acceptodds
Under review as a conference paper at ICLR 2027

Beyond Feasibility: Solution-Landscape Preserving Generation of Mixed-Integer Linear Programs

Abstract

Real-world Mixed-Integer Linear Programming (MILP) instances are scarce and costly to collect, limiting the training and evaluation of learning-based MILP solvers. Existing instance generation methods mainly preserve structural similarity, computational difficulty, feasibility, and boundedness. However, these properties do not ensure solution-landscape preservation: a generated instance may remain structurally similar and feasible while substantially altering the representative solutions and their distribution compared with those of the source instance. We propose Solution-Landscape Preserving Generation of Mixed-Integer Linear Programs (SLP-MILP), a controllable generation framework that explicitly incorporates source solution behavior into the generation process. SLP-MILP constructs solution-aware constraint groups based on the activation relationships between solutions and constraints as well as the constraint structure and regenerates selected groups through solution-conditioned masked generative reconstruction under an explicit structural modification budget. For each regenerated constraint direction, its right-hand side is determined by aligning the constraint-response profile with that of the original constraint on the source solution-landscape. A separate conformal retention safeguard prevents the aligned constraints from discarding excessive source probability mass. Experiments on multiple MILP families demonstrate the effectiveness of SLP-MILP in preserving solution-landscape characteristics while maintaining structural similarity.

Primary area:
optimization
Keywords:
MILP, Solution-Landscape Preservation, Instance Generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hohQOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses MILP instance generation, arguing that existing generators (G2MILP, ACM-MILP, MILP-FBGen) preserve structural similarity, solving hardness, or feasibility, but not the solution landscape — the distribution of representative feasible and near-optimal solutions. The authors formalize a conditioned solution distribution πP\pi_P (a multi-temperature Gibbs mixture over the feasible set, weighted by normalized objective quality), approximated by a solver-assisted LNS pool. The proposed SLP-MILP: (i) builds diverse source solution pools via perturbed-objective LNS; (ii) groups constraints by combining topological overlap with solution-activity correlations; (iii) regenerates selected constraint groups under a modification budget η\eta using a solution-conditioned masked VAE with a spike-and-slab decoder; and (iv) sets RHS values by aligning multi-quantile slack-response profiles with the source constraint, subject to a one-sided split-conformal retention safeguard guaranteeing at least 1−ϵf1-\epsilon_f of source solution mass remains feasible. The evaluation on MIK/CA/IP/IS uses eight newly proposed solution-landscape metrics (Chamfer, SymNN, support/active Jaccard, MMD, SWD, correlation shift, objective shift) plus structural similarity, and a downstream variable-fixing task. SLP-MILP wins on most landscape metrics, keeps competitive structural similarity (except on MIK, where it is notably lower), and achieves the best downstream fixing result (327 variables fixed, 49.3% time reduction).

Critical strengths
  1. Novel and well-argued problem formulation. To my knowledge this is the first work to explicitly target solution-distribution preservation in MILP instance generation. The observation that F(P′)≠∅⇏πP′≈πPF(P')\neq\emptyset \not\Rightarrow \pi_{P'}\approx\pi_P — e.g., a generator may preserve only trivial witness solutions such as the zero vector in MIS-style formulations — is correct and consequential: it changes the evaluation target for the whole instance-generation line of work (G2MILP, ACM-MILP, MILP-FBGen), and the downstream variable-fixing experiment shows the preserved landscape has measurable practical value.

  2. Principled, guaranteed safeguards. The one-sided split-conformal retention correction (Appendix B.3) yields a distributionally valid bound Pr⁡[ai′⊤X≤b′]≥1−ϵf\Pr[a_i'^\top X \le b']\ge 1-\epsilon_f under exchangeability, and decoupling direction generation (VAE) from tightness calibration (response-profile alignment) is technically clean and clearly differentiated from prior purely generative pipelines.

Critical weaknesses
  1. Partially circular evaluation protocol. Both source and generated solution pools are constructed by the authors' own LNS/solver-sampling procedure, and all headline metrics are proposed by the authors with no external validation or ground truth. Because identical pipeline settings are used on both sides, the metrics may reward any method that induces similar sampling behavior. In particular, there is no comparison against a trivial baseline that perturbs only the RHS while retaining ≥1−ϵf\ge 1-\epsilon_f of source mass (leaving AA and cc unchanged), which would plausibly also score well on several landscape metrics. This makes the central empirical claim hard to falsify as presented.

  2. Limited evidence for the motivating use case. Structural similarity degrades on MIK (0.887 vs. 0.991 for MILP-FBGen at η=0.1\eta=0.1); all four families are standard synthetic generators of modest size (MIK: 387 variables; at most 1500 variables overall); MILP-FBGen results come from the authors' own reimplementation and score anomalously poorly (e.g., Chamfer 2.093 on MIK vs. 0.127 for SLP-MILP), raising reproduction-fidelity questions; and the pipeline's cost — pool construction via a commercial solver — is acknowledged as the main limitation but never quantified against baselines. The motivating scenario (scarce, large, complex real-world industrial instances) is never tested, and no sensitivity analysis is given for ϵf\epsilon_f, the temperature mixture, or pool size KK.

