acceptodds
Under review as a conference paper at ICLR 2027

Learn Feasibility Once, Optimize All Objectives: Derivative-Free Diffusion Models for Chance-Constrained Programming

Abstract

Chance-constrained programs (CCPs) optimize decisions under uncertainty by limiting the probability of constraint violation. Despite advances in both traditional and learning-based approaches, challenges remain in optimizing non-convex or non-smooth objectives and adapting to different objectives under fixed chance constraints. In this paper, we propose a Derivative-free Diffusion-based framework that Disentangles constraint modeling from objective optimization, termed DOpt. We learn the chance-feasible structure once, independently of any particular objective, by training a risk-conditioned diffusion model solely on constraint-filtered decisions and freezing it as a reusable prior for post-specified objectives. At inference time, we propose an annealed, particle-based Feynman–Kac correction along the frozen reverse diffusion process to optimize post-specified objectives using only function evaluations. This enables derivative-free optimization of non-convex and non-smooth objectives without objective-specific retraining. We prove that the correction preserves feasibility when this property holds for the frozen prior, and derive an optimization-error bound separating learned-prior coverage, finite-particle approximation, and finite-temperature effects. Experiments on linear Gaussian CCPs, objective-transfer tasks, and chance-constrained economic dispatch demonstrate effective optimization across smooth and non-smooth objectives, including non-convex cases, and objective generalization under fixed chance constraints without retraining.

Primary area:
optimization
Keywords:
Diffusion Model, Chance Constrained Programming, Learning to Optimize

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_NTBAOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents D3Opt, a framework for chance-constrained programming that disentangles constraint modeling from objective optimization. In the first stage, feasible decision data are generated using constraint information only (no objective evaluation), and a risk-conditioned diffusion model is trained and frozen. In the second stage, a post-specified objective — possibly non-convex, non-smooth, and accessible only through function values — is introduced at inference time via an annealed, particle-based Feynman–Kac correction along the frozen reverse diffusion, reweighting and resampling particles by objective values to concentrate mass on low-objective regions. Theoretically, the authors prove that the correction preserves the support of the frozen prior (feasibility preservation) and derive an optimization-error bound decomposing into coverage, finite-particle, and finite-temperature terms. Experiments cover linear Gaussian CCPs, objective transfer with the same frozen checkpoint (smooth convex and non-convex non-smooth objectives), chance-constrained economic dispatch with valve-point effects, multi-risk adaptation, and robust waveform design

Critical strengths
  1. Data generation, diffusion training, and frozen inference never touch the objective; the FK potentials depend only on objective values at denoised predictions and telescope to the desired terminal Boltzmann tilt. The same checkpoint transfers to ASQ (smooth convex) and MW4 (non-convex non-smooth) without retraining, directly validating the decoupling claim. 2.Comparisons span convex exact references (SOC-CVX), SAA variants, GGDOpt, DiffOPT, and classical methods (DE, PWL-MILP, sphere bounding, Bernstein); tasks include objective transfer, multi-risk levels, ablations, a non-smooth practical application, and a non-convex objective. The frozen checkpoint is reused across four risk levels with all returned solutions satisfying the analytic chance constraint.
Critical weaknesses

1.Coverage error is the only term in the decomposition that does not vanish with particles or temperature, yet the paper provides no way to measure or bound it. The choice of proposal distribution strongly affects coverage (the gap between 0.8549 without FKC and ~0.098 for the ASQ optimum shows the prior is heavily biased toward suboptimal regions), and this most sensitive component receives no sensitivity analysis or diagnostic. 2.Feynman–Kac correction and SMC are mature tools, and Skreta et al. 2025 already applied them to diffusion inference steering. The increment here is the combination of an objective-free prior with telescoping potentials. An ablation showing why evaluating potentials at denoised predictions is superior to evaluating them at particle states would substantiate the design choice.

Questions for authors

How can the coverage gap be estimated? For example, can the paper report the prior's probability mass near the global optimum sublevel set for each task, or at least ablate the proposal distribution (uniform vs. others) to demonstrate sensitivity?

2 comments
AuthorsOpenReview ↗

For W1: Coverage error and sensitivity to the training-data proposal.

