geometryE2E-CRO geometry
50.49080.49100.49760.50720.5096
80.37420.37570.37990.38570.3874
100.30780.30860.31300.31870.3224
120.24710.24770.25190.25720.2625

BCRC is within 0.02–0.15 pp of the oracle at every budget and closer than all baselines. Since all rules share the same deployment, the difference comes from the learned geometry.

AuthorsOpenReview ↗

Response to Reviewer 4v6x (Part 3/4)

W3. Notation in Sections 3.1–3.2 is hard to follow.

Thank you for pointing this out. We have reorganized Section 3.1 so that each object is introduced with its role, and updated the notation table in Appendix C (Table 5). We use U(x)\mathcal U(x) to denote a general uncertainty set, Uθ(x;s)\mathcal U_\theta(x;s) to denote the learned parameterized nested family, and Uθ(x)\mathcal U_\theta(x) to denote its largest budget-feasible member selected for deployment. The oracle α⋆(T)\alpha^\star(T) is a fixed benchmark and does not depend on θ\theta or xx. By contrast, αˉθ(x)\bar\alpha_\theta(x) is the model probability outside the deployed set for input xx. Its average is the proxy we minimize in Eq. (8). The notation follows three stages:

  1. Target (Eqs. 2–6). A rule outputs a decision z(x)z(x) and a loss certificate r(x)r(x), and fails if ϕ(Y,z(X))>r(X)\phi(Y,z(X))>r(X); α⋆(T)\alpha^\star(T) is the smallest failure probability over all rules that satisfy the budget. Through Eq. (5), each uncertainty set gives a decision and a certificate, and Proposition D.1 shows that describing rules this way does not change the optimal value.
  2. Deployment rule (Eq. 7). For a given θ\theta, the budget selects the size sθ(x)s_\theta(x) and hence the deployed set Uθ(x)\mathcal U_\theta(x); Eθ(x,y)E_\theta(x,y) records whether this rule fails.
  3. Learning and certification (Eqs. 8–12). We learn θ\theta through Eq. (8), and Eq. (12) certifies the failure risk of the learned rule on independent data.
Reviewer_ntPfOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes budget-first conformal risk control (BCRC), which enforces budget constraints in a contextual robust optimization setting. This is done via a budget-aware learning procedure that leverages a Lagrangian-based surrogate to address discontinuities. Empirically, they evaluate their approach in both classic contextual robust settings with continuous uncertainty sets, and discrete settings in a machine learning context involving selection.

Critical strengths
  • Generality: The generality between continuous and discrete settings is a nice element of the work. Specifically, it handles the standard continuous contextual robust optimization setting while remaining adaptable to discrete settings, with interesting downstream applications (vision/LLMs) where selecting among candidates via a relatively generic method is a strong point. The generality outweighs the scope limitations I discuss below, which is why I gave it a slightly favorable score.
  • Results: The method consistently outperforms truncated baselines that meet the same budget requirement. This matters because a method designed for the proposed problem should at least beat baselines not designed for it.
Critical weaknesses
  • Scope: Overall, I do think the paper studies a reasonably interesting problem in robust decision-making. However, it seems very limited in scope; for example, the method is designed around a narrow use case (hard budget constraints), and the baselines aren't designed for this setting and are somewhat crudely adapted via truncation. Even in this favorable setting, the proposed method is consistently better, but only marginally, against methods that aren't designed for the task. Moreover, the authors note that this comes at a notable cost to coverage, and as far as I could tell, they don't even compare coverage post-truncation across methods. Beyond this, in the continuous setting, the authors present C(⋅)C(\cdot) as a generic measure but empirically test only one metric (average certificate). This matters because the method sacrifices many commonly desired traits in robust decision-making (coverage and failure risk, the focus of most prior work) for a new metric the method itself was designed for, and I am not entirely convinced of its utility.
Questions for authors

Both questions/remarks are for clarification and do not materially affect my assessment.

