acceptodds
Under review as a conference paper at ICLR 2027

Online Conditional Robust Optimization with Conformal Robustness Control

Abstract

We study online conditional robust optimization, where decisions are made by minimizing worst-case loss over prediction sets before the true labels are observed. Existing online conformal prediction methods can ensure long-run decision robustness under arbitrary distribution shifts, but may yield overly conservative decisions because predictive coverage is sufficient but not necessary for robustness. To address this issue, we propose Online Conformal Robustness Control (OCRC), which formulates this task as an online constrained learning problem, minimizing average worst-case loss on past observations subject to an explicit robustness constraint. OCRC dynamically updates prediction-set thresholds directly from feedback on past robustness violations. We establish long-run robustness control for arbitrary data sequences and asymptotically exact marginal robustness under stochastic data settings. Experiments on synthetic and real-world data show that OCRC maintains the target robustness level while achieving substantially lower decision risks than coverage-based methods.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal Prediction, Robust Optimization, Online Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CZYiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper applies the ACI online conformal prediction idea to the online CRO constrained problem. Instead of updating prediction coverage, this paper uses robustness-violation updating control. Since the violation is not monotone in the multi-dimensional parameter θt\theta_t, the paper adds an external scalar qtq_t to the certificate: qtq_t carries the robustness control through an ACI-style update, while θt\theta_t is updated separately to shrink the certificate. The paper provides some theorems and runs experiments on synthetic and real data.

Critical strengths

The introduction of the external scalar q to decouple the terms is interesting. Directly applying ACI-style updating has a non-nested issue, but this design brings the monotone pattern back.

Critical weaknesses
  • The paper is difficult to follow.

    • At the very beginning, the paper presents an abstract problem without a friendly, example-based articulation to help readers understand the application or scenario. Without this, readers cannot imagine the potential pipeline behind this and do not know how to logically and correctly connect the prediction set, the decision, the certificate, the coverage, and the violation rate.
    • The notation definition is confusing. For example, when the paper first introduces HσH_\sigma and κ\kappa in lines 188 and 195, it does not immediately explain their role, leaving readers unsure what they represent. Additionally, the authors could clearly define the Lipschitz and smoothness conditions in Assumption 2.
  • The experiments are too simple to fully support the claims.

    • The experimental comparison (better than coverage-based ACI and PID) is not that informative or apple-to-apple, because the proposed method changes the controlled event. Specifically, in the synthetic dataset, the baselines adapt a scalar radius around a fixed center and covariance shape, but the proposed method learns them, which conflates the effects of introducing qtq_t.
    • Algorithm 1 explicitly mentions updating θ\theta, but the real dataset experiment fixes the MLP.
Questions for authors
  • In lines 13-14, what do you mean by robustness in conformal prediction? I remember that regular conformal prediction doesn't mention robustness, except the efficiency (prediction set size) and the validity (coverage). Without clear articulation, this may confuse readers about robustness in regular conformal prediction versus OCRC.

  • Lines 428 and 429: Why for pretraining? Section 5.2 defines μθ=MLP\mu_\theta = MLP, and Algorithm 1 updates all of θ\theta, so the MLP weights should be learned online. Otherwise, the algorithms, the theorem's ground, and the real implementation will not match.

  • For the experiments, which certificate do we report (rtr_t or btb_t)? Can we report both?

4 comments

Response to Reviewer CZYi

Thank you for your helpful comments on the exposition and experimental protocol. We have clarified the problem setup, notation, and online learning procedure, and added a comparison using the same prediction-set family. Line numbers refer to the revised manuscript.

Weaknesses

W1(a). At the very beginning, the paper presents an abstract problem without a friendly, example-based articulation to help readers understand the application or scenario. Without this, readers cannot imagine the potential pipeline behind this and do not know how to logically and correctly connect the prediction set, the decision, the certificate, the coverage, and the violation rate.

Response. Thank you for highlighting the need for a concrete example. The revised Introduction explains the full pipeline through factory electricity scheduling (lines 51–57). A prediction set represents plausible hourly load profiles. The factory selects a schedule that minimizes the worst-case cost over this set and reports that cost as its certificate. Coverage asks whether the entire realized load profile lies in the prediction set; robustness asks whether the realized cost of the selected schedule is at most its certificate. For example, a load spike during an unused hour may cause a coverage failure without changing the schedule’s cost. Thus, a failure of coverage need not imply a robustness violation.

W1(b). The notation definition is confusing. For example, when the paper first introduces HσH_\sigma and κ\kappa in lines 188 and 195, it does not immediately explain their role, leaving readers unsure what they represent.

Response. Thank you for pointing out this ambiguity. Section 3.2 now explains both quantities when they are introduced. We define the margin vv as the realized loss minus the current certificate and introduce Hσ(v)=H(v/σ)H_\sigma(v)=H(v/\sigma) as a smooth approximation to the violation indicator 1v>0\mathbf{1}\\{v > 0\\}. The scale σ\sigma controls the width of the transition around zero, and smoothing provides a gradient through which robustness feedback can guide updates of θ\theta (lines 207–213). The coefficient κ\kappa determines the weight of the integrated violation penalty in the downstream objective (lines 214–221).

W1(c). Additionally, the authors could clearly define the Lipschitz and smoothness conditions in Assumption 2.

Response. Thank you for this suggestion. We have clarified the variables to which these conditions apply in Assumption 2 (lines 276–277). For ϕ(y,z)\phi(y,z), Lipschitz continuity and smoothness are with respect to zz, with yy held fixed. For zθ(x)z_\theta(x) and rθ(x)r_\theta(x), they are with respect to θ\theta, with xx held fixed. Smoothness means that the scalar maps have Lipschitz gradients and that the decision map has a Lipschitz Jacobian.

W2(a). The experimental comparison (better than coverage-based ACI and PID) is not that informative or apple-to-apple, because the proposed method changes the controlled event. Specifically, in the synthetic dataset, the baselines adapt a scalar radius around a fixed center and covariance shape, but the proposed method learns them, which conflates the effects of introducing qtq_t.

Response. Thank you for identifying this potential confounding factor. We added a comparison in which all methods use the same oracle center and fixed shape and adapt only the prediction-set scale (Appendix B.1, lines 936–966, Figure 5). The robustness target is 90%.

MethodCoverageRobustnessMean certificate
OCRC25.18%90.70%5.02
ACI89.92%98.38%5.76
Conformal PID89.74%98.24%5.71

With prediction-set flexibility held fixed, OCRC achieves robustness closest to the 90% target and a smaller mean certificate: 5.02, compared with 5.76 for ACI and 5.71 for Conformal PID. Its lower coverage is consistent with directly controlling downstream violations. These results show that the improvement persists when all methods use the same prediction-set family.

W2(b). Algorithm 1 explicitly mentions updating θ\theta, but the real dataset experiment fixes the MLP.

Response. Thank you for pointing out the need to clarify the online updates. All methods update their MLPs during online evaluation. OCRC updates its parameters through the downstream objective, while ACI and Conformal PID take squared-error gradient steps for their mean predictors and separately update their coverage thresholds. We have clarified these updates in the revised experimental description (lines 391–395, Figures 3–4).

Questions

Q1. In lines 13-14, what do you mean by robustness in conformal prediction? I remember that regular conformal prediction doesn't mention robustness, except the efficiency (prediction set size) and the validity (coverage). Without clear articulation, this may confuse readers about robustness in regular conformal prediction versus OCRC.

Response. Thank you for asking us to clarify this distinction. Here, robustness refers to the downstream decision: a decision is robust on a given round if its realized loss does not exceed the reported risk certificate. Prediction-set coverage is sufficient for this property, but outcomes outside the prediction set can also produce losses below the certificate. We now state this distinction explicitly in the abstract (lines 11–17):

We study online conditional robust optimization, where each decision minimizes the worst-case loss over a prediction set before the outcome is observed. The resulting worst-case loss serves as a risk certificate, and a decision is robust on a given round if its realized loss is at most that certificate. Online conformal methods can ensure long-run decision robustness through prediction-set coverage, but this approach can be conservative because outcomes outside the prediction set may also yield losses below the certificate.

Q2. Lines 428 and 429: Why for pretraining? Section 5.2 defines μθ=MLP\mu_\theta=MLP, and Algorithm 1 updates all of θ\theta, so the MLP weights should be learned online. Otherwise, the algorithms, the theorem's ground, and the real implementation will not match.

Response. Thank you for highlighting this ambiguity. Pretraining provides a common initialization to reduce early instability, and the MLP weights continue to be updated online. Load-Climate-Open uses the first 40% of observations for standardization and pretraining, as clarified in Section 5.3 (lines 484–485). In the DEPI notebook, the first 40% are used to fit the MLP, the next 10% for validation, and the final 50% for online evaluation.

During online evaluation, after observing each outcome, OCRC takes a gradient step on its downstream objective to update θt\theta_t, including the MLP weights and covariance parameters used for the next decision. This is the update specified in Algorithm 1 (lines 223–231).

Q3. For the experiments, which certificate do we report (rtr_t or btb_t)? Can we report both?

Response. Thank you for this suggestion. Each method reports the threshold it deploys: OCRC reports bt=rt+qtb_t=r_t+q_t, whereas ACI and Conformal PID report the CRO certificate rtr_t. We have clarified this convention in the captions of Figure 3 (lines 480–482) and Figure 4 (lines 514–515).

For a direct comparison within the same OCRC run, the fixed-center experiment gives a mean CRO certificate of 4.38 and a mean deployed certificate of 5.02, computed from the logged certificates and offsets. The figures report btb_t because this is the threshold deployed by OCRC.

Thanks for the authors’ revision.

  • I have a concern about the modeling updates. Actually, I checked the code submission: the code mean() returns the global network network.forward(x), while the optimizer updates a deep copy self.ellipsoid.network, so OCRC's MLP does not train. This is contrary to your claim that “All methods update their MLPs”.
  • I checked the revised PDF but didn't find clarification for κ\kappa and Assumption 2: Assumption 2 (lines 276–277) is identical to the original, and lines 214–221 do not explain what κ\kappa weighs.
  • You reported both values for B.1 (4.38 vs 5.02). Could you also report OCRC's mean rtr_t alongside its mean rt+qtr_t+q_t for the main experiments? This would show the size of the added offset in each experiment.

Thank you for checking the revised manuscript and supplementary code. We address your three follow-up points below.

Q1. Online MLP updates

I have a concern about the modeling updates. Actually, I checked the code submission: the code mean() returns the global network network.forward(x), while the optimizer updates a deep copy self.ellipsoid.network, so OCRC's MLP does not train. This is contrary to your claim that “All methods update their MLPs”.

