acceptodds
Under review as a conference paper at ICLR 2027

Unify ML4MILP: Disentangle, Recombine, and Reevaluate to Benchmark Neural Primal Heuristics with Solver-backed Search

13 views

Abstract

Machine learning for (mixed) integer linear programming (ML4MILP) is rapidly expanding from solver-internal assistance toward end-to-end primal solution generation. Yet progress remains difficult to compare and interpret: neural prediction, solver-backed completion, and incumbent refinement are often entangled within monolithic pipelines, while datasets, training styles, solver backends, and evaluation protocols vary substantially across studies. We introduce ML4MILP-Bench, a modular benchmark and systematic analysis framework that decomposes learning-based MILP primal heuristics into three reusable stages. A *Learner* maps a standardized bipartite graph to variable-wise soft scores; an *Enforcer* converts these predictions into a first complete solution; and a *Refiner* further improves a feasible incumbent under fixed budgets. Within the unified taxonomy and re-implementation, we streamline, recombine, and systematically probe 8 learners, 6 enforcers, and 3 refiners via re-modularized learning components and re-encapsulated solver backends. Across experiments on diverse settings, we separately assess raw neural predictions and their interactions with solver-backed components to discern where performance gains originate and which design principles remain effective under controlled ablation. We further envision ML4MILP-Bench as a living benchmark to maintain up-to-date reproductions and empirically grounded insights.

Primary area:
datasets and benchmarks
Keywords:
mixed integer linear programming, neural combinatorial optimization, benchmark

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_m7obOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is motivated by the fact that progress in solving MILP is hard to judge when a) a combination of different methods is used (i.e., for a combination of Method A + search trick + Gurobi solver, it is difficult to predict whether the gain came from the neural network method, search trick, or solver), b) different data is used (i.e., different instance sizes, generators, and benchmark reference solutions differ or are not released), c) metrics are inconsistent (i.e., primal gap often calculated differently or infeasible outputs are handled differently), and d) comparison is based on old solvers. To benchmark the MILP-solving pipeline, the authors split each learning-based pipeline into three swappable stages: learner, enforcer, and refiner. They re-implemented everything with a shared GNN backbone, the same data (set cover, independent set, combinatorial auction, facility location), and current solvers (Gurobi 13, SCIP 10, COPT 8). They then tested many Learner × Enforcer × Refiner combinations.

Critical strengths

The major positive points about the paper are a) it addresses the real problem of fair comparison, b) uses up-to-date solvers, c) systematic experimentation, and d) reports even the negative results.

Critical weaknesses

The paper suffers from several limitations: a) the problems studied are easy (up to 100 instances), b) very few mixed-integer LP instances studied, c) forcing everything onto one backbone and one hyperparameter setup may not compare the results fairly for methods reported often with different optimized parameters, d) claims are not backed statistically as results are not generated with multiple seeds, e) checkpoints trained on set cover were applied directly to WA, BIP and MIPLIB which may not generalize well and f) several interesting extensions such as RL methods like RL-SPH, and solver-internal learning (branching, cuts) are left out.

Questions for authors

The detailed comments are given below:

  1. The main tests use 100 synthetic instances per task, and Gurobi solves several of them in seconds. On such easy instances, learned guidance has little room to help. The benchmark proposed needs to be demonstrated on medium and hard problems to prove its usefulness. Further, the conclusion that the solvers make the difference added by learners obscure may not hold on hard problems.
  2. Very few mixed-integer problems were covered, despite the paper’s title. FMIP’s continuous-variable component was disabled, and only pure-binary MIPLIB instances were used.
  3. Forcing everything onto one backbone and one hyperparameter setup may not compare the reported methods fairly. DDIM’s guidance was disabled, Apollo was simplified (no UEBO training), and Local Branching does only one solve. ConPaS and CoCo-MILP got near-zero validity or absurd gaps on several tasks. The differences in performance reported in the paper may be caused by implementation or tuning issues rather than by the actual limitations of the methods being compared, and the paper doesn’t check against the original papers’ reported numbers.
  4. The main tables in the paper report no confidence intervals. Results should be generated with multiple random seeds for statistical validity.
  5. Checkpoints trained on set cover were applied directly to WA, BIP, and MIPLIB. That’s a very weak test of transfer.
  6. Learned LNS, RL methods like RL-SPH, and solver-internal learning (branching, cuts) are left out. The paper justifies this, but it limits the generalizability of the conclusions.
8 comments
AuthorsOpenReview ↗

Author Response to Reviewer m7ob (page 1)

Dear Reviewer m7ob,

Thank you for the time and effort reviewing our paper. We appreciate your recognition of the importance of fair comparison, systematic experimentation, and transparent reporting of negative results. Below, we respond point by point to the weaknesses (W) and questions (Q) you have kindly raised, grouping closely related points for clarity.


W1/Q1: Instance difficulty and the validity of conclusions on harder problems.

A: Thank you for raising this important concern. We agree that instance difficulty matters when interpreting the benefits of learned guidance. We would appreciate the opportunity to clarify the benchmark settings and the scope of our conclusions.

1. The number of test instances is distinct from instance size and difficulty.

“100 instances” denotes the test-set size for each problem family, over which performance is aggregated. It does not indicate the number of variables or constraints within an instance (an analogy is the distinction between the number of test images and their resolution). Appendix F, Table 9 reports the actual dimensions:

DatasetTest instancesAvg. variablesAvg. constraints
SC1002,0003,000.0
MIS1001,5005,941.9
CA1001,500578.8
CFL1005,0505,151.0
WA10061,00064,304.6
BIP1001,083195.0
Selected MIPLIB3960,88841,439.3

The selected MIPLIB instances span 214–364,417 variables and 64–188,879 constraints. Moreover, our 1500-variable MIS instances match the graph size designated “Hard” in Gasse et al. [1], although difficulty also depends on instance structure, solver configuration, and computational budget.

2. The benchmark combines established problem families and evaluation conventions.

Gasse et al. [1] study SC/MIS/CA/CFL using 20 instances per difficulty level and five seeds. Guided Diffusion [2] evaluates 100 instances per family on SC/MIS/CA/CFL. Predict-and-Search [3] uses 100 test instances each for MIS/CA/BIP/WA, and DiffILO [4] uses 100 each for SC/MIS/CA. Our selection combines these established families to support comparable evaluation across methods developed under different protocols.

Regarding the Gurobi performance, our results show that the suite includes settings where optimization remains nontrivial under the evaluated budgets: Gurobi retains a mean BKS-relative gap of 0.14% on SC after 1000 seconds, and 26.023% and 6.881% on BIP after 100 and 1000 seconds, respectively.

3. Fast solver performance does not eliminate the challenge of reliable neural prediction.

We acknowledge that Gurobi handles some instances quickly. Our benchmark measures the feasibility and quality of neural predictions before solver intervention. With 1,000 samples, DDIM and FMIP find feasible solutions on only 36% and 50% of CFL instances, respectively, and several learners find none (Table 2). This leaves substantial room to improve neural primal heuristics on established benchmarks.

4. Our goal is to identify effective learned guidance and its interaction with search.

Our analyses explicitly examine both better raw predictions and better solver guidance. For example, Sec. 3.1 identifies the feasibility-quality strengths of generative learners and constraint-aware unsupervised learning, Sec. 3.2 examines why prediction calibration and Learner-Enforcer compatibility matter, and Sec. 3.3 identifies Refiner-dependent learner rankings. On BIP, the learned pipelines reduce primal integral by approximately up to 59% relative to Gurobi, demonstrating substantial benefits within our own evaluation.

Solver-backed search can narrow differences in final solution quality across learners for certain tasks, backends, and budgets. Meaningful differences in anytime performance can persist. Unify ML4TSP [5] reports a related finding. Pre-MCTS solution quality has limited predictive value for post-MCTS quality, and learned heatmaps improve search. These interactions motivate modular evaluation to identify effective learning–search combinations.

Remark. We value the contributions of existing learning-based methods and see considerable potential to improve learned guidance. Our framework supports reproduction and fair, transparent evaluation so researchers can identify each method’s strengths and build on complementary advances.

[1] Exact Combinatorial Optimization with Graph Convolutional Neural Networks. NeurIPS, 2019.

[2] Effective Generation of Feasible Solutions for Integer Programming via Guided Diffusion. KDD, 2024.

[3] A GNN-Guided Predict-and-Search Framework for Mixed-Integer Linear Programming. ICLR, 2023.

[4] Differentiable Integer Linear Programming. ICLR, 2025.

[5] Unify ML4TSP: Drawing Methodological Principles for TSP and Beyond from Streamlined Design Space of Learning and Search. ICLR, 2025.

AuthorsOpenReview ↗

Author Response to Reviewer m7ob (page 2)

W2&5 / Q2&5: Mixed-integer coverage, FMIP's continuous-variable component, and SC-only zero-shot transfer versus in-domain adaptation.

A: Thank you for raising these concerns about mixed-integer coverage and in-domain adaptation. First and foremost, we have clarified in the new abstract that ML4MILP-Bench is a modular benchmark and analysis framework for neural primal heuristics, with a primary empirical focus on binary ILPs and support for mixed-integer settings, consistent with our statement in Sec. 2.1 in the initial submission.

In the past week, following your suggestion, we have conducted in-distribution training on WA and BIP, both of which contain integer and continuous variables. In particular, we implemented FMIP's joint continuous-integer flow in the benchmark and enabled its continuous-variable component in both experiments. The new FMIP models therefore learn integer and continuous variables jointly. For Base-SL, PS-SL, ConPaS, and CoCo-MILP, we train on integer-variable targets and let PS/Gurobi assign the continuous variables, following the original prediction setups. We trained each learner for 100 epochs per target task. Note that DiffILO’s loss penalizes expected constraint violations under a distribution over binary assignments, and guided DDIM likewise supports learning only for binary ILPs. So they are not trained on MILP data, but we retain the evaluation for both by transferring their SC-trained checkpoints directly to WA and BIP. The SC-transfer rows serve as a zero-shot baseline alongside the new target-trained models. All learner rows use PS with a 1000s Gurobi budget.

WA Results

LearnerSettingGap@10s(%)Gap@100s(%)Gap@1000s(%)Obj.@1000s
GurobiSolver only13.8850.0400.013707.910
Base-SLSC transfer23.9380.1010.036708.070
Base-SLi.i.d.1.6560.0960.037708.080
PS-SLSC transfer23.8470.1020.036708.070
PS-SLi.i.d.1.7970.0850.029707.663
ConPaSSC transfer19.7592.9342.871728.080
ConPaSi.i.d.1.3140.0910.033708.050
CoCoSC transfer25.8040.1020.036708.070
CoCoi.i.d.1.2680.1000.039708.090
FMIPSC transfer24.9240.0990.037708.080
FMIPi.i.d.1.2150.1200.033708.050
OPTFMSC transfer22.9780.1760.126708.700
DiffILOSC transfer24.8210.0990.036708.070
DDIMSC transfer22.9590.1040.036708.070

BIP Results

LearnerSettingGap@10s(%)Gap@100s(%)Gap@1000s(%)Obj.@1000s
GurobiSolver only61.85626.0236.88113.294
Base-SLSC transfer28.7108.653-0.73012.532
Base-SLi.i.d.29.3827.074-1.70912.366
PS-SLSC transfer27.4758.708-0.76712.530
PS-SLi.i.d.33.1669.8421.15412.704
ConPaSSC transfer34.88913.1540.30612.579
ConPaSi.i.d.33.6848.907-0.73912.528
CoCoSC transfer27.2378.624-0.75112.534
CoCoi.i.d.37.47614.7851.89812.642
FMIPSC transfer31.0539.490-0.64512.547
FMIPi.i.d.31.0669.908-0.72912.488
OPTFMSC transfer27.4028.677-0.76312.530
DiffILOSC transfer27.3838.704-0.71012.538
DDIMSC transfer29.8687.425-1.82912.427

Analysis. The clearest short-budget gains from in-distribution training occur on WA. At 100 s, four of the five learners have lower reported gaps, including ConPaS, whose gap falls from 2.934% to 0.091%. On BIP, three of the five matched learners improve at 1,000 s, and all eight SC-transfer pipelines reach lower final gaps than standalone Gurobi. For the jointly trained FMIP model on BIP, the gap falls from 31.066% at 10 s to −0.729% at 1,000 s within a single PS solve on the same 100 instances. Target-domain training can improve the guidance provided to PS, and the gains within the solve show how substantially final solution quality depends on solver-backed completion and search.

Remark. Considering current results, we note that raw learner results show considerable room to improve feasibility and objective quality even on our standard binary ILPs. As such, we believe the binary ILP study, the zero-shot transfer evaluations, and the new in-distribution WA/BIP results together properly support the revised scope now. Also, the framework provides common interfaces for MILP training and evaluation, including joint continuous–integer learners, and for incorporating additional learner and solver components. The scope clarification and the new WA/BIP experiments are also highlighted in red in our revised manuscript (Abstract, p. 1, lines 19–20 and Sec. 3.2, p. 9, lines 457–468 and Table 6). We sincerely hope these additions and discussions address your concerns about mixed-integer coverage and in-domain adaptation.

AuthorsOpenReview ↗

Author Response to Reviewer m7ob (page 3)

W3/Q3: Reproduction fidelity, simplified components, and anomalous baseline results.

A: Thank you for raising this concern. We appreciate the opportunity to clarify our standardization choices and the interpretation of the stage-wise results.

1. Backbone choice and preservation of method-specific components.

  • Variable–constraint bipartite graphs are widely used in learning-based MILP optimization. Most methods considered here (e.g. PS [1], ConPaS [2], DiffILO [3], CoCo-MILP [4], FMIP [5]) build on closely related GNN backbones, with method-specific extensions. Much of their innovation concerns training paradigms, learning objectives, and interactions with solver-backed search. This architectural commonality motivates a shared foundation for examining these differences while preserving each method’s specialized mechanisms. Related benchmarks, like RL4CO [6], ML4TSPBench [7], and ML4CO-Bench-101 [8], likewise standardize models and/or data to facilitate comparable evaluation.

