acceptodds
Under review as a conference paper at ICLR 2027

SCoR: Unsupervised State-Conditioned Refinement for Integer Linear Programming

Abstract

Integer linear programming (ILP) is a general framework for combinatorial optimization. Unsupervised approaches learn to solve ILPs without solver-generated supervision, but their updates often rely on aggregate signals that obscure the prediction-dependent states of individual constraints and their effects across variables. We propose State-Conditioned Refinement (SCoR), which incorporates localized constraint feedback into unsupervised ILP optimization. Given a relaxed prediction, SCoR encodes normalized constraint residuals and propagates them to incident variables through sparse attention over the bipartite incidence graph. This feedback and the current optimization state parameterize a bounded positive diagonal preconditioner that adapts the base update to heterogeneous local conditions. We characterize the refinement's boundedness and descent compatibility and show that, under a matched update budget, variable-wise preconditioning admits a broader allocation set than uniform scaling and can yield a strictly larger local-model decrease. SCoR preserves the underlying unsupervised objective and can be integrated with different optimization schemes. Experiments on standard benchmarks and real-world instances demonstrate improvements across representative unsupervised ILP optimizers.

Primary area:
optimization
Keywords:
Integer linear programming, Neural Combinatorial Optimization, Graph Neural Network, Unsupervised Learning, State-Conditioned Refinement

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zWV4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes State-Conditioned Refinement (SCoR) for unsupervised integer linear programming (ILP). SCoR incorporates localized constraint feedback into the refinement process by propagating constraint residual information to incident variables and adaptively scaling variable-wise updates. Experiments on standard binary ILP benchmarks demonstrate improvements over existing unsupervised solvers.

Critical strengths
  1. The paper is well written and easy to follow. The experimental results are promising.

  2. SCoR serves as an additional refinement module that can be integrated into different unsupervised ILP optimization frameworks.

Critical weaknesses
  1. The proposed method appears to be primarily based on residual-aware attention over a bipartite graph and learned diagonal preconditioning. The methodological novelty beyond combining these existing techniques is not sufficiently clear.

  2. The theoretical analysis is not closely connected to the actual model design. In particular, Proposition 4.2 characterizes the potential advantage of variable-wise update allocation, but the proposed model does not explicitly optimize for such an allocation. Therefore, the theoretical results provide limited justification for the effectiveness of the learned refinement mechanism.

Questions for authors
  1. The computational overhead of SCoR is not sufficiently discussed. Could the authors provide more details on the training and inference times, particularly in comparison with the base methods?

  2. Could the authors compare SCoR against simpler preconditioning strategies, such as gradient-based diagonal scaling or residual-based scaling, to better demonstrate the necessity of the proposed architecture?

2 comments

We thank the reviewer for the feedback on the design, theory, and computational cost. We added simpler scaling baselines with comparable parameter counts for the learned alternatives and measured the additional training and prediction time.

Response to Weakness 1 and Question 2. Methodological contribution and simpler alternatives

SCoR learns how to allocate a base update across variables using the current states of their incident constraints. Figure 1 illustrates why this matters on CA. Aggregate violation decreases by 3.4%, yet 19.7% of the evaluated constraint residuals worsen. The learned scaling head uses local constraint feedback to respond to these different constraint states when deciding how much of the base update each variable receives. This relationship is learned jointly with the predictor through the existing unsupervised objective in both DiffILO and PUMA.

We compared SCoR-DiffILO with fixed gradient-magnitude scaling, a learned gradient-logit controller, and mean residual aggregation. The gradient-logit controller uses the current logit, marginal probability, and normalized gradient. Mean aggregation retains the residual-conditioned values and activity gate and gives incident constraints equal weights. We widened the first hidden layer in both learned alternatives to bring their parameter counts within 0.53% of the full refinement module and 0.06% of the full model (Appendix C.9, Table 14).

MethodSC Obj ↓CA Obj ↑
DiffILO124.6597281.68
Gradient-magnitude scaling124.6097236.99
Gradient-logit controller, parameter matched124.7196509.82
Mean residual aggregation, parameter matched124.6296427.74
SCoR-DiffILO124.5697316.01

SCoR achieves the best mean objective on SC and CA under the shared search protocol. On SC, fixed gradient scaling and mean aggregation also improve upon DiffILO. On CA, the full model improves on DiffILO while each simpler alternative yields a lower objective. The mean-aggregation comparison supports learning the relative weights of incident constraints with comparable model capacity.

Response to Weakness 2. Connection between Proposition 4.2 and the learned mechanism

Proposition 4.2 asks what can be gained by redistributing a fixed amount of update magnitude. It shows that moving part of the update from a weaker-gradient coordinate to a stronger-gradient coordinate with remaining capacity can yield a strictly larger decrease in the local linear model. This provides a rationale for variable-wise scaling. The performance of a particular learned controller is an empirical question.

In SCoR, the scales are learned through the refined base objective in Eq. (17). Holding the initial prediction and base direction in Eq. (11) fixed gives

=d_i^{(0)}\frac{\partial\mathcal L_{\mathrm{base}}(z^{\mathrm{ref}};\mathcal G,\xi_{\mathrm{ref}})}{\partial z_i^{\mathrm{ref}}}.$$ The scale of each variable receives a training signal through its effect on the refined prediction. The controller determines the effective budget $B=\sum_i\eta_i|d_i^{(0)}|$ as part of the update. The discussion after Proposition 4.2 and in Appendix B.1 explains this distinction between the fixed-budget analysis and the finite-step training objective. The experiments above assess the resulting controller against simpler alternatives under the shared search protocol. **Response to Question 1. Computational cost** Training and neural refinement require no solver calls or solver-generated solution labels. Table 12 in Appendix C.7 reports the incremental costs relative to DiffILO and PUMA. | Method | SC params | CA params | IS params | SC time (h) | CA time (h) | IS time (h) | |---|---:|---:|---:|---:|---:|---:| | DiffILO | 177,620 | 390,644 | 390,644 | 15.748 | 10.587 | 10.234 | | SCoR-DiffILO | 198,133 | 439,157 | 439,157 | 16.344 | 10.849 | 10.949 | | PUMA | 147,254 | 390,710 | 390,710 | 16.230 | 14.271 | 12.282 | | SCoR-PUMA | 164,055 | 440,279 | 440,279 | 16.435 | 14.654 | 13.067 | SCoR increases the parameter count by 11.4%–12.7% and **total training time by 1.3%–7.0%**, including the refinement computation. The increases on SC, CA, and IS are 3.8%, 2.5%, and 7.0% for DiffILO and 1.3%, 2.7%, and 6.4% for PUMA. At test time, the measured prediction-stage overhead, including refinement, was at most 1 second per instance. This corresponds to at most 0.1% of the common 1,000-second limit on downstream solver search. --- We have also updated the manuscript. Changes are marked in blue in the revised PDF. We hope our response resolves your concerns, and we are open to any further discussions.

I thank the author for the detailed response to my concerns. It has addressed some of my concerns. However, I think that my concerns for the theoretical justification still exist. The improvement from DiffILO is also limited. At this stage I prefer to maintain my score.

Reviewer_Jwa9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SCoR, a state-conditioned refinement module for unsupervised integer linear programming. SCoR uses normalized constraint residuals and sparse attention on the variable-constraint graph to learn a bounded positive diagonal preconditioner that adapts the base update to local conditions. Experiments with DiffILO and PUMA show improvements on several ILP benchmarks and real-world instances.

Critical strengths
  1. SCoR provides a clear and modular mechanism that uses state-dependent constraint feedback to allocate update magnitudes across variables while preserving the underlying unsupervised objective.

  2. The paper provides rich tables and figures covering multiple unsupervised ILP optimizers, benchmark families, and real-world instances.

Critical weaknesses
  1. Dependence on the downstream search protocol. The paper provides pre-search diagnostics and Δ-only comparisons, but the gains are stronger under the confidence-guided partial search used in the main experiments and are not consistent across inference protocols. It remains unclear how much of the final solver improvement comes from SCoR itself and how much comes from the particular downstream search strategy.

  2. Unclear cost effectiveness. SCoR adds residual processing, graph attention, variable-wise scaling, and additional training and inference computation. The paper does not clearly quantify this overhead or show that the performance gains justify the added cost.

Questions for authors
  1. The paper includes pre-search and Δ-only comparisons, but the main results use confidence-guided partial search. Can you provide matched comparisons under both inference protocols and clarify how much of the final improvement comes from SCoR itself versus the downstream search strategy?

  2. What is the end-to-end training and inference overhead of SCoR relative to DiffILO and PUMA, including memory and solver-assisted computation? Does the 1,000-second budget include all prediction and refinement overhead, and do the gains remain under matched end-to-end budgets?

  3. How stable and general are the reported gains across independent training seeds and problem families? Please report variation across runs and clarify whether the conclusions extend beyond the benchmark settings where the improvements are strongest.

2 comments

We thank the reviewer for the questions about the source, cost, and stability of the gains. We compare SCoR with each backbone under both search protocols and report what changes in the sampled predictions before the solver runs.

Response to Weakness 1 and Question 1. Refinement gains under both search protocols

In the main comparison, all prediction-guided methods, including PaS, CoCo-MILP, DiffILO, PUMA, and the SCoR variants, use the same confidence-guided partial search settings and solver time limit within each task. We also evaluate DiffILO, PUMA, and their SCoR variants under the original full-variable Δ\Delta-only protocol. The following table brings together the mean final objectives from Tables 1 and 7.

MethodSC, Δ\Delta-only ↓SC, partial search ↓CA, Δ\Delta-only ↑CA, partial search ↑
PaSn/a124.57n/a97186.62
CoCo-MILPn/a124.66n/a97273.99
DiffILO125.27124.6597265.4297281.68
SCoR-DiffILO125.21124.5697306.2397316.01
PUMA125.23124.6297267.9897235.62
SCoR-PUMA125.18124.5697280.7597262.44

SCoR improves the mean final objective over its backbone in all eight paired comparisons. The size of the gain varies with the task, backbone, and protocol. For example, on CA, SCoR improves DiffILO by 34.33 under partial search and by 40.81 under Δ\Delta-only search. On SC, partial search improves the objectives of both base predictors. On CA, Δ\Delta-only search yields higher objectives for both PUMA and SCoR-PUMA. The paired gains therefore hold across protocols even where the settings yield different absolute objectives.

The prediction diagnostics show where an improvement is already visible before search. With SCoR-DiffILO, the mean feasible-sample rate rises from 31.05% to 54.69% on SC and from 48.48% to 56.82% on CA. Among instances with feasible samples from both predictors, the best feasible sampled objective improves on 71 of 98 SC instances and 91 of 97 CA instances. These measurements show that refinement improves sampled predictions on SC and CA without any solver search. The paired solver results show that the final-objective gains persist under both evaluated search protocols (Figure 8 and Appendix C.3).

Response to Weakness 2 and Question 2. Computational overhead and budget accounting

The following table reports trainable parameter counts and total training wall-clock times in hours (Appendix C.7, Table 12).

MethodSC (params / h)CA (params / h)IS (params / h)
DiffILO177,620 / 15.748390,644 / 10.587390,644 / 10.234
SCoR-DiffILO198,133 / 16.344439,157 / 10.849439,157 / 10.949
PUMA147,254 / 16.230390,710 / 14.271390,710 / 12.282
SCoR-PUMA164,055 / 16.435440,279 / 14.654440,279 / 13.067

The parameter count increases by 11.4%–12.7%. Total training time increases by 1.3%–7.0%, including gradient computation and constraint attention for refinement. On SC, CA, and IS, the increases are 3.8%, 2.5%, and 7.0% for DiffILO and 1.3%, 2.7%, and 6.4% for PUMA. The new parameters add approximately 0.07–0.20 MB of FP32 weight storage. Peak GPU memory remains unmeasured. Training and neural refinement require no solver calls.

In our measurements, SCoR added at most 1 second per instance to the prediction stage, including refinement. Prediction, refinement, candidate sampling, and seed selection all precede the 1,000-second solver search. The gains above therefore come with a small measured prediction-stage overhead at equal solver time. Whether they persist at equal end-to-end wall-clock time remains to be evaluated.

Response to Question 3. Stability across random seeds and scope of generalization

For each SCoR variant and task, we trained two models independently using seeds 1 and 2. Each trained model was evaluated with inference seeds 1, 2, and 3 on the same 100 test instances, with the search settings and solver budget fixed. We report the mean and, where available, the sample standard deviation of the six test-set mean objectives. The DiffILO and PUMA entries come from the original runs and serve as fixed references (Appendix C.8, Table 13).

MethodSC Obj ↓CA Obj ↑IS Obj ↑
DiffILO124.65097281.680684.140
SCoR-DiffILO124.558 ± 0.00897314.495 ± 1.073684.140 ± 0.000
PUMA124.62097235.620684.140
SCoR-PUMA124.560 ± 0.01097264.732 ± 2.657684.140 ± 0.000

Both SCoR variants retain better mean objectives than their fixed backbone references on SC and CA. The spread across the evaluated configurations is small relative to those mean differences. On IS, SCoR-DiffILO attains the same test-set mean in all six configurations at the reported precision, consistent with the small differences among methods on the standard IS benchmark.

Both variants also improve on their respective backbones on larger instances of SC, CA, and IS (Table 9). In the real-world evaluation, we added two held-out mine production scheduling instances to the two CVS instances. Table 10 shows that SCoR-DiffILO improves the objective on reblock354 from −39,280,200 to −39,280,400 relative to both Gurobi and DiffILO, complementing the improvement from −96 to −97 on cvs16r128-89. These four tests assess applicability to held-out instances within two families, with training conducted separately for each family (Appendix C.5).


We have also updated the manuscript. Changes are marked in blue in the revised PDF.

We hope our response resolves your concerns, and we are open to any further discussions.

Reviewer_YKJ5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SCoR (State-Conditioned Refinement), a framework for unsupervised integer linear programming (ILP) optimization. It addresses a limitation of existing methods: aggregating constraint violations into global optimization objectives may obscure the evolving states of individual constraints and their heterogeneous effects on decision variables.

SCoR computes normalized constraint residuals from the current relaxed prediction and propagates them to incident variables through sparse attention. This local feedback, together with structural representations, prediction states, and gradients of the base objective, is used to parameterize a bounded positive diagonal preconditioner that adaptively scales variable-wise updates.

The paper theoretically characterizes the boundedness, descent compatibility, and allocation flexibility of the refinement. SCoR is integrated with two unsupervised methods, DiffILO and PUMA, and evaluated on three ILP benchmarks, larger instances, and selected MIPLIB problems. The results demonstrate improvements in prediction quality and solver-assisted optimization, particularly on SC and CA.

Critical strengths

Strength 1. The paper identifies a meaningful limitation of aggregate constraint feedback in unsupervised ILP optimization.

The distinction between global optimization guidance and variable-specific constraint states is well motivated. Variables with similar gradients may face different local constraint conditions, whose importance evolves with the prediction. Explicitly incorporating such state information into the refinement process is a meaningful research direction.

Strength 2. SCoR provides a modular refinement mechanism while preserving the underlying unsupervised objectives.

By learning a bounded positive diagonal preconditioner rather than an unconstrained correction direction, SCoR incorporates local constraint feedback without replacing the base optimization objective. Its integration with DiffILO and PUMA demonstrates flexibility across different unsupervised optimization frameworks.

Strength 3. The experiments provide encouraging evidence of improved prediction quality beyond final solver-assisted performance.

The evaluation covers multiple benchmarks, two unsupervised backbones, size generalization, and both Gurobi and SCIP.

Appendix C.3 shows notable prediction-level improvements: the feasible-sample rate on SC increases from 31.05% to 54.69%, with corresponding gains on CA. These results suggest that refinement improves the predicted solution distribution before downstream solver search. The component ablations and original-search comparisons are also informative.

Critical weaknesses

Weakness 1. The necessity of the state-conditioned attention mechanism remains insufficiently established.

The current experiments do not establish why the full attention mechanism is necessary compared with simpler alternatives. For example, a scaling controller based only on gradients and logits, or mean/sum aggregation of constraint residuals, may provide similar variable-wise adaptation.

Table 2 reinforces this concern. On CA, removing Constraint Feedback retains most of the improvement (97308.49 versus 97316.01 for the full model). On SC, removing State Conditioning yields nearly identical performance (124.57 versus 124.56). These results suggest task-dependent contributions rather than consistent benefits from the complete mechanism.

Without comparisons against simpler, parameter-matched refinement baselines, it remains unclear whether the gains arise from state-conditioned constraint reasoning or increased model capacity.

Weakness 2. The theoretical results provide limited justification for the effectiveness of the learned mechanism.

Proposition 4.1 establishes descent compatibility and bounded updates, but these properties follow primarily from positive diagonal scaling and the sign-preserving tanh transformation, independently of the proposed attention mechanism or the quality of the learned preconditioner.

Moreover, descent compatibility does not guarantee finite-step objective decrease or improved discrete solution quality. The use of different stochastic samples for the update direction (ξ₀) and the refined loss (ξ_ref) further limits the interpretation of this guarantee for the expected objective.

Proposition 4.2 establishes greater allocation flexibility than uniform scaling, but this result largely follows from feasible-set inclusion. The learned controller neither explicitly optimizes the proposed allocation objective nor enforces the matched-budget constraint.

Thus, the theory demonstrates the potential flexibility of variable-wise scaling without establishing why the learned state-conditioned preconditioner should outperform simpler alternatives.

Weakness 3. The current constraint-state representation does not fully capture probabilistic feasibility risk.

SCoR represents each constraint using the positive residual of the relaxed marginal prediction, [aj⊤p−bj]+[a_j^\top p-b_j]_+. However, feasibility of the marginal prediction does not imply feasibility of samples drawn from the predicted distribution.

For example, consider x1+x2≤1x_1+x_2\leq 1 with independent Bernoulli marginals p1=p2=0.5p_1=p_2=0.5. The marginal residual is zero, suppressing the corresponding constraint message, despite a 25%25\% probability of sampled violation.

More generally, E[[aj⊤x−bj]+]\mathbb{E}[[a_j^\top x-b_j]_+] can be positive even when [aj⊤p−bj]+=0[a_j^\top p-b_j]_+=0. The representation also suppresses feedback from satisfied constraints approaching their boundaries.

This raises concerns about whether marginal residuals adequately characterize probabilistic constraint states. Comparisons with signed slack, expected sampled violation, or estimated violation probability would help justify this design choice.

Questions for authors

Q1. Core novelty and necessity of state-conditioned feedback.

What constitutes the main methodological novelty of SCoR beyond combining constraint residual features, sparse attention, and learned diagonal preconditioning?

Could the authors compare SCoR against simpler alternatives, including gradient-magnitude-based scaling, learned controllers using only gradients and logits, and mean/sum residual aggregation without attention?

Parameter-matched baselines would help determine whether the gains arise from state-conditioned constraint reasoning or increased model capacity.

A clearer discussion of how SCoR differs from prior neural correction methods, particularly Apollo-MILP, would also strengthen the novelty argument.

Q2. End-to-end computational fairness and efficiency.

Could the authors report training wall-clock time, GPU memory usage, additional parameter counts, and inference latency for DiffILO, PUMA, and their SCoR variants, including the overhead of gradient computation and constraint attention?

Do the reported 1,000-second budgets include neural prediction, refinement, candidate sampling, and seed selection, or only downstream solver time?

Solution-quality-versus-time curves under matched end-to-end computational budgets would provide stronger evidence of practical efficiency.

Q3. Interpretation and scope of the theoretical guarantees.

Proposition 4.1 establishes a descent-compatible direction but does not guarantee a finite-step decrease in the objective. Under what additional smoothness assumptions or step-size conditions could a one-step descent guarantee be established?

Furthermore, when the update direction is computed using ξ₀ while the refined loss uses independently sampled ξ_ref, how should the theoretical result be extended or interpreted with respect to the expected objective?

For Proposition 4.2, could the authors compare learned scaling against uniform scaling and simple gradient-based allocations under strictly matched effective update budgets, thereby connecting the theoretical advantage to the actual behavior of the learned controller?

Q4. Adequacy of the current constraint-state representation.

Given that marginal feasibility does not imply sampled feasibility, why is positive marginal residual preferred over signed slack, expected sampled violation, or violation probability?

Could the authors experimentally compare these representations, or at least provide a discussion of their relative merits, particularly when relaxed predictions are nearly feasible but sampled assignments frequently violate constraints?

Additionally, how sensitive is the instance-wise residual normalization to the total number of constraints and the proportion of violated constraints? Does it remain stable under changes in instance size?

Q5. Generalization and real-world applicability.

Could the authors extend the evaluation beyond the two held-out CVS instances to additional MIPLIB families with more diverse constraint structures and coefficient distributions?

How could SCoR be extended to general integer and continuous variables, for which the current Bernoulli-based residual representation may no longer be appropriate? Do the authors have a corresponding extension strategy?

Additional evaluations on heterogeneous real-world optimization problems would be needed to support broader claims of transferability.

4 comments

We thank the reviewer for the detailed feedback. We have added simpler scaling baselines and computational measurements, and expanded the discussion of what the theoretical results establish about the learned refinement.

Response to Weakness 1 and Question 1. State-conditioned update allocation and simpler alternatives

SCoR learns how to distribute a base update across variables using the current states of their incident constraints. The base gradient combines contributions from multiple constraints into a single signal per variable. Constraint-to-variable feedback gives the scaling head access to the individual constraint states and their local associations with each variable. The predictor and scaling head learn jointly through the existing unsupervised objective, as demonstrated with DiffILO and PUMA.

We agree that the contribution of each component varies by task. On CA, removing constraint feedback retains much of the improvement over DiffILO (97308.49 versus 97316.01 for full SCoR). On SC, removing state conditioning changes the objective from 124.56 to 124.57. The full model still achieves the best mean objective among the ablations on both tasks (Section 5.4).

To test whether simpler scaling rules can account for the gains, we added fixed gradient-magnitude scaling and two learned alternatives. The gradient-logit controller uses the current logit, its marginal probability, and the normalized gradient. Mean residual aggregation retains the residual-conditioned value transformation and activity gate, with equal weights across incident constraints. We widened the first hidden layer in each learned alternative, bringing their parameter counts within 0.53% of the full refinement module and 0.06% of the full model (Appendix C.9, Table 14).

MethodSC Obj ↓CA Obj ↑
DiffILO124.6597281.68
Gradient-magnitude scaling124.6097236.99
Gradient-logit controller, parameter matched124.7196509.82
Mean residual aggregation, parameter matched124.6296427.74
SCoR-DiffILO124.5697316.01

SCoR achieves the best mean objective on both tasks under the shared search protocol. Fixed gradient scaling and mean residual aggregation improve upon DiffILO on SC, while all three added alternatives fall below DiffILO on CA. Across both tasks, the comparison with the similarly sized mean-aggregation control supports learning how to weight incident constraints. The gradient-logit controller also provides a more restricted comparison than the existing w/o Constraint Feedback ablation, whose scaling head still receives the backbone's structural variable representations.

Apollo-MILP alternates neural prediction with solver-based trust-region correction at inference and uses the resulting reference solutions to guide variable fixing. It trains its predictor on solver-generated solution pools. SCoR learns a bounded logit update during unsupervised training and applies that refinement before downstream solver search. Training and neural refinement require no solver calls. Section 2.1 now clarifies where Apollo-MILP performs its correction.

Response to Question 2. Computational overhead and budget accounting

Table 12 in Appendix C.7 reports trainable parameter counts and total training wall-clock times. Each entry below gives the parameter count and training time in hours.

MethodSC (params / h)CA (params / h)IS (params / h)
DiffILO177,620 / 15.748390,644 / 10.587390,644 / 10.234
SCoR-DiffILO198,133 / 16.344439,157 / 10.849439,157 / 10.949
PUMA147,254 / 16.230390,710 / 14.271390,710 / 12.282
SCoR-PUMA164,055 / 16.435440,279 / 14.654440,279 / 13.067

Across the six comparisons, SCoR increases the parameter count by 11.4%–12.7% and total training time by 1.3%–7.0%. The timings include the gradient computation for refinement, residual processing, constraint attention, and evaluation of the refined loss. The added parameters correspond to approximately 0.07–0.20 MB of FP32 weights. Peak GPU memory remains unmeasured.

In our measurements, SCoR added at most 1 second per instance to the prediction stage, including refinement. This is at most 0.1% of the 1,000-second solver limit. Neural prediction, refinement, candidate sampling, and seed selection precede the solver search. The reported curves therefore compare solution quality at equal solver time. A matched end-to-end comparison would count all pre-solver computation toward a common wall-clock budget and remains to be evaluated.

Response to Weakness 2 and Question 3. Interpretation of the theoretical guarantees

Proposition 4.1 follows from the positive scales and the bounded base direction. This property lets the controller change how the update is distributed while preserving descent compatibility with the sampled gradient and bounding the movement of each logit. The guarantee applies to the sampled base objective in logit space. The effect on discrete solution quality is assessed empirically.

For a fixed draw ξ0\xi_0, let ℓ(z)=Lbase(z;G,ξ0)\ell(z)=\mathcal L_{\mathrm{base}}(z;\mathcal G,\xi_0) denote a differentiable surrogate with g(0)=∇ℓ(z(0))g^{(0)}=\nabla\ell(z^{(0)}). If its gradient is LL-Lipschitz along the update segment for L>0L>0 and g(0)≠0g^{(0)}\ne 0, the descent lemma gives a strict decrease at z(0)+tδzz^{(0)}+t\delta z whenever

0<t<−2⟨g(0),δz⟩L∥δz∥22.0<t<\frac{-2\langle g^{(0)},\delta z\rangle}{L\|\delta z\|_2^2}.

For the implemented step t=1t=1, the bound on the scales alone is insufficient to ensure this inequality.

The independent draw ξref\xi_{\mathrm{ref}} gives an unbiased evaluation of the expected loss at the updated prediction conditional on ξ0\xi_0. Under analogous smoothness for F(z)=Eξ[Lbase(z;G,ξ)]F(z)=\mathbb E_\xi[\mathcal L_{\mathrm{base}}(z;\mathcal G,\xi)], a sufficient alignment condition for expected decrease at small positive steps is ⟨∇F(z(0)),Eξ0[δz]⟩<0\langle\nabla F(z^{(0)}),\mathbb E_{\xi_0}[\delta z]\rangle<0. Since the controller and bounded direction transform the sampled gradient nonlinearly, even an unbiased raw gradient is insufficient to establish this alignment. Appendix B.1 gives the derivations.

Proposition 4.2 identifies when redistributing update magnitude can improve the predicted decrease of the local linear model under a fixed budget. In SCoR, the effective budget is set by the learned scales, which are trained through the finite-step refined loss in Eq. (17). Holding the initial prediction and base direction fixed gives

=d_i^{(0)}\frac{\partial\mathcal L_{\mathrm{base}}(z^{\mathrm{ref}};\mathcal G,\xi_{\mathrm{ref}})}{\partial z_i^{\mathrm{ref}}}.$$ Each scale thus receives a training signal through its effect on the refined prediction. Figure 1 shows how the learned controller redistributes update magnitudes relative to uniform scaling at a matched budget. The additional scaling baselines instead compare final objectives at a common solver budget. A performance comparison of the learned, uniform, and gradient-based allocations under strictly matched effective update budgets remains open (Section 4.3 and Appendix B.1). **Response to Weakness 3 and Question 4. Marginal residuals and sampled feasibility** The example captures a limitation of the feedback signal. For $x_1+x_2\leq 1$ with independent Bernoulli marginals $p_1=p_2=0.5$, the marginal residual is zero even though samples violate the constraint with probability 25%. More generally, if $\mathbb E[x]=p$, Jensen's inequality gives $$[a_j^\top p-b_j]_+\leq\mathbb E_x\bigl[[a_j^\top x-b_j]_+\bigr].$$ We use marginal residuals because a sparse matrix-vector product provides a deterministic, constraint-level signal for allocating the update. Eq. (7) sets a constraint's value message to zero when its marginal residual is zero. A violation of that constraint in a sampled assignment can still contribute to the base gradient and refined training loss through the sampled violation terms retained from DiffILO and PUMA (Appendices A.2 and A.3). Signed slack would preserve information about how close satisfied constraints are to their boundaries, provided the gate retains their messages. Expected sampled violation reflects both the frequency and magnitude of violations, while violation probability captures their frequency. Either quantity could be estimated using the base samples, with accuracy depending on the number of draws. With the current feedback, SCoR-DiffILO increases the mean feasible-sample rate from **31.05% to 54.69% on SC and from 48.48% to 56.82% on CA** before solver search. The relative merits of the alternative representations remain to be tested (Figure 8 and Appendix C.3). The instance normalization depends on the composition of the constraint set. If $q>0$ is the fraction of rows with a positive row-normalized residual and $\mu_+$ is their mean, the instance mean is $q\mu_+$ and $\bar\nu_j=\nu_j/(q\mu_+)$ whenever the denominator floor is inactive. At a fixed prediction, replicating every row equally leaves the normalization unchanged. Adding satisfied rows lowers $q$ and can amplify the existing positive normalized residuals. The numerical floor keeps the denominator positive, and the activity gate bounds the resulting gate values. Section 4.2 now distinguishes these two ways in which the number of constraints can increase.

Response to Question 5. Generalization and real-world applicability

We extended the MIPLIB 2017 evaluation from capacitated vertex separator (CVS) instances to the multi-period mine production scheduling (MINE) family used in PUMA. Following its MINE split, we train on reblock420, reblock166, mine-166-5, and mine-90-10 and test on reblock115 and reblock354. Table 10 in Appendix C.5 reports the following objectives at 1,000 seconds of solver search. Lower is better.

InstanceGurobiDiffILOSCoR-DiffILO
cvs16r106-72−81−81−81
cvs16r128-89−96−96−97
reblock115−36,800,603−36,800,603−36,800,603
reblock354−39,280,200−39,280,200−39,280,400

Table 10 shows an improvement on one held-out instance in each family and a tie on the other. In particular, the objective on reblock354 is 200 lower than for both references. These results provide evidence of applicability within the two evaluated families, each with its own training split and two held-out instances.

The current implementation operates on binary ILPs. Bounded general integer variables can be encoded with binary variables and the original bounds enforced in the resulting formulation, as discussed in DiffILO's Remark 1. This would let SCoR operate on the reformulated instance at the cost of additional variables. We have yet to evaluate this extension. For continuous decisions, the Bernoulli outputs would need to be replaced by a prediction mechanism for continuous values and paired with a compatible unsupervised objective. Constraint residuals could then be computed at the resulting mixed prediction, with state-conditioned scaling applied to the optimization signal supplied by that objective.


We have also updated the manuscript. Changes are marked in blue in the revised PDF.

We hope our response resolves your concerns, and we are open to any further discussions.

Thank you for the detailed rebuttal and additional experiments, which address several of my concerns. However, the necessity of certain components remains insufficiently established, particularly given the marginal improvements in some ablations. Additionally, the theoretical justification and constraint-state representation still have limitations. Therefore, I will maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.