Response. Thank you for identifying this bug in the previously submitted code. As you noted, mean() called the global pretrained MLP, while the optimizer held the parameters of a separate copy. As a result, the MLP used by OCRC did not update online in that version. The figures and results in the revised manuscript were generated with the corrected implementation, in which mean() calls the same MLP instance that the optimizer updates. We have updated the supplementary code for inspection. We also reran both real-data main experiments and verified that the MLP weights of OCRC, ACI, and Conformal PID each change during online evaluation.

Q2. The role of κ\kappa and the definitions in Assumption 2

I checked the revised PDF but didn't find clarification for κ\kappa and Assumption 2: Assumption 2 (lines 276–277) is identical to the original, and lines 214–221 do not explain what κ\kappa weighs.

Response. You are right that the revised PDF still omits both explanations.

(a) The role of κ\kappa. Immediately after Eq. (3), we will explain that κ>0\kappa>0 weights the integrated smoothed constraint residual ∫0qgt(θ,q′),dq′\int_0^q g_t(\theta,q')\\,\mathrm{d}q' relative to the base certificate ft(θ)=rθ(Xt)f_t(\theta)=r_\theta(X_t) in the primal objective. It therefore scales the contribution of robustness feedback to the update of θ\theta. Appendix B.3 illustrates this role empirically. Larger values of κ\kappa yield smaller offsets qtq_t after the trajectory stabilizes. This trend is consistent with Theorem 1, which shows that the optimal offset qκ∗q_\kappa^* of the penalized problem tends to zero as κ→∞\kappa\to\infty.

(b) Assumption 2. We will add the following definitions immediately after the assumption. A map FF is LL-Lipschitz if ∣F(u)−F(v)∣≤L∣u−v∣\\|F(u)-F(v)\\|\le L\\|u-v\\| for all u,vu,v in its domain. A differentiable map FF is ss-smooth if ∣DF(u)−DF(v)∣op≤s∣u−v∣\\|DF(u)-DF(v)\\|_{\mathrm{op}}\le s\\|u-v\\| for all u,vu,v in its domain. Here DFDF is the Jacobian for the vector-valued decision map, and the derivative norm reduces to the Euclidean norm of the gradient for scalar-valued maps. These conditions apply to z↦ϕ(y,z)z\mapsto\phi(y,z) uniformly over yy, and to θ↦zθ(x)\theta\mapsto z_\theta(x) and θ↦rθ(x)\theta\mapsto r_\theta(x) uniformly over xx, with the constants stated in Assumption 2.

Q3. Mean base and deployed certificates in the main experiments

You reported both values for B.1 (4.38 vs 5.02). Could you also report OCRC's mean rtr_t alongside its mean rt+qtr_t+q_t for the main experiments? This would show the size of the added offset in each experiment.

Response. The table reports both certificates from the same OCRC run for each main experiment, averaged over all online evaluation rounds. The last column gives the mean offset actually used when deploying the certificate.

Main experimentMean rtr_tMean rt+qtr_t+q_tMean qtq_t
Stationary simulation4.27485.18940.9146
Temporally dependent simulation4.61295.46140.8485
Load-Climate-Open−0.1375−0.09370.0438
DEPI5.43125.67490.2437

On Load-Climate-Open, for example, the offset raises the mean certificate by 0.0438 in the experiment's loss units, from −0.1375 to −0.0937. In each row, the two certificates evaluate the same sequence of decisions.

Reviewer_N7VLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies controlling robustness in an online setup. The authors propose OCRC that achieves this by introducing a scalar threshold qtq_t and optimizing θt\theta_t to minimize a proposed loss function. This work derives theory on the convergence of the robustness and the loss function under some assumptions. Experiments on synthetic and real datasets show that OCRC achieves robustness closer to the target level compared to baselines.

Critical strengths
  1. Theorem 2 is strong, as it holds for any formulation of θt\theta_t, and the convergence rates seem sufficiently informative.
  2. The empirical results favor the proposed OCRC: it achieves a smaller robustness certificate than the coverage-based baselines and achieves robustness closer to the target.
Critical weaknesses
  1. My main reservation is the novelty of the proposed approach - it appears to be a combination of CRO (Chenreddy et al. 2022), CRC (Hu et al. 2026), and ACI (Gibbs & Candes, 2021) or RRC (Feldman et al. 2023). Furthermore, the authors' claim that the downstream robustness violations "cannot be handled by the scalar mechanism used in these methods" seems too strong, as Rolling Risk Control (Feldman et al., 2023) can be employed to control robustness in this setting. Specifically, RRC can be employed with the loss $ e_t=q_t+1 { \phi(Y_t,z_{\theta_t}(X_t)) > r_{\theta_t}(X_t) } $ and the corresponding update qt+1=[qt+γ(et−α)]+q_{t+1} = [q_t+\gamma(e_t-\alpha)]^{+}
  2. Regarding the theory, Theorem 2 provides a guarantee only on the smoothed indicator rather than the hard one. Moreover, Theorem 5 requires Assumption 5, which might be too restrictive: it considers the supremum over all sequences of the parameters θ,q\theta, q. The left-hand side of line 326 can be positive even as ε→0\varepsilon \rightarrow 0, which contradicts Assumption 5. Furthermore, this work does not provide a theoretical guarantee about controlling the robustness, even over an infinite time horizon. Additionally, Theorem 3 requires the loss function to be convex, which is a strong assumption, and even the authors acknowledge that it is nontrivial to construct and verify.
Questions for authors
  1. Where do gains come from? What is the contribution of optimizing θ\theta to the risk certificate, and what is the empirical benefit of using qtq_t? Can the authors conduct some ablation studies to clarify this point?
  2. How does RRC (Feldman et al. 2023) perform in this setup compared to your method?
  3. Please compare your method to CRO (Chenreddy et al. 2022) (offline) + RRC and to Online CRO + RRC.
4 comments

Response to Reviewer N7VL

Thank you for your thoughtful comments on the distinction between decision optimization and calibration. All line numbers below refer to the revised manuscript.

