acceptodds
Under review as a conference paper at ICLR 2027

Pretrain-Then-Calibrate: Enhancing Optimality-Feasibility Tradeoff in Safe Contextual Optimization

26 views

Abstract

Contextual optimization, or decision-focused learning, aims to learn context-based decision rules for downstream decision-making tasks. In addition to maintaining statistical efficiency like in conventional supervised learning, this problem entails delicate interplays with optimization tractability and the effectiveness of data-optimization integration. In this paper, we consider safe contextual optimization where, in addition to an optimality criterion, the decision rule also satisfies a high-probability-enforced risk criterion. Natural attempts to integrate chance-constraint methods such as robust optimization, with established contextual optimization pipelines such as predict-then-optimize, suffers from a high price of optimality in exchange for an adequate feasibility guarantee. This motivates us to propose a pretrain-then-calibrate framework to significantly enhance the optimality-feasibility tradeoff. Specifically, our framework translates any heuristic decision rules into rigorous statistically feasible rules via what we call a feasibility calibrator, constructed from a careful interpolation against a safe reference rule to possess certifiable joint feasibility-optimality guarantees. At a high level, our approach exploits the low-dimensional interpolation from readily obtainable good heuristics, and is advantageously model-agnostic, modular and consequently implementation-friendly. In synthetic and real-data experiments on operations management and portfolio optimization, pretrain-and-calibrate results in solutions significantly more optimal than a range of benchmarks including conformal and (distributionally) robust optimization methods without sacrificing feasibility, thus achieving demonstrably improved optimality-feasibility tradeoffs.

Primary area:
optimization
Keywords:
contextual optimization, data-driven optimization, chance constraint programming

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.60 over 5 reviews, above 70% of the others.
Reviewer_8KPKOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a pretrain-then-calibrate methodology to address contextual optimization problems under chance constraint satisfaction. The authors propose an interpolation method where the new predictor is formed as a convex combination of a pretrained decision rule and a fixed feasible decision rule. To determine the appropriate convex combination parameter, they introduce four calibration methods that select the parameter over a validation set. The paper provides calibration guarantees for these methods and presents experimental results demonstrating the effectiveness of their approach in mitigating the feasibility-optimality tradeoff.

Critical strengths

The paper addresses an important challenge currently faced by the contextual optimization community: designing contextual optimization methods with strong feasibility guarantees. The paper proposes a clear and intuitive approach that achieves a favorable trade-off between feasibility and optimality. To calibrate the decision rule, the paper introduces four distinct calibration methods and provides theoretical feasibility guarantees for each resulting rule. The experimental results were run on both synthetic and real data, and indicate that performance depends on the specific calibration rule selected which requires empirical testing to choose the best rule for a given context; the method overall offers a superior compromise on the feasibility-optimality trade-off. Furthermore, the approach is agnostic to the choice of the pretrained decision rule, requiring only the calibration step in the validation phase making the method compatible with most of the existing DFL methods proposed in the literature.

Critical weaknesses

While the paper's strength is that the proposed framework is agnostic, the paper criticizes existing methodologies that deal with contextual optimization with chance-constraints and views them as overly-conservative methods. However, the paper doesn’t offer a substitute approach, it rather proposed method that relies on pre-trained decision rules existing in the literature to build an interpolation rule to mitigate the feasibility-optimality tradeoff. Therefore, the proposed method is a calibration layer rather than a totally new approach to deal with the underlying problem. This should be clarified in the positioning of the paper.

The decision policy introduced by the paper relies heavily on the pretrained decision rule and the fixed feasible rule. However, the paper does not detail these 2 policies. From the pretrained decision side, although agnostic, the paper would greatly benefit from theoretical results that benchmarks the optimality and feasibility guarantees of existing methods compared to the interpolated policy to show the improvement that this method provided. From the fixed feasibility rule w0, the choice of such a rule and how to obtain such a rule for our application has not been discussed by the paper. Additionally, the quality of such a rule with respect to both metrics and how the choice of such policy affects the method are not discussed theoretically.

In the experimental section, only a single decision rule is used (although the model varies from linear to NN), which limits the empirical demonstrating that the method is model-agnostic. Although additional methods are introduced in the appendix, the main text would benefit significantly from an ablation study across diverse DFL approaches to highlight this advantage.

Minor comments :

The main idea in the paper claims to offer a rule with rigorous feasibility guarantees (page 2 line 90), yet the text also suggests that strict feasibility is not explicitly sought (page 5 line 242). The framing and terminology should be revised for consistency and precision.

Lemma 1 does not appear to be utilized or referenced anywhere in the main analysis; its necessity should be clarified .

In most applications, being able to control the tradeoff between feasibility and optimality might be preferred.Beyond \alpha, can this method be adapted so that the decision maker has more control over this tradeoff and can set a threshold for either optimality or feasibility to fit their particular applications?

The title safe contextual optimization might be a bit misleading

Questions for authors

1- Could the authors elaborate on how practitioners should construct the fixed feasible rule w₀ in practice? How sensitive is the empirical optimality to the choice of w₀, for instance when comparing an overly conservative rule against a nearly optimal choice?

2- How does the quality of w₀ across feasibility and optimality metrics impact the performance of the overall interpolated policy, and can this relationship be formally characterized theoretically?

3- To strengthen the model-agnostic claims, could the main text include an ablation study benchmarking diverse DFL frameworks as the base pretrained rule, rather than evaluating a single decision rule across different network architectures?

4- The paper characterizes existing chance-constrained contextual optimization techniques as overly conservative. Could you clarify this point further? Also could you discuss whether those baseline methods can be integrated into your framework, and address if applying this calibration layer helps alleviate their inherent conservatism?

5- Could the authors reconcile the claim of rigorous feasibility guarantees (page 2, line 90) with the subsequent remark indicating that strict feasibility is not explicitly sought (page 5, line 242)?

6- What specific role does Lemma 1 play within the theoretical framework?

7- Beyond adjusting \alpha, can this calibration framework be extended to grant decision-makers explicit control by enforcing targeted thresholds on either feasibility or optimality?

8- Given that the framework relies on existing pretrained decision rules rather than replacing them, would it be more accurate to position this work as a modular calibration layer? Could the framing in the introduction be updated accordingly?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 8KPK (part 1)

We sincerely appreciate your detailed review and thoughtful comments on our paper.

Response to W1 (methodology contribution):

We have clarified our contribution in the revised introduction. We politely mention that our whole pretrain-then-calibrate pipeline is exactly the substitute (and indeed better) approach to deal with this problem. Please read our general response on method contribution on the novelty, strength and practicality of our framework for more details.