  • The authors discuss two end-to-end approaches they compare against, labeled E2E, but I believe the approach from Chenreddy & Delage should be labeled CRO. Can the authors clarify this?
  • Not necessarily a question, but Figure 1 doesn't have enough context, and most of the terms are not defined where they are used, so I think this either needs (1) more clarity or (2) complete definitions.
5 comments
AuthorsOpenReview ↗

Response to Reviewer ntPf (Part 1/3)

We sincerely thank the reviewer for the careful assessment and constructive feedback. To address the main concerns, we (a) clarified practical settings with pointwise budgets, (b) added stronger baselines and an oracle benchmark, and (c) reported post-truncation coverage and clarified the certificate metric. We respond to each point in detail below.


W1. Hard budgets are a narrow use case, and the baselines are crudely adapted via truncation.

In robust decision-making, larger uncertainty sets provide greater protection but often lead to more conservative decisions. Risk-first methods prescribe a coverage or robustness level and accept the resulting deployment cost. However, this order may not be suitable when each decision must satisfy a fixed operational limit. Related motivations appear in Backward Conformal Prediction (Gauthier et al., 2025), which starts from a set-size constraint, and Zhou & Zhu (2026), which discusses the difficulty of prescribing a robustness level in advance. Pointwise budgets arise in several practical settings:

  • In robust portfolio selection (Goldfarb & Iyengar, 2003), the manager chooses an allocation that performs well under the worst returns in the uncertainty set, and the worst-case loss of this allocation is often required to stay below a fixed limit.
  • In robust generator dispatch (Bertsimas et al., 2013), the operator schedules generation to cover the worst demand in the uncertainty set, but only a limited amount of reserve capacity is available in each period.
  • In human or model review, a clinician may examine at most TT candidate diagnoses per case, and an LLM verifier may run at most TT checks per query.

A marginal coverage or robustness guarantee does not ensure that every output satisfies a pointwise budget. Truncation can enforce this constraint, which is why we include truncated baselines. To test whether BCRC's advantage comes from the truncation protocol, we additionally evaluate the frozen baseline set families using the same budget-induced deployment construction as BCRC (W2). This comparison isolates the benefit of learning the set family specifically for the budget-induced rule.

AuthorsOpenReview ↗

Response to Reviewer ntPf (Part 2/3)

W2. Even in this favorable setting, the gains are marginal, and coverage after truncation is not compared.

We thank the reviewer for this comment. We have added the three results below. When all methods use the same budget, BCRC has the lowest failure risk at every budget in the continuous task, and its coverage is similar to or higher than that of the baselines.

Comparisons without truncation. To examine whether BCRC's gains arise from the truncation protocol, we apply the same budget-induced deployment construction to the set families learned by CRC, CRO, and E2E-CRO, selecting the largest budget-feasible set for each input. We also include a Direct Decision Network that learns z(x)z(x) by minimizing a smooth surrogate of the exceedance risk P{ϕ(Y,z(X))>T}\mathbb P\lbrace\phi(Y,z(X))>T\rbrace with r(x)≡Tr(x)\equiv T. Since the synthetic data-generating distribution is known, we compute the budget-constrained oracle risk and each method's exact conditional exceedance risk, averaged over the same 7,500 test inputs (5 seeds ×\times 1,500).

Table R1. Exact conditional exceedance risk (lower is better).

TTOracleBCRCCRC geometryCRO geometryE2E-CRO geometryDirect Decision Network
50.49080.49100.49760.50720.50960.4929
80.37420.37570.37990.38570.38740.3773
100.30780.30860.31300.31870.32240.3192
120.24710.24770.25190.25720.26250.2659

Across all four budgets, BCRC achieves significantly lower exact conditional exceedance risk than every learned comparator (hierarchical paired bootstrap; all 95% CIs exclude zero). These improvements persist when the baseline geometries use the same budget-induced deployment construction, indicating that the gains are not driven by the original truncation protocol.

Size of the gains. In the continuous task, the strongest adapted geometry baseline (CRC geometry) is already close to the oracle, with a gap of only 0.48–0.68 percentage points (pp). BCRC closes 74–97% of this gap, although the absolute improvements are small. On ImageNet, BCRC reduces failure risk by 2.0–3.0 pp on average compared with the truncated conformal baselines across 45 settings. At T=5T=5, its average failure risk is 8.2%, compared with 10.7–12.3% for these baselines.