Weaknesses

W1(a). My main reservation is the novelty of the proposed approach - it appears to be a combination of CRO (Chenreddy et al. 2022), CRC (Hu et al. 2026), and ACI (Gibbs & Candes, 2021) or RRC (Feldman et al. 2023).

Response. Thank you for raising this point. A smaller valid certificate can arise in two ways: learning a better decision can reduce its underlying upper-tail loss, while tighter calibration can reduce the gap between the certificate and the loss quantile for a given decision. We clarify this distinction in revised Section 3.1 (lines 162–173). OCRC couples these two tasks by using robustness feedback to calibrate an external certificate offset while updating the decision parameters θt\theta_t through a downstream objective. The scalar update follows the familiar ACI or Rolling RC form. Our contribution lies in coupling this update with decision learning to improve decision efficiency while controlling robustness violations.

W1(b). Furthermore, the authors' claim that the downstream robustness violations "cannot be handled by the scalar mechanism used in these methods" seems too strong, as Rolling Risk Control (Feldman et al., 2023) can be employed to control robustness in this setting. Specifically, RRC can be employed with the loss et=qt+1ϕ(Yt,zθt(Xt))>rθt(Xt)e_t=q_t+1\phi(Y_t,z_{\theta_t}(X_t)) > r_{\theta_t}(X_t) and the corresponding update qt+1=[qt+γ(et−α)]+q_{t+1}=[q_t+\gamma(e_t-\alpha)]^+

Response. Thank you for suggesting this baseline. We agree that RRC can control downstream robustness in this setting. We interpret qtq_t in the proposed update as the conformal threshold indexing the prediction set and denote it by τt\tau_t to distinguish it from OCRC’s external certificate offset. In CRO+RRC, the current prediction set determines both the CRO decision and its certificate rtr_t, and the violation indicator records whether the realized loss exceeds that certificate. Updating τt\tau_t using this indicator can control downstream robustness, as our new experiment demonstrates. This feedback updates only the radius through its log scale. The other parameters of the decision rule require a separate learning rule, such as an MSE update for the center. Decision efficiency therefore also depends on these parameters and how they are learned. We now make this distinction in Related Work (lines 121–124). Appendix B.4 explains how each method learns its decision parameters (lines 1118–1123).

W2(a). Regarding the theory, Theorem 2 provides a guarantee only on the smoothed indicator rather than the hard one. Moreover, Theorem 5 requires Assumption 5, which might be too restrictive: it considers the supremum over all sequences of the parameters θ,q\theta,q. The left-hand side of line 326 can be positive even as ε→0\varepsilon\to0, which contradicts Assumption 5.

Response. Thank you for identifying these issues. We agree that Theorem 2 controls smoothed violations rather than hard violations. We also agree that taking a supremum over all parameter sequences is problematic: such sequences can keep the margin arbitrarily close to zero, so the near-zero frequency need not vanish as ε→0\varepsilon\to0.

We have removed this frequency requirement and revised Assumption 5 and Theorem 5 to establish a guarantee for a buffered certificate. For a horizon TT, choose a deterministic δT>0\delta_T > 0 before observing the data, with σ→0\sigma\to0, δT→0\delta_T\to0, and δT/σ→∞\delta_T/\sigma\to\infty; for example, δT=σ\delta_T=\sqrt{\sigma}. The reported certificate is b~t=rt+qt+δT\widetilde b_t=r_t+q_t+\delta_T, where rtr_t is the CRO worst-case loss and qtq_t is OCRC’s online offset. The buffer affects only the reported threshold; Theorem 2 and the algorithm continue to use the smoothed margin relative to rt+qtr_t+q_t (revised lines 354–359). Whenever the realized loss exceeds the buffered certificate, its smoothed violation is at least Hσ(δT)→1H_\sigma(\delta_T)\to1. Theorem 2 therefore yields the hard-violation bound in revised Theorem 5 without any frequency assumption on near-zero margins (lines 360–365).

W2(b). Furthermore, this work does not provide a theoretical guarantee about controlling the robustness, even over an infinite time horizon.

Response. Thank you for asking us to clarify the robustness guarantee. A decision is robust on round tt if its realized loss ϕ(Yt,zt)\phi(Y_t,z_t) does not exceed the reported certificate. For the buffered certificate b~t=rt+qt+δT\widetilde b_t=r_t+q_t+\delta_T, revised Theorem 5 bounds the fraction of rounds on which this condition fails by

α+qT+1/(γT)Hσ(δT),\frac{\alpha+q_{T+1}/(\gamma T)}{H_\sigma(\delta_T)},

where qT+1q_{T+1} is the final offset and γ\gamma is its update step size. Under the default schedule, the numerator tends to α\alpha and the denominator tends to one. Thus, cumulative robustness is at least 1−α−o(1)1-\alpha-o(1) for every sequence satisfying the standing assumptions (revised lines 360–365). Theorem 2 separately bounds the average smoothed violation relative to the unbuffered certificate rt+qtr_t+q_t (lines 289–298).