Response to W2 (dependence of decision policy on the pretrained decision rule / feasible rule):

We provide a practical guideline for choosing the two rules at the end of Section 3. We recommend selecting wprew_{\mathrm{pre}} for strong downstream objective performance, using either task-only training (with a feasibility heuristic if needed); rigorous feasibility is not required at this stage. In contrast, w0w_0 should provide readily verifiable safety through problem-specific knowledge, without good objective performance. These complementary roles allow calibration to exploit the pretrained rule’s objective performance as well as provide feasibility.

We emphasize that the flexibility in choosing pretrained decision rules is a key advantage of our framework. Our method does not prescribe a particular pretraining algorithm or decision-rule class, allowing practitioners to leverage expressive models such as neural networks, to get an objective-driven decision rule. Importantly, our calibration guarantees are independent of the complexity of the pretrained decision-rule class. This enables us to obtain statistical feasibility guarantees for flexible, objective-driven decision rules without requiring model-specific feasibility analysis, model retraining, or downstream robust optimization. Our experiments with linear, neural, and RF-SAA rules further demonstrate this flexibility.

Regarding theoretical comparisons, our Theorems 1 and 2 already establish statistical feasibility guarantees, with Theorem 2(ii) additionally providing a pathwise optimality guarantee. Rather than claiming universal theoretical dominance over existing optimization approaches, which would require method-specific assumptions, we evaluate their relative performance through extensive empirical comparisons in Section 4.3 and Appendix E. These results demonstrate favorable optimality–feasibility tradeoffs across different applications, highlighting the simplicity, generality, and practical effectiveness of our framework.

Regarding the construction of the safe reference rule w0w_0, and its impact on objective performance, we refer the reviewer to our general response.

Response to W3 (limited use of a single decision rule):

We understand this concern as involving two complementary aspects: (1) whether our framework can accommodate decision rules obtained from different pretraining methodologies, and (2) how our complete pipeline compares with existing contextual and decision-focused optimization approaches.

Regarding the first aspect, our framework can accommodate decision rules from different pretrained methodologies. We have added a new ablation study in Appendix D.5. Specifically, we consider task-only neural training, prediction-first learning (MSE followed by optimization), and decision-focused learning (task-based learning/SPO+). We use the resulting decision rules as pretrained inputs to the same calibration framework, without retraining or modifying the pretrained models. The results demonstrate effective feasibility calibration across these distinct pretraining methodologies, providing direct empirical support for our model-agnostic claim.

Regarding the second aspect, our framework outperforms existing contextual and decision-focused optimization approaches empirically. Section 4.3 compares our complete pipeline against CPO, LRO, CPP, and Trim-DRO, while Appendix E provides additional comparisons with SPO-RC+ and ICRC-based robust optimization. These experiments demonstrate the practical competitiveness of our framework against alternative contextual optimization pipelines.

Finally, we emphasize that our theoretical calibration guarantees are independent of the specific pretraining algorithm and the complexity of its hypothesis class. Together with the new pretraining-pipeline ablation and existing benchmark comparisons, these results support both the generality and practical effectiveness of our approach.

AuthorsOpenReview ↗

Response to Reviewer 8KpK (Part 2)

Regarding the minor comments, we list them as M1-M4 respectively.

Response to M1:

The two statements refer to different stages of our framework: during pretraining, we do not require rigorous feasibility guarantees, allowing us to prioritize objective performance; the subsequent calibration stage then establishes statistical feasibility guarantees. We have revised the wording on page 5 using blue color to clarify this distinction and ensure consistency.

Response to M2:

We thank the reviewer for pointing this out. Lemma 1 establishes the nesting property, which is essential to the conformal calibration guarantee (Theorem 1) and the DKW calibration guarantee (Theorem 2). We have added explicit references to Lemma 1 in the revised manuscript to clarify its role in these results.

Response to M3:

Our framework follows the standard chance-constrained programming formulation, where the risk tolerance α\alpha serves as a natural and interpretable parameter for controlling the optimality–feasibility tradeoff. Decision makers can specify their desired feasibility level through α\alpha, and our calibration procedure seeks a low-cost decision rule satisfying the corresponding statistical feasibility requirement. Different prespecified values of α\alpha allow practitioners to explore different tradeoffs within the same framework.

We also appreciate the suggestion of specifying an optimality threshold instead. This leads to a related satisficing-style formulation, where one seeks to maximize feasibility subject to an acceptable cost level (Chen & Sim, 2009; Long et al., 2022) . While this is a meaningful alternative, it requires a different calibration objective and corresponding statistical analysis. We view it as an interesting direction for future research, while the present work focuses on the standard chance-constrained formulation.

Response to M4:

We thank the reviewer for this suggestion. We have revised the title to “Pretrain-Then-Calibrate: Enhancing the Optimality–Feasibility Tradeoff in Chance-Constrained Contextual Optimization” to better reflect the probabilistic nature of our feasibility guarantees.

Response to Q1:

Yes, we have added related discussion. Please see our general response on w0w_0 discussion, especially on how to choose them using domain knowledge and in a data-driven way, its sensitivity, and its impact on objective performance.

Response to Q2:

The impact of w0w_0 on the objective performance depends on the interpolation path. Theorem 2(ii) provides a formal guarantee relative to the best feasible candidate along this path, while Appendix D.7 empirically examines sensitivity to different reference rules. Please see our general response on w0w_0 for further discussion.

Response to Q3:

Please see our previous response on weakness 3.

Response to Q4:

Existing (distributionally) robust optimization approaches often enforce feasibility through learned uncertainty sets, followed by worst-case optimization. While these approaches provide feasibility protection, they can be overly conservative because protecting against all realizations within an uncertainty set may impose stronger requirements than the original chance constraint. This can lead to unnecessary sacrifices in objective performance, as illustrated by our empirical comparisons in Section 4.3 and Appendix E.

Our framework takes a fundamentally different approach. Instead of starting from an overly conservative decision and attempting to improve its optimality, we start from an objective-driven pretrained decision rule with potentially strong optimality performance but without feasibility guarantees. We calibrate this rule toward a safe reference decision, adjusting it sufficiently to achieve statistical feasibility guarantees while preserving objective performance. Our approach directly calibrates constraint-satisfaction probabilities, without requiring learned uncertainty sets or downstream robust optimization. This is the key mechanism through which we aim to mitigate excessive conservatism.

Regarding integration, our model-agnostic framework can in principle accommodate decision rules produced by existing RO/DRO or contextual optimization methods. However, applying safety-oriented calibration to an already conservative rule does not necessarily improve its objective performance, since our interpolation moves toward greater feasibility instead of reversing existing conservatism. Our primary contribution is therefore an optimization pipeline that avoids introducing unnecessary conservatism during pretraining.

Response to Q5:

We have fixed this issue in our revised paper and in our previous response.

Response to Q6:

Please see our previous response on minor comments.

Response to Q7:

Please see our previous response on minor comments.

Response to Q8:

Please see our previous response on weakness 1.