Coverage after truncation. Table R2 reports test coverage of the budget-feasible rules in the continuous task (five-seed averages).

Table R2. Test coverage under the same pointwise budget (higher is better).

TTBCRCCRC + trunc.CRO + trunc.E2E-CRO + trunc.
50.4020.4040.4070.363
80.5690.5080.5300.507
100.6370.5630.5940.567
120.7150.6120.6500.621

Without truncation, the forward methods reach coverage of 0.81–0.91, but their sets exceed the budget on 29–82% of inputs. Under the same budget, BCRC has similar coverage at T=5T=5 and higher coverage at T=8,10,12T=8,10,12.


W3. Utility of the evaluation metrics.

We thank the reviewer for raising this concern. The loss certificate r(x)r(x) is the worst-case loss of a robust decision over its uncertainty set, rather than a metric introduced specifically for BCRC. RA-DPO (Kiyani et al., 2025) and CRC (Hu et al., 2026) minimize the expected certificate E[r(X)]\mathbb E[r(X)] subject to robustness constraints. BCRC instead imposes a pointwise deployment limit and minimizes the failure risk under that limit.

Beyond certificate-only budgets. To evaluate a nontrivial decision-dependent cost, we additionally consider portfolio rebalancing with C(z,r,x)=r+κ∥z−z0(x)∥1\mathcal C(z,r,x)=r+\kappa\Vert z-z_0(x)\Vert_1. Here, transaction costs and the robustness certificate share the same budget. Across κ∈{0.25,1}\kappa\in\lbrace 0.25,1\rbrace and T∈{5,8,10,12}T\in\lbrace 5,8,10,12\rbrace, BCRC maintains zero budget violations and achieves lower exact conditional common-cap exceedance risk than the adapted CRC, CRO, E2E-CRO, and Direct Decision Network baselines. For example, at κ=1,T=5\kappa=1,T=5, BCRC incurs an average trading cost of 0.658 and an average certificate of 4.197, while achieving a risk of 0.5454, close to the oracle value of 0.5445 (Direct DN: 0.5473; CRC geometry: 0.5512).

AuthorsOpenReview ↗

Response to Reviewer ntPf (Part 3/3)

Q1. Should the approach of Chenreddy & Delage be labeled CRO?

We agree that Chenreddy & Delage (2024), End-to-end Conditional Robust Optimization, belongs to the broader Conditional Robust Optimization (CRO) framework. Following Hu et al. (2026), we use CRO for the two-stage predict-then-calibrate baseline of Sun et al. (2023), and E2E for the end-to-end approach of Chenreddy & Delage (2024). Our experiments implement a single E2E baseline primarily following Chenreddy & Delage (2024), with Yeh et al. (2025) cited as related work rather than a separate baseline. We have renamed this method E2E-CRO and clarified the naming and references in the revised manuscript.


Q2. Figure 1 lacks context; terms are not defined where they are used.

We thank the reviewer for pointing this out. We have revised the caption as follows (in blue).

Revised Figure 1 caption. Overview of BCRC under a hard budget TT. Left: For each input xx, BCRC learns a nested family of candidate sets Uθ(x;s)\mathcal U_\theta(x;s). The parameter θ\theta controls the set family, and larger ss produces larger sets. The budget selects the largest feasible set Uθ(x)\mathcal U_\theta(x) whose deployment cost C(U,x)\mathcal C(\mathcal U,x) does not exceed TT. Right: The plot shows empirical failure risk based on the failure indicator EθE_\theta, along with its upper bound computed from independent certification data (Proposition 4.2). BVR denotes the budget-violation rate.


Thank you again for your time and constructive feedback, and we would be glad to answer further questions. We hope these additions address your concerns.

Comments on the paper

1 comment

As the discussion period is approaching its end, please take a moment to review the authors’ responses and share any remaining thoughts or updates to your assessment. Thanks for your thoughtful reviews and discussion!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.