W2(c). Additionally, Theorem 3 requires the loss function to be convex, which is a strong assumption, and even the authors acknowledge that it is nontrivial to construct and verify.

Response. Thank you for highlighting the scope of the convexity assumption. Theorem 3 characterizes the algorithm in the classical online convex–concave setting. Its saddle-regret benchmark compares the accumulated objective of the online primal–dual iterates with fixed primal and dual choices in hindsight. Convexity in the primal variable allows the proof to convert the primal gradient steps into a sublinear bound on this comparison (lines 308–317).

We acknowledge that convexity can be difficult to ensure for flexible prediction sets. We therefore also analyze the nonconvex case under stationary i.i.d. sampling. Revised Theorem 4 establishes an efficiency guarantee by showing that the average expected squared gradient of the population value function Φ\Phi along the iterates tends to zero. For each fixed decision parameter, Φ\Phi is the expected penalized objective maximized over the certificate offset (lines 323–348).

Questions

Q1. Where do gains come from? What is the contribution of optimizing θ\theta to the risk certificate, and what is the empirical benefit of using qtq_t? Can the authors conduct some ablation studies to clarify this point?

Response. Thank you for asking us to clarify the roles of decision learning and calibration. Optimizing θt\theta_t changes the deployed decision and can reduce its upper-tail loss, while calibration can tighten the reported certificate for a given decision. Revised Section 3.1 makes this distinction explicit through the conditional VaR benchmark (lines 162–173). The offset qtq_t directly adjusts the reported threshold, and κqt\kappa q_t acts locally as a multiplier in the primal gradient, as explained in Remark 1.

To examine the empirical effect of the offset, Appendix B.2 compares OCRC with a baseline that fixes qt=0q_t=0 and uses a standard nonnegative multiplier to weight the smoothed violation penalty. Both methods update the center and shape online from matched initializations. OCRC can also raise the reported certificate directly through its offset, allowing it to reduce next-round violations while the decision parameters continue to adapt. The baseline construction is given in lines 970–986. The following table reports the run with better initialization in Figure 6(b).

Method with better initializationCumulative robustnessMean certificate
OCRC90.20%5.19
Multiplier only89.88%5.32

Both methods finish within 0.20 percentage points of the 90% target in this run. OCRC achieves a smaller mean certificate and less variation in moving-average robustness. With poorer initialization, the multiplier baseline recovers more slowly, remains below the cumulative robustness target, and incurs larger certificates during recovery. We report these results in revised Appendix B.2 (Figure 6(a), lines 1019–1023).

Q2. How does RRC (Feldman et al. 2023) perform in this setup compared to your method?

Response. Thank you for the suggestion. We have added a direct comparison with CRO+RRC using robustness feedback over the same 4,353 Load-Climate-Open rounds, with a target robustness of 90% (revised Appendix B.4, Figure 9, lines 1109–1116).

MethodRobustnessMean certificateMean realized loss
CRO+RRC (robustness)90.54%−0.0599−0.2972
OCRC91.04%−0.0937−0.3236

Both methods achieve robustness close to the 90% target. OCRC attains a lower mean certificate and a lower mean realized loss, measured in standardized units, while achieving slightly higher robustness.

Q3. Please compare your method to CRO (Chenreddy et al. 2022) (offline) + RRC and to Online CRO + RRC.

Response. Thank you for suggesting these comparisons. The new experiment evaluates Online CRO+RRC with two feedback rules (revised Appendix B.4, lines 1082–1093). Both variants start from the same pretrained MLP, update its center online using MSE, and adapt the radius using either miscoverage or downstream violation feedback. Holding the learning rule and prediction-set family fixed, switching to downstream feedback changes robustness from 97.86% to 90.54% and reduces the mean certificate from 0.2407 to −0.0599. OCRC updates the decision parameters through the downstream objective and achieves a mean certificate of −0.0937 at 91.04% robustness (lines 1113–1116).

Thank you for the detailed revision. Theorem 5 and Lemmas 2, 3, 4 address the concerns I raised regarding the proofs. However, my main concern regarding the novelty of OCRC remains, and the new results strengthen it. Therefore, I keep my score.

  1. In the setting of Section 5.1-5.2 or Appendix B.1 (Gaussianity with linear loss), the conditional-VaR-optimal decision is already provided by CRO with the oracle mean and radius Φ−1(1−α)\Phi^{-1} (1-\alpha), so a scalar update from a robustness feedback is anticipated to be approximately optimal. Please run the ACI/PID/RRC baselines with the violation indicator and add their performance to Figures 1, 2, and 5. Right now, it is unclear whether the gains from the synthetic experiment come from OCRC or from the change in the feedback signal.