Reviewer_BF4COpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how to improve the safety of pretrained contextual decision rules while preserving objective performance. It interpolates a pretrained rule with a known safe reference and uses independent calibration data to select the adjustment. By assuming that the constraint function is convex in the decision for each realization---a condition natural in inventory and portfolio applications---the authors establish a key structural property: constraint-satisfaction events are nested along the interpolation path. Among the paper's four calibrators---conformal (C), DKW (D), and normalized/unnormalized Gaussian (N/U)---C exploits nesting to control expected violation via a conformal quantile argument, while D uses it to construct simultaneous confidence bounds for high-confidence feasibility. Experiments across several applications assess feasibility and objective performance.

Critical strengths
  1. The framework is simple, modular, and computationally light, providing a practical way to add statistical feasibility control to different pretrained rules without modifying their training procedures.

  2. The experiments cover multiple model classes, synthetic and real data, and several optimization tasks. The sensitivity and ablation studies provide useful evidence of practical applicability.

Critical weaknesses
  1. The methodological advance appears limited. Although the conformal and DKW calibrators exploit the interpolation construction and the resulting nesting structure, their arguments remain relatively direct. The Gaussian calibrators apply existing validation schemes.

  2. The evidence for the central claim of a better optimality--feasibility tradeoff remains incomplete. Comparisons at different achieved reliability levels leave unclear how close the methods are to an efficient frontier and how predictably conservatism can be adjusted while preserving the stated guarantees.

Questions for authors

Q1. The role of convexity and nesting.

Assumption 1 requires the constraint function g(w,z)g(w,z) to be convex in ww for every realization zz, a condition underlying the nesting property used by calibrators C and D.

(a) Could this requirement be relaxed to a weaker geometric condition on the feasible sets, such as star-shapedness with respect to the safe reference decision? Would this cover additional practically relevant applications?

(b) Could the authors discuss natural nonconvex constraints for which nesting fails? For calibrator D, a generic union bound would retain high-confidence feasibility but introduce an additional log⁡p\log p term under the square root in the calibration margin. Could weaker structural properties reduce the resulting conservatism and its impact on objective performance? Separately, could calibrator C be adapted to retain its expected-violation guarantee without full nesting?

Q2. The precision and efficiency of the optimality--feasibility tradeoff.

In the linear newsvendor experiment in Table 1, D/N/U attain test-pass rates of 98.5--100% at 95% nominal confidence, with mean violation rates of 6.11--7.76%. Their costs are higher than C, whose mean violation rate is closer to the 10% target (9.48%), but whose test-pass rate is only 70.75%.

(a) Could the authors characterize how predictably conservatism and objective performance can be adjusted while retaining the stated feasibility guarantee? The sensitivity analysis in β\beta is helpful, but lowering the nominal confidence also weakens that guarantee.

(b) At a given test-pass rate, how close is the achieved cost to the lowest attainable with the same pretrained rule, interpolation path, and calibration sample size? Could the authors plot cost against test-pass rate for the reported calibrators to assess their relative efficiency and the scope for further improvement toward an efficient frontier?

4 comments
AuthorsOpenReview ↗

Response to reviewer BF4C (part 1)

Thank you for recognizing our significance and strengths, and for the thoughtful comments. We address your comments and questions as follows.

Response to W1

Our approach is easy to implement, yet with theoretical guarantees essentially at the same level as existing chance-constraint methods. It is easy to use in practice and computationally lightweight. The simplicity of the approach, including the simplicity of the theory, is our main advantage. Please see our general response on method contribution on the novelty, simplicity and strength of our framework for more details.

Response to W2

We would like to clarify that our central claim of an enhanced optimality–feasibility tradeoff is justified relative to existing contextual and chance-constrained optimization approaches, including robust optimization, DRO, and conformal predict-then-optimize methods with the same type of theoretical guarantees and superior performance across various empirical studies.

We also point out the theoretical difficulty of attaining the theoretically optimal efficient frontier due to the inherent difficulty in identifying the optimal objective in safe contextual optimization problems. We refer more details to our discussion on path-wise optimality v.s. global optimality in our general response.

Comparing our approaches and the existing literature in Section 4.3 (Table 2), we evaluate them under the same target violation probability (α=0.10)(\alpha=0.10), with both objective performance and achieved violation rates reported. Our approach achieves better objective performance by avoiding excessive conservatism. This illustrates the improvement in the optimality–feasibility tradeoff that our framework aims to achieve.

AuthorsOpenReview ↗

response to reviewer BF4C (part 2)

Response to Q1a:

(a) Yes, our requirement can be relaxed to a weaker geometric condition as the star-shapeness with respect to the safe reference decision.

First, we note that Gaussian calibrators (N/U) do not require convexity of the constraint function or the resulting nesting property for their feasibility guarantees. In general, it suffices for the feasible sets {w:g(w,z)≤0}\{w:g(w,z)\leq 0\} to be star-shaped with respect to the safe reference decision w0(x)w_0(x). We have added the relaxations In Appendix C.3, and discussed the star-shaped conditions such that our theoretical guarantees are valid. Under this condition, Lemma 1 still holds, and Theorems 1 and 2 remain valid.

Compared with the convexity condition, the star-shapeness condition with respect to the safe reference decision covers additional practically relevant applications. For example, when we have cardinality constraints, in inventory or resource allocation problems with only a few changes to the baseline safe reference rules, a portfolio with cash and a cardinality limit on risky holdings. However, we adopt convexity throughout the main body, because it is a standard, verifiable structural assumption in optimization that naturally covers a broad range of applications, including all four applications considered in our paper, whose constraint functions are affine or max-affine.

Response to Q1b:

(b) For general nonconvex constraints, nesting may fail. As the reviewer points out, high-confidence feasibility can still be established through a Hoeffding union bound, with a margin of O(log⁡(p/β)/nval)O(\sqrt{\log(p/\beta)/n_{\mathrm{val}}}). To ensure the nesting property, we can weaken the convexity assumption to the star-shaped constraint set with respect to gg. We provide more details in Appendix C.3. For conformal calibration, one possible extension is to construct monotone satisfaction indicators Y~ij=min⁡k≥jYik\widetilde Y_{ij}=\min_{k\geq j}Y_{ik} and apply conformal risk control to these conservative indicators, retaining the expected-violation guarantee.

Response to Q2a: We have added Figure 7 and a brief discussion in Appendix D.7 to compare mean test cost directly against observed test-pass rate.

(a) Adjusting conservatism while retaining the guarantee. We agree that lowering nominal confidence weakens the feasibility guarantee. Figure 6 examines this sensitivity. The validation-size analysis in Figure 5 provides a complementary comparison at fixed target risk and nominal confidence: for DKW, the certification margin is proportional to nval−1/2n_{\mathrm{val}}^{-1/2} at fixed confidence, so increasing calibration sample size reduces the margin while preserving the finite-sample guarantee. The Gaussian guarantees retain the approximation-error terms stated in Theorem 2.