Questions for authors
  1. (Assessment) How was the MILP-FBGen reimplementation validated? Its landscape scores are 5–20× worse than everything else (e.g., Chamfer 2.093 on MIK vs. 0.127 for SLP-MILP), which raises reproduction-fidelity questions for Table 2.
  2. (Assessment) Would a trivial "loose-RHS" baseline — perturbing or relaxing bib_i only, retaining 1−ϵf1-\epsilon_f source mass with A,cA, c unchanged — already achieve good scores on your eight metrics? If not, why not? This would establish that the VAE reconstruction, rather than the retention safeguard, drives the gains.
  3. (Clarification) What is the sensitivity of the results to ϵf\epsilon_f, the temperature mixture (βr,αr)(\beta_r, \alpha_r), and pool size KK? Only η\eta is swept.
  4. (Clarification) What is the total per-instance cost (pool construction + grouping + generation + RHS calibration) relative to G2MILP/ACM-MILP, and how does it scale to instances with tens of thousands of variables?
  5. (Clarification) In the variable-fixing task, is G2MILP's infeasible reduced problem a property of the method or of the fixing-threshold protocol (ρ=0.70\rho = 0.70)?
3 comments
AuthorsOpenReview ↗

Comments on Questions

Q1:Reproduction of MILP-FBGen

We acknowledge that our reproduction of MILP-FBGen did not attain the best performance, and we have added a supplementary clarification.

1.Since no official MILP-FBGen implementation was available, we used an independent implementation based on the published method and pseudocode. To check reproduction fidelity on the shared MIK dataset, at η=0.1\eta=0.1, our structural similarity is 0.991 versus 0.804 in the original paper, with 100% feasibility in both. Gurobi limitations prevent a direct solving-time comparison, but the reported MILP-FBGen-to-G2MILP solving-time ratio matches ours under matched conditions.

2.MILP-FBGen enforces only feasibility and objective boundedness: from a feasible solution, it reconstructs constraint coefficients via diffusion and generates RHS and objective coefficients from selected primal and dual solutions, guaranteeing at least one original primal solution remains feasible. It does not guarantee retention of other original solutions, the distance of new feasible solutions from source solutions, or preservation of near-optimal ordering. Since objective coefficients are also generated, even small changes can substantially alter the feasible set.


Q2:Performance on trivial "loose-RHS" baseline

We agree with the reviewer’s comment and have supplemented the experiments with a “loose-RHS” baseline.

1.Following the reviewer’s suggestion, we keep AA fixed and retain the 1−ϵf1-\epsilon_f source-solution quality, relaxing only bib_i. We conducted comparative experiments with four relaxation levels, with detailed results provided in https://anonymous.4open.science/r/files4rebuttal-5259/Sup._of_ablation_study.png, and their visualization is provided in https://anonymous.4open.science/r/files4rebuttal-5259/Metrics_on_Loose-RHS.pdf.

2.At small relaxation, metrics change little and some improve; at larger relaxation, all metrics except Supp. Jac. deteriorate substantially. Constraint relaxation only modestly changes the relative distance between new and old support solutions, so Supp. Jac. may even improve. Feasible-region changes also alter active constraints and the objective, while solution-distance metrics remain stable early but drift noticeably at large relaxation. We therefore attribute the gains to the combined effect of VAE reconstruction, response-profile alignment, and conformal calibration.


Q3:Sensitivity analysis of the relevant parameters

We agree with the reviewer and have performed a sensitivity analysis on the relevant parameters.

1.We first conducted sensitivity experiments on the three parameters ϵf\epsilon_f, βr\beta_r, and$\alpha_r$, with the results shown in https://anonymous.4open.science/r/files4rebuttal-5259/Sensitivity%20Analysis.png.

2.Results show sensitivity to ϵf\epsilon_f and the temperature mixture, with trade-offs across metrics. Increasing ϵf\epsilon_f from 0.05 to 0.20 generally improves Chamfer distance, support Jaccard, active-constraint Jaccard, and SWD, but raises objective shift (1.453 to 10.633); SymNN and MMD vary non-monotonically. Similarly, emphasizing higher-β\beta components tends to improve support- and active-constraint-related metrics, whereas more weight on β=0\beta=0 favors solution-position metrics. We therefore use β=(0,1,4)\beta=(0,1,4) with uniform mixture weights as a balanced configuration, not as the best setting for every metric.

3.We reported the results in https://anonymous.4open.science/r/files4rebuttal-5259/Sample_Size%26Methods.png. Sampling methods perform similarly for large source solution pools. For small pools, performance hinges on matching the dataset’s instance-solution distribution: matches improve metrics, mismatches cause severe degradation, and this effect weakens as the pool grows.


Q4:Computation Costs

In accordance with the reviewer’s comment, we have collected statistics on the solving cost.

The solving-cost statistics are reported in https://anonymous.4open.science/r/files4rebuttal-5259/Computation_Costs.pdf, which presents the total cost of each method across different datasets. When the problem scales to tens of thousands of variables, solution-pool construction is mainly determined by instance-solving difficulty. Convex instances allow fast construction even at high dimensionality, whereas generative-model training time grows with the number of variables as masked nodes increase.


Q5:Downstream task

We clarify that G2MILP’s infeasibility stems from a fixed threshold; the instances become feasible at threshold 0.9. To compare with predict-and-research (threshold ≤70), we excluded the ρ=0.9\rho=0.9 instances.


We thank the reviewer for the constructive comments. We hope our responses address your concerns, and we wish you all the best!


[1]Yahong Zhang,Chenchen Fan,Donghui Chen,et al. MILP-FBGen: LP/MILP instance generation with feasibility/boundedness[C]. ICML, 2024.