Thank you for your comments on the coverage error. We provide a quantitative diagnostic using unguided diffusion sampling. For the linear Gaussian CCP with n=8n=8 and ρ=0.1\rho=0.1, we draw 10,00010{,}000 raw samples and compute the minimum objective among those. Let IN,ρ\mathcal I_{N,\rho} denote their indices. Under the almost-sure feasibility assumption in Proposition 2, we have

0≤Δθ,ρ∣fρ⋆∣≤min⁡i∈IN,ρf(Xi)−fρ⋆∣fρ⋆∣almost surely.0\le\frac{\Delta_{\theta,\rho}}{|f_\rho^\star|}\le\frac{\min_{i\in\mathcal I_{N,\rho}}f(X_i)-f_\rho^\star}{|f_\rho^\star|}\qquad\text{almost surely}.

Thus, this statistic gives a finite-sample upper bound on the relative coverage error; the true gap can be smaller. For CCED-VPE-10, we additionally report the best returned cost from the w/o FKC evaluations and its gap to the numerical DE reference.

Unguided-sampling diagnostics under different training-data proposals at ρ=0.1\rho=0.1. For Linear Gaussian, fmin⁡f_{\min} is the minimum feasible objective among 10,00010{,}000 raw samples, and freff_{\mathrm{ref}} is the exact SOC optimum. For CCED-VPE-10, fmin⁡f_{\min} is the best returned cost from the w/o FKC evaluations, and freff_{\mathrm{ref}} is the DE reference with exact projection. Relative gaps are 100(fmin⁡−fref)/∣fref∣100(f_{\min}-f_{\mathrm{ref}})/|f_{\mathrm{ref}}|.

TaskTraining-data proposalReference freff_{\mathrm{ref}}fmin⁡f_{\min} ↓Relative gap (%) ↓
Linear GaussianN(0,I8)\mathcal N(0,I_8) (original proposal)−0.6586−0.588710.61
Linear GaussianStudent-t5t_5 (covariance I8I_8)−0.6586−0.592110.10
CCED-VPE-10Uniform (original proposal)104399.055106146.7151.67
CCED-VPE-10Truncated Student-t5t_5104399.055106200.5871.73

The tested proposal replacements give similar observed gaps within each task: Student-t5t_5 gives a slightly tighter Gaussian bound, while the original uniform proposal gives a slightly lower CCED reference gap. Appendix D.2 (Table 9) also reports a controlled proposal-sensitivity experiment. For each proposal, we regenerate the training data and retrain the prior, keeping the retained sample count, network, training budget, and FKC settings fixed. The final mean objectives are −0.6519-0.6519 vs. −0.6521-0.6521 on Linear Gaussian and 104823.577104823.577 vs. 104904.758104904.758 on CCED-VPE-10. These results show similar final objective values under the tested proposal replacements. The appendix also includes objective-value histograms of the training data and unguided outputs as a descriptive diagnostic.

Regarding ASQ, it's an objective-transfer experiment. The value 0.85490.8549 reflects the mean quality of unguided returned solutions and does not directly measure the coverage gap. The results shows that, with the same frozen checkpoint, FKC improves the ASQ mean objective from 0.85490.8549 to 0.10350.1035 and the best objective from 0.32090.3209 to 0.09990.0999, compared with the optimum of 0.09790.0979. This ablation demonstrates the role of FKC in directing sampling toward low-objective regions while keeping the feasibility model fixed.

AuthorsOpenReview ↗

For W2: Objective Evaluation at Denoised Predictions

Thank you very much for this helpful suggestion. To examine this design choice, we compare objective evaluation at denoised predictions, f(X^k)f(\widehat X_k), with direct evaluation at noisy particle states, $f(X_k) $. Within each task, we retain the original checkpoint, risk level, particle count, reverse-diffusion steps, inverse-temperature schedule, and ESS resampling rule. Both variants use the same telescoping FK construction, terminal decision definition, and feasibility correction and selection protocol. Only the intermediate objective-evaluation input is changed. Each variant is evaluated over 100 inference runs.

Objective-evaluation ablation under the original experimental protocol. Bold indicates the lower mean and standard deviation within each task.