Response to Q2b:

(b) Cost versus test-pass rate. We understand this question as asking whether the reported calibrators incur different costs at comparable levels of empirical feasibility. The new Figure 6 directly examines this comparison using the linear-newsvendor subset of the original confidence sweep, holding the pretrained rules, interpolation paths, target risk, and calibration samples fixed across calibrators. The figure plots mean test cost against the fraction of repetitions whose test violation rate meets the target. The right panel enlarges the high test-pass region. C appears as a single reference point because its criterion does not depend on nominal confidence. The dashed line joins non-dominated tested settings and summarizes the best observed cost–feasibility tradeoffs within the evaluated family. The Gaussian calibrators generally lie closer to this envelope than DKW in this experiment, indicating room for lower cost at comparable observed pass rates.

Thank you for the detailed response and additional analyses. I am satisfied with the clarifications. My original rating of 3 still reflects my overall assessment, so I will keep it unchanged.

AuthorsOpenReview ↗

Thank you!

Dear reveiwer BF4C, we are glad that you are satisfied with our clarifications and hope our response may address your concerns. Thank you again for your time and effort in reviewing our work!

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

The paper proposed a Pretrain-then-calibrate framework for decision making for downstream tasks, which not only contains the statistical efficiency but also enhance the optimality-feasibility tradeoff.

Critical strengths
  1. The paper is well-written with easy read introduction and rigorous statistical guarantee supports.
  2. The problem setup section is clear in statistical way with clear definition of the population where the expectation is taken over.
  3. The experiments part is sufficient with synthetic, real-world datasets, and additional applications in appendix.
Critical weaknesses
  1. The introduction is too long now for read. It combines the introduction and related work together. IMHO, for ease the readers' burden, it would be helpful to separate them to two sections. Also a simple illustrative figure would also help for understanding the setting too.
  2. Since the related work is merged in the introduction, I think the novelty of the paper should have further illustration. As I know that in Zhu's paper (https://arxiv.org/abs/2510.07750) they already proposed the calibrator with conformal risk control.
  3. In section 2.2, there is missing full name of the DKW calibrator which only shows in the method section in abbreviation way.
  4. The method contribution is marginal, with only providing different calibrators to select the jj. Also in the abstract, the main selling point is the theoretical guarantee, but in the paper, the experimental results take up more pages than the theory.
Questions for authors
  1. In the first paragraph of the method section, the authors only introduce the g(w,z)≤0g(w, z)\leq 0 as a constraint. But what does the g(⋅)g(\cdot) function means here? Does it mean an goal or anything? I would kindly ask that more illustration of constrains. Same as the J\mathcal{J}, in this para only says that it is a field generated by training data.
  2. Is the convexity assumption of ω\omega is feasible or easy to meet in applications? Is there any sensitivity analysis about that? Also in the paper, the author assume that there is an exist ω0\omega_0 which is almost surely satisfying the constraint. Does it reliable in real-world application? and How restrictive are these conditions in practice?
  3. In the current version, jj is selected through the grid search. What about jj is a smooth parameter through (0,1)(0, 1)?
3 comments
AuthorsOpenReview ↗

Response to Reviewer pCqe (part 1)

We sincerely appreciate your detailed review and thoughtful comments on our paper. We clarify your questions and comments as follows.

Response to W1:

In the revised manuscript, we have streamlined the introduction part by moving the less relevant literature discussions to Appendix and placing greater emphasis on the motivation, methodological insights, and contributions of our framework. More specifically, we retain discussions of the most relevant literature to motivate the research problem and highlight the gaps in existing approaches, while deferring a more comprehensive and separately organized literature review in Appendix A.

Additionally, we have added an illustrative figure to the Introduction that visually presents our pretrain-then-calibrate pipeline, including the pretrained and safe reference rules, interpolation path, and feasibility calibration. We believe this figure helps readers better understand the key ideas and advantages of our methodology.

Response to W2:

While Zhou & Zhu (2026) (https://arxiv.org/abs/2510.07750) also employ conformal risk control, their methodology is fundamentally different from ours. Their framework calibrates the robustness levels of policies generated through uncertainty-set-based robust optimization, requiring separate robust optimization solves for different candidate robustness levels. In contrast, our framework directly constructs a family of decision rules through convex interpolation and calibrates their constraint-satisfaction probabilities, without learning uncertainty sets or repeatedly solving robust optimization problems.

This distinction is central to our methodological contribution: we enable flexible, objective-driven decision-rule learning, including expressive neural networks, with rigorous statistical feasibility guarantees, while avoiding the computational burden and potential conservatism associated with downstream robust optimization. Thus, our novelty lies in the interpolation-based calibration framework but not the conformal risk control technique itself. We have also explicitly discussed Zhou & Zhu (2026) and these methodological differences in Appendix A of our paper.

To further substantiate these distinctions, we have added a new experimental comparison with an ICRC-based contextual robust optimization pipeline in Appendix E (Table 9). We evaluate both approaches on contextual fractional knapsack and alloy allocation problems under control and stress settings. The results show that ICRC-based robust optimization is substantially more conservative, with operational violation rates far below the 10% target. Our framework achieves higher rewards in both knapsack settings and lower procurement costs in several alloy settings, while maintaining violation rates below the target. Moreover, our calibrated decision rules achieve approximately two orders of magnitude faster deployment, as they avoid per-context robust optimization solves. These findings provide additional empirical evidence for the improved optimality–feasibility tradeoffs and computational advantages of our pipeline.

Response to W3:

We would like to clarify that the full name, Dvoretzky–Kiefer–Wolfowitz (DKW), has already been provided in the Introduction (page 2, lines 103–104).

Response to W4:

We politely point out that the pretrain-then-calibrate pipeline is our method contribution. have further provided clarifications on our method contribution in the general response.

Besides the clarification, we have followed your suggestions to revise the abstract to balance the contents between the theoretical guarantees and the experimental results. Regarding the balance between theory and experiments, we believe they play complementary roles and are both important in our framework. Our theoretical results establish feasibility and pathwise optimality guarantees, with detailed proofs provided in the appendix. Meanwhile, extensive experiments are essential to demonstrate that these guarantees can be achieved without excessive conservatism in practice. The evaluations across different decision-rule classes, applications, and synthetic and real-world datasets demonstrate our pipelines are general, and achieve favorable optimality–feasibility tradeoffs compared with existing approaches. We therefore believe that the experimental evidence is an important part of validating the practical value of our theoretically grounded framework.

AuthorsOpenReview ↗

response to reviewer pCqe (part 2)

Response to Q1:

g(w,z)≤0g(w,z)\leq 0 represents the constraint instead of an objective function. Concrete examples are provided in Section 4.1, including the stockout constraint in newsvendor (g(w,z)=z−w)(g(w,z)=z-w) and the downside-risk constraint in portfolio optimization (g(w,z)=−w⊤z−0.01)(g(w,z)=-w^\top z-0.01). In the newsvendor problem, a warehouse manager wants to keep their ordering quantity to be larger than the demand to reduce the event of stockout, i.e., g(w,z)=z−w≤0g(w, z)=z - w\leq 0 represents the ordering quantity is larger than the demand.