AuthorsOpenReview ↗

Comments on Weaknesses

W1:Partially circular evaluation protocol

We acknowledge that the current evaluation protocol indeed exhibits a degree of circularity. We agree with the reviewer’s comment and have added a comparison experiment with a simple "Loose-RHS" baseline to validate the effectiveness of the current method.

Following the reviewer’s suggestion, we fix the specified term and retain the 1−ϵf1-\epsilon_f source-solution quality, relaxing only the other term. Four relaxation levels were compared; see https://anonymous.4open.science/r/files4rebuttal-5259/Sup._of_ablation_study.png for details and https://anonymous.4open.science/r/files4rebuttal-5259/Metrics_on_Loose-RHS.pdf for visualization.

With mild relaxation, metrics change little and some improve; with stronger relaxation, all metrics except Supp. Jac. deteriorate substantially. Constraint relaxation only slightly changes the relative distance between original and updated support solutions, so Supp. Jac. can even improve. Feasible-region changes also affect active constraints and the objective, while solution-distance metrics remain stable early but drift noticeably under strong relaxation. We thus attribute the gains to the joint effect of VAE reconstruction, response-profile alignment, and conformal calibration.


W2:Limited evidence for the motivating use case.

We agree with the reviewer and have clarified the relevant scenarios accordingly.

1.On MIK, SLP-MILP achieves a structural similarity score of 0.887 at η=0.1\eta=0.1, compared with 0.991 for our MILP-FBGen reimplementation. This gap also appears at the other edit ratios reported in Table 1. Because training primarily targets the similarity of solution-landscape structures rather than the structure of the instances themselves, a drop in structural similarity is expected.

2.We acknowledge that the current validation of SLP-MILP is limited to standard datasets of moderate size. To assess the proposed method on larger-scale datasets, we selected datasets from MIPLIB for supplementary experiments. Since the instances uploaded to MIPLIB are all single-instance problems, to preserve a certain degree of solution diversity, we selected all instances in the “air” group under “Collection” as the initial training set.

3.Since no official MILP-FBGen implementation was available, we used an independent implementation based on the published method and pseudocode; on the shared MIK dataset at η=0.1\eta=0.1, our structural similarity is 0.991 versus 0.804 in the original paper, with 100% feasibility in both, and although Gurobi limitations prevent a direct solving-time comparison, the reported MILP-FBGen-to-G2MILP solving-time ratio is similar to ours under matched conditions.

4.The solving-cost statistics are reported in https://anonymous.4open.science/r/files4rebuttal-5259/Computation_Costs.pdf, which presents the total cost of each method across different datasets.

5.We acknowledge that the current paper does not evaluate its motivating scenario: scarce, large-scale, complex real-world industrial instances. We have therefore added a case study on the “air” group from MIPLIB, and the results are consistent with the second point under Weakness 2.

6.Following the reviewer’s suggestion, we provide a sensitivity analysis for ϵf\epsilon_f, the temperature mixture (βr,αr)(\beta_r,\alpha_r), and pool size KK. The results are consistent with the third response in “Comments on Questions.” On this basis, we further investigate the solution-space distribution of the problem itself, as shown in https://anonymous.4open.science/r/files4rebuttal-5259/Feature%20of%20datasets.pdf. Results show that when the sampling scheme is similar to the solution-space distribution, even a small number of samples can yield reasonably good results; if the distributions differ, reducing the number of samples may degrade the performance of the method. On the one hand, this indicates that the solution landscape is indeed closely related to the instance itself; on the other hand, it also suggests for future research that if the original solution-space distribution can be better approximated, the framework proposed in this paper can generate better problem instances.


We thank the reviewer for the comments. We hope our responses help address the weaknesses of the current experiments, and we wish you all the best!


[1] Ambros Gleixner,Gregor Hendel,Gerald Gamrath,et al. MIPLIB 2017: data-driven compilation of the 6th mixed-integer programming library[J]. Mathematical Programming Computation,2021,Vol.13(3): 1-48.

Thank you again for your time and helpful feedback. As the discussion deadline is approaching, we would appreciate it if you could review our responses and additional results when you have a chance. Please let us know if any concerns remain or if further clarification would be helpful.

Thank you for your consideration.

Reviewer_8T1SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SLP-MILP (Solution-Landscape Preserving Generation of Mixed-Integer Linear Programs), a framework for generating mixed-integer linear programs that preserve the solution landscape of the source instance beyond structural similarity and feasibility. It constructs source solution pools through solver-assisted neighborhood search, groups constraints using structural and solution-activity relationships and reconstructs selected groups with a conditional variational autoencoder. Generated right-hand sides of the constraints are determined through response-profile alignment followed by a conformal retention safeguard. Experiments compare against 3 generators on 4 MILP families using structural and solution-distribution metrics. A downstream independent-set experiment reports a 49.3% reduction in solving time with a 3.55% objective degradation

Critical strengths

The paper identifies a meaningful limitation of existing generation objectives. Structural similarity and the survival of a feasible witness do not ensure that generated instances to preserve useful solution patterns. Explicitly defining a conditioned solution distribution gives this distinction a concrete formulation and motivates a more informative evaluation of generated MILPs. The empirical results provide encouraging evidence across several problem families. SLP-MILP improves many solution-landscape metrics against the evaluated baselines. For example, it reduces CA Chamfer distance from the strongest baseline's 0.093 to 0.072 and increases IS support Jaccard from 0.632 to 0.872. These results suggest that conditioning generation on source solution behavior can preserve information missed by structural objectives alone

Critical weaknesses

The evaluation does not isolate the additional value of generation or establish its computational benefit. The downstream task already requires a source solution pool, however Table 3 does not compare against fixing variables directly from that pool. Therefore it is unclear how much benefit comes from generating and solving additional instances rather than from the source solutions already collected. The reported timing also does not clearly account for source-pool construction, generation, and solving generated instances. The 49.3% reduction consequently demonstrates faster solving of the reduced source problem, but does not establish a net computational saving. Moreover, the ablations omit response-profile alignment and the conformal correction, leaving the contribution of these central components uncertain.