  2. The real-data comparison does not isolate where the contribution of OCRC comes from. In Appendix B.4, robustness-feedback CRO+RRC already achieves approximately 90% of OCRC's certificate reduction relative to coverage-based calibration. The remaining gap is obtained from a single run against a baseline restricted to an isotropic shape. To obtain a better understanding of the gaps, please run (i) the additive-offset variant I proposed in my review with the same set family as OCRC (ii) the offline CRO+RRC, and report results over multiple seeds

  3. Theorem 5 (the hard-violation guarantee) applies only to the buffered certificate, which is not evaluated in the experiments, while the offset-RRC update gives an unbuffered guarantee for the deployed certificate. Furthermore, the experiments do not satisfy the assumptions required by Theorems 3,4.

  4. In the Load-Climate-Open dataset, are all covariates available when the schedule of the six-hour is chosen?

Reviewer_7CsaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for sequential robust decision-making. The main idea is to control downstream robustness violations directly, rather than relying on the coverage of prediction sets.

The authors introduce an external scalar threshold and jointly optimize it with the parameters of the prediction set using an online primal-dual algorithm.

The paper provides theoretical results on long-run robustness control and convergence. The results suggest that the method can achieve lower risk certificates than coverage-based methods while maintaining a similar robustness level.

Critical strengths
  1. The observation that predictive coverage is sufficient but not necessary for downstream robustness seems to be motivated.

  2. The introduction of an external threshold is intuitive and provides a useful way to separate robustness control from the optimization of prediction sets.

  3. The paper considers both theoretical properties and empirical evaluation, including applications to electricity consumption and retail demand.

Critical weaknesses

x. The robustness guarantees need further clarification

Theorem 2 establishes control over the smoothed violation function, while Theorem 5 translates this result into a bound on actual violations under an additional assumption. Assumption 5 imposes a strong condition on the frequency of observations with robustness margins close to zero. It is not clear whether this assumption can be expected to hold for arbitrary or adversarial data sequences.

Furthermore, the abstract claims asymptotically exact marginal robustness, whereas the presented results mainly establish upper bounds on long-run violation frequencies. They seem to be mismatched.

x. The convergence analysis in Theorem 4 appears incomplete.

The proof of Theorem 4 relies on the stochastic gradient descent-ascent result from Lin et al. (2020). However, the cited theorem concerns stationarity defined through the Moreau envelope in the general nonconvex-concave setting, rather than the gradient of the original value function used in Theorem 4.

The cited analysis also assumes a bounded dual domain, whereas the external threshold in OCRC is defined over a non-negative, unbounded domain. It is not clear that the stated convergence result follows directly from the cited theorem.

x. Experimental comparisons might not fully isolate the contribution of OCRC

ACI and Conformal PID use prediction sets based on fixed conditional means and covariances, whereas OCRC jointly optimises a more flexible parameterisation of the prediction sets. Thus, the observed improvements may come partly from the additional flexibility in prediction-set optimisation rather than from the proposed robustness-control mechanism alone. The authors could consider methods using the same prediction model and prediction-set parameterisation, while changing only the calibration mechanism.

Questions for authors

Please try to address the major concerns above.

Some minor comments:

x. The synthetic experiment uses Gaussian observations and an unbounded linear loss, which do not directly satisfy the boundedness assumption used in the theoretical analysis. This discrepancy should be discussed more carefully.

x. The parameter convergence statement in Theorem 1 appears to require uniqueness of the optimal solution. Without uniqueness, convergence to the set of optimal solutions may be a more appropriate statement.

x. The experimental protocol could provide more details about hyperparameter selection, model initialization and the chronological use of data for pretraining.

3 comments

Response to Reviewer 7Csa

Thank you for your careful assessment of the theoretical results and experimental comparisons. Your comments helped us clarify the guarantees and strengthen the evaluation. All line numbers below refer to the revised manuscript.

Weaknesses

W1(a). Theorem 2 establishes control over the smoothed violation function, while Theorem 5 translates this result into a bound on actual violations under an additional assumption. Assumption 5 imposes a strong condition on the frequency of observations with robustness margins close to zero. It is not clear whether this assumption can be expected to hold for arbitrary or adversarial data sequences.

Response. Thank you for pointing out that an adversarial sequence can keep the margins arbitrarily close to zero. We have removed the original near-boundary frequency assumption and revised Theorem 5 to establish hard-violation control for a buffered certificate.

For a run of length TT, we choose a deterministic buffer δT>0\delta_T > 0 before the run, with σ→0\sigma\to0, δT→0\delta_T\to0, and δT/σ→∞\delta_T/\sigma\to\infty as T→∞T\to\infty; for example, δT=σ\delta_T=\sqrt{\sigma}. The reported certificate is

b~t=rθt(Xt)+qt+δT,\widetilde b_t=r_{\theta_t}(X_t)+q_t+\delta_T,

where rθt(Xt)r_{\theta_t}(X_t) is the worst-case loss over the prediction set and qtq_t is OCRC's online offset. The buffer is added only to the reported certificate; the algorithm continues to update using the original, unbuffered margin (lines 354–359).

Whenever the realized loss exceeds b~t\widetilde b_t, the original margin exceeds δT\delta_T, so the smoothed violation is at least Hσ(δT)→1H_\sigma(\delta_T)\to1. Theorem 2 therefore yields an upper bound of α+o(1)\alpha+o(1) on the fraction of hard violations under the default setting. This result applies to arbitrary sequences, including adversarial sequences, subject to the standing assumptions, such as bounded loss, without any restriction on the frequency of near-zero margins (lines 360–365).