Regarding T\mathcal T, it denotes the sigma-field generated by the training data, a standard probability notation that allows us to condition on the pretrained decision rule when establishing statistical guarantees using independent validation data. This formulation follows the standard convention in optimization (Boyd & Vandenberghe, 2004, Section 4.1.1) and chance-constrained programming (Shapiro et al., 2009, Chapter 4),

Response to Q2:

Convexity is a standard structural assumption in optimization (Boyd & Vandenberghe, 2004). In our framework, convexity of the decision space ensures that interpolated decisions remain admissible, while convexity of the constraint function ensures the nesting property established in Lemma 1. Importantly, all four applications considered in our paper satisfy these conditions, with affine or max-affine constraint functions. The convexity assumption of w is used to ensure the feasibility of all convex combinations. In our revision, we have further relaxed the convexity condition to the so-called star-shaped conditions with details in Appendix C.3. Since convexity is a structural assumption instead of a tunable parameter, conventional parameter sensitivity analysis is not directly applicable. Please see our general response regarding the existence and practical construction of the safe reference rule w0w_0. For the nonconvex constraint discussion, please read our response to reviewer BF4C about their Q1.

Response to Q3:

We adopt the grid search for the index jj over finite parameter candidates for implementation convenience. Since calibration involves only a one-dimensional search, grid evaluation is computationally inexpensive (101 candidates in our experiments) and provides a unified implementation for all four calibrators. For the conformal and DKW calibrators, it is possible to do a continuous extension. By the nesting property in Lemma 1, we can define the first-satisfaction threshold for each validation observation as

τi=inf⁡λ∈[0,1]:g(wλ(xival),zival)≤0.\tau_i=\inf\\{\lambda\in[0,1]:g(w_\lambda(x_i^{\mathrm{val}}),z_i^{\mathrm{val}})\leq 0\\}.

Such infimum can be found by binary search in general. Under continuity along the interpolation path, the empirical feasibility function is the empirical CDF of these thresholds. Consequently, conformal calibration can be implemented through order statistics, while DKW calibration can leverage a uniform confidence bound over the continuous parameter space. Thus, discretization is not essential for the feasibility guarantees of these two calibrators.

Dear reviewer pCqe,

Thank you again for your time and effort on reviewing our paper. The insightful feedback indeed help our paper’s clarification and improvement. We have been trying to addressing your concerns on
(1) the readability of introduction; (2) the comparison (including direct numerical experiments) with Zhu's paper (https://arxiv.org/abs/2510.07750); (3) our method motivation and contribution; (4) convex constraint assumption clarification; (5) other notation and terminology explanation.

As the discussion phase comes to an end (November 18 AoE), we appreciate your feedback on our responses and see if your concerns have been addressed. We would be happy to discuss if you have further questions or need further clarification.

Thank you for your time, effort and consideration again!

Reviewer_8ptzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies safe contextual optimization where, in addition to an optimality criterion, the decision rule also satisfies a high-probability-enforced risk criterion. Naive approaches that integrate chance-constraint methods such as robust optimization, with established contextual optimization pipelines such as predict-then-optimize, suffers from a high price of optimality in exchange for an adequate feasibility guarantee. The paper provides an approach to improve this tradeoff.

Critical strengths

The paper proposes a simple approach that effectively reduces constraint violations in the experiments. The calibration step is easy to implement and is computationally cheap.

Critical weaknesses

It is unclear how the performance of their algorithm depends on the choice of the safe policy. Their approach also assumes that a policy that is feasible almost surely is known, which may not be the case in some applications. The algorithm only considers policies on a segment. They provide an optimality guarantee on this segment, but it is not clear why the segment would contain good policies compared with other feasible alternatives. The theoretical results seem very incremental, as they mostly rely on existing calibration methods.

Questions for authors

Can the authors explain how their method depends on the choice of the safe policy? Also, does their method apply when it is not possible to obtain an almost-surely feasible policy, or when such a policy is not known? Can the authors explain why the convex combinations between the safe policy and the learned policy are a good choice in terms of cost? Can we bound how much we lose compared with other feasible policies? Can the authors compare their DKW and Gaussian calibrators with exact-binomial calibration on the same segment, using the same target risk and nominal confidence levels?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 8ptz (part 1)

Thank you for the detailed review and the list of thoughtful comments. In the following, we would like to address your concerns on weakness and questions.

Response to Weakness (W1):

We have further relaxed the almost-sure safety assumption and provided data-driven constructions of reference rules (Appendix C.1), together with a new sensitivity analysis on the choice of w0w_0 (Appendix D.7). While our optimality guarantee is explicitly pathwise (Theorem 2(ii)), our main contribution lies in providing a simple framework that equips expressive pretrained rules with rigorous statistical feasibility guarantees. Please see our general response for further discussion of the methodological contribution, the choice of w0w_0, and pathwise versus global optimality.

Response to Q1 (dependence on the choice of the safe policy):

Our method does not depend heavily on the choice of the safe policy. Please see our general response on w0w_0 discussion, especially on their choices and multi-safe reference rule sensitivity analysis.

Response to Q2 (extension beyond the almost-surely feasible policy or unknown policy):

Our method applies when it is not possible to obtain an almost-surely feasible policy or under an unknown policy. Please see our general response on w0w_0 discussion on relaxing surely safe assumptions and data-driven safe reference rule construction.

Response to Q3 (justification of convex combination):

Our convex interpolation is a deliberate methodological choice that exploits the geometry of convex decision space while preserving the objective performance of pretrained decision rules.

First, the pretrained rule wprew_{\mathrm{pre}} is obtained through objective-driven training, potentially achieving good cost performance but without rigorous feasibility guarantees. In contrast, the reference rule w0w_0 provides a feasible endpoint, although it may be conservative.

Second, our convex interpolation is specifically tailored to the convex constraint and decision space. By constructing wλ=(1−λ)wpre+λw0w_\lambda=(1-\lambda)w_{\mathrm{pre}}+\lambda w_0, we obtain the nesting property established in Lemma 1: moving toward w0w_0 monotonically improves constraint satisfaction. This provides a principled one-dimensional path toward feasibility.

Convex interpolation in decision space is computationally lightweight: once we have got a pretrained decision rule, we can fix the pretrained model. Depending on various chance constraints and confidence requirements, we can directly adjust the pretrained rule on the decision space to satisfy feasibility, without retraining the original model

Third, our calibration procedures aim to achieve the desired statistical feasibility guarantee while preserving objective performance. The conformal calibrator selects the smallest adjustment passing its calibration criterion, while the DKW and Gaussian calibrators explicitly minimize validation cost among statistically certified candidates. This allows us to leverage the good objective performance of the pretrained rule while making only the adjustments needed to achieve the desired feasibility guarantee, as reflected in the pathwise optimality guarantee of Theorem 2(ii).

Overall, our constraint-geometry-aware calibration strategy is designed to balance statistical feasibility and objective performance.

Response to Q4 (bound loss compared with other feasible policies)):

Obtaining the optimality gap for the general decision set cannot generally be tractably guaranteed without additional structural assumptions, which is why we pursue a computationally tractable approach with both statistical feasibility and pathwise optimality guarantees. We also note that our calibration procedure achieves a better optimality-feasibility tradeoff compared with other baseline methods in the literature empirically. Please see our general response on pathwise v.s. global optimality for more details.

AuthorsOpenReview ↗

Response to Reviewer 8ptz (part 2)

Response to Q5 (exact-binomial calibration):

Yes, we performed additional experiments using an exact-binomial calibrator (EB), which provides finite-sample statistical feasibility guarantees, and our empirical results show that the EB calibrator is more conservative than the Gaussian calibrator. For each candidate (j), let Kj=∑i=1nval(1−Yij)K_j=\sum_{i=1}^{n_{\mathrm{val}}}(1-Y_{ij}) denote its validation violation count. We certify candidate jj if

P(Binomial(nval,α)≤Kj)≤β.P(Binomial(n_{\mathrm{val}},\alpha)\le K_j)\le\beta.

The EB ensures simultaneous feasibility certification over all pp candidates. We select the lowest-validation-cost certified candidate, reverting to the safe reference rule if none passes. For a matched comparison, we conducted paired experiments under the manuscript's synthetic newsvendor and portfolio protocols, using neural and RF–SAA pretrained rules under both control and stress DGPs. All calibrators use the same pretrained rule, safe reference, 101 interpolation candidates, and validation/test samples within each repetition. We retain nval=1,200n_{\mathrm{val}}=1{,}200, 10,000 test observations, α=0.10\alpha=0.10, and nominal confidence 1−β=0.951-\beta=0.95. Each row pools 400 repetitions across the two DGPs.

The following table reports mean objective / test violation rate (%) / test-pass rate (%). Lower objective is better for newsvendor, while higher return is better for portfolio.

Application / RuleDKWGaussian NGaussian UEB
Newsvendor / Neural1.688 / 5.58 / 1001.633 / 6.89 / 1001.616 / 7.33 / 1001.662 / 6.15 / 100
Newsvendor / RF–SAA2.080 / 5.91 / 1001.979 / 7.23 / 1001.990 / 7.05 / 1002.023 / 6.61 / 100
Portfolio / Neural1.202 / 6.35 / 1001.305 / 7.42 / 99.251.324 / 7.60 / 981.261 / 6.98 / 99.75
Portfolio / RF–SAA1.047 / 6.39 / 1001.154 / 7.85 / 98.51.169 / 8.04 / 97.751.100 / 7.13 / 100

The Gaussian calibrators achieve better mean objective performance than exact-binomial calibration in all four pooled settings, generally allowing greater utilization of the prescribed violation budget. Meanwhile, EB improves upon DKW in objective performance across all four settings and retains empirical test-pass rates of 99.75–100%. These results illustrate the empirical optimality–feasibility tradeoffs among different validation procedures. The exact-binomial method provides a distribution-free finite-sample guarantee for general finite candidate families, whereas the Gaussian calibrators exploit dependence among candidates through their joint covariance structure, achieving less conservative objective performance in these experiments.

Dear reviewer 8ptz,

Thank you again for your time and effort on reviewing our paper. The insightful feedback indeed help our paper’s clarification and improvement. We have been trying to addressing your concerns on
(1) the performance dependence on safe reference rule; (2) the assumption of the safe reference rule (we already relaxed them to some extent); (3) our method motivation and contribution; (4) the so-called pathwise v.s. global optimality; (5) additional calibrator scheme, including their numerical performance.

As the discussion phase comes to an end (November 18 AoE), we appreciate your feedback on our responses and see if your concerns have been addressed. We would be happy to discuss if you have further questions or need further clarification.

Thank you for your time, effort and consideration again!

Reviewer_eU1jOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies safe contextual optimization, where the constraint is unknown and determined by the contextual variable. The authors propose a pretrain-then-calibrate framework that interpolates between any pretrained decision rule and some feasible reference rule, reducing feasibility calibration to a one-dimensional problem. Then, they establish feasibility and pathwise optimality guarantees and evaluate the method through numerical experiments.

Critical strengths

For the safe contextual optimization problem, the authors propose a new interpolation method, which not only can i) incorporate and calibrate existing contextual optimization methodologies to improve feasibility, but also is easy to implement in real life.

Critical weaknesses

My main concern is about potential conservatism. Given that g is generally unknown, an always-feasible reference rule is either hard to find or very conservative. This could be especially problematic in high-dimensional settings. For example, in a newsvendor problem with multiple products, if the demands of different products have very large support but different tail distributions, then the reference point itself might be very conservative, which also affects the performance of the interpolation.

Questions for authors
  1. In practice, how should the safe reference rule w_0 be constructed? Also, if there are multiple different rules, which reference rule should one select? For example, in portfolio construction, instead of cashing out all securities, one can also buy option straddles to avoid the risk.

  2. Given the number of training data sample size, what is the best choice of p?

  3. The optimality in Theorem 2 is not defined for the general decision set, and I wonder what the optimality gap could be for the general decision set.

1 comment
AuthorsOpenReview ↗

Response to Reviewer eU1j

Thank you for your recognition of our significance and strengths, and your detailed review. We address your comments and questions as follows.

Response to Weakness (W1):

We would like to clarify some misunderstanding of what is known and unknown: First, our framework assumes the constraint function g(w,z)g(w,z) is known and evaluable, while the joint distribution of (x,z)(x,z) is unknown, as in data-driven optimization literature (Im et al., 2025).

Second in many applications, a safe reference rule can be constructed using problem-specific structural information, such as in Section 4.1: (1) known demand-support bounds in newsvendor: the stockout constraint in newsvendor (g(w,z)=z−w)(g(w,z)=z-w); an all-cash allocation in portfolio optimization: the downside-risk constraint in portfolio optimization (g(w,z)=−w⊤z−0.01)(g(w,z)=-w^\top z-0.01), or a zero-allocation rule in resource allocation (Appendix E). Please see more discussion on the safe reference rule in our general response. Regarding conservatism, we first point out that our numerical experimental results (including the new results; see Section 4, Appendix E) show that our approach outperforms all benchmarks uniformly and significantly (Table 2, 10, 11) by evaluating linear, neural, and RF–SAA rules and compare against multiple contextual, conformal, and robust optimization approaches across newsvendor, portfolio, fractional-knapsack, and alloy-allocation problems. So, in terms of practical performance, there is strong evidence that our approach is not conservative.