Questions for authors
  1. In equation (63), how are the calibration solutions sampled relative to the weighted distribution ? It would be great to clarify the temperature parameters, estimation of the component normalizers, and why the implemented sampling satisfies the assumptions in Appendix B.3

  2. I have a concern on equation (59), is it intended as a marginal guarantee over calibration randomness, or as a retained-mass guarantee for each generated instance ? How is this distinction reflected in the reported experiments ?

  3. Which computational stages are included in Table 3 ? Are results available from the existing experiments for variable fixing directly from the source solution pool, and for removing response-profile alignment or conformal correction ?

3 comments
AuthorsOpenReview ↗

Comments on Questions

Q1:Clarification of Calibration-Solution Sampling, Parameter Estimation, and Sampling.

We agree with the reviewer and have clarified the sampling procedure, parameter estimation, and theoretical guarantee.

1.After completing solver-assisted search and diversity filtering, we freeze the retained solution pool and uniformly randomly partition it into disjoint training, calibration, and source test subsets of prespecified sizes. Each retained solution therefore has the same probability of assignment to the calibration subset, regardless of its weight in Equation (63). Those weights are used for empirical response-profile estimation in Appendix B.2. The conformal threshold in Appendix B.3 is computed from the unweighted calibration scores.

2.Equation (63) uses three inverse temperatures, (β1,β2,β3)=(0,2,8)(\beta_1,\beta_2,\beta_3)=(0,2,8), with equal mixture weights, α1=α2=α3=1/3\alpha_1=\alpha_2=\alpha_3=1/3. For each component, we compute its normalizing constant under the uniform empirical base measure on the frozen pool SPS_P as Z^P,r=∣SP∣−1∑x∈SPexp⁡[−βrf~P(x)]\widehat Z_{P,r}=|S_P|^{-1}\sum_{x\in S_P}\exp[-\beta_r\widetilde f_P(x)]. This is an exact finite-pool sum. We normalize each component separately before averaging the three component distributions. The β=0\beta=0 component is uniform over the pool, while the β=2\beta=2 and β=8\beta=8 components increasingly favor solutions with lower normalized objective values.

3.Conditional on the frozen pool, the realized training subset, and the generation randomness fixed before calibration, the score function is fixed. Under the uniform random partition, the calibration solutions in random permutation order and one solution selected uniformly from the source test subset jointly form an ordered sample without replacement from the remaining pool. Applying the same score function to all these solutions yields jointly exchangeable scores, which is the condition needed for the rank argument in Appendix B.3. We have revised Appendix B.3 to state the resulting 1−ϵf1-\epsilon_f marginal joint retention guarantee for a uniformly selected source test solution under this finite-pool randomization.


Q2:Calibration marginal guarantee

We accept the reviewer’s suggestion and have provided clarification on the calibration marginal guarantee.

1.Equation (59) gives a marginal joint retention guarantee. Under the finite-pool protocol, the probability is taken over the random allocation of calibration solutions and the uniform selection of one source test solution, conditional on the frozen source pool, the realized training subset, and the generation randomness fixed before calibration. The guarantee therefore applies to a candidate whose constraint directions and provisional right-hand sides are fixed before calibration. For a final instance obtained from a particular calibration split, the realized fraction of source test solutions satisfying all modified constraints can fall below 1−ϵf1-\epsilon_f. We have revised Appendix B.3 to make the conditioning and the sources of randomness explicit.

2.We illustrate this distinction using 800 random splits per setting. For ϵf\epsilon_f = (0.01,0.05,0.10,0.20), the mean retention rates are 99.13%, 96.70%, 91.50%, and 83.94%, respectively, all above their nominal targets. Nevertheless, 25%, 33%, 24%, and 17% of splits fall below those targets. Thus, the reported results distinguish average retention from single-instance variability: favorable marginal retention can coexist with below-target outcomes for individual calibrated instances, which Equation (59) does not rule out. etailed results are shown in https://anonymous.4open.science/r/files4rebuttal-5259/Retention_violin.png.


Q3:Downstream tasks

In response to the reviewer’s comment, we have clarified the computational stages of the downstream task and conducted supplementary experiments.

1.Table 3 involves four main computational stages: instance generation, problem solving, fixed-variable selection, and solver-based subproblem solving.

2.Detailed experimental results are provided in https://anonymous.4open.science/r/files4rebuttal-5259/Variable-fixing_task.png. For different fixed-variable counts, we compare SLP-MILP with two baselines: direct variable fixing from the source solution pool and prediction via predict-and-research. The comparative results demonstrate the superiority of the proposed method on this downstream task.

3.For the results obtained by removing response-profile alignment or conformal calibration, we performed consistent experiments and comparisons in the ablation study, which demonstrate the positive effect of these two modules on performance. Detailed results are provided in https://anonymous.4open.science/r/files4rebuttal-5259/Ablation_study.png.


We thank the reviewer for the constructive comments. We hope our responses address your concerns, and we wish you all the best!


AuthorsOpenReview ↗

Comments on Weaknesses

W1:Value of generation not isolated

We agree with the reviewer and have performed experiments on the generated instances.

We use the downstream predict-and-research task to verify the incremental benefit of the generated instances, with detailed results shown in the table below. The results clearly demonstrate the improvement in solving time achieved by adding generated instances to the base training set.

MethodTraining SetG2MILPACM-MILPMILP-FBGenSLP-MILP
Time(s)154.4124.8130.1117.7113.7

W2:Missing source-pool fixing baseline

We agree with the reviewer and compare the proposed method against direct variable fixing using the solution pool.

To verify SLP-MILP against direct sampling from the source solution pool for downstream variable fixing, for each fixed-variable count we sample an equal number of variables from the source pool and use predict-and-research to predict the same number. We then solve with COPT under the fixed-variable conditions and record the solving time and objective value. The results demonstrate the superiority of SLP-MILP on this downstream task. The specific results are provided in https://anonymous.4open.science/r/files4rebuttal-5259/Variable-fixing_task.png.


W3:Missing ablations for alignment and conformal calibration

We agree with the reviewer and have supplemented the relevant ablation studies.

We have added the ablation results for response-profile alignment and conformal calibration to the ablation study section; these results are reported in https://anonymous.4open.science/r/files4rebuttal-5259/Results_on_ablation_study.png. The experiments indicate that these two modules contribute substantially to performance.


We thank the reviewer for the comments. We hope our responses help address the weaknesses of the current experiments, and we wish you all the best!


Thank you again for your time and helpful feedback. As the discussion deadline is approaching, we would appreciate it if you could review our responses and additional results when you have a chance. Please let us know if any concerns remain or if further clarification would be helpful.

Thank you for your consideration.

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

This paper proposes SLP-MILP, a learning-based framework for generating Mixed-Integer Linear Programming (MILP) instances while preserving the solution-landscape characteristics of source instances. Unlike existing approaches that primarily focus on structural similarity, computational hardness, or feasibility, the proposed method explicitly incorporates source solution information into the generation process.

Experiments on four MILP families demonstrate improvements over several representative baselines on a range of solution-landscape similarity metrics. The authors also evaluate a downstream variable-fixing task, reporting improvements in solving efficiency while maintaining relatively high solution quality.

Critical strengths
  1. The paper identifies an important limitation of existing MILP instance generation methods: preserving feasibility or structural similarity does not necessarily preserve the distribution of meaningful feasible and near-optimal solutions.
  2. The proposed framework combines solution-aware constraint grouping, conditional generative reconstruction, and response-profile-based RHS alignment. These components are reasonably motivated and form a coherent generation pipeline.
Critical weaknesses
  1. The downstream evaluation is promising but limited. The variable-fixing experiment in Table 3 is conducted solely on the Independent Set problem. Furthermore, the lack of testing on other downstream tasks fails to demonstrate the broad applicability of the problems generated by SLP-MILP.
  2. The evaluation focuses exclusively on four problems, CA, IP, IS, and MIK, which feature relatively simple and general structures. The authors should also conduct testing on MILP datasets that contain non-standard structures, such as those found in the MIPLIB collection.
  3. SLP-MILP requires obtaining diverse source solutions before generating new problems, which incurs significant computational overhead when generating large-scale, highly difficult-to-solve MILP instances. Existing approaches for MILP generation do not require this step.
  4. The authors argue that the motivation for generating MILP instances is to address the data scarcity issue faced by learning-based MILP solving methods. However, the current frontier in MILP solving methods is represented not by learning-based approaches, but rather by LLM-based methods, such as LLM4Branch and LLM-LNS.
  5. Existing papers on MILP generation, such as G2MILP and ACM-MILP have focused on downstream tasks like hyperparameter tuning, initial basis prediction, and predict-and-search. Why did SLP-MILP choose a different task: variable fixing? In my view, the role of SLP-MILP within the variable fixing task is essentially that of a GNN designed to predict the solution to a problem, much like the predictive component in predict-and-search. It seems to me that the "SLP-MILP + variable fixing" framework should be compared directly against predict-and-search.
  6. The ablation study in Table 4 focuses on solution-pool perturbation strategies and solution-aware constraint grouping. However, several other important components are not independently evaluated. In particular, the paper does not provide ablations that remove solution-conditioned generative features, response-profile alignment, or conformal calibration. Consequently, it is difficult to determine whether the observed improvements primarily arise from the conditional generative model, the RHS alignment procedure, or the retention safeguard.
Questions for authors
  1. Since the conformal guarantee requires exchangeable calibration samples, how is this assumption justified under the sequential solver-assisted sampling and diversity-filtering procedure? Can the authors report empirical source-mass retention rates on independently sampled evaluation solutions?
  2. How sensitive is SLP-MILP to the number and diversity of source solutions? In particular, how does performance change when the source solution collection budget is substantially reduced?
  3. What are the computational costs of the complete pipeline, including source solution collection, constraint grouping, model training, and generation? How do these costs compare with those of the baselines?
  4. How much of the improvement comes specifically from response-profile alignment and conformal calibration? Would applying these two components to constraint directions generated by G2MILP or ACM-MILP produce comparable results?
4 comments
AuthorsOpenReview ↗

Comments on Questions

Q1:Sample exchangeability assumption and independently evaluated solution-quality retention rate

We thank the reviewer for pointing out that the exchangeability assumption should be stated in relation to the sampling protocol.

1.We clarify that source solutions are first sampled via LNS and solver-assisted search, and then filtered using a distance-based method to ensure diversity, forming a finite, diversity-oriented solution pool. The retained solutions are subsequently randomly assigned to the training, calibration, and evaluation subsets according to prespecified proportions. Therefore, our calibration results are interpreted conditionally on the resulting pool. We have also revised the scope of the conformal guarantee in Appendix B.3, it guarantees the exchangeability of calibration samples within the retained solution pool.

2.In response to the reviewer's comment, taking ηmod=0.1\eta_{\mathrm{mod}}=0.1, ϵf=0.1\epsilon_f=0.1 as an example, we report the empirical source-mass retention rates on independently sampled evaluation solutions.

Family|# instance|# solutions/instance|rate before conf.(%)|rate after conf.(%)|
MIK200509292(-)
CA200506490(↑\uparrow)
IP200508590(↑\uparrow)
IS200507190(↑\uparrow)

Q2:Sensitivity analysis of source solution quantity and diversity

We agree with the reviewer and have added experiments analyzing the sensitivity of SLP-MILP to the number and diversity of source solutions.

In response to the reviewer’s request, we supplemented the sensitivity analysis of SLP-MILP with respect to the number and diversity of source solutions. We varied the diversity of source solutions using Gaussian, uniform, and random sampling, respectively. For each setting, we reported results for sampling sizes from 24,000 to 600. Detailed results are shown in https://anonymous.4open.science/r/files4rebuttal-5259/Sample_Size%26Methods.png. The corresponding supplementary content in the paper is provided in Appendix C.8. Analysis of the table shows that the proposed SLP-MILP method achieves good performance under different source-solution quantities and diversity conditions. However, when the number of samples decreases substantially (24,000 →\rightarrow 600), solution quality declines, but the percentage decrease in the evaluation metrics is much smaller than that in the sample size.


Q3:Computational cost of the pipeline

We agree with the reviewer and have added the computational cost of the pipeline together with a comparison against the baselines.

We recorded the runtime of each module during the experiments, and the results are shown in https://anonymous.4open.science/r/files4rebuttal-5259/Computation_Costs.pdf. The corresponding supplementary content in the paper is provided in Appendix C.2. To ensure fairness in the experimental validation, we unified the training batch size, number of epochs, and other parameters across different methods, and removed the early-stopping strategy used in the SLP-MILP experiments. In particular, we set the number of samples for SLP to 24,000. In each case, training time accounts for the largest proportion. Solution-pool generation also takes a relatively long time, but this is acceptable.


Q4:Ablation on response-profile alignment and conformal calibration

We agree with the reviewer and have added ablation studies for response-profile alignment and conformal calibration.

  1. We supplemented the ablation results for response-profile alignment and conformal calibration in the ablation study section, as shown in the table below. |Method|Chamfer↓|Act. Jaccard↑|Obj. shift↓| |---|---:|---:|---:| |SLP-MILP|0.072| 0.858|8.437| |w/o response-profile alignment|0.364|0.408|10.74| |w/o conformal calibration|0.157|0.572|8.857|

The experiments demonstrate that these two modules yield substantial performance gains. The complete ablation results are shown in https://anonymous.4open.science/r/files4rebuttal-5259/Results_on_ablation_study.png, with the corresponding content presented in Section 5.2 of the paper.

2.To further verify the roles of the two components, we applied them to G2MILP and ACM-MILP, respectively. The experimental results show that both modules improve the performance of the methods. However, because the generative model training did not account for solution-related features of the instances, it is also evident that the improvements brought by these modules on metrics related to the solution-landscape distribution are limited. The complete experimental results are shown in https://anonymous.4open.science/r/files4rebuttal-5259/Sup._of_ablation_study.png, with the corresponding content provided in Appendix C.7 of the paper.


We thank the reviewer for the constructive comments. We hope our responses address your concerns, and we wish you all the best!


AuthorsOpenReview ↗

Comments on Weaknesses

W1:The downstream evaluation is promising but limited

We agree with the reviewer’s comment. To verify the effectiveness of SLP-MILP on different downstream tasks, we have also supplemented the variable-fixing downstream experiments on three other datasets. The results are shown in https://anonymous.4open.science/r/files4rebuttal-5259/supp_variable_fixing_task.png.


W2:The set of problem families evaluated is limited

In accordance with the reviewer’s comment, we selected datasets from MIPLIB for supplementary experiments.

1.Because the instances available in MIPLIB are all single-instance problems, we selected all instances in the “air” group under “Collection” as the initial training set to preserve a certain level of solution diversity. The dataset configuration is provided in https://anonymous.4open.science/r/files4rebuttal-5259/Inf._of_air.png.

2.Since MIPLIB contains limited instances, we sampled 200 solutions per instance into the source pool and evaluated SLP-MILP on difficult problems using the same pipeline and settings as in the main text. To exclude differences from reproduction accuracy, we compare only with G2MILP and ACM-MILP. Results are shown in https://anonymous.4open.science/r/files4rebuttal-5259/Metrics_of_air.png; the proposed method retains certain advantages on large-scale problems.


W3:Potentially High Cost of Source Solution Collection

We acknowledge that SLP-MILP incurs additional sampling costs; however, neural network training dominates the workflow time. Sampling cost depends mainly on problem nature and varies little with instance graph size, whereas training cost grows with the number of nodes. Based on the timing comparison in https://anonymous.4open.science/r/files4rebuttal-5259/Computation_Costs.pdf, we consider the resulting overhead acceptable.


W4:The motivation regarding learning-based solvers should be updated

We acknowledge that the motivation for learning-based solvers requires updating. In accordance with the reviewer’s suggestion, we have surveyed LLM-based methods and added a discussion of the motivation for LLM-based approaches to the Introduction and Related Work sections. The corresponding references have also been supplemented accordingly.


W5:Downstream task selection and baseline comparison are insufficient

In accordance with the reviewer’s comment, we have supplemented the experiments on downstream tasks.

1.We use the variable-fixing downstream task to test whether solution-landscape information preserved by generated instances provides reliable variable preferences for reducing the search space and accelerating optimization. We transfer high-confidence assignments inferred from high-quality generated-instance solutions to the corresponding source instances and compare this with direct fixing from source solutions, which better reveals whether the landscape information is beneficial.

2.The reviewer’s concern about the similarity between SLP-MILP and GNN-based methods for prediction problems is indeed worth considering. We have therefore added a direct comparison between “SLP-MILP + variable fixing” and predict-and-research, and further included the downstream predict-and-research task to further validate the performance of SLP-MILP. The specific results are provided in https://anonymous.4open.science/r/files4rebuttal-5259/Variable-fixing_task.png and https://anonymous.4open.science/r/files4rebuttal-5259/Isolated%20value.png, respectively.


W6:Insufficient Ablation of Key Components

We agree with the reviewer and have supplemented the ablation studies on key components.

We first added ablations for response-profile alignment and conformal calibration (results: https://anonymous.4open.science/r/files4rebuttal-5259/Ablation_study.png). Then we added a “Loose-RHS” experiment that relaxes only the constraints to validate the VAE-based conditional generative model (results: https://anonymous.4open.science/r/files4rebuttal-5259/Sup._of_ablation_study.png). These results confirm that each component positively affects performance.


We thank the reviewer for the comments. We hope our responses help address the weaknesses of the current experiments, and we wish you all the best!


[1]Huigen Ye, Hua Xu, An Yan, et al. Large Language Model-driven Large Neighborhood Search for Large-Scale MILP Problems[C]. ICML, 2025.

[2]Zhinan Hou, Xingchen Li, Yankai Zhang, et al. LLM4Branch: Large Language Model for Discovering Efficient Branching Policies of Integer Programs[C]. ICML, 2026.

[3]Qingyu Han, Linxin Yang, Qian Chen, et al. A GNN-Guided Predict-and-Search Framework for Mixed-Integer Linear Programming[C]. ICLR, 2023.

[4]Ambros Gleixner,Gregor Hendel,Gerald Gamrath,et al. MIPLIB 2017: data-driven compilation of the 6th mixed-integer programming library[J]. Mathematical Programming Computation,2021,Vol.13(3): 1-48.

Regarding the downstream tasks, the authors have only evaluated their method on additional datasets in their rebuttal, without providing comparisons against other baselines. More importantly, they have not conducted experiments on the same downstream tasks considered in previous MILP generation work, such as Initial Basis Prediction, Predict and Search, and Hyperparameter Tuning.

I also remain concerned about the performance of SLP-MILP on real-world instances, as its generalization capability may be fundamentally limited by the expressive power of the GNN architecture it employs.

Furthermore, I still have concerns regarding the novelty of this work. Specifically, the constraint grouping strategy and GNN decoder design adopted in SLP-MILP, including the use of masked selected groups, appear to be consistent with those in the existing ACM-MILP framework. The rebuttal has not sufficiently addressed these concerns or clarified the methodological novelty of the proposed approach.

Therefore, I have decided to maintain my original score.

AuthorsOpenReview ↗

Supplementary comments

We thank the reviewer for the comments on our paper and fully understand the concerns regarding the current downstream tasks and the novelty of our method. We provide further explanations from the following aspects.

A1: Downstream tasks.

1.We currently reported only the performance of our method versus each baseline on a single dataset for the variable-fixing task, as well as the performance of our method on multiple datasets. We now additionally report the complete aggregated results, as shown in https://anonymous.4open.science/r/files4rebuttal-5259/Sup._var-fix_full.png. Experimental results show that our method can fix more reliable variables under the same ρ\rho value and achieves better results in the variable-fixing task. To verify that the solving advantage of our method is not solely attributable to fixing more variables, we relaxed ρ\rho in the table so that each baseline fixes the same number of variables as SLP-MILP. The results demonstrate the significant advantage of our method.

2.The Predict-and-Search task has already been reported, here we provide supplementary results, as shown in https://anonymous.4open.science/r/files4rebuttal-5259/Pre._and_rese._full.png.

3.Following the reviewer’s suggestion, we added downstream tasks on initial basis prediction and COPT-based hyperparameter tuning. The results are shown in https://anonymous.4open.science/r/files4rebuttal-5259/Results_on_IBP.png and https://anonymous.4open.science/r/files4rebuttal-5259/Hyper%20tuning.png.


A2:Real-world instances.

Regarding experiments on real-world and large-scale problems, the current version has included experiments on the “air” collection from MIPLIB 2017, as shown in https://anonymous.4open.science/r/files4rebuttal-5259/Metrics_of_air.png. However, this dataset is only described as an airline scheduling problem and does not specify the provenance of individual instances. Considering the reviewer’s concern about real-world problems and the need for method comparison, we selected MIRP, the same real-world maritime freight dataset used in the downstream tasks of the FBGen-MILP paper, and conducted experiments on downstream tasks. The results are reported in https://anonymous.4open.science/r/files4rebuttal-5259/Real-world%20instance.png.


A3:Novelty of the paper.

1.The main novelty of this paper lies in the perspective of instance generation. Conventional reference metrics characterize instance structure, such as similarity, solving difficulty, and feasibility. Instead, we generate instances from the perspective of solution-landscape distribution, which describes instance solving behavior. While obtaining better solution-landscape distributions, our method also preserves favorable structural characteristics of instances.

2.Methodologically, tailored to the characteristics of solution-landscape distributions, we specifically designed the instance sampling and constraint grouping strategies. The advantages of our method are demonstrated in the ablation study, as shown in https://anonymous.4open.science/r/files4rebuttal-5259/Ablation_full.png. We are also aware of the methods used in other papers; however, for the problem studied in this paper, the similarity requirement of solution-landscape distributions must be ensured, making the VAE framework more suitable. We report and compare VAE and Duffision method's results on metrics and downstream tasks, as shown in https://anonymous.4open.science/r/files4rebuttal-5259/Metrics_on_VAE_and_Duff..png and https://anonymous.4open.science/r/files4rebuttal-5259/Downstream_on_VAE_and_Duff..png, respectively.


We thank the reviewer for the follow-up comments. We hope our responses help to address the reviewer’s concerns, and we sincerely appreciate the reviewer’s time and feedback.


[1] D. J. Papageorgiou, G. L. Nemhauser, J. Sokol, M.-S. Cheon, A. B. Keha. “MIRPLib – A library of maritime inventory routing problem instances: Survey, core model, and benchmark results.” European Journal of Operational Research, 235(2):350-366, 2014.

[2] Zhenan Fan,Xinglu Wang,Oleksandr Yakovenko,et al. Smart Initial Basis Selection for Linear Programs[C]//40th International Conference on Machine Learning, ICML 2023. 2023.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Our Responses

Summary of Our Responses

We thank the reviewers for their careful reviews and constructive suggestions. We have revised the manuscript accordingly, and the main clarifications and additional experiments are summarized below.

  1. Clarified the calibration protocol and theoretical guarantee: We explained the sampling, diversity filtering, partitioning, and parameter estimation procedures, revised Appendix B.3 to clarify the finite-pool marginal retention guarantee and its limitations. (Reviewer zQep: Q1; Reviewer 8T1S: Q1–Q2)

  2. Isolated the contributions of key components: We added ablations for response-profile alignment and conformal calibration, tested them with other generators, and compared against a Loose-RHS baseline to separate solution-landscape preservation from constraint relaxation alone. (Reviewer zQep: Q4, W6; Reviewer 8T1S: W3; Reviewer hohQ: Q2, W1)

  3. Strengthened downstream evaluation: We compared against direct source-pool variable fixing and a prediction-based baseline, extended variable-fixing experiments to other datasets, and evaluated the incremental value of generated instances for downstream learning. (Reviewer zQep: W1, W5; Reviewer 8T1S: Q3, W1–W2)

  4. Expanded evaluation to larger-scale instances: We added a case study using the MIPLIB “air” collection to assess performance beyond the original benchmark families. We also clarified the scope of the evidence for large-scale applications. (Reviewer zQep: W2; Reviewer hohQ: W2)

  5. Examined sensitivity and sampling requirements: We evaluated calibration levels, temperature mixtures, source-pool sizes, and sampling schemes. Results reveal trade-offs across metrics and show that performance depends on both source-pool size and its alignment with the solution-space distribution. (Reviewer zQep: Q2; Reviewer hohQ: Q3, W2)

  6. Reported computational costs: We provided a breakdown of solution-pool construction, model training, and generation costs, together with baseline comparisons under matched training settings. (Reviewer zQep: Q3, W3; Reviewer hohQ: Q4)

  7. **Clarified baseline reproduction and updated the motivation:**We documented our independent reproduction of MILP-FBGen and its comparison limitations, distinguished feasibility from solution-landscape preservation, and updated the discussion of recent LLM-based MILP methods. (Reviewer zQep: W4; Reviewer hohQ: Q1, W2)


The additional results and illustrations are summarized below.

ItemPurposeLink
Calibration retentionDistinguish marginal retention from variability across individual calibration splitsResults
Component ablationsIsolate the effects of response-profile alignment and conformal calibrationResults
Supplementary ablationsEvaluate component transfer and the Loose-RHS baselineResults
Loose-RHS visualizationExamine how constraint relaxation affects landscape-preservation metricsIllustration
Variable-fixing comparisonsCompare with direct source-pool fixing and a prediction-based baselineResults
Additional downstream datasetsEvaluate variable fixing on three additional datasetsResults
Value of generated instancesAssess the incremental benefit of generation for downstream learningResults
MIPLIB case studyEvaluate performance on the larger-scale “air” collectionResults
Hyperparameter sensitivityExamine calibration levels and temperature mixturesResults
Source-pool sensitivityExamine source-solution quantity and sampling schemesResults
Solution-space distributionsExplain the relationship between sampling schemes and source-pool requirementsIllustration
Computational costsReport pipeline runtime and comparisons with baseline methodsResults

We have incorporated the clarifications and additional experiments into the revised manuscript. We hope these responses address the reviewers’ concerns. We welcome any further suggestions and sincerely appreciate the reviewers’ time and thoughtful feedback.

Dear Reviewers,

The authors have submitted their rebuttal. Please review it and share your comments. If anything is unclear or requires clarification, please let the authors know as soon as possible. The author–reviewer discussion period ends on 18 November.

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.