acceptodds
Under review as a conference paper at ICLR 2027

LEAP: LLM-Empowered Adaptive Portfolio Discovery for Instance-Aware MILP Configuration

Abstract

We study instance-aware solver configuration for mixed-integer linear programs (MILPs). Modern MILP solvers expose numerous tunable parameters, and different instances within the same problem family can benefit from different configurations. Portfolio-based approaches exploit this variation by discovering a set of complementary configurations in advance and selecting an appropriate configuration from this portfolio for each new instance. However, portfolio discovery requires expensive solver evaluations over large configuration spaces, while conventional configurators primarily rely on numerical performance feedback without directly leveraging solver parameter semantics or problem-family knowledge. To address this limitation, we propose LEAP (LLM-Empowered Adaptive Portfolio Discovery), which integrates semantic guidance into numerical search. An LLM Initializer proposes starting configurations using problem descriptions and parameter semantics, while an LLM Refiner uses portfolio performance and search feedback to adaptively focus or expand a working subspace within a fixed base configuration space, guiding an existing model-based configurator to search for configurations that improve the current portfolio. At deployment, a lightweight GNN selector, with no LLM calls or configuration search, selects the configuration with the lowest predicted solving cost for each incoming instance. Experiments on MILP benchmarks demonstrate solving-time reductions of up to 43.2% and 94.7% over the default SCIP and Gurobi configurations, respectively. Ablations demonstrate the benefits of knowledge-guided initialization and adaptive working-subspace refinement for discovering higher-quality configuration portfolios.

Primary area:
optimization
Keywords:
Mixed-Integer Linear Programming, LLM

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_Eg4dOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LEAP, an LLM-guided framework for instance-aware MILP solver configuration. It combines LLM-based initialization, adaptive search-space refinement, SMAC-based portfolio discovery, and GNN-based configuration selection. Experiments across 12 datasets and two MILP solvers demonstrate promising improvements in solving efficiency over default configurations and existing baselines.

Critical strengths
  1. The paper addresses a practically important problem in MILP solver configuration and presents a coherent framework integrating LLM guidance with numerical optimization.
  2. The evaluation covers 12 datasets and two widely used MILP solvers, providing encouraging evidence of practical effectiveness.
  3. The ablation studies support the contributions of LLM-guided initialization and adaptive search-space refinement.
Critical weaknesses
  1. The methodological novelty is moderate, as the framework largely combines established configuration and selection techniques. The unique benefits of LLM guidance over simpler heuristic alternatives remain insufficiently demonstrated.
  2. The computational comparisons do not fully account for additional initialization and portfolio evaluation costs, potentially affecting the fairness of the reported efficiency improvements.
  3. Some performance improvements are relatively small, and stronger statistical analysis would help establish their reliability.
Questions for authors
  1. How does LEAP compare with non-LLM heuristic initialization and search-space refinement methods?
  2. Can the authors provide comparisons under matched total computational budgets, including initialization and portfolio evaluation costs?
  3. Are the smaller performance improvements statistically significant across different random seeds?
3 comments

We thank the reviewer for the encouraging assessment and helpful suggestions. We appreciate the opportunity to strengthen the comparisons and presentation of our results, and respond to each point below.

Response to Weaknesses:

W1. Methodological novelty and benefits of LLM guidance. Thank you for this comment. LEAP’s methodological contribution centers on coupling problem-informed initialization with portfolio-aware search-space refinement throughout iterative configuration discovery. In related MIP configuration work, Hydra-MIP discovers complementary configurations through repeated solver evaluations, whose cost limits exploration of large parameter spaces under finite budgets. Lawless et al. (2025) provide effective cold-start separator configurations, but their family-level selection remains constrained by the generated candidates and cannot fully exploit differing configuration preferences across instances. LEAP addresses these limitations by using problem-family knowledge and parameter semantics to propose strategically distinct starting configurations across multiple solver components, then using portfolio performance and search history to focus or expand subsequent numerical search. Semantic guidance thus shapes both the initial portfolio and where subsequent searches seek additional complementary gains.

To quantify the benefits of LLM guidance over simpler heuristic alternatives, we added two controls that separately replace LLM initialization and refinement. LEAP achieves lower mean test solving time with GNN selection than both controls across all 14 evaluated solver–dataset settings. The gains are substantial in several settings: on Gurobi mix_2, LEAP improves over LHS-Init from 52.9s to 22.5s (57.5% improvement) with the same number of initial candidates; on Gurobi rhs_3, it improves over Staged-Refine from 39.8s to 22.9s (42.5% improvement) with LLM initialization retained. These comparisons quantify the additional benefits of the LLM components relative to the evaluated non-LLM alternatives. Our response to Question 1 provides the control definitions, experimental protocol, and complete results.

W2: Discovery-cost accounting. Thank you for highlighting the fairness of the computational comparisons. We added experiments that include initialization and inter-round evaluations within LEAP’s 10-hour total wall-clock discovery budget, correspondingly reducing its SMAC search time. LEAP achieves the lowest mean test solving time among the compared methods in 12 of 14 settings, including one tie. Appendix A.3 also provides the stage-by-stage timing breakdown and supplementary CPU evaluation costs. Our response to Question 2 presents the experimental protocol, complete performance results, and cost-accounting details.

W3: Reliability of smaller improvements. Thank you for highlighting the reliability of smaller performance improvements. Appendix E.1 now reports mean ± standard deviation across three configuration-discovery repetitions for the main-table results. Our response to Question 3 discusses the improvement magnitudes and variability, including both substantial gains and settings with close mean performance.

Response to Questions:

Q1: Non-LLM heuristic controls. Thank you for suggesting comparisons with non-LLM heuristic initialization and search-space refinement methods. We added LHS-Init and Staged-Refine, each replacing one LLM component. LEAP achieves lower mean test solving time with GNN selection than both controls across all 14 solver–dataset settings. Given the time and computational constraints during rebuttal, these experiments cover all Hard and BenLOC benchmarks in the paper. We will extend the evaluation to the remaining benchmarks and provide full reproducibility details in the final version.

LHS-Init uses SMAC’s Latin hypercube sampling to generate the same number of ten distinct initial candidates within the same master configuration space, replacing only the LLM Initializer while retaining the LLM Refiner.

Staged-Refine retains LLM initialization and adapts the alternating optimization of parameter blocks in BliStrTune (Jakubův & Urban, 2017) [1] to LEAP’s refinement between discovery rounds. Parameters are partitioned beforehand into two blocks, primarily covering cuts and presolving (A), and branching and heuristics (B). The first round searches the full space; subsequent rounds alternate between the two blocks, fixing inactive parameters to their values in the previous round’s SMAC incumbent. The portfolio is still updated after each round, and SMAC is reinitialized in the next round to optimize complementarity with the updated portfolio.

Both controls retain the same master configuration space, solver default, portfolio update mechanism, and SMAC settings. They follow the original protocol of five SMAC rounds with two hours per round; initialization and evaluations between rounds remain outside this SMAC search budget. The tables report test solving time with GNN selection, in seconds, as mean ± standard deviation over three random seeds; lower is better.

SCIP

DatasetLEAPLHS-InitStaged-Refine
rhs_3290.7 ± 29.3323.2 ± 20.4347.0 ± 40.6
obj_2364.3 ± 11.5388.3 ± 9.2379.1 ± 19.6
mix_1125.3 ± 14.2151.2 ± 12.8132.4 ± 24.3
mix_2264.4 ± 3.1276.4 ± 21.4274.5 ± 7.1
mix_386.0 ± 1.3100.0 ± 28.2103.0 ± 21.3
nn52.9 ± 9.167.7 ± 7.653.7 ± 5.2
load217.7 ± 9.2225.1 ± 13.8243.1 ± 21.7

Gurobi

DatasetLEAPLHS-InitStaged-Refine
rhs_322.9 ± 8.626.6 ± 6.439.8 ± 14.9
obj_21.5 ± 0.61.9 ± 0.11.62 ± 0.10
mix_112.0 ± 0.412.8 ± 1.212.5 ± 1.3
mix_222.5 ± 11.352.9 ± 19.723.5 ± 2.3
mix_36.6 ± 1.37.9 ± 0.48.5 ± 1.4
nn4.6 ± 0.95.4 ± 1.95.9 ± 3.5
load0.8 ± 0.10.9 ± 0.10.9 ± 0.2

The LHS-Init comparison supports the value of LLM-proposed initialization with the same candidate count, while Staged-Refine supports the benefit of LLM-guided refinement relative to the evaluated block-search rule, with LLM initialization retained. These replacement experiments complement the existing component-removal ablations and show that the benefits of both LLM components carry through to performance with the deployed selector. We would also welcome the reviewer’s suggestions for other relevant non-LLM heuristics to explore in future work.

[1] Jakubův, Jan, and Josef Urban. "BliStrTune: hierarchical invention of theorem proving strategies." Proceedings of the 6th ACM SIGPLAN Conference on Certified Programs and Proofs. 2017.

Q2: Discovery costs and budget comparison. Thank you for highlighting the importance of including initialization and full discovery-set evaluations when assessing discovery efficiency. We conducted additional experiments in which initialization evaluations, SMAC search, and inter-round evaluations jointly receive a 10-hour wall-clock discovery budget. LEAP achieves the lowest mean test solving time with GNN selection among the compared methods in 12 of 14 settings, including one tie, with improvements of up to 93.8% over Global-SMAC and 61.0% over w/o Both. In one further setting, SCIP obj_2, LEAP is within 0.6% of the strongest baseline. In the remaining setting, Gurobi rhs_3, LEAP takes 26.5s compared with 23.0s for the strongest baseline, DGCAC.

Given the time and computational resources available during rebuttal, these experiments cover all five Hard and both BenLOC datasets, using both solvers. We train GNN selectors for the portfolios discovered under this budget. Baselines retain their main-experiment results: Global-SMAC and w/o Both retain their full 10-hour SMAC search allocations, and w/o Both also retains its additional evaluations. Thus, LEAP achieves these results after deducting evaluation time and using less SMAC search time, while the baselines’ search and evaluation resources remain unchanged.

The tables report test solving time in seconds, as mean ± standard deviation over three repetitions. Lower is better; bold marks the lowest mean in each row.

SCIP

DatasetGlobal-SMACDGCACw/o BothLEAP (10h)
rhs_3317.4 ± 39.3327.1 ± 27.8313.1 ± 12.8260.9 ± 32.1
obj_2364.5 ± 2.0398.4 ± 20.6383.7 ± 10.5366.7 ± 23.9
mix_1146.3 ± 32.1154.2 ± 36.4155.7 ± 10.5117.2 ± 17.8
mix_2259.2 ± 1.3264.0 ± 5.5285.8 ± 3.1258.5 ± 11.1
mix_395.3 ± 3.191.4 ± 33.2111.8 ± 34.772.9 ± 17.2
nn72.3 ± 8.565.6 ± 6.758.9 ± 2.754.0 ± 6.3
load241.9 ± 9.0348.6 ± 101.9252.3 ± 19.3229.8 ± 1.2

Gurobi

DatasetGlobal-SMACDGCACw/o BothLEAP (10h)
rhs_325.8 ± 8.923.0 ± 13.525.1 ± 10.926.5 ± 2.3
obj_225.7 ± 0.128.3 ± 1.84.1 ± 5.91.6 ± 0.0
mix_112.2 ± 0.013.6 ± 0.012.3 ± 0.412.2 ± 4.5
mix_245.7 ± 17.147.8 ± 9.526.6 ± 32.118.0 ± 10.5
mix_38.6 ± 0.18.8 ± 2.313.1 ± 12.65.2 ± 0.9
nn5.2 ± 0.75.1 ± 1.25.1 ± 1.33.9 ± 0.3
load1.3 ± 0.43.5 ± 4.00.9 ± 0.40.5 ± 0.0

Note that SMAC search uses previous evaluation feedback to adaptively select subsequent candidates; in contrast, during initialization and inter-round evaluation, the candidate configurations are already fixed, and runs of a given configuration on different instances require no feedback from one another. These evaluations are therefore naturally amenable to batch parallelism, which we exploit using 20 parallel single-thread CPU workers, allowing the evaluations to proceed concurrently and shortening their elapsed time.

Q3: Variability and smaller performance improvements. Thank you for highlighting the reliability of smaller improvements. Appendix E.1 now reports the results corresponding to the main tables as mean ± standard deviation across three configuration-discovery repetitions. LEAP’s overall performance gains span both solvers and diverse tasks: it achieves lower mean test solving time than Default in all 24 settings, with improvements of up to 43.2% for SCIP and 94.7% for Gurobi.

The advantages over other configuration methods also extend across most settings. LEAP achieves lower mean solving time than DGCAC and Global-SMAC in 21/24 settings each. For example, on Gurobi obj_2, LEAP achieves 1.5 ± 0.6 seconds compared with 25.7 ± 0.1 seconds for Global-SMAC, an improvement of approximately 94.2%. Other comparisons are close: on SCIP obj_2, LEAP achieves 364.3 ± 11.5 seconds versus 364.5 ± 2.0 seconds for Global-SMAC. Across the ablations, LEAP achieves the lowest mean test solving time in 21/24 settings. On SCIP load, LEAP achieves 217.7 ± 9.2 seconds compared with 241.1 ± 15.7 seconds without refinement, an improvement of approximately 9.7%. On Gurobi load, both LEAP and w/o Refine achieve 0.8 ± 0.1 seconds, corresponding to improvements over Default of 94.7% and 94.6%, respectively. The variant retaining LLM initialization therefore already delivers substantial gains on this benchmark.

Reviewer_Tk75OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LEAP (LLM-Empowered Adaptive Portfolio Discovery), an instance-aware MILP solver configuration framework that combine LLM guidance with iterative portfolio discovery. An LLM initializer proposes configurations using problem descriptions and parameter semantics, while an LLM refiner adapts the working subspace searched by SMAC. A GNN subsequently selects a configuration for each incoming instance without additional LLM calls or configuration search. The experiments cover 12 datasets with SCIP and Gurobi, including component ablations, test-set portfolio evaluation, and transfer to larger instances. The proposed method gives solving-time reductions of up to 43.2% and 94.7% over the respective defaults, with improvements over unguided portfolio discovery across most settings.

Critical strengths

The framework provides a practical integration of semantic guidance and measured solver performance. The integration of LLM remain within an explicit admissible configuration space, and portfolio admission depends on empirical performance. Expensive discovery occurs offline while deployment uses a lightweight selector. This design makes the approach applicable to settings where many related MILPs must be solved repeatedly. The experimental evaluation provides evidence of practical effectiveness. LEAP improves over default on all 24 solver–dataset pairs and achieves the lowest reported mean solving time on 9 datasets for each solver. The component ablation and test-set portfolio results provide evidence beyond comparisons against defaults. In particular the test-set per-instance-best results show that improvements persist independently of selector errors

Critical weaknesses
  1. Appendix A.1 explicitly excludes the evaluation of 10 initial configurations and subsequent full discovery-set evaluations from the 10 hour search budget. Appendix A.3 reports 3.76 additional elapsed hours for these evaluations on SCIP load, using 20 parallel workers. Variants without initialization do not seem to receive the same 10-candidate evaluation stage. Therefore, comparisons against unguided portfolio discovery and Global-SMAC combine differences in guidance with differences in evaluation resources. The separate cost reporting is useful, but a comparison accounting for total solver evaluations or CPU time is needed to establish discovery efficiency.

  2. The 4-way ablation shows that the chosen LLM components can help, but there is no matched control using ten non-LLM initial candidates or a numerical working-subspace adaptation rule. Therefore, it remains unclear how much benefit comes from semantic knowledge, candidate diversification or generic search-space adaptation. Moreover, the largest headline improvement is nearly matched by simpler variants: on Gurobi load, LEAP achieves 94.7% improvement, compared with 94.6% without refinement and 94.0% without either LLM component. This does not negate the broader gains, but makes the marginal contribution important to quantify. Main-table selector results also lack uncertainty estimates, limiting interpretation of smaller differences

Questions for authors
  1. What are the total solver-evaluation counts and CPU costs for LEAP and its principal baselines including evaluations outside SMAC ? Do existing cost records support comparable performance conclusions under a common total discovery budget ?
  2. Were any controls using the same number of non-LLM initial configurations or a numerical subspace-refinement rule evaluated? If such results already exist, how do they compare with the LLM components?
  3. What varies across the 3 random seeds—solver randomn, SMAC search, GNN training, and independently generated LLM responses ? Can the existing runs provide paired uncertainty estimates for LEAP's improvements over the single-component ablations ?
3 comments

We thank the reviewer for the careful evaluation and detailed feedback. These comments have helped us strengthen the experimental evidence and clarify our methodology. We address each point below.

Response to Weaknesses:

W1: Discovery-cost accounting. Thank you for highlighting the importance of accounting for additional initialization and portfolio evaluation costs when assessing discovery efficiency. We added experiments that include these evaluations within LEAP’s 10-hour total discovery budget and correspondingly reduce its SMAC search time. LEAP achieves the lowest mean test solving time among the compared methods in 12 of 14 settings. We also supplement the timing breakdown with CPU evaluation costs in Appendix A.3. Our response to Question 1 provides the experimental protocol, complete performance results, and cost-accounting details.

W2: Marginal contributions of LLM guidance. Thank you for highlighting the importance of quantifying the marginal contributions of LLM guidance. We added both requested replacement controls, and LEAP achieves lower mean test solving time than both across all 14 evaluated solver–dataset settings. The gains are substantial in several settings: on Gurobi mix_2, LEAP improves over LHS-Init from 52.9s to 22.5s (57.5% improvement) with the same number of initial candidates; on Gurobi rhs_3, it improves over Staged-Refine from 39.8s to 22.9s (42.5% improvement) with LLM initialization retained. These comparisons quantify the additional benefits of LLM initialization and refinement relative to the evaluated non-LLM alternatives. Question 2 provides the control definitions, experimental protocol, and complete results.

The four-way ablation shows broad performance gains: LEAP achieves the lowest mean test solving time in 21 of 24 solver–dataset settings, including one tie. Some gains are substantial: on Gurobi mix_3, LEAP improves over w/o Init from 15.0s to 6.6s (56.0% improvement); on Gurobi rhs_3, it improves over w/o Refine from 34.8s to 22.9s (34.2% improvement). In two of the three remaining settings, SCIP bnd_3 and obj_1, LEAP is within 2.9% and 1.9% of the best ablation, respectively; in the third, Gurobi obj_2, the difference is 0.3s (1.5s versus 1.2s). The close improvements over Default on Gurobi load—94.7%, 94.6%, and 94.0% for LEAP, w/o Refine, and w/o Both—likewise illustrate that the marginal gains depend on the benchmark. Thus, the component-level evidence combines broad overall performance with substantial gains in several settings and close comparisons in others. Appendix E.1 reports the complete selector results as mean ± standard deviation, and Question 3 clarifies the sources of randomness.

Response to Questions:

Q1: Discovery costs and budget comparison. Thank you for highlighting the importance of including initialization and full discovery-set evaluations when assessing discovery efficiency. We conducted additional experiments in which initialization evaluations, SMAC search, and inter-round evaluations jointly receive a 10-hour wall-clock discovery budget. LEAP achieves the lowest mean test solving time with GNN selection among the compared methods in 12 of 14 settings, with improvements of up to 93.8% over Global-SMAC and 61.0% over w/o Both. In one further setting, SCIP obj_2, LEAP is within 0.6% of the strongest baseline. In the remaining setting, Gurobi rhs_3, LEAP takes 26.5s compared with 23.0s for the strongest baseline, DGCAC.

Given the time and computational resources available during rebuttal, these experiments cover all five Hard and both BenLOC datasets, using both solvers. We train GNN selectors for the portfolios discovered under this budget. Baselines retain their main-experiment results: Global-SMAC and w/o Both retain their full 10-hour SMAC search allocations, and w/o Both also retains its additional evaluations. Thus, LEAP achieves these results after deducting evaluation time and using less SMAC search time, while the baselines’ search and evaluation resources remain unchanged.

The tables report test solving time in seconds, as mean ± standard deviation over three repetitions. Lower is better; bold marks the lowest mean in each row.

SCIP

DatasetGlobal-SMACDGCACw/o BothLEAP (10h)
rhs_3317.4 ± 39.3327.1 ± 27.8313.1 ± 12.8260.9 ± 32.1
obj_2364.5 ± 2.0398.4 ± 20.6383.7 ± 10.5366.7 ± 23.9
mix_1146.3 ± 32.1154.2 ± 36.4155.7 ± 10.5117.2 ± 17.8
mix_2259.2 ± 1.3264.0 ± 5.5285.8 ± 3.1258.5 ± 11.1
mix_395.3 ± 3.191.4 ± 33.2111.8 ± 34.772.9 ± 17.2
nn72.3 ± 8.565.6 ± 6.758.9 ± 2.754.0 ± 6.3
load241.9 ± 9.0348.6 ± 101.9252.3 ± 19.3229.8 ± 1.2

Gurobi

DatasetGlobal-SMACDGCACw/o BothLEAP (10h)
rhs_325.8 ± 8.923.0 ± 13.525.1 ± 10.926.5 ± 2.3
obj_225.7 ± 0.128.3 ± 1.84.1 ± 5.91.6 ± 0.0
mix_112.2 ± 0.013.6 ± 0.012.3 ± 0.412.2 ± 4.5
mix_245.7 ± 17.147.8 ± 9.526.6 ± 32.118.0 ± 10.5
mix_38.6 ± 0.18.8 ± 2.313.1 ± 12.65.2 ± 0.9
nn5.2 ± 0.75.1 ± 1.25.1 ± 1.33.9 ± 0.3
load1.3 ± 0.43.5 ± 4.00.9 ± 0.40.5 ± 0.0

Timing and CPU costs. In Appendix A.3, Tables 9–10 provide stage-by-stage timing for portfolio discovery, label generation, and selector training in the main experiments across all 12 datasets and both solvers. Table 11 reports the cumulative CPU-hours for evaluating the ten initial configurations and the five round-end configurations in the additional 10-hour experiments on the seven Hard and BenLOC datasets. Table 12 separately reports online inference time.

Note that SMAC search uses previous evaluation feedback to adaptively select subsequent candidates; in contrast, during initialization and inter-round evaluation, the candidate configurations are already fixed, and runs of a given configuration on different instances require no feedback from one another. These evaluations are therefore naturally amenable to batch parallelism, which we exploit using 20 parallel single-thread CPU workers. This parallelism makes the actual elapsed evaluation time substantially shorter than the cumulative CPU-hours in Table 11, which sum computation across individual solver runs.

Q2: Non-LLM baselines. Thank you for suggesting these controls. We evaluated both: LHS-Init uses the same number of ten non-LLM initial configurations, and Staged-Refine replaces the LLM Refiner with a numerical subspace-refinement rule. LEAP achieves lower mean test solving time with GNN selection than both controls across all 14 evaluated solver–dataset settings. Each control replaces one LLM component while retaining the other. Given the time and computational constraints during rebuttal, these experiments cover all Hard and BenLOC benchmarks in the paper.

LHS-Init uses SMAC’s Latin hypercube sampling to generate the same number of ten distinct initial candidates within the same master configuration space, replacing only the LLM Initializer while retaining the LLM Refiner.

Staged-Refine retains LLM initialization and adapts the alternating optimization of parameter blocks in BliStrTune (Jakubův & Urban, 2017) [1] to LEAP’s refinement between discovery rounds. Parameters are partitioned beforehand into two blocks, primarily covering cuts and presolving (A), and branching and heuristics (B). The first round searches the full space; subsequent rounds alternate between the two blocks, fixing inactive parameters to their values in the previous round’s SMAC incumbent. The portfolio is still updated after each round, and SMAC is reinitialized in the next round to optimize complementarity with the updated portfolio.

Both controls retain the same master configuration space, solver default, portfolio update mechanism, and SMAC settings. They follow the main-experiment protocol of five SMAC rounds with two hours per round; initialization and evaluations between rounds remain outside this SMAC search budget. The tables report test solving time with GNN selection, in seconds, as mean ± standard deviation over three random seeds; lower is better.

SCIP

DatasetLEAPLHS-InitStaged-Refine
rhs_3290.7 ± 29.3323.2 ± 20.4347.0 ± 40.6
obj_2364.3 ± 11.5388.3 ± 9.2379.1 ± 19.6
mix_1125.3 ± 14.2151.2 ± 12.8132.4 ± 24.3
mix_2264.4 ± 3.1276.4 ± 21.4274.5 ± 7.1
mix_386.0 ± 1.3100.0 ± 28.2103.0 ± 21.3
nn52.9 ± 9.167.7 ± 7.653.7 ± 5.2
load217.7 ± 9.2225.1 ± 13.8243.1 ± 21.7

Gurobi

DatasetLEAPLHS-InitStaged-Refine
rhs_322.9 ± 8.626.6 ± 6.439.8 ± 14.9
obj_21.5 ± 0.61.9 ± 0.11.62 ± 0.10
mix_112.0 ± 0.412.8 ± 1.212.5 ± 1.3
mix_222.5 ± 11.352.9 ± 19.723.5 ± 2.3
mix_36.6 ± 1.37.9 ± 0.48.5 ± 1.4
nn4.6 ± 0.95.4 ± 1.95.9 ± 3.5
load0.8 ± 0.10.9 ± 0.10.9 ± 0.2

The LHS-Init comparison supports the value of LLM-proposed initialization with the same candidate count, while Staged-Refine supports the benefit of LLM-guided refinement relative to the evaluated block-search rule, with LLM initialization retained. These replacement experiments complement the existing component-removal ablations and show that the benefits of both LLM components carry through to performance with the deployed selector. We would also welcome the reviewer’s suggestions for other relevant non-LLM heuristics to explore in future work.

[1] Jakubův, Jan, and Josef Urban. "BliStrTune: hierarchical invention of theorem proving strategies." Proceedings of the 6th ACM SIGPLAN Conference on Certified Programs and Proofs. 2017.

Q3: Sources of randomness and uncertainty. Thank you for asking us to clarify the sources of randomness and provide paired uncertainty estimates. The three seeds are used for SMAC search; solver random settings, the GNN training seed, and LLM initialization configurations remain fixed across repetitions for each solver–dataset setting. LLM refinement is invoked separately using feedback from each run. The reported standard deviations therefore reflect variation across three configuration-discovery repetitions. Appendix E.1 (“Test Solving Performance”) reports the complete test solving results for LEAP and its single-component ablations as mean ± standard deviation across these three repetitions.

Reviewer_9p5tOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

In this work, the authors propose LEAP, an LLM-empowered adaptive portfolio discovery framework for instance-aware MILP configuration. LEAP integrates problem-family knowledge and solver parameter semantics into numerical search to discover complementary configurations. The discovered portfolio is then used to select a configuration for each new instance without additional LLM calls or configuration search. The authors evaluate LEAP with SCIP and Gurobi on 12 MILP datasets. The results show solving-time improvements over the default configurations, while ablations examine initialization and adaptive refinement.

Critical strengths

LEAP combines problem-family knowledge and solver parameter semantics with numerical search to discover complementary configurations.

Experiments with SCIP and Gurobi on 12 datasets show consistent improvements over the default configurations, supported by component ablations.

Critical weaknesses
  1. Limited methodological novelty. LEAP mainly combines existing portfolio search, SMAC, LLM guidance, and GNN selection. The contribution is primarily a practical integration of established components.

  2. Limited transfer evaluation. LEAP repeats portfolio discovery and selector training for each solver and problem family. The only transfer experiment uses larger instances from the same family, leaving cross family generalization untested.

Questions for authors
  1. The paper reports detailed offline costs only for the load dataset. Could you report the total numbers of solver evaluations and the complete offline costs for discovery, label generation, and selector training across the other datasets and both solvers?

  2. On the same final LEAP portfolio, how does the GNN selector compare with the single configuration selected on the validation set? Please report the corresponding test performance and selection regret, and identify cases where the PI gains are not realized by the deployed selector.

  3. How well does LEAP transfer across problem families or solