acceptodds
Under review as a conference paper at ICLR 2027

Constrained Graph Diffusion for Mixed-Integer Optimization

Abstract

This paper proposes a novel learning-based approach to approximately solve instances of mixed-integer optimization problems. These problems are computationally challenging, as they require jointly determining discrete and continuous decisions while satisfying complex combinatorial constraints. We introduce Constrained Graph Diffusion (CGD), a learning-based framework that approximately solves recurring instances of such problems by learning a conditional distribution over their discrete decisions. CGD uses a graph-based diffusion model and incorporates constraint information directly into the reverse diffusion process, steering intermediate predictions toward the feasible region throughout generation. By operating on continuous relaxations of the discrete variables, CGD defines a differentiable constrained generation pathway up to terminal discrete recovery. Once the discrete decision is recovered and fixed, a numerical optimizer solves the remaining continuous problem, avoiding online combinatorial search over the binary variables while retaining numerical optimization for continuous completion. We evaluate CGD on AC-OPF with branch switching and discrete portfolio optimization, demonstrating substantial improvements in feasibility and solution quality over learning-based baselines while achieving speedups of up to over state-of-the-art MIP solvers on large instances.

Primary area:
generative models
Keywords:
Constrained Generation, Diffusion Models, Mixed-Integer Optimization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7rBMOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Constrained Graph Diffusion (CGD) for mixed-integer optimization. The main idea is to use a diffusion model to predict integer variables and then solve the remaining continuous problem with a numerical solver. To improve feasibility, the authors introduce constraint correction during the diffusion process. The authors test their method on AC-OPF with branch switching and portfolio optimization. The experiments show improvements in solution quality and feasibility compared with several learning-based baselines, as well as significant speedups over Gurobi.

Critical strengths

Good experimental results: The authors test the method on two different optimization problems. The results show good improvements in solution quality and feasibility. The ablation study also shows the benefit of applying constraint correction during the diffusion process.

Critical weaknesses
  1. Missing recent baselines: This paper does not compare with some recent methods, such as [1] and [2]. Especially, the latter also predicts integer variables first and then solves the continuous problem. I think these papers are closely related to this work, and the authors should compare with them or explain why they are not included.

  2. Limitation of the proposed method: I have some concerns about the proposed framework. The authors first predict the integer variables and then solve the continuous problem. However, even if the continuous problem is solved exactly, the final solution is not guaranteed to be optimal for the original mixed-integer problem. Once the integer variables are fixed, there is no way to update them based on the continuous optimization results. I think this is a fundamental limitation of this framework, especially when the integer and continuous variables are strongly coupled.

  3. Potential issue in the AC-OPF formulation: In Appendix F.1, constraint (47c) is written as

−Δθijzij≤∠(ViVj∗)≤Δθijzij.-\Delta\theta_{ij}z_{ij}\leq \angle(V_iV_j^*)\leq\Delta\theta_{ij}z_{ij}.

When zij=0z_{ij}=0, this constraint forces the phase angle difference to be zero, even though the corresponding line is disconnected. I think this is incorrect, since the phase angle constraint should be relaxed when the line is disconnected. If this constraint is also used in the implementation, it may change the feasible region of the original problem and affect the experimental results. The authors should clarify this issue.

Reference: [1] FMIP: Joint Continuous-Integer Flow For Mixed-Integer Linear Programming, ICLR 2026 [2] Unsupervised Neural Langevin Sampler for Mixed Integer Linear Programming, ICML 2026

Questions for authors

Please see "Critical Weakness".

4 comments
AuthorsOpenReview ↗

Authors' response

We thank the reviewer for the helpful review and, in particular, for acknowledging the quality of the results produced by the proposed method, the importance of the ablation study and the clarity of the presentation. Below, we address the reviewer's concern.

Weaknesses.

Missing recent baselines. The paper does not compare with methods such as FMIP (Joint Continuous-Integer Flow For Mixed-Integer Linear Programming) and Unsupervised Neural Langevin Sampler for Mixed Integer Linear Programming.

Response. We thank the reviewer for highlighting these recent works. However, both FMIP [1] and the Unsupervised Neural Langevin Sampler [2] are specifically designed for mixed-integer linear programming (MILP). FMIP [1] uses objective and feasibility guidance from the linear formulation during flow matching. The Langeving Sampler [2] learns the integer decisions through Langevin dynamics and recovers the continuous variables with a linear program.

CGD instead addresses a more general class of mixed-integer optimization problems, including nonconvex MINLPs. AC-OTS involves nonlinear, nonconvex constraints, and fixing the binary transmission-switching decisions leaves a nonconvex problem. The portfolio benchmark also has quadratic objectives and constraints. Applying either method would require substantial effort adapting its formulation and constraint-handling mechanisms to accommodate nonlinear optimization.

We therefore focus our experimental comparisons on methods that can be applied to the considered benchmarks, including DIFUSCO, learning-based prediction baselines, and controlled ablations isolating the contribution of constraint-aware diffusion. We have included these two references in the related-work discussion to clarify the distinctions and acknowledge their contributions.

Limitation of the proposed method: I have some concerns about the proposed framework. The authors first predict the integer variables and then solve the continuous problem. However, even if the continuous problem is solved exactly, the final solution is not guaranteed to be optimal for the original mixed-integer problem. Once the integer variables are fixed, there is no way to update them based on the continuous optimization results. I think this is a fundamental limitation of this framework, especially when the integer and continuous variables are strongly coupled.

Response. On the optimality and coupling of discrete and continuous decisions. We agree that fixing the discrete decisions and subsequently optimizing the continuous variables does not guarantee global optimality for the original mixed-integer problem, nor do we make such a claim. Finding and certifying a global optimum for a nonconvex MIP can be computationally prohibitive, which motivates our focus on efficiently high-quality solutions.

CGD accounts for the coupling between discrete and continuous decisions before fixing a discrete assignment. In AC-OTS, the augmented-Lagrangian correction uses the nonlinear AC constraints and their gradients with respect to the relaxed switching decisions, to refine the discrete network topology during reverse diffusion. Our experiments show that CGD achieves low objective gaps, high observed downstream feasibility while substantially reducing solve times.

Potential issue in the AC-OPF formulation.

Response. The reviewer is correct. As written, the phase-angle constraint in Eq. (47c), enforces a zero phase-angle difference when a transmission line is disconnected. We corrected the equation of the AC-OTS formulation in the revised manuscript by restoring the big-M relaxation, analogous to the already correct DC-OTS formulation in Appendix F.3. Our implementation already relaxes the phase-angle constraint when the corresponding line is disconnected, so the feasible set used in the experiments and the reported results are unaffected.

[1] Hongpei Li, Hui Yuan, Han Zhang, Jianghao Lin, Dongdong Ge, Mengdi Wang, and Yinyu Ye. “FMIP: Joint Continuous-Integer Flow For Mixed-Integer Linear Programming.” ICLR, 2026.

[2] Yixin Huang, Shengyu Feng, and Yiming Yang. “Unsupervised Neural Langevin Sampler for Mixed Integer Linear Programming.” ICML, 2026.

AuthorsOpenReview ↗

Authors' comment

Thank you again for your review. We wanted to follow up on our clarification of CGD’s relationship to FMIP and the Unsupervised Neural Langevin Sampler, and how our constraint corrections account for discrete–continuous coupling without guaranteeing global optimality. We also restored the big-M relaxation in Eq. (47c) and clarified that our implementation already relaxed the phase-angle constraint for disconnected lines, so the reported results are unaffected. Please let us know if any part of our explanation or the corrected formulation needs further clarification. If these responses address your concerns, we would appreciate it if you would reconsider your rating in light of the revised manuscript.

Thank you for your clarification. I will increase my score to 3 and maintain my highest confidence score.

Thank you very much!

Reviewer_Cg5jOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CGD targets recurring instances of mixed-integer optimization problems. It uses a graph-based diffusion model to generate discrete decisions, applies constraint corrections to relaxed predictions at every reverse denoising step, and performs terminal binary recovery, followed by numerical optimization for continuous completion. The key contributions of this work is proposing that guiding each denoising step with constraints produces better solutions than repairing them only at the end, while reducing the need for online combinatorial search.

Critical strengths

1, CGD feeds corrected predictions back into subsequent denoising steps, allowing the model to readjust the entire decision structure. This is a well-motivated approach.

2, The overall ablation design is well targeted and structural.

Critical weaknesses

1, The method primarily integrates existing components. Its theoretical results rely on exact convex projection and do not directly cover the approximate, nonconvex correction used in the AC-OPF experiments.

2, Some reported Hamming distances and exact reconstruction rates are mathematically inconsistent. For the 9-bus case, the mean normalized Hamming distance must be at least 30.57/12≈2.5530.57/12\approx2.55%, whereas the table reports 1.71%.

3, Generalization to unseen problem sizes and offline training costs are insufficiently evaluated.

Questions for authors

1, Can the model generalize to unseen network sizes or topologies?

2, When data generation and training costs are included, how many instances must be solved before the method achieves an overall computational benefit?

4 comments
AuthorsOpenReview ↗

Authors' response

We thank the reviewer for the constructive feedback and for recognizing the motivation behind our approach and the relevance of our ablation studies. We address the concerns below.

Weaknesses.

The method primarily integrates existing components.

Response. We agree that CGD builds upon established components. As it is common in machine learning, combining existing techniques can be a methodological contribution when the combination enables a new capability. CGD brings together diffusion, constraint correction, and numerical optimization to generate high-quality solutions to mixed-integer nonlinear problems without online combinatorial search. Our ablations show that applying constraint corrections throughout generation improves solution quality over both unconstrained diffusion and correction only at the final step. We believe that CGD provides a valuable contribution to the community of generative models for optimization.

The theoretical results of the method rely on exact convex projection and do not directly cover the approximate, nonconvex correction used in the AC-OPF experiments.

Response. We acknowledge that Propositions 1–2 characterize exact projections onto convex feasible sets and do not directly establish equivalent guarantees for the approximate, nonconvex correction employed in AC-OTS. However, extending these guarantees to general nonconvex settings is extremely difficult, since projections onto nonconvex sets can have multiple minmizers and finding globally optimal projections can itself be computationally expensive. Our theoretical results establish the foundations of the proposed projection-based framework, characterizing the ideal projection mechanism under well-defined assumptions. The approximate correction employed in AC-OTS builds upon this foundation to extend the underlying principle to challenging nonlinear applications. Its effectiveness is established empirically through controlled ablation studies, demonstrating substantial improvements in discrete prediction quality, downstream feasibility and objective value.

Correction of the Hamming Distance metric.

Response. Thank you for catching the inconsistency. We recomputed the Hamming distances and corrected four entries for AC-OTS (Table 1) and five for DC-OTS (Table 5). The exact reconstruction rates, generated solutions, downstream results, and relative performance of the methods are unchanged.

Questions.

Can the model generalize to unseen network sizes or topologies?

Response. CGD is designed as a general framework for mixed-integer optimization. Its graph-based architecture operates on node and edge representations and supports different network sizes. We evaluate it across the 9, 197, 500-buses, as well as two portfolio optimization benchmarks with 50 and 150 assets. For each benchmark, we train a separate model per network and test it on held-out instances with different operating conditions and parameters.

As already mentioned in the Conclusion section of the manuscript, the development of foundation diffusion models capable of handling unseen network sizes or topologies for mixed-integer optimization problems represents an interesting direction for future research.

When data generation and training costs are included, how many instances must be solved before the method achieves an overall computational benefit?

Response. We agree that offline computational costs are important to assessing the overall efficiency of the proposed method. CGD is designed for recurring instances of mixed-integer optimization problems, where data generation and model training represent one-time costs that can be amortized over subsequent optimization instances.

To provide a more comprehensive assessment of the computational requirements of CGD, we have additionally included in the manuscript the offline costs associated with data generation and model training in the following table.