Next, we would like to provide intuition for why our approach is not conservative,which is precisely what motivates our calibration framework. Our approach starts from an objective-driven pretrained rule, which may lack rigorous feasibility guarantees, and constructs the interpolation path. Our calibrators seek a statistically feasible decision along this path while preserving objective performance. Therefore, the quality of the final calibrated rule depends on the available interpolation path, not solely on the conservatism of w0w_0. Our framework is specifically designed to mitigate the conservatism of the safe reference rule.

Response to Q1:

In practice, safe reference rules can be constructed using domain knowledge or data-driven feasibility certification (Appendix C.1.1). When multiple references are available, including hedged portfolios such as option straddles if their feasibility can be certified, our framework supports calibration across multiple interpolation paths (Appendix C.2). We have also added a sensitivity analysis in Appendix D.7 to investigate how reference selection affects performance. Please see our general response on w0w_0 discussion on (1) constructions of them and (2) multi safe reference rule sensitivity analysis .

Response to Q2:

The number of grid points pp is primarily a discretization parameter in the calibration stage. Its choice is therefore more closely related to the validation sample size nvaln_{\mathrm{val}} and the properties of the interpolation path than to the training sample size nn.

There is a tradeoff between discretization accuracy and statistical estimation error. For a uniform grid, assuming the cost function is LcL_c-Lipschitz, the discretization error relative to the best feasible decision on the continuous path is at most LcD/(p−1)L_cD/(p-1), where D=E[∣wpre(X)−w0(X)∣∣T]D=\mathbb E[|w_{\mathrm{pre}}(X)-w_0(X)|\mid\mathcal T]. Meanwhile, Theorem 2(ii) provides a uniform cost estimation bound of order O(log⁡p/nval)O(\sqrt{\log p/n_{\mathrm{val}}}). Combining these terms suggests the following upper bound:

E(p)≲LcDp−1+Bclog⁡pnval,\mathcal E(p)\lesssim\frac{L_cD}{p-1}+B_c\sqrt{\frac{\log p}{n_{\mathrm{val}}}},

where BcB_c bounds the absolute cost. Minimizing this surrogate bound suggests a grid size of order p=Θ(nvallog⁡nval)p=\Theta(\sqrt{n_{\mathrm{val}}\log n_{\mathrm{val}}}). This provides a theoretical guideline.

Importantly, the feasibility guarantees of our conformal and DKW calibrators are independent of pp, thanks to the nesting property. In particular, increasing pp does not increase the DKW feasibility margin. The Gaussian calibrators, in contrast, involve additional dimension-dependent approximation errors. In practice, we recommend choosing a grid sufficiently fine that further refinement offers negligible improvement relative to statistical uncertainty and computational cost. We use p=101p=101 throughout our experiments as a computationally convenient choice, rather than claiming it is universally optimal.

Response to Q3:

Obtaining the optimality gap for the general decision set cannot generally be tractably guaranteed without additional structural assumptions, which is why we pursue a computationally tractable approach with both statistical feasibility and pathwise optimality guarantees. Please see our general response on pathwise v.s. global optimality for more details.

Comments on the paper

6 comments
AuthorsOpenReview ↗

General Response 1: Methodological Motivation and Contribution (part 1).

We thank the reviewers for their thoughtful comments. We would like to answer some common questions in our general response as below.

  1. Methodological Motivation and Contribution. For reviewers eU1j (W1), 8ptz (W1), pCqe (W4), BF4C (W1), and 8KPK (W1).

In the revision, we have streamlined the introduction to highlight our contributions better. Let us explain in more detail here.

In this paper, we focus on safe contextual optimization, i.e., seeking a decision rule that not only performs well according to an optimality criterion, but also satisfies constraints that represent a risk criterion subject to uncertainty, using chance constraints. Our goal is to obtain an approach that is both theoretically grounded (i.e., rigorous statistical feasibility guarantees), and have good optimality-feasibility balance (i.e., allowing reliable use of expressive decision-rule learning).

To explain our positioning clearly, we first elaborate the challenges in the landscape of the present literature:

Challenges of Existing Approaches:

First, chance-constrained optimization is generally computationally intractable. Even for simple (such as linear) safety conditions, the associated chance-constraints can be non-convex. This means that generally, finding optimal solution, or nearly optimal solution with theoretical optimality guarantee for chance-constrained optimization is challenging, even if the randomness distribution is completely known (Nemirovski & Shapiro, 2006; Luedtke et al., 2010; Pagnoncelli et al., 2009; Ahmed & Shapiro, 2014).

Second, even if the goal is confined to obtaining a feasible solution, existing approaches are by-and-large conservative. The majority of the literature focuses on replacing the intractable chance constraint by more tractable convex constraints. These latter replacing constraints are carefully designed so that the resulting solution is guaranteed to be feasible (with statistical confidence). However, the main existing approaches to do this are conservative:

  • Safe convex approximation or robust optimization: This replaces the chance constraints with an inner approximation that is convex. A prime example is CVaR approximation (Nemirovski & Shapiro, 2006). This approach is intimately related to robust optimization (Bertsimas et al., 2011). However, when data are present, the parameters in the CVaR approximation need to be calibrated, with statistical confidence, to ensure that the resulting feasible region is indeed a valid inner approximation of the original chance constraint with high confidence. This can give rise to high conservativeness, as shown in Hong et al., (2014) and our numerics in Section 4.3.
  • Learning-based enhancement of robust optimization: More recent approaches learn the shape and size of ambiguity sets in robust optimization from data (e.g., Hong et al., 2021, Ben-Tal et al., 2013, Delage & Ye, 2010). This helps reduce some conservatism in 1). However, fundamentally, using robust optimization is conservative because this paradigm separates the objective function from uncertainty set construction: By using an ambiguity set that is a confidence region on the uncertain parameter, the resulting solution is guaranteed to be feasible with at least the same level of confidence. However, this translation from the confidence guarantee in containing the true parameter value to solution feasibility is not tight. The resulting solution feasibility often holds with a confidence level much higher than the nominal, making it very conservative. This is documented in Lam & Qian (2019).
  • Sampling-based approaches: Another popular approach is scenario optimization, which uses samples to approximate the constraints. The basic approach is by sampling enough randomness and replacing the chance constraint by sampled constraints (Campi & Garatti, 2018), or refined methods using a subset of constraints (Campi & Garatti, 2011, Luedtke & Ahmed, 2008). These approaches have statistical feasibility guarantees that depend on the dimension of the decision space. This suggests an intrinsic conservativeness, since the approaches themselves often only contain a low-dimensional hyperparameter, e.g., the number of scenarios. This conservative behavior motivates us to consider our pretrain-then-optimize scheme that operates by using a low-dimensional calibration parameter.
AuthorsOpenReview ↗

General Response 2: Methodological Motivation and Contribution (part 2).

Third, extending these established chance-constraint methods to contextual settings presents even more conservativeness. The major approaches in contextual optimization, including predict-then-optimize (Elmachtoub and Grigas, 2022), integrated estimation-optimization (Qi et al., 2026), and end-to-end learning (Zhang et al., 2026), all focus on the objective level. When uncertainty is present in the constraint, however, these techniques do not automatically ensure feasibility and the guarantees associated with the established chance-constrained methods and contextual-optimization methods are incompatible. A native adoption of the uncertainty set in the contextual setting, e.g., incorporating the context in the set with a high probability conditional coverage (Patel et al., 2024; Wang et al., 2023; Im et al., 2025), which is only a sufficient condition for decision feasibility, and therefore conservative. Hence, achieving statistical feasibility might come with a high cost of optimality.

Our Contributions to Resolve Existing Challenges:

In terms of motivation, our pretrain-then-calibrate framework is designed to convert powerful objective-driven decision rules into also rigorously feasible rules, without sacrificing much on the optimality side. For example, end-to-end neural decision maps (Rychener et al., 2023); generative decision models (Babier et al., 2024); and piecewise-affine decision rules (Zhang et al., 2026). These heuristics include end-to-end decision rule training that directly leverage modern neural network architectures and other advanced AI models.

As explained above, converting these objective-driven end-to-end decision rules into rigorous feasibility guaranteed rules, however, is challenging due to the intrinsic conservativeness of existing chance-constraint methods as well as the translation of these methods into contextual settings. Our paper therefore takes a novel path of NOT refining the training algorithm of these end-to-end rules, BUT conducting a light-weighted adjustment directly on the decision space to equip a pretrained rule with high-confidence feasibility guarantees. The pretrained rule in particular can be obtained by the existing powerful objective-driven methods. In this way, the price on optimality to trade off with feasibility is much smaller than that suggested in the existing literature.

In terms of overall strengths, our approach is easy to implement, yet with theoretical guarantees essentially at the same level as existing chance-constraint methods. It is easy to use in practice and computationally lightweight since (1) our methodology is entirely modular and can be applied simply as an add-on to the pretrained; (2) once we have got a pretrained decision rule, we can fix the pretrained model. Depending on various chance constraints and confidence requirements, we can directly adjust the pretrained rule on the decision space to satisfy feasibility, without retraining the original model or completing a new robust optimization solving per new context. The updated Appendix E demonstrates two orders of magnitude faster running time than the ICRC-based robust baseline in the reported experiments.

In terms of theory, note that the statistical confidence of feasibility that we achieve in our approach (Theorem 2(i)) is the same type of feasibility guarantee in the non-contextual chance constraint literature (Campi & Garatti, 2008; Luedtke & Ahmed, 2008; Campi & Garatti, 2011; 2018; Hong et al., 2021; Lam & Qian, 2019), as discussed in Theorem 2.

To this end, we politely point out that we view the simplicity of the approach, including the simplicity of the theory, as our main advantage. Our numerical experimental results (including the new results; see Section 4, Appendix E) show that our approach outperforms all benchmarks uniformly and significantly (Table 2, 10, 11) by evaluating linear, neural, and RF–SAA rules and compare against multiple contextual, conformal, and robust optimization approaches across newsvendor, portfolio, fractional-knapsack, and alloy-allocation problems. This numerical superiority comes with a simple implementation, and a theoretical guarantee that is easily derivable. To us, this combination of advantages can only be a strength rather than a weakness of our approach. We hope this could address the weakness concern from reviewers 8ptz and bf4c.

AuthorsOpenReview ↗

General Response 3: Discussion on the safe reference rule (part 1)

  1. Discussion on the safe reference rule w0w_0

For reviewers eU1j (W1, Q1), 8ptz (W1, Q1, Q2), pCqe (Q2), and 8KPK (W2).

In the revision, we have relaxed the current assumptions regarding the safe reference rule, as well as adding discussions on constructing the rules and their sensitivity analyses. Let us describe each point in detail.

(1) First, the safe reference rule and convexity assumptions are sufficient but not necessary conditions for the statistical feasibility guarantees of our methods. For candidates fixed before validation, the Gaussian certification argument does not require convexity or an almost-surely safe reference: it controls the probability of accepting any population-infeasible candidate, accounting for the Gaussian approximation error (Appendix C, proof of Theorem 2(i)). A safe reference ensures that the procedure can still return a feasible decision when no candidate passes certification. Without such a fallback, the procedure should instead report that no candidate is certified. The conformal and DKW calibrators additionally exploit the nesting structure; we discuss the role of convexity in our reviewer-specific responses.

(2) We have formally relaxed the almost-sure safety assumption in the revised Appendix C.1 and the convexity assumption in Appendix C.3. Specifically, it suffices to have a reference rule with a known population low-violation rate with ε<α\varepsilon<\alpha, which is more general in practice. Following similar analyses as the main body, we retain the same guarantees for conformal, DKW and Gaussian calibration. Their guarantees only require constructing modified nested violation indicators and allocating the remaining risk budget γ=α−ε\gamma=\alpha-\varepsilon in their corresponding procedures for conformal calibration and DKW calibration.

(3) We would like to note that safe reference rules are naturally available in many practical optimization problems through domain knowledge, and constructing such a rule is often substantially easier than finding an optimal feasible decision. Importantly, w0w_0 only needs to provide a safe reference without good objective performance. A simple sufficient condition is the availability of a known support set for the uncertainty, over which a decision can be verified to satisfy the constraints. Concrete examples in our paper include all-cash allocation in portfolio optimization, zero allocation in resource allocation, known demand-support bounds in newsvendor, and guaranteed-content procurement in alloy allocation. Such constructions do not require solving the original chance-constrained optimization problem. Moreover, leveraging known feasible references is an established design principle in constrained and safe learning (e.g., Tordesillas et al., 2023; Schneider & Kuhn, 2026; Khezeli & Bitar, 2020).

(4) We also add the discussions of constructing data-driven safe reference rules and provide practical guidance. We have further strengthened Appendix C.1 by adding a new discussion of data-driven safe reference construction in Appendix C.1.1. When the underlying distribution is unknown, an approximately safe reference can be constructed using feasibility-oriented procedures. We provide two examples: a simple newsvendor rule based on the maximum historical demand and a scenario-based linear program for alloy allocation. Such guarantees can be combined with independent calibration, with an additional failure probability of at most δ0\delta_0 for DKW calibration. Moreover, in our real-world newsvendor experiments (Appendix D.6), reference rules constructed from historical demand observations exhibit zero violations across the reported validation and test datasets, illustrating their practical applicability.

These results demonstrate that our framework can accommodate more general, approximately safe reference rules while preserving rigorous statistical feasibility guarantees, further extending its applicability beyond the original almost-surely safe setting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.