TaskObjective evaluationRepeatFvalMean ↓FvalStd ↓
Linear GaussianNoisy-state100−0.64920.0031
Linear GaussianDenoised (ours)100−0.65190.0015
ASQNoisy-state1000.12310.0034
ASQDenoised (ours)1000.10350.0018
MW4Noisy-state1000.21780.0102
MW4Denoised (ours)1000.18340.0018
CCEDNoisy-state100106244.87153.281
CCEDDenoised (ours)100104823.58102.065

The table shows that denoised evaluation yields lower mean objectives and smaller standard deviations on all four tasks. Relative to noisy-state evaluation, the mean objective decreases by 0.0027 on Gaussian and by approximately 15.9%, 15.8%, and 1.34% on ASQ, MW4, and CCED, respectively. These results support using denoised predictions for objective evaluation in finite-particle optimization under the tested settings.

We hope our responses address your concerns, and we would be very grateful if you could raise your score, as this would mean a great deal to us.

Reviewer_idxhOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes D³Opt, a derivative-free diffusion framework for chance-constrained programming that separates constraint learning from objective optimization. It trains a risk-conditioned diffusion prior on objective-independent, constraint-filtered decisions and freezes the model for subsequent objectives. During inference, an annealed Feynman–Kac particle correction uses objective values to reweight and resample trajectories. The telescoping weights yield a terminal Boltzmann tilt of the learned prior. The analysis establishes feasibility preservation under an almost-surely feasible prior and, under additional particle-error assumptions, separates optimization error into coverage, finite-particle, and finite-temperature contributions. Experiments examine linear Gaussian CCPs, objective transfer, economic dispatch, and an additional robust waveform-design task.

Critical strengths
  1. A useful separation of reusable feasibility learning and objective-specific optimization. The objective is excluded from data generation, training, and checkpoint selection, making the proposed reuse mechanism explicit rather than merely omitting the objective from the network input. The value-only inference procedure accommodates nonsmooth or black-box objectives. This addresses a meaningful setting in which constraints remain fixed while operational objectives change, and the ASQ/MW4 experiments provide controlled demonstrations of checkpoint reuse.

  2. An interpretable particle construction and conditional error decomposition. The telescoping FK weights clearly identify the terminal tilted distribution. The theoretical decomposition explicitly exposes the dependence on prior coverage rather than implying that reweighting can recover regions missing from the prior. Together with exact optimization references in the linear Gaussian and objective-transfer tests, this provides a useful basis for interpreting the method’s behavior and limitations.

Critical weaknesses
  1. Insufficient technical positioning relative to existing methods. Section 2 identifies classical CCP and diffusion-based approaches. However, these discussions remain too high-level to establish the incremental contribution clearly. A systematic comparison should distinguish training-data dependence on the objective, constraint handling, derivative requirements, frozen-model reuse, and theoretical guarantees. A detailed comparison table or expanded appendix would help separate the new components from the combination of established mechanisms.

  2. Insufficiently specified data generation and unverified coverage of near-optimal regions. Section 4.1 explains the objective-free motivation, but the problem-specific proposal distributions and their acceptance behavior are insufficiently specified. Proposition 2 introduces a coverage gap, but does not connect it to proposal design, finite-scenario screening, dataset size, or diffusion approximation. Consequently, the analysis does not establish whether near-optimal feasible regions remain represented with meaningful probability.

  3. Missing end-to-end computational accounting and matched-budget comparisons. The paper reports sample counts, training settings, hardware, and a default population of 512 particles, but not data-generation, training, or per-objective inference times. Candidate rejection and repeated particle propagation and objective evaluation may incur substantial costs. Without total evaluation counts and wall-clock comparisons at comparable solution quality and feasibility, the practical advantage cannot be assessed.

  4. Limited evidence of scalability and insufficiently targeted baseline comparisons. There is no systematic dimension-scaling study within a common problem family. Although multiple learning-based and conventional baselines are included, further comparisons should isolate the value of the proposed particle correction, for example against directly comparable derivative-free/FK guidance using the same frozen prior. Broader tests against conventional CCP and derivative-free solvers should track optimality gaps, per-run feasibility, data-acceptance rates, and computational resources as dimension grows.

Questions for authors

See Critical Weaknesses

6 comments
AuthorsOpenReview ↗

For W1: Technical positioning relative to existing methods

Thank you for suggesting a more systematic technical comparison. We have added a table (Table1 in revised manuscript) comparing GGDOpt, DiffOPT, and D$^3$Opt in terms of training-data dependence on the objective, risk conditioning, derivative requirements, objective-independent prior reuse, and annealed trajectory reweighting.

MethodObjective-independent training dataRisk-conditioned feasibility priorDerivative- and density-evaluation-free optimizationObjective-independent prior reuseAnnealed trajectory reweighting
GGDOpt✗✓✗✗✗
DiffOPT✓✗✗✓✗
D³Opt✓✓✓✓✓

The comparison distinguishes what is shared from what D$^3$Opt adds. DiffOPT already separates prior learning from objective optimization and supports derivative-free inference. It starts from supplied feasible samples, without specifying an objective-independent data-generation procedure, and requires additional density evaluation. GGDOpt generates training decisions by solving objective-dependent optimization problems. Although its guidance objective can be changed, the resulting prior retains this dependence through its training data. D$^3$Opt instead constructs decision–risk pairs using only chance-constraint information, learns a risk-conditioned feasibility prior, and freezes it for subsequent objectives. Its annealed FK correction introduces objective values through incremental trajectory weights along the reverse diffusion process. Thus, the contribution concerns the concrete construction and analysis of a reusable CCP solver, beyond objective–constraint separation alone.

The theoretical scope is also clarified. Proposition 1 establishes feasibility preservation under an almost-surely feasible prior. Under the stated assumptions, including the SMC particle-error assumption, Proposition 2 bounds the expected objective gap of the weighted terminal particle approximation relative to the CCP optimum. It separates coverage, finite-particle, and finite-temperature contributions without requiring objective differentiability.

AuthorsOpenReview ↗

For W2: Data generation and near-optimal coverage.

Thank you for highlighting the need to specify the data-generation procedure and examine the representation of near-optimal feasible regions. We have expanded Appendix C.1, Constraint-Only Data Generation, to clarify the problem-specific proposal distributions and acceptance criteria. The linear Gaussian experiments use ν=N(0,In)\nu=\mathcal N(0,I_n); ASQ and MW4 reuse the Gaussian training dataset at n=8n=8. Robust waveform design uses N(0,I48)\mathcal N(0,I_{48}), while CCED samples each generator output independently from Unif[Pimin⁡,Pimax⁡]\mathrm{Unif}[P_i^{\min},P_i^{\max}]. Target risk levels are sampled independently of candidate decisions from Unif[0,0.5]\mathrm{Unif}[0,0.5].

Each dataset contains 10,000 accepted decision–risk pairs. Generating these datasets required 33,279 candidates for Linear Gaussian, 993,760 for CCED-VPE-10, and 29,427 for Robust Waveform, corresponding to empirical acceptance rates of 30.05%, 1.006%, and 33.98%, respectively.

These 10,000 training pairs are pooled across the sampled risk levels. Proposal design and finite-scenario screening determine the accepted-data distribution, while a finite dataset can miss low-probability near-optimal regions. Increasing dataset size improves the chance of representing regions with positive mass under this distribution; their subsequent representation in the learned prior also depends on diffusion approximation. Proposition 2 retains the resulting prior-coverage gap as an explicit term in the optimization-error bound.

To provide an observable coverage diagnostic, we draw 10000 unguided, unprojected samples from the frozen prior at ρ=0.1\rho=0.1 on Linear Gaussian. Let fmin⁡,Nf_{\min,N} denote the minimum objective among feasible samples. Under the almost-sure feasibility assumption of Proposition 2, the definition of the essential infimum gives

0≤Δθ,ρ=fθ,ρ⋆−fρ⋆≤fmin⁡,N−fρ⋆almost surely.0\le\Delta_{\theta,\rho}=f_{\theta,\rho}^{\star}-f_\rho^\star\le f_{\min,N}-f_\rho^\star \quad \text{almost surely}.

The observed minima are −0.5887 and −0.5921 under the normal and Student-t5t_5 proposals, respectively. Their relative gaps to the exact SOC optimum are 10.61% and 10.10%, providing conditional upper bounds on the relative coverage gap. The actual gap can be smaller. This diagnostic concerns the coverage-gap term and does not directly estimate near-optimal probability mass.