W1(b). Furthermore, the abstract claims asymptotically exact marginal robustness, whereas the presented results mainly establish upper bounds on long-run violation frequencies. They seem to be mismatched.

Response. Thank you for pointing out this mismatch. We have revised the abstract to distinguish long-run control of smoothed violations from the upper bound on hard violations obtained with a vanishing buffer (lines 20–23):

For arbitrary data sequences, we establish long-run control of smoothed violations and an upper bound on hard violations when the reported certificates include a vanishing buffer.

W2(a). The proof of Theorem 4 relies on the stochastic gradient descent-ascent result from Lin et al. (2020). However, the cited theorem concerns stationarity defined through the Moreau envelope in the general nonconvex-concave setting, rather than the gradient of the original value function used in Theorem 4.

Response. Thank you for identifying this gap. The cited SGDA result controls the gradient of a Moreau envelope, whereas Theorem 4 concerns the ordinary gradient of the population value function

Φ(θ)=max⁡q≥0ℓ(θ,q),\Phi(\theta)=\max_{q\ge0}\ell(\theta,q),

where ℓ\ell is the expected penalized objective. We have added the argument connecting these two gradients.

First, the expected smoothed violation is strictly decreasing in qq, which yields a unique maximizing offset. Its variation is bounded by the uniform Lipschitz variation of the loss margin. Differentiating the objective at this maximizer then shows that Φ\Phi has a Lipschitz gradient with constant KT=O(κ/σ)K_T=O(\kappa/\sigma) (Lemma 3, Eq. (9) and its derivation in lines 817–829).

Next, let λ=1/(2sT)\lambda=1/(2s_T), where sTs_T is the joint smoothness bound. If pp is the proximal point at θ\theta, the Moreau-envelope gradient satisfies

g=(θ−p)/λ=∇Φ(p).g=(\theta-p)/\lambda=\nabla\Phi(p).

Consequently,

∣∇Φ(θ)∣≤(1+λKT)∣g∣.\\|\nabla\Phi(\theta)\\|\le(1+\lambda K_T)\\|g\\|.

Because KT/sT=O(σ)K_T/s_T=O(\sigma), the squared multiplicative factor tends to one, preserving the convergence rate (lines 911–915). Revised Theorem 4 thus bounds the average squared ordinary gradient along the iterates under i.i.d. sampling, zero-offset initialization with expected smoothed violation at most the target, and unprojected primal updates (lines 323–348).

W2(b). The cited analysis also assumes a bounded dual domain, whereas the external threshold in OCRC is defined over a non-negative, unbounded domain. It is not clear that the stated convergence result follows directly from the cited theorem.

Response. Thank you for highlighting the mismatch between the dual domains. Bounding the iterates alone would not rule out a larger value of the maximization defining Φ\Phi outside their range. Revised Lemma 2 addresses both the maximizers and the iterates.

Let Wθ=ϕ(Y,zθ(X))−rθ(X)W_\theta=\phi(Y,z_\theta(X))-r_\theta(X) denote the realized loss minus the CRO certificate before adding the offset. The uniform loss bound implies Wθ≤2BϕW_\theta\le2B_\phi. For

q>βσ=max⁡0,2Bϕ−Hσ−1(α),q > \beta_\sigma=\max\\{0,2B_\phi-H_\sigma^{-1}(\alpha)\\},

every sample satisfies Hσ(Wθ−q)<αH_\sigma(W_\theta-q) < \alpha. The expected objective is therefore strictly decreasing in this region, so all maximizers lie in [0,βσ][0,\beta_\sigma].

The sample update also decreases qq above βσ\beta_\sigma, and can overshoot this threshold by at most γ(1−α)\gamma(1-\alpha), where γ\gamma is the offset step size. Under the default schedule, this yields a common bound QQ independent of TT. Restricting the maximization to [0,Q][0,Q] leaves Φ\Phi unchanged, and projection onto this interval leaves the iterates unchanged. The revised SGDA argument therefore uses this fixed compact dual domain (lines 766–785).

W3. ACI and Conformal PID use prediction sets based on fixed conditional means and covariances, whereas OCRC jointly optimises a more flexible parameterisation of the prediction sets. Thus, the observed improvements may come partly from the additional flexibility in prediction-set optimisation rather than from the proposed robustness-control mechanism alone. The authors could consider methods using the same prediction model and prediction-set parameterisation, while changing only the calibration mechanism.

Response. Thank you for suggesting a comparison with matched prediction-model flexibility. We have made two changes. In the real-data experiments, ACI and Conformal PID now update the same type of mean-predictor MLP using squared prediction error. Their downstream certificates are computed from the prediction sets after scalar coverage calibration, giving the baselines the same mean-model capacity as OCRC (lines 391–395).

We have also added a comparison in Appendix B.1 that fixes the center and shape for all three methods and allows each method to adjust a single scale. At a 90% robustness target, the results are as follows (Figure 5, lines 959–966).

MethodCoverageRobustnessMean certificate
OCRC25.18%90.70%5.02
ACI89.92%98.38%5.76
Conformal PID89.74%98.24%5.71

With the same prediction-set family, OCRC achieves 90.70% robustness and a mean certificate of 5.02. The coverage-based baselines exceed the robustness target by more than eight percentage points and produce larger mean certificates. Thus, the observed advantage persists when prediction-set flexibility is matched.

Questions and minor comments

Q1. The synthetic experiment uses Gaussian observations and an unbounded linear loss, which do not directly satisfy the boundedness assumption used in the theoretical analysis. This discrepancy should be discussed more carefully.

Response. Thank you for pointing out this discrepancy. Revised Assumption 1 explicitly requires ∣ϕ(y,z)∣≤Bϕ|\phi(y,z)|\le B_\phi uniformly over outcomes and feasible decisions. This bounds the absolute value of the worst-case certificate by BϕB_\phi and the unbuffered loss gap before adding qtq_t by 2Bϕ2B_\phi (lines 274–275).

In the stationary simulation, however, ϕ(y,z)=−y⊤z\phi(y,z)=-y^\top z and the Gaussian outcomes have unbounded support. We now explicitly state that this simulation is an empirical evaluation outside the scope of the uniform loss assumption (lines 399–406).

Q2. The parameter convergence statement in Theorem 1 appears to require uniqueness of the optimal solution. Without uniqueness, convergence to the set of optimal solutions may be a more appropriate statement.

Response. Thank you for pointing out the issue with the parameter-convergence statement. We have revised Theorem 1 to state that the distance from the penalized minimax solution θκ∗\theta_\kappa^* to the original feasible set tends to zero. Specifically, this set is D=θ∈Θ:g(θ,0)≤0D=\\{\theta\in\Theta:g(\theta,0)\le0\\}, where g(θ,0)g(\theta,0) is the constraint evaluated at zero offset. The theorem also establishes convergence of the optimal values to the original constrained optimum and qκ∗→0q_\kappa^*\to0 as κ→∞\kappa\to\infty (lines 251–265).

Q3. The experimental protocol could provide more details about hyperparameter selection, model initialization and the chronological use of data for pretraining.

Response. Thank you for requesting these implementation details. For pretraining, we use Adam with a learning rate of 10−310^{-3} and early stopping, retaining the checkpoint with the lowest validation mean squared error. The online step sizes remain fixed within each run. The notebooks specify the method-specific step sizes and initial prediction-set scales.

All real-data methods are initialized from copies of the same pretrained MLP, as clarified in the revised experimental description (lines 391–395). Pretraining and validation precede online evaluation in chronological order. Load-Climate-Open uses the first 40% of the data for standardization and pretraining (lines 484–485). DEPI uses the first 40% for MLP fitting, the next 10% for validation, and the final 50% for online evaluation.

Comments on the paper

1 comment

General Response

We thank the reviewers for their careful and constructive feedback. We revised the manuscript to clarify the decision objective, make the theoretical guarantees more precise, and test which parts of OCRC account for the empirical gains. The main revisions are summarized below.

Decision objective and exposition (Abstract, Introduction, and Section 3). We added a factory electricity-scheduling example that connects the prediction set, the decision, and the risk certificate (lines 51–57). We also introduced conditional value at risk as a benchmark. A smaller valid certificate can come from learning a decision with lower upper-tail loss or from calibrating the certificate more tightly for a given decision (lines 162–173). OCRC couples these two tasks by using downstream robustness feedback while optimizing the parameters of the decision rule.

Robustness guarantees (Assumption 5 and Theorem 5). We replaced the condition on the frequency of near-zero margins with a deterministic buffer δT\delta_T, chosen before the run. For example, δT=σ\delta_T=\sqrt{\sigma} vanishes as the smoothing scale σ\sigma goes to zero while remaining large relative to σ\sigma. The revised theorem bounds the fraction of rounds on which the realized loss exceeds the buffered certificate rt+qt+δTr_t+q_t+\delta_T by α+o(1)\alpha+o(1) under the default schedule (lines 354–365). We revised the abstract to distinguish this guarantee from Theorem 2’s control of smoothed violations (lines 20–23).

Optimization guarantees (Theorems 1 and 4 and Appendix A). Theorem 1 now states that the penalized solutions approach the original feasible set (lines 251–265). For the stationary nonconvex setting, we rewrote the proof of Theorem 4 to justify a compact threshold domain that preserves the value function and the algorithm’s iterates (lines 766–785). The proof then converts the Moreau-envelope gradient bound into a bound on the ordinary gradient of the population value function (lines 911–915).

Experimental protocol and ablations (Section 5 and Appendices B.1–B.2). In both real-data experiments, all methods start from the same pretrained MLP, and ACI and Conformal PID now update their MLPs online using squared prediction error (lines 391–395). We clarified the chronological pretraining split in Load-Climate-Open and stated which certificate each method reports. We added a fixed-center comparison with the same prediction-set family across methods, and a comparison with a conventional multiplier to assess the effect of letting the offset directly adjust the certificate.

Comparison with Rolling Risk Control (Appendix B.4, Figure 9). We added two online CRO+RRC baselines that use the same online MSE update for the center and the same scalar radius update, with different feedback signals. On Load-Climate-Open, switching from miscoverage to downstream violations brings robustness from 97.86% to 90.54%, close to the 90% target. OCRC attains 91.04% robustness with a lower mean certificate (−0.0937 versus −0.0599) and lower mean realized loss (−0.3236 versus −0.2972), in standardized units (lines 1109–1116). The revised Related Work explains that RRC can control downstream violations by tuning the radius. The remaining decision parameters require a separate update rule. OCRC couples robustness control with learning the multidimensional parameters that determine the decision (lines 121–124).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.