Computational costs for dataset generation and CGD training, reported in CPU solver-hours and GPU hours, respectively.

TaskProblem sizeCGD training costData-generation cost
AC-OTS9 buses0.89 GPU-h1,6811{,}681 CPU solver-h
AC-OTS197 buses36.01 GPU-h4,9244{,}924 CPU solver-h
AC-OTS500 buses21.62 GPU-h2,5372{,}537 CPU solver-h
Portfolio50 assets4.70 GPU-h1.017 CPU solver-h
Portfolio150 assets10.31 GPU-h1.646 CPU solver-h

Determining a precise break-even number requires accounting for the relative costs of CPU and GPU resources which goes beyond the scope of our work. We therefore report the offline computational costs alongside the end-to-end inference runtimes and matched-gap solver comparisons in Table 3.

We emphasize that these offline costs are incurred once, whereas the computational savings accumulate over repeated optimization instances, which is precisely the setting motivating the proposed framework.

AuthorsOpenReview ↗

Authors' comment

Thank you again for your review. We wanted to follow up on our clarification of the exact convex projection assumptions underlying the theory and the empirical evidence for approximate corrections in nonconvex AC-OTS. We also corrected the Hamming-distance entries, reported offline data-generation and training costs, and clarified that generalization to unseen network sizes or topologies remains future work. Please let us know if any part of our explanation or the reported costs needs further clarification. If these responses address your concerns, we would appreciate it if you would reconsider your rating in light of the revised manuscript.

Thank you for your detailed response and clarifications. Your responses have addressed most of my concerns. I will increase my overall rating accordingly.

Thank you very much!

Reviewer_37MgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a diffusion-based approach for mixed-inter linear programs, which is called CGD. Specifically, this approach predicts integer variables, and integrate the constraint-related information into the diffusion process. Within the process, numerical optimization solvers are used to restore continuous variables. Both theoretical materials and experimental results are provided.

Critical strengths
  1. The proposed approach decomposites discrete and continuous variables, which is a well-motivated approach, especially for prohibitively large mixed-integer problems.

  2. The authors provided plenty of theoretical supports for the proposed framework.

  3. The numerical results are promising, particularly on large-scale AC-OPF instances, where the proposed method achieves significant acceleration.

Critical weaknesses
  1. The novelty is limited. The idea of incorporating constraint projections into diffusion models for MILPs has been well-explored by other works such as [1], [2]. The proposed method seems to be an incremental to the existing works.

  2. The generality is limited as well. The author state that the proposed framework is for general MILPs, but it requires problem specific information for constraint correction. Besides, the dataset for AC-OPF seems to be generated seperately for different settings, and the perturbation does not significantly varies. It is unclear whether CGD can work on other general MILPs.

  3. The nuerical experiments lack comparisons with other diffusion baselines and simple problem-specific heuristics such as [3]. Besides, the reported performance on AC-OPF appears to be extremely good given that multiple denoising and constraint-correction steps. It is not clear if the training and testing is fair and proper.

[1] Christopher, Jacob K., Stephen Baek, and Ferdinando Fioretto. "Constrained synthesis with projected diffusion models." Advances in Neural Information Processing Systems 37 (2024): 89307-89333.

[2] Hong, Seong-Hyun, et al. "Unsupervised training of diffusion models for feasible solution generation in neural combinatorial optimization." arXiv preprint arXiv:2411.00003 (2024).

[3] Sun, Zhiqing, and Yiming Yang. "Difusco: Graph-based diffusion solvers for combinatorial optimization." Advances in neural information processing systems 36 (2023): 3706-3731.

Questions for authors
  1. Please justify what is the fundamental advances of the proposed framework comparing to previous diffusion-based works. How does its performance change under their experimental settings?

  2. How well does CGD generalize to unseen/structurally different problems? Diffusion models on CO shows proper generalization power. Is it possible to see how do them perform on the AC-OPF dataset (only predict binary variables)?