We complement this diagnostic with the controlled proposal-sensitivity experiment in Appendix D.2 and Table 9. For each proposal, we regenerate the data and retrain the prior while keeping the retained dataset size, architecture, training budget, and FKC settings fixed, thereby isolating the effect of proposal design. The final mean objectives are −0.6519 versus −0.6521 on Linear Gaussian and 104823.577 versus 104904.758 on CCED-VPE-10. These results show similar final mean objectives under the tested proposal replacements. Figure 1 in Appendix D.3 additionally compares the objective-value distributions of training data and projected unguided outputs as a descriptive diagnostic.

AuthorsOpenReview ↗

For W3: Computational accounting and runtime comparisons.

Thank you for suggesting a more complete accounting of the computational costs. We have added a computational-cost analysis in Appendix C.6, separating data generation and training from online objective optimization. At each dimension, we train a separate feasibility prior; data generation and training are performed once for that prior, which is then reused across subsequent objectives. The costs are reported below. Total sampling time covers 100 runs at ρ=0.1\rho=0.1.

Stagen=8n=8n=16n=16n=32n=32
Data generation time (min)0.911.061.24
Training time (h)1.552.433.17
Total sampling time over 100 runs (min)0.08850.10330.1532

To put the online cost in context, the following table compares sampling or solving time per returned solution on Linear Gaussian at n=8n=8 and ρ=0.1\rho=0.1, together with the corresponding objective statistics. Each stochastic method is evaluated over 100 runs. All runtimes are measured on the same workstation equipped with an NVIDIA RTX 5090 GPU.

MetricSOC_CVXGGDOpt First-orderGGDOpt Second-orderDiffOPT GradientDiffOPT IISD³Opt
fmeanf_{\mathrm{mean}}−0.6586−0.6483−0.6491−0.5591−0.5614−0.6519
fstdf_{\mathrm{std}}00.00510.00560.02780.02520.0015
Average time per returned solution (s)0.32140.04900.05710.05610.06260.0531

D³Opt takes 0.0531 seconds per returned solution, slightly above first-order GGDOpt's 0.0490 seconds and below second-order GGDOpt's 0.0571 seconds and both DiffOPT variants. Its particle reweighting and resampling support objective-value-only optimization with a frozen, reusable prior. Under the reported configurations, D³Opt achieves the lowest mean objective and standard deviation among the compared learning-based methods, with online runtime comparable to these baselines.

Reviewer_FB2xOpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a diffusion-based framework for solving chance-constrained programs (CCPs). Existing mathematical programming-based methods developed for CCPs often suffer from the curse of dimensionality and the underlying non-convexity, which results in scalability issues. Motivated by this, the paper extends recent diffusion-based frameworks for constrained optimization problems to the chance-constrained setting, proposing D$^3$Opt, which stands for Derivative-free Diffusion-based Framework that Disentangles constraint modeling from objective optimization. As the name suggests, the main features of the framework are threefold. First, it requires only a zeroth-order oracle for the objective function. Second, it is based on a diffusion process designed to learn the feasible set given a risk level ρ\rho. Third, the diffusion model for feasible sets is trained without objectives, based on which it may take any objective function once the constraint modeling is done.

Critical strengths

The paper provides a new framework for solving chance-constrained programs. In particular, the paper proposes a constraint modeling framework based on diffusion, which can adapt to varying levels of the risk level ρ\rho.

Critical weaknesses
  • The proposed framework is advantageous when we need to repeatedly solve many instances with distinct objectives over the same set of feasible solutions. However, the motivation behind using chance constraints is to deal with varying constraints. Conceptually, it makes more sense to consider the setting where the feasible set itself changes across different instances, which is the case when the underlying distribution shifts for a new instance.
  • The list of baseline methods is limited, missing many relevant mathematical programming and optimization-based ones. For example, the paper does not test ALSO-X / ALSO-X+ by Jiang and Xie, which provides a convex approximation less conservative than CVaR. Moreover, the SAA-MIP benchmark seems outdated.
  • It is difficult to claim sufficient novelty of the proposed method in comparison with prior diffusion-based frameworks for constraint modeling. DiffOPT, due to Kong et al., also disentangles constraint modeling from objectives, and thus, the statement "However, the extent to which these distributions can be reused across objectives depends on their construction" after Q1 is not true for DiffOPT. Moreover, the sampling part is very similar to FK steering by Singhal et al., which also runs with just a zeroth-order oracle.
Questions for authors
  • What is a novel component of the proposed framework compared with the existing diffusion-based frameworks for constraint modeling? Doesn't this simply combine DiffOPT and FK Steering?
  • Can the method be extended to take situations where the set of feasible solutions also changes?
  • Does the proposed method perform better than ALSO-X and ALSO-X+?
10 comments
AuthorsOpenReview ↗

Part1(W1 & Q2): Motivation and adaptation to changing feasible sets.

Thank you for raising this point. The basic motivation for chance constraints is to control constraint-violation probability under uncertainty. This principle underlies the classical formulations of Charnes and Cooper (1959, 1962) [1, 2] and is also emphasized in the recent review by Cui et al. (2025) [3]: constraints are required to hold at a prescribed probability level. In particular, the classical formulation already considers random variables with known distributions. A fixed uncertainty distribution therefore remains a meaningful CCP setting: its realizations vary, and decisions must account for the resulting risk of constraint violations.

Within this setting, repeated optimization across objectives has practical value. Different costs or decision preferences may be considered under the same uncertainty model and reliability requirements. D³Opt addresses this need by learning feasibility without using any objective, allowing the frozen prior to support subsequently specified objectives. Moreover, changing ρ\rho defines a family of chance-feasible sets, which our risk-conditioned prior is designed to model. The multi-risk experiments (Table 10) and ASQ/MW4 objective-transfer experiments (Table 3) examine these two forms of reuse without retraining.

We agree that adapting to shifts in the uncertainty distribution or changes in constraint structure is an important problem. Extending our framework to these settings is a direction for future work. The present paper focuses on cross-objective reuse and risk-level adaptation under a shared uncertainty model.

reference: [1] Charnes, A., and Cooper, W. W. (1959). Chance-Constrained Programming. Management Science, 6(1), 73–79. [2] Charnes, A., and Cooper, W. W. (1962). Chance Constraints and Normal Deviates. Journal of the American Statistical Association, 57(297), 134–148. [3] Cui, Y.-W., Guo, S.-Y., Wang, X., and Xiao, X.-T. (2025). A Brief Review of Recent Advances on Chance Constrained Programs. Journal of the Operations Research Society of China.

AuthorsOpenReview ↗

Part2(W2 & Q3): Optimization baselines and comparison with ALSO-X and ALSO-X+

Our original evaluation includes SAA-CVaR, SAA-MIP, SAA-PDCA(Wang et al., 2023), and SAA-SNSCO(Zhou et al., 2024), covering both classical and more recent optimization methods. We appreciate the suggestion and have added comparisons with ALSO-X and ALSO-X+(Jiang and Xie, 2022) on the linear Gaussian CCP and CCED in Table2 and Table4.

On Gaussian, D$^3$Opt achieves a lower mean objective of −0.6519-0.6519, compared with −0.6406-0.6406 for ALSO-X and −0.6384-0.6384 for ALSO-X+, and a smaller standard deviation of 0.00150.0015. On CCED, both ALSO methods with exact feasibility projection match the projected DE reference cost of 104399.055104399.055, compared with D$^3$Opt's 104823.577104823.577. D$^3$Opt is approximately 0.407%0.407\% above this reference and achieves the lowest mean cost among the evaluated learning-based methods. These results are reported and discussed in the revised Gaussian and CCED experiments.

AuthorsOpenReview ↗

Part3(W3&Q1): Contribution relative to DiffOPT and FK Steering

Thank you for raising this important question. We would like to clarify our contribution is a CCP framework combining objective-independent, risk-conditioned feasibility learning with an annealed, derivative-free FK construction for optimizing post-specified objectives. This allows a single frozen feasibility model to support different objectives and risk levels, with objective preferences introduced only during inference.

This design addresses specific limitations of existing approaches and obtains better experiment results. GGDOpt supports risk conditioning, but its training solutions depend on the objective and its inference uses objective derivatives. DiffOPT already supports objective-independent prior reuse, but its derivative-free method relies on iterative importance sampling with importance weights proportional to πβ(x)/Qk(x)\pi_\beta(x)/Q_k(x). This requires approximating the proposal density Qk(x)Q_k(x), while the target density also involves pθ(x)p_\theta(x), which is not directly available from a standard score-based diffusion model. Such density evaluations or approximations introduce additional approximation considerations. While our framework addresses CCPs by constructing decision--risk training pairs directly through objective-independent proposals and chance-constraint screening, then optimize over the learned prior using only objective values, without explicitly evaluating the learned-prior or proposal densities. The comparison table added to Section 2 clarifies these distinctions.

Compared with FK Steering, our contribution is the CCP-specific feasibility-learning procedure and its coupling with annealed FK optimization. We introduce an increasing inverse-temperature schedule into the telescoping potentials along the frozen reverse diffusion. This schedule is designed to limit premature weight concentration by gradually strengthening objective preference, while retaining the terminal target proportional to pθ(x∣ρ)exp⁡[−βKf(x)]p_\theta(x\mid\rho)\exp[-\beta_K f(x)]. To isolate the effect of annealing, we compare annealed (✓) and constant-temperature (✗) FKC using the same frozen prior, inference settings, and terminal βmax⁡\beta_{\max}. Only the temperature schedule changes.

βmax⁡\beta_{\max}AnnealingCost Mean ↓Cost Std ↓
5✓104977.214124.773
5✗105131.978161.594
10✓104823.577102.065
10✗105362.210134.097
20✓104768.454104.858
20✗105899.777760.602

Annealing yields lower mean costs and standard deviations at all three tested βmax⁡\beta_{\max}, supporting this design on CCED-VPE-10. Beyond this controlled comparison, the ASQ/MW4 transfer results (Table3 of the revised manuscript) and multi-risk experiments (AppendixD.4) demonstrate frozen-model reuse across objectives and tested risk levels. Our analysis connects this construction to the underlying CCP: Proposition 1 establishes feasibility preservation when the prior is almost surely feasible; under the stated assumptions, Proposition 2 bounds the expected objective error of the weighted terminal particles relative to the original CCP optimum through prior-coverage, finite-particle, and finite-temperature terms.

We have also revised the Introduction to explicitly acknowledge DiffOPT's reuse capability and clarify that the example of objective-dependent training-data construction refers to GGDOpt.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Manuscript Revisions and Additional Results

We thank the chairs and all reviewers for their constructive feedback. We have revised the manuscript and provided additional analyses in our responses. Revisions and additions are highlighted in blue in the updated manuscript. The main changes are summarized below.

  1. Clarified the contribution and scope. Sections 1–2 and Table 1 clarify the differences from DiffOPT and GGDOpt. Section 6 identifies adaptation to changing constraints and uncertainty distributions as future work. (FB2x: W1/Q2, W3/Q1; idxh: W1)

  2. Added ALSO-X/ALSO-X+ comparisons on Linear Gaussian, ASQ, MW4, and CCED-VPE-10 in Sections 5.1–5.3 and Tables 2–4. (FB2x: W2/Q3; idxh: W4)

  3. Expanded data-generation details in Appendix C.1, specifying proposals, risk sampling, screening rules, scenario counts, and training-set sizes. (idxh: W2)

  4. Added computational-cost reporting in Appendix C.6 and Tables 6–7, separating data generation and training from online optimization. (idxh: W3/W4)

  5. Added dimension-scaling experiments at n=8,16,32n=8,16,32 in Appendix D.1 and Table 8, using separately trained priors and exact SOC references. (idxh: W4)

  6. Added proposal-sensitivity experiments and distribution diagnostics in Appendices D.2–D.3, Table 9, and Figure 1. (idxh: W2; NTBA: W1/Q1)

  7. Provided additional evidence in the responses: controlled annealing and objective-evaluation ablations, together with a conditional coverage-gap diagnostic based on unguided samples. (FB2x: W3/Q1; idxh: W2/W4; NTBA: W1/Q1, W2)

We sincerely appreciate the time and effort that the chairs and reviewers have devoted to evaluating our work. We hope the revised manuscript and additional results address your questions and concerns.

Good job. I recommend acceptance and have raised my score.

I also sincerely appreciate the authors’ rigor and care—qualities that are especially valuable in the current era of automated research.

We are deeply grateful for your recognition and encouragement. Your support means a great deal to us and strengthens our commitment to pursuing research with dedication, rigor, and care.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.