  • As detailed in Appendix G, our backbone follows DiffILO, retaining alternating bipartite message passing, graph-wise normalization, and multi-layer readout. We also preserve CoCo-MILP’s competitive layer and contrastive learning, OPTFM’s separate pretraining and downstream prediction pipeline [9], and the newly added optional joint continuous–discrete modeling for FMIP. Appendix H.2 documents method-specific learning rates, objectives, etc., drawing on the original settings. Enforcer defaults similarly follow FMIP’s published configurations for PS, ND, Apollo, and PMVB.

  • For Apollo [10], we adopt the simplified downstream implementation used in FMIP. Our empirical focus places learning at the Learner stage: improving the feasibility and quality of initial variable predictions and examining their value under subsequent solver intervention. Accordingly, we evaluate representative post-processing procedures without separately comparing their additional learning components. This keeps the investigation focused and proportionate to a conference paper. Our guiding principle is to combine a shared architectural foundation with specialized training mechanisms and established method-specific parameter choices.

2. Interpretation of the seemingly anomalous baseline results.

  • ConPaS and CoCo-MILP were developed as prediction–search pipelines in which neural predictions interact closely with a powerful solver. Table 2 evaluates raw learner outputs before solver intervention, while the subsequent enforcer and refiner experiments assess their contribution to feasibility recovery and solution improvement. These stages together characterize each method’s capabilities.

  • For example, with Gurobi-based LB refinement on MIS (Table 4), ConPaS and CoCo-MILP achieve PIs of 0.89 and 0.83, respectively, outperforming DDIM [11] (1.19) and FMIP (1.40) under the same refinement configuration. With the PS enforcer on SC (Table 3), CoCo-MILP achieves a PI of 853.56, compared with 1035.82 for DDIM with the same enforcer and 973.68 for standalone Gurobi. These results demonstrate the value of their predictions when paired with compatible search procedures.

  • This distinction motivates our benchmark. Methods sharing a variable-logit interface can eventually reach similarly small gaps after substantial solver processing, while differing considerably in raw prediction quality, enforcer compatibility, and anytime performance. Our stage-wise evaluation makes these distinct contributions visible.

In general, we value all the contributions of the included methods, and we sincerely hope this explains how our evaluation supports fair comparison while recognizing their different strengths and intended roles.

[1] A GNN-Guided Predict-and-Search Framework for Mixed-Integer Linear Programming. ICLR, 2023.

[2] Huang et al. Contrastive Predict-and-Search for Mixed Integer Linear Programs. ICML, 2024.

[3] Differentiable Integer Linear Programming. ICLR, 2025.

[4] CoCo-MILP: Inter-Variable Contrastive and Intra-Constraint Competitive MILP Solution Prediction. AAAI, 2026.

[5] FMIP: Joint Continuous-Integer Flow for Mixed-Integer Linear Programming. ICLR, 2026.

[6] Berto et al. RL4CO: An Extensive Reinforcement Learning for Combinatorial Optimization Benchmark. KDD, 2025.

[7] Unify ML4TSP: Drawing Methodological Principles for TSP and Beyond from Streamlined Design Space of Learning and Search. ICLR, 2025.

[8] ML4CO-Bench-101: Benchmark Machine Learning for Classic Combinatorial Problems on Graphs. NeurIPS, Datasets and Benchmarks Track, 2025.

[9] OPTFM: A Scalable Multi-View Graph Transformer for Hierarchical Pre-Training in Combinatorial Optimization. NeurIPS, 2025.

[10] Apollo-MILP: An Alternating Prediction-Correction Neural Solving Framework for Mixed-Integer Linear Programming. ICLR, 2025.

[11] Effective Generation of Feasible Solutions for Integer Programming via Guided Diffusion. KDD, 2024.

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

This paper introduces ML4MILP-Bench, a unified benchmark that decomposes neural MILP solving into Learner, Enforcer, and Refiner stages. It systematically evaluates different component combinations and provides insights into neural guidance and solver-backed optimization.

Critical strengths
  1. The modular framework provides a clear and unified taxonomy for neural MILP methods.
  2. Extensive experiments cover diverse learners, enforcers, and refiners.
  3. The proposed FQS metric jointly considers feasibility and solution quality.
  4. The use of modern solver backends improves benchmarking relevance and reproducibility.
Critical weaknesses
  1. Evaluation on real-world mixed-integer problems remains limited, particularly for instances containing continuous variables.
  2. Single-seed testing raises concerns about the statistical reliability of the conclusions.
  3. The benchmark excludes several important method families, such as reinforcement learning and learned LNS.
  4. The weighting of feasibility and solution quality in FQS lacks sufficient justification.
Questions for authors
  1. Can the framework be evaluated on more real-world mixed-integer benchmarks?
  2. Can the authors provide multi-seed results with statistical significance analysis?
  3. How easily can the framework incorporate RL-based and learned LNS methods?
  4. How sensitive are method rankings to the formulation of FQS?
8 comments
AuthorsOpenReview ↗

Author Response to Reviewer ptCZ (page 1)

Dear Reviewer ptCZ,

Thank you for your careful review and constructive suggestions. We appreciate your positive assessment of the modular framework, the breadth of the component-wise experiments, and the emphasis on evaluating feasibility and solution quality with modern solver backends. Below, we respond point by point to the paired weaknesses (W) and questions (Q) you have raised.


W1/Q1: Supplementary Mixed-integer Results.

A: Thank you for asking us to clarify the empirical scope and strengthen the mixed-integer evidence.

  • First, we have revised the abstract and introduction to state explicitly and accurately that ML4MILP-Bench is a modular benchmark and systematic analysis framework for neural primal heuristics, with a primary empirical focus on binary ILPs and supported extensions to mixed-integer settings, consistent with Sec. 2.1.

  • Next, we add in-distribution training experiments on WA and BIP in the main text. We trained Base-SL, PS-SL, ConPaS, and CoCo-MILP on the integer variables of each target task and implemented FMIP’s joint continuous–discrete architecture for training on both variable types. This follows a common design in PS, ConPaS, and CoCo-MILP, where the learner predicts the binary variables and the solver assigns the continuous variables. Each learner was trained for 100 epochs per task. Note DiffILO’s and Guided DDIM likewise formulates their learning problems for binary ILPs only, so they are exclude from training, while we evaluate both learners through zero-shot SC transfer, applying their SC-trained checkpoints directly to WA and BIP and using PS/Gurobi to complete the solutions. All learner rows use PS with a 1000s Gurobi budget.

WA

LearnerSettingGap@10s(%)Gap@100s(%)Gap@1000s(%)Obj.@1000s
GurobiSolver only13.8850.0400.013707.910
Base-SLSC transfer23.9380.1010.036708.070
Base-SLi.i.d.1.6560.0960.037708.080
PS-SLSC transfer23.8470.1020.036708.070
PS-SLi.i.d.1.7970.0850.029707.663
ConPaSSC transfer19.7592.9342.871728.080
ConPaSi.i.d.1.3140.0910.033708.050
CoCoSC transfer25.8040.1020.036708.070
CoCoi.i.d.1.2680.1000.039708.090
FMIPSC transfer24.9240.0990.037708.080
FMIPi.i.d.1.2150.1200.033708.050
OPTFMSC transfer22.9780.1760.126708.700
DiffILOSC transfer24.8210.0990.036708.070
DDIMSC transfer22.9590.1040.036708.070

BIP

LearnerSettingGap@10s(%)Gap@100s(%)Gap@1000s(%)Obj.@1000s
GurobiSolver only61.85626.0236.88113.294
Base-SLSC transfer28.7108.653-0.73012.532
Base-SLi.i.d.29.3827.074-1.70912.366
PS-SLSC transfer27.4758.708-0.76712.530
PS-SLi.i.d.33.1669.8421.15412.704
ConPaSSC transfer34.88913.1540.30612.579
ConPaSi.i.d.33.6848.907-0.73912.528
CoCoSC transfer27.2378.624-0.75112.534
CoCoi.i.d.37.47614.7851.89812.642
FMIPSC transfer31.0539.490-0.64512.547
FMIPi.i.d.31.0669.908-0.72912.488
OPTFMSC transfer27.4028.677-0.76312.530
DiffILOSC transfer27.3838.704-0.71012.538
DDIMSC transfer29.8687.425-1.82912.427

Analysis. The clearest short-budget gains from in-distribution training occur on WA. At 100 s, four of the five learners have lower reported gaps, including ConPaS, whose gap decreases from 2.934% to 0.091%. On BIP, the gains vary across learners, with three of the five matched learners improving at 1000s. Zero-shot transfer also provides effective guidance under PS. All eight SC-transfer pipelines reach lower final BIP gaps than standalone Gurobi. The improvement within a single solve illustrates the contribution of search. On the same 100 BIP instances, the i.i.d. FMIP pipeline reduces its gap from 31.066% at 10s to −0.729% at 1000s. Thus, we agree that target-domain training can improve the guidance provided to PS, while final solution quality still depends substantially on solver-backed completion and search.

Remark. Provided the current results, it can be observed that even on the standard binary ILPs in our study, the raw learner results show considerable room to improve feasibility and objective quality. In this sense, we believe the binary ILP study, zero-shot transfer results, and new in-distribution WA/BIP experiments together support the revised claims. We have also clarified the binary-ILP scope and added the in-distribution WA/BIP results in our revised manuscript, with both revisions highlighted in red (Abstract, p. 1, lines 19–20 and Sec. 3.2, p. 9, lines 457–468 and Table 6). Beyond that, the framework provides common interfaces for MILP training and evaluation and for incorporating additional learners and solver components. We sincerely hope the clarified scope and additional experiments address your concern.

AuthorsOpenReview ↗

Author Response to Reviewer ptCZ (page 2)

W2/Q2: Multi-seed reliability and statistical significance.

A: Thank you for raising this concern. To our knowledge, neural combinatorial optimization (NCO) does not yet have a universally adopted protocol for statistical significance testing across solver evaluations of this kind. We agree that multi-seed results provide a direct descriptive view of solver stability under stochastic inference. We therefore reran all Table 2 evaluation settings with seeds 0, 1, and 2, covering 8 learners, 4 tasks, both Rounding (R) and Sampling (S), and the same 100 test instances (192 runs in total).

The tables below report the three-seed mean ± standard deviation for Obj, Gap, IFR, FQS, and Time. Obj is averaged over all test instances, Gap is averaged over feasible outputs only. IFR and FQS are percentages, and Time is seconds per instance. "--" indicates that Gap is not applicable or unavailable because there is no feasible output.

SC

LearnerEnforcerObjGap(%)IFR(%)FQS(%)Time(s)
Base-SLR84.66±0.008.83±0.003.00±0.002.77±0.000.420±0.134
PS-SLR78.80±0.00—0.00±0.000.00±0.000.343±0.004
ConPaSR17.08±0.00—0.00±0.000.00±0.000.339±0.002
CoCoR341.94±0.00296.49±0.0099.00±0.0041.64±0.000.495±0.009
OPTFMR38.80±0.00—0.00±0.000.00±0.000.248±0.022
DiffILOR123.70±0.0046.16±0.0067.00±0.0048.57±0.000.343±0.009
DDIMR97.37±0.4313.21±0.5877.00±6.2468.49±5.379.543±1.025
FMIPR91.38±0.216.14±0.2580.33±5.5175.81±5.089.531±1.084
Base-SLS90.53±0.148.09±0.3512.67±1.1511.75±1.111.015±0.047
PS-SLS114.34±0.6430.94±0.2249.67±4.0439.12±3.230.981±0.032
ConPaSS38791.01±9.2645261.39±15.82100.00±0.0014.05±0.000.991±0.021
CoCoS278.87±0.29223.42±0.37100.00±0.0046.00±0.021.114±0.012
OPTFMS770.99±34.09734.71±45.265.33±2.081.70±0.650.872±0.043
DiffILOS121.41±0.0840.55±0.09100.00±0.0074.61±0.040.973±0.043
DDIMS97.47±0.3512.86±0.36100.00±0.0089.21±0.259.869±1.565
FMIPS91.29±0.195.57±0.2799.67±0.5894.54±0.3210.465±0.500

MIS

LearnerEnforcerObjGap(%)IFR(%)FQS(%)Time(s)
Base-SLR680.71±0.00—0.00±0.000.00±0.000.089±0.011
PS-SLR678.13±0.00—0.00±0.000.00±0.000.087±0.011
ConPaSR238.80±0.0065.64±0.0095.00±0.0063.14±0.000.088±0.012
CoCoR771.22±0.00—0.00±0.000.00±0.000.112±0.017
OPTFMR659.56±0.00—0.00±0.000.00±0.000.168±0.019
DiffILOR636.51±0.006.96±0.00100.00±0.0093.69±0.000.104±0.022
DDIMR681.83±0.080.34±0.0198.67±1.5398.33±1.532.943±0.651
FMIPR682.19±0.070.29±0.0198.67±1.1598.38±1.162.941±0.752
Base-SLS672.51±0.38—0.00±0.000.00±0.000.205±0.017
PS-SLS661.80±0.43—0.00±0.000.00±0.000.217±0.018
ConPaSS245.59±0.2564.10±0.04100.00±0.0066.88±0.010.213±0.016
CoCoS742.03±0.85—0.00±0.000.00±0.000.241±0.019
OPTFMS651.36±1.98—0.00±0.000.00±0.000.303±0.003
DiffILOS638.01±0.026.74±0.00100.00±0.0093.87±0.000.229±0.019
DDIMS681.83±0.100.34±0.01100.00±0.0099.66±0.013.018±0.748
FMIPS682.25±0.110.28±0.02100.00±0.0099.72±0.023.065±0.760

CA

LearnerEnforcerObjGap(%)IFR(%)FQS(%)Time(s)
Base-SLR19083.95±0.0013.24±0.001.00±0.000.89±0.000.060±0.016
PS-SLR16094.01±0.0028.31±0.002.00±0.001.60±0.000.061±0.016
ConPaSR0.00±0.00100.00±0.00100.00±0.0059.06±0.000.059±0.015
CoCoR61903.77±0.00—0.00±0.000.00±0.000.084±0.022
OPTFMR2720.65±0.0090.82±0.0012.00±0.007.29±0.000.124±0.021
DiffILOR15323.81±0.0031.26±0.0098.00±0.0077.04±0.000.059±0.015
DDIMR21619.96±35.803.00±0.1290.00±4.0087.41±3.792.853±0.723
FMIPR21685.21±22.892.68±0.0988.67±0.5886.38±0.602.906±0.751
Base-SLS18295.19±31.4619.15±0.5419.00±0.0016.17±0.060.101±0.022
PS-SLS16170.49±107.42—0.00±0.000.00±0.000.101±0.019
ConPaSS1650.00±7.5692.59±0.03100.00±0.0060.41±0.010.099±0.022
CoCoS55881.98±55.53—0.00±0.000.00±0.000.124±0.027
OPTFMS16184.39±267.06—0.00±0.000.00±0.000.160±0.025
DiffILOS15659.62±1.6829.68±0.01100.00±0.0079.37±0.000.099±0.020
DDIMS21642.00±16.512.83±0.0799.67±0.5896.96±0.532.909±0.738
FMIPS21684.82±17.382.64±0.0699.33±1.1596.81±1.072.928±0.727
AuthorsOpenReview ↗

Author Response to Reviewer ptCZ (page 3)

CFL

LearnerEnforcerObjGap(%)IFR(%)FQS(%)Time(s)
Base-SLR11151.41±0.00—0.00±0.000.00±0.000.085±0.020
PS-SLR10457.62±0.00—0.00±0.000.00±0.000.113±0.022
ConPaSR0.00±0.00—0.00±0.000.00±0.000.101±0.007
CoCoR53533.21±0.00—0.00±0.000.00±0.000.124±0.000
OPTFMR3514.39±0.00—0.00±0.000.00±0.000.179±0.006
DiffILOR18210.24±0.0062.25±0.0085.00±0.0057.28±0.000.094±0.001
DDIMR12691.58±41.0516.83±0.6424.33±1.5321.05±1.233.280±0.255
FMIPR12300.91±18.0913.51±1.2124.33±1.5321.60±1.473.026±0.770
Base-SLS11499.71±9.03—0.00±0.000.00±0.000.276±0.013
PS-SLS12412.53±23.27—0.00±0.000.00±0.000.280±0.026
ConPaSS0.00±0.00—0.00±0.000.00±0.000.276±0.016
CoCoS54243.98±26.41—0.00±0.000.00±0.000.304±0.011
OPTFMS20194.55±165.70—0.00±0.000.00±0.000.340±0.006
DiffILOS18088.75±18.8159.55±0.0693.00±0.0063.39±0.020.271±0.004
DDIMS12710.73±26.7116.43±0.5834.00±4.3629.52±3.873.520±0.134
FMIPS12382.18±24.7011.82±0.4554.67±4.1649.18±3.913.811±0.253

We additionally calculate approximate, pointwise 95% Student’s tt confidence intervals for the mean performance across evaluation seeds:

CI_{95 \\%} = \bar{x}\pm t_{0.975,2}\frac{s}{\sqrt{3}}= \bar{x}\pm 2.484s,

where ss denotes the sample standard deviation across the three seeds. Here, the sample size is three seed-level observations. These intervals quantify uncertainty in mean performance over evaluation randomness, conditional on the checkpoint and test set.

The following table illustrates these intervals for the sampling Gap/Obj./FQS of DDIM and FMIP on all four tasks, together with DiffILO on CFL, where it achieves the highest mean FQS. FQS and interval endpoints are expressed in percentage points.

TaskLearnerObj:mean ± SDObj:95% CIGap (%):mean ± SDGap:95% CIFQS (%):mean ± SDFQS:95% CI
SCDDIM97.47 ± 0.35[96.60, 98.34]12.86 ± 0.36[11.97, 13.75]89.21 ± 0.25[88.59, 89.83]
SCFMIP91.29 ± 0.19[90.82, 91.76]5.57 ± 0.27[4.90, 6.24]94.54 ± 0.32[93.75, 95.33]
MISDDIM681.83 ± 0.10[681.58, 682.08]0.34 ± 0.01[0.32, 0.36]99.66 ± 0.01[99.64, 99.68]
MISFMIP682.25 ± 0.11[681.98, 682.52]0.28 ± 0.02[0.23, 0.33]99.72 ± 0.02[99.67, 99.77]
CADDIM21642.00 ± 16.51[21600.99, 21683.01]2.83 ± 0.07[2.66, 3.00]96.96 ± 0.53[95.64, 98.28]
CAFMIP21684.82 ± 17.38[21641.65, 21727.99]2.64 ± 0.06[2.49, 2.79]96.81 ± 1.07[94.15, 99.47]
CFLDDIM12710.73 ± 26.71[12644.38, 12777.08]16.43 ± 0.58[14.99, 17.87]29.52 ± 3.87[19.91, 39.13]
CFLFMIP12382.18 ± 24.70[12320.82, 12443.54]11.82 ± 0.45[10.70, 12.94]49.18 ± 3.91[39.47, 58.89]
CFLDiffILO18088.75 ± 18.81[18042.02, 18135.48]59.55 ± 0.06[59.40, 59.70]63.39 ± 0.02[63.34, 63.44]

Analysis. Across the 64 learner–task–decoder FQS entries, 52 have a standard deviation of at most 1.00 percentage point. The task-level FQS picture is preserved in the three-seed aggregates: DDIM and FMIP form the leading group on SC and MIS, and DiffILO has the highest FQS on CFL, followed by FMIP and DDIM. We interpret these as consistent and broad performance tiers. These results support the conclusion that, with the trained checkpoints held fixed, most solver outputs are stable across evaluation seeds and the broad task-dependent performance ranking is retained. We have also updated these results in our revised manuscript (pages 7--8, Sec. 3.1, lines 378--386, Fig. 3, and Appendix I.4, table 19, revisions are highlighted in red).

Reviewer_g1RqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ML4MILP-Bench, a modular benchmark and toolkit for learning-based MILP solving. It decomposes existing pipelines into three stages—Learner, Enforcer, and Refiner—and explicitly separates exact solver backends from method definitions. The paper standardizes the variable–constraint bipartite representation, supervision assets, solver interfaces, evaluation metrics, and provides code-level examples and an anonymous repository. Empirically, it retrains and recombines multiple learners, enforcers, and refiners on several binary ILP benchmarks, with additional generalization tests on mixed-integer and MIPLIB instances.

Critical strengths
  1. Useful decomposition and engineering effort. The Learner–Enforcer–Refiner abstraction is sensible and practically useful. The paper makes a substantial engineering contribution by standardizing data I/O, graph representation, solver wrappers, training/evaluation interfaces, and modular recombination.

  2. Explicit separation of solver backends. Treating Gurobi/SCIP/COPT as configurable backends rather than part of method names is a meaningful step toward controlled comparison. This helps clarify ambiguous labels such as “PS+Gurobi” and enables backend-controlled ablations.

  3. Evaluation perspective is valuable. The emphasis on raw neural prediction quality before solver intervention, together with IFR/SFR/IViol and FQS, highlights an important blind spot in existing evaluations: low average constraint violation does not imply instance-wise feasibility.

  4. Stage-wise empirical observations are informative. The paper reports several practically relevant findings, e.g., enforcer effectiveness is task- and learner-dependent, solver backends can change rankings, and refiners may quickly wash out learner-conditioned differences. These observations are useful even if they are not fully causal.

Critical weaknesses
  1. Novelty is moderate. The main contribution is primarily infrastructural and organizational rather than methodological. The Learner–Enforcer–Refiner decomposition is reasonable, but conceptually not deeply novel. The paper’s value depends heavily on the quality, completeness, and reproducibility of the released toolkit.

  2. Some background claims are overly strong. The limitations attributed to prior work—especially “ambiguous taxonomy,” “fragmented implementations,” and “outdated solver backends”—are directionally valid but sometimes stated too absolutely. Using older solver versions is often a natural consequence of publication time or licensing constraints, not necessarily an evaluation flaw. These should be framed as comparability barriers rather than deficiencies of individual papers.

  3. Fairness of unified retraining is not fully established. Sharing a unified GNN backbone and fixed training budgets may systematically disadvantage methods whose original strengths come from method-specific architectures or training objectives. The paper should clarify tuning budgets and provide at least a capacity or hyperparameter sensitivity analysis. Otherwise, the comparisons are better described as “performance under the ML4MILP-Bench backbone” rather than a definitive ranking of original methods.

  4. Metrics still contain selection bias. Averaging objective/gap only over feasible solutions can favor methods that solve only easy instances. IFR alone does not fully correct this. The paper should report feasible coverage, penalized all-instance gap, and per-instance win/loss statistics. FQS also needs clearer conventions for infeasible cases and sensitivity analysis.

  5. Scope is narrower than the title suggests. The main experiments focus on binary ILP benchmarks, while mixed-integer and MIPLIB results are relatively limited and mostly based on direct transfer without fine-tuning. The paper should either narrow its claims to binary ILP primal heuristics or strengthen the mixed-integer evidence.

Questions for authors
  1. How were hyperparameters tuned across learners, enforcers, and refiners? Were tuning budgets comparable? Can the authors provide sensitivity analysis for the shared backbone capacity and key enforcer hyperparameters?

  2. Which results correspond to full reproductions, and which are bench-native adaptations? Could the authors add an explicit reproduction-status column to the main comparison tables?

  3. Can the authors report both feasible-only gap and all-instance penalized gap, together with feasible coverage and paired significance tests for the main comparisons?

  4. Since the main experiments are binary ILP, will the authors narrow the title/claims, or add a mixed-integer main experiment with fine-tuning or multi-task training?

8 comments
AuthorsOpenReview ↗

Author Response to Reviewer g1Rq (page 1)

Dear Reviewer g1Rq,

Thank you for the time and care you have devoted to reviewing our paper and for your constructive feedback. We deeply appreciate your initial positive assessment of our work. Your comments help us identify where the scope, experimental comparisons, and reproducibility of the benchmark need clearer explanation and stronger support.

Below, we respond point by point to the weaknesses you raised (W) and your questions (Q), grouping closely related points for clarity.


W1: Moderate novelty and the reproducibility of the toolkit.

A: Thank you for your comment. We broadly agree that our primary contribution is infrastructural and empirical, consistent with the purpose of a paper in this category.

  • From the outset, we intentionally followed an established research direction exemplified by RL4CO [1], Unify ML4TSP [2], PredictiveCO [3], ML4CO-Bench-101 [4], and the MIS benchmarking study [5], etc. These works demonstrate the scientific value of modular implementations, controlled comparisons, and systematic examination of algorithmic components. Developing a meaningful taxonomy, implementing interoperable components, and designing experiments that disentangle their contributions are themselves technically nontrivial. We pursue this direction for ILP primal heuristics, where feasibility enforcement and solver-backed search make component attribution particularly important.

  • We also agree that the toolkit must substantiate this contribution. So, our initial submission already included an anonymous repository for a quick inspection of the implementation and reproduction scripts, complemented by detailed configurations in Appendix H and usage examples in Appendix J. The full codebase and documentation will be made publicly available after the review process.

  • Beyond infrastructure, our experiments provide scientific insights into learning–search interactions. A few illustrative examples: (i) generative learners often favor solution quality, while explicitly penalizing constraint violations, as in DiffILO, promotes feasibility, (ii) contrastive or ranking-based supervision does not necessarily yield scores suited to probability-dependent enforcers, and (iii) strong raw predictions do not consistently translate into greater solver acceleration, thus motivating jointly considering learning objectives, prediction interfaces, and downstream search. These design insights can inform future end-to-end learning approaches to constrained optimization, which our modular framework facilitates developing, evaluating, and iteratively improving.

[1] Berto et al. RL4CO: An Extensive Reinforcement Learning for Combinatorial Optimization Benchmark. KDD, 2025.

[2] Li et al. Unify ML4TSP: Drawing Methodological Principles for TSP and Beyond from Streamlined Design Space of Learning and Search. ICLR, 2025.

[3] Geng et al. Benchmarking PtO and PnO Methods in the Predictive Combinatorial Optimization Regime. NeurIPS, 2024.

[4] Ma et al. ML4CO-Bench-101: Benchmark Machine Learning for Classic Combinatorial Problems on Graphs. NeurIPS, 2025.

[5] Böther et al. What's Wrong with Deep Learning in Tree Search for Combinatorial Optimization. ICLR, 2022.


W2: Positioning of prior work and solver-version comparability.

A: Thank you for this constructive suggestion. We agree that some wording is overly strong. As you correctly pointed out, our intention was to summarize barriers to cross-study comparability instead of criticizing shortcomings of individual papers.

  • Specifically, regarding solver versions, our concern mainly includes releases that were already several years old when the respective papers appeared. For example, Table 6 documents that BTBS-LNS (2025) used Gurobi 9.5.0 and SCIP 7.0.3, both released in 2021. That said, they do not diminish these works' technical contributions, which we fully acknowledge. Our goal is to complement them with controlled evaluations using updated backends, clarifying how learning-based gains depend on the solver setting.

  • Following your advice, we have revised the relevant wording to avoid overstatement: “ambiguous taxonomy” to “inconsistent terminology across studies,” “fragmented implementations” to “heterogeneous implementations and data interfaces,” and “outdated solver backends” to “solver-version differences and their implications for comparability.” We have also qualified the statement about older backends overstating gains, clarifying that improvements demonstrated with one solver version warrant reevaluation before being generalized to others. These revisions are highlighted in red in our updated manuscript.

AuthorsOpenReview ↗

Author Response to Reviewer g1Rq (page 2)

W3/Q1/Q2: Fairness of unified retraining, parameter sensitivity, and reproduction status.

A: Thank you for raising these important questions. We agree that the backbone choice, parameter-selection protocol, and reproduction status should be explicit. Our aim is to enable controlled comparisons of learning and search components, with conclusions interpreted within the evaluated configurations.

  • Backbone choice and method-specific components. Our choice builds on substantial architectural commonality in prior work. For example, PS [2] and ConPaS [3] use variable–constraint bipartite graphs and closely related GNN predictors. ConPaS explicitly adopts the GCN architecture of PS while introducing contrastive training. This motivates using a common architectural foundation to examine differences in supervision, training objectives, and subsequent search.

    As detailed in Appendix G, our shared backbone follows DiffILO [1], retaining alternating bipartite message passing, graph-wise normalization, and multi-layer readout. We standardize compatible components while retaining method-specific mechanisms, including CoCo-MILP’s competitive layer and contrastive learning [4], OPTFM’s separate pretraining and downstream prediction pipeline [7], optional joint continuous–discrete modeling for FMIP [5], and CISP pretraining for guided diffusion [6]. In Appendix H.2, we have reported method-specific learning rates, training caps, and objectives, basically following their respective original settings.

  • Enforcer and refiner configurations. For overlapping tasks, our enforcer defaults follow the published configurations in FMIP [5, Table 6], which evaluates the same four downstream procedures: PS, ND, PMVB, and Apollo. Within each task, these configurations are held fixed across learners. Refiner parameters are likewise fixed across learners and documented in Appendix H.4, and each refiner receives the same cached incumbent for a given learner–instance pair.

  • Additional enforcer sensitivity analysis. Following your suggestion and considering the tight timeline of rebuttal, we evaluated PS-SL, OPTFM, DiffILO, and FMIP on SC and MIS, using 100 instances per task and a 100s Gurobi budget per instance. Each learner–enforcer combination uses the same five configurations: the original default and four controlled perturbations. Here, pp denotes PMVB’s success-probability parameter. Parameters not being varied remain at their defaults.

    EnforcerDefault: SC / MISAdditional configurations
    PS: (k0,k1,Δ)(k_0,k_1,\Delta)(0.30,0.04,0.20)(0.30,0.04,0.20) / (0.30,0.20,0.30)(0.30,0.20,0.30)Scale Δ\Delta by 0.50.5 or 1.51.5; separately scale (k0,k1)(k_0,k_1) jointly by 0.50.5 or 1.51.5
    ND: (K,α)(K,\alpha)(50,0.10)(50,0.10) / same(50,0.05)(50,0.05), (50,0.20)(50,0.20), (25,0.10)(25,0.10), (100,0.10)(100,0.10)
    PMVB: (p,τ)(p,\tau)(0.70,0.90)(0.70,0.90) / same(0.70,0.80)(0.70,0.80), (0.70,0.95)(0.70,0.95), (0.60,0.90)(0.60,0.90), (0.90,0.90)(0.90,0.90)
    Apollo: (k0,k1,Δ,K)(k_0,k_1,\Delta,K)(0.30,0.04,0.20,2)(0.30,0.04,0.20,2) / (0.30,0.20,0.30,2)(0.30,0.20,0.30,2)Scale Δ\Delta by 0.50.5 or 1.51.5; separately set K=1K=1 or 44

    The following table reports penalized Gap@100 (%)↓ as default [minimum–maximum] across the five configurations. Brackets indicate sensitivity ranges, not confidence intervals.

    EnforcerLearnerSCMIS
    NDPS-SL1.50 [1.50–90.25]0.17 [0.03–0.60]
    NDOPTFM6.50 [6.50–98.40]15.06 [1.09–99.08]
    NDDiffILO1.71 [1.71–61.21]1.08 [0.42–2.33]
    NDFMIP0.89 [0.89–88.80]0.08 [0.01–0.17]
    PSPS-SL0.23 [0.16–0.29]0.00 [0.00–0.00]
    PSOPTFM0.24 [0.18–0.24]0.00 [0.00–0.00]
    PSDiffILO0.24 [0.18–0.29]0.00 [0.00–0.06]
    PSFMIP0.28 [0.21–0.28]0.00 [0.00–0.00]
    PMVBPS-SL0.22 [0.22–0.40]0.02 [0.01–0.03]
    PMVBOPTFM0.86 [0.70–0.86]0.00 [0.00–0.00]
    PMVBDiffILO20.60 [14.01–21.87]4.42 [3.76–4.63]
    PMVBFMIP2.11 [1.36–2.70]0.07 [0.04–0.09]
    ApolloPS-SL0.86 [0.86–1.87]0.00 [0.00–0.00]
    ApolloOPTFM0.91 [0.91–2.20]0.00 [0.00–0.00]
    ApolloDiffILO0.89 [0.89–2.09]0.00 [0.00–0.03]
    ApolloFMIP0.77 [0.77–2.04]0.00 [0.00–0.00]
AuthorsOpenReview ↗

Author Response to Reviewer g1Rq (page 3)

Analysis:

  • These results provide several useful checks on the original configurations. PS maintains small final gaps across the tested settings, PMVB preserves the final-gap ordering of all four learners on both tasks, and the ND and Apollo defaults attain the lowest final gaps among the tested configurations for every learner on SC.

  • We also note that sensitivity depends on the enforcer and task. ND is substantially more sensitive, and reducing its fixing ratio improves results on MIS. Moreover, similar final gaps can conceal differences in anytime performance: for example, FMIP–PS on MIS retains a reported final gap of 0.00%, while PI@100 ranges from 1.34 to 9.36. These observations support our emphasis on learner–enforcer interactions and anytime evaluation, while identifying settings that benefit from task-specific adjustment.

Remarks: Intended scope and continued development.

  • We also respectfully clarify that our central contribution is a systematic, modular framework, together with an initial investigation of useful design principles through broadly applicable, literature-informed configurations and representative component recombinations. We believe this combination provides a substantive contribution within the scope of a conference benchmark paper.
  • Although we have deliberately controlled the manuscript’s length, exhaustively expanding grid searches across architectures, training regimes, and search parameters would substantially lengthen an already extensive paper and shift its emphasis toward a comprehensive configuration manual. That said, we fully appreciate the need for targeted sensitivity analyses to substantiate the reported conclusions, which is why we have added the experiments above.
  • Our broader vision is to make such investigations straightforward to conduct within the framework, so we warmly welcome researchers to extend the evaluations around their particular methods, applications, and computational budgets, and intend to support this through continued maintenance and documentation.

We hope this clarifies our intended scope and helps establish a shared understanding of how the benchmark can support rigorous comparisons and cumulative progress. We would also appreciate your consideration of our careful attention to the reproducibility of prior work, as well as the additional empirical efforts we have undertaken, particularly at this stage. We have also accordingly incorporated these results in the updated manuscript (Page 9, Sec. 3.2, lines 448--463 and Appendix I.6, tables 24--25, revisions are highlighted in red).

[1] Geng et al. Differentiable Integer Linear Programming. ICLR, 2025.

[2] Han et al. A GNN-Guided Predict-and-Search Framework for Mixed-Integer Linear Programming. ICLR, 2023.

[3] Huang et al. Contrastive Predict-and-Search for Mixed Integer Linear Programs. ICML, 2024.

[4] Pu et al. CoCo-MILP: Inter-Variable Contrastive and Intra-Constraint Competitive MILP Solution Prediction. AAAI, 2026.

[5] Li et al. FMIP: Joint Continuous-Integer Flow for Mixed-Integer Linear Programming. ICLR, 2026.

[6] Zeng et al. Effective Generation of Feasible Solutions for Integer Programming via Guided Diffusion. KDD, 2024.

[7] Yuan et al. OPTFM: A Scalable Multi-View Graph Transformer for Hierarchical Pre-Training in Combinatorial Optimization. NeurIPS, 2025.


W4/Q3: Selection bias, all-instance evaluation, and FQS conventions.

A: Thank you for these helpful suggestions. We agree that feasible-only gaps should be interpreted together with coverage and all-instance metrics. Following your advice, we first extended the evaluation of all 64 configurations in Table 2, using the original checkpoints and decoding settings (0.5-threshold rounding or 1000-candidate sampling). We independently checked all 6400 returned solutions at a feasibility tolerance of 10−510^{-5}.

For a penalty cap CC, each feasible output contributes its signed gap capped above at CC, and each infeasible output contributes CC. PG100 is the primary result, while PG1000 shows sensitivity to a larger cap. We also report per-instance wins/ties/losses (W/T/L) against Base-SL with the same enforcer: feasibility takes priority, then we compare uncapped gaps when both outputs are feasible. This keeps high-gap feasible outputs distinct from infeasible ones when PG100 gives both a value of 100%.

All-instance Penalized Gap Comparison

All gap, PG, and FQS values are percentages. Each metric cell shows value (rank), ranked among the eight learners within the same task and enforcer: lower gap/PG and higher FQS are better. Ties share a competition rank (e.g., 1, 2, 2, 4). A method with no feasible output has no feasible-only gap or rank, and — in W/T/L marks the Base-SL comparison baseline. R is rounding and S is 1000-candidate sampling. CoCo denotes CoCo-MILP, and DDIM denotes Guided DDIM.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to AC: Rebuttal Summary and Reviewer Feedback

Dear Area Chair,

We sincerely thank you and the reviewers for your time and care. As the author-reviewer discussion enters its final 24 hours, we would like to summarize the feedback received and the additional evidence and revisions provided during rebuttal.

In the past fortnight, we have presented point-by-point responses to every concern and question raised, conducted supplementary experiments, clarified the empirical scope and implementation choices, and incorporated the new evidence into our revised manuscript. The anonymous repository also supports inspection of our reproducible implementation.

To facilitate your final recommendation, we summarize below the strengths recognized by the reviewers and main efforts we devoted during rebuttal.


I. Strengths Recognized and Current Status

  • Reviewer g1Rq (initial rating: 3, stating "no substantive concerns about the paper's soundness") describes the Learner-Enforcer-Refiner decomposition as "sensible and practically useful" and recognizes a "substantial engineering contribution". The reviewer also states that "Evaluation perspective is valuable" and "Stage-wise empirical observations are informative".
  • Reviewer ptCZ (initial rating: 3, maintained after discussion) notes that the framework provides "a clear and unified taxonomy for neural MILP methods", that "Extensive experiments cover diverse learners, enforcers, and refiners", and that modern solver backends improve "benchmarking relevance and reproducibility". Following our rebuttal, the reviewer expressed satisfaction with our responses and reaffirmed the recommendation for acceptance.
  • Reviewer m7ob (initial rating: 2, increased to 3 after discussion) recognizes that our paper "addresses the real problem of fair comparison" and "reports even the negative results", highlighting the paper's "systematic experimentation". Following our rebuttal, the reviewer expressed satisfaction with our revisions and raised the score to 3.

We deeply appreciate the consensus reached toward acceptance in these assessments.


II. Main Concerns and Our Responses

The following summary consolidates the shared concerns and our updates.

Main concernClarifications and/or additional evidenceCorresponding reviewer responsesLocation in the revised manuscript
1. In-domain adaptation for MIPAction: Clarified the primary binary-ILP scope and added WA/BIP in-distribution training for five learners, including FMIP’s joint continuous-integer architecture. Finding: Compared against SC transfer, with gains depending on the task and learner.g1Rq: W5/Q4, ptCZ: W1/Q1, m7ob: W2&5/Q2&5Abstract, Sec. 3.2 (Table 6), Appendix I.1 (Tables 17–18).
2. Multi-seed tests & confidence intervalsAction: Added configuration-seed evaluations covering all 64 raw-learner settings. Reported means, standard deviations, and representative approximate 95% confidence intervals. Finding: Broad performance patterns are retained.ptCZ: W2/Q2, m7ob: W4/Q4Sec. 3.1 (Fig. 3), Appendix I.4 (Table 20).
3. Metric choice & FQS sensitivityAction: Added all-instance penalized gaps and per-instance comparisons, and tested α=0.5/1/2 for FQS sensitivity. Finding: Leading learners remain unchanged across weights in all eight settings. FQS agrees with each penalized-gap variant in seven of eight settings.g1Rq: W4/Q3, ptCZ: W4/Q4Sec. 3.1 (Table 3), Appendices I.2 and I.4 (Tables 19–20).
4. Reproducibility & parameter fairnessAction: Explained the shared architectural foundation, retained method-specific mechanisms, and documented adaptations. Added sensitivity experiments with 4 learners, 4 enforcers, 2 tasks, and 5 configurations, identifying both stable settings and task-dependent sensitivity.g1Rq: W3/Q1/Q2, m7ob: W3/Q3Appendices G--H, Sec. 3.2 (Table 5), Appendix I.6 (Tables 25–26).
5. Broader method coverageAction: Added imitation-learning and RL-based LNS implementations through a shared interface, explicitly identifying their status as available for further training and evaluation. Explained RL-SPH availability and the scope for solver-internal learning.ptCZ: W3/Q3, m7ob: W6/Q6Appendix A.2, anonymous repository.

We respectfully hope that the positive ratings shared by the reviewers, with our additional experiments, detailed clarifications, and manuscript revisions, could support a favorable recommendation of our work. We genuinely appreciate your time and consideration.

Sincerely,
Authors of Submission 3182

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.