4 comments
AuthorsOpenReview ↗

Authors' response (1/3)

We thank the reviewer for the helpful review, and, in particular, for acknowledging that the paper introduces a well-motivated decomposition of discrete and continuous variables in mixed-integer non linear programs, that the experiments are supported by theoretical findings and that the numerical results are promising, particularly in large scale, nonconvex AC OPF instances. In what follows we will: 1) provide a detailed answer to your questions and 2) report the revised part of the paper in blue.

Weaknesses.

The novelty is limited. The idea of incorporating constraint projections into diffusion models for MILPs has been well-explored by other works such as [1], [2]. The proposed method seems to be an incremental to the existing works.

Response. We believe this is not correct. Please note that, while it is true that constraints have been incorporate into diffusion models and we are well aware of this space, no previous approach (including Christopher et al. [1] and Hong et al. [2]) incorporates constraints into diffusion for solving mixed-integer optimization problems.

Christopher et al. [1] employ projections for constrained generation (trajectories, images) but not for optimizaiton problems; while Hong et al. [2] focus specifically on discrete combinatorial optimization problems, not on Mixed integer problems and encourages feasibility through training-time methods, not, as we do, through sampling time projection-based methods.

Our approach goes substantially beyond these approaches. Its central contribution is a unified framework that decomposes mixed-integer optimization into constraint-aware generation of discrete decisions and numerical optimization of the remaining continuous variables. This decomposition fundamentally changes the role of diffusion as it allow us to focus diffusion for the (difficult) discrete component while preserving numerical optimization for the continuous subproblem. This enables the framework to address challenging mixed-integer nonlinear problems, including large-scale AC-OPF with transmission switching, which are outside the scope of the cited approaches.

The generality is limited as well. The author state that the proposed framework is for general MILPs, but it requires problem specific information for constraint correction. Besides, the dataset for AC-OPF seems to be generated separately for different settings, and the perturbation does not significantly vary. It is unclear whether CGD can work on other general MILPs.

Response. Generality and data generation. We would like to clarify that CGD is formulated as a general framework for solving instances of mixed-integer optimization tasks, encompassing both MILPs and MINLPs, rather than a method specifically designed for AC-OTS. The AC-OTS and portfolio experiments demonstrate its applicability to two fundamentally different problem classes. AC-OTS involves combinatorial decisions over network edges, subject to graph connectivity and nonlinear physical constraints, whereas portfolio optimization involves asset-selection decisions subject to cardinality and quadratic constraints, followed by continuous allocation through a convex quadratic program. These applications differ substantially in their combinatorial structure, constraint formulations, and continuous optimization problems, demonstrating the flexibility of the proposed framework. The experimental results achieved on both the AC-OTS and portfolio benchmarks suggest that CGD can accommodate other classes of MIPs that go beyond the settings studied in the paper.

[1] Jacob K. Christopher, Stephen Baek, and Ferdinando Fioretto. “Constrained Synthesis with Projected Diffusion Models.” NeurIPS, 2024

[2] Seong-Hyun Hong, Hyun-Sung Kim, Zian Jang, Deunsol Yoon, Hyungseok Song, and Byung-Jun Lee. “Unsupervised Training of Diffusion Models for Feasible Solution Generation in Neural Combinatorial Optimization.” arXiv preprint, 2024.

AuthorsOpenReview ↗

Authors' response (2/3)

AC-OTS data generation. Regarding the data generation for AC-OTS, note that constructing datasets for the AC OTS problem requires careful consideration of two competing objectives: (i) ensuring that the perturbed instances remain feasible and (ii) generating sufficiently diverse optimal branch switching decisions, without which the learning task would become trivial. This is particularly challenging for AC-OTS, where nonlinear, nonconvex power-flow equations and strict operational limits restrict the admissible range of system parameters. Importantly, limited perturbation to the problem parameters do not necessarily imply limited solution diversity: due to the nonlinearity and nonconvexity of AC-OPF, even small changes in parameters, such as system demand, can lead to substantially different solutions, as previously documented by Zhang et al. [1]. The additional combinatorial nature of transmission switching further allows changes in operating conditions to alter the resulting network topology. Our dataset generation therefore balances physically meaningful perturbations, feasibility, and diversity of switching decisions, as detailed in Appendix F.2. To reflect these considerations, we have added the following paragraph to the section on data generation for AC-OTS.

[1] Learning to solve the ac optimal power flow via a lagrangian approach, Zhang et al., 2022 North American Power Symposium (NAPS)

Data Generation

This data-generation procedure balances two important considerations: (i) preserving the feasibility of the resulting optimization instances and (ii) inducing sufficiently diverse optimal switching decisions, without which the learning task would become trivial. This is particularly challenging for AC-OTS, where nonlinear, nonconvex power-flow equations and strict operational limits restrict the admissible parameter ranges. Nevertheless, even relatively small parameter perturbations can substantially affect the resulting solutions due to the nonlinear nature of the problem. The selected perturbation ranges are therefore intended to generate physically meaningful instances while promoting diversity in the discrete switching decisions.

The numerical experiments lack comparisons with other diffusion baselines and simple problem-specific heuristics such as DIFUSCO(sun2023difusco).Besides, the reported performance on AC-OPF appears to be extremely good given that multiple denoising and constraint-correction steps. It is not clear if the training and testing is fair and proper.

Response. Baselines and computational efficiency.

We would like to clarify that comparisons with DIFUSCO (sun2023difusco), a state-of-the-art graph-based diffusion solver, are already described in Section 6 (Baselines and Evaluation) and Appendix E.1, with numerical results reported in Tables 1–2. Additionally, we include two controlled ablations: unconstrained CGD and CGD with final-step projection.These comparisons isolate the benefits of incorporating constraint information throughout the reverse diffusion trajectory, as the experimental results demonstrate.

Regarding the computational performance on AC-OTS, the substantial speedups arise from two key aspects of the proposed framework. First, CGD replaces a computationally expensive projection onto the discrete feasible set with an approximate continuous correction.

Second, and more importantly, CGD avoids the combinatorial search required by the original mixed-integer nonlinear problem by generating the discrete decisions and subsequently solving only the remaining continuous AC-OPF problem. As reported in Table 3, the entire diffusion sampling procedure, including constraint corrections, requires only 0.13, 0.16, and 0.11 seconds for the 9-, 197-, and 500-bus benchmarks, respectively. The majority of CGD's end-to-end runtime is instead devoted to solving the continuous optimization subproblem induced by the predicted discrete decisions.

Regarding experimental fairness, all methods are evaluated on the same held-out test instances, with an 80/10/10 training, validation, and test split. Hyperparameters are selected using the validation set, and the reported CGD runtimes include both diffusion sampling and downstream numerical optimization. Furthermore, comparisons with Gurobi are conducted at matched objective gaps, as described in Section 6 and Table 3.

AuthorsOpenReview ↗

Authors' response (3/3)

To address the reviewer concern and provide a more comprehensive assessment of the computational requirements, we have added the following table to the manuscript, detailing the offline computational costs associated with data generation and CGD model training.

Computational costs for dataset generation and CGD training. Data-generation costs are reported in CPU solver-hours, while training is reported in GPU hours.

TaskProblem sizeCGD training costData-generation cost
AC-OTS9 buses0.89 GPU-h1,6811{,}681 CPU solver-h
AC-OTS197 buses36.01 GPU-h4,9244{,}924 CPU solver-h
AC-OTS500 buses21.62 GPU-h2,5372{,}537 CPU solver-h
Portfolio50 assets4.70 GPU-h1.017 CPU solver-h
Portfolio150 assets10.31 GPU-h1.646 CPU solver-h

We emphasize that these offline costs are incurred only once and can be amortized over recurring optimization instances.Once the dataset has been generated and the CGD model is trained, subsequent instances can be solved efficiently, benefiting from the computational speedups reported in Table 3.

An analysis on the trade-off between training and data generation cost, CGD generation time and quality of the decisions produced by CGD goes beyond the scope of this work. Consequently, the offline costs and end-to-end runtimes are reported separately to enable a transparent assessment of the computational trade-offs.

Questions.

Please justify what is the fundamental advances of the proposed framework comparing to previous diffusion-based works. How does its performance change under their experimental settings?

Response. The fundamental contribution of CGD is the development of a constraint-aware diffusion framework for mixed-integer optimization, which jointly addresses the combinatorial structure of discrete decisions and their interaction with downstream continuous optimization. Unlike previous diffusion-based approaches, which primarily focus on constrained data synthesis, purely combinatorial optimization, or mixed integer linear programming, CGD introduces a framework that (i) generates relaxed binary decisions while incorporating problem-specific feasibility corrections throughout the reverse diffusion trajectory, (ii) recovers discrete decisions only at the terminal step, and (iii) solves the remaining continuous optimization problem using a numerical solver. Importantly, this framework allows us to address mixed-integer nonlinear problems without performing online combinatorial search.

Regarding performance under the experimental settings of the cited approaches, we emphasize that these works consider constrained data generation, discrete combinatorial optimization problems, such as ATSP and scheduling and MILPs. Consequently, these benchmarks fall outside the primary scope of CGD, and direct comparisons would require adapting the proposed framework to their specific formulations. Our experiments demonstrate the applicability of CGD to two substantially different classes of mixed-integer optimization problems, including large-scale nonconvex AC-OTS, which is not addressed by the cited approaches.

How well does CGD generalize to unseen/structurally different problems? Diffusion models on CO shows proper generalization power. Is it possible to see how do them perform on the AC-OTS dataset (only predict binary variables)?

Response. Table 1 reports the requested binary-only comparison on AC-OTS. Specifically, we evaluate DIFUSCO sun2023difusco, a state-of-the-art graph-based diffusion solver, adapted to predict the discrete decisions and measure the Hamming distance and exact reconstruction against reference solutions. We also evaluate feasibility and objective value afrer solving the resulting continuous AC-OPF problems. Across all network sizes, CGD outperforms DIFUSCO on discrete prediction quality and downstream performance, demonstrating the importance of incorporating application-specific constraint guidance throughout generation.

Regarding generalization, we test CGD on held-out instances generated by varying the operating conditions and network parameters. Generalization across structurally different networks, is an interesting direction for future work, as it is already reported in the conclusion paragraph of the main paper. By operating on node and edge representations, the graph-based denoiser can accommodate different network structures, suggesting a promising direction for future work.

AuthorsOpenReview ↗

Authors' comment

Thank you again for your review. We wanted to follow up on our clarification of CGD’s contribution to mixed-integer nonlinear optimization, its relationship to prior diffusion methods, and the scope of our generalization evaluation. We also clarified the AC-OTS data-generation and held-out evaluation protocols, highlighted the DIFUSCO comparisons and controlled ablations, and reported offline data-generation and training costs. Please let us know if any part of our explanation or evaluation needs further clarification. If these responses address your concerns, we would appreciate it if you would reconsider your rating in light of the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Authors' comment

Dear Reviewers,

Thank you for your constructive feedback. We revised the manuscript and addressed your questions in our individual replies.

  • Related work and baselines. We explain how CGD differs from FMIP and the Unsupervised Neural Langevin Sampler and point to the DIFUSCO comparisons in Tables 1 and 2.
  • Scope of the theory. We clarify that the projection guarantees apply to exact convex projections. We build on that foundation by using approximate corrections in nonconvex settings and evaluating their effectiveness empirically.
  • Data and computational costs. We explain how AC-OTS data generation balances feasibility and switching diversity (Appendix F.2) and report offline dataset-generation and training costs (Table 7).
  • Corrections to the results tables and formulation. We corrected the Hamming-distance entries in Tables 1 and 5 and restored the big-M relaxation in the AC-OTS phase-angle constraint (Eq. 47c).

We hope our responses and revisions address your concerns, and we would appreciate it if you could revisit your assessment of our work in light of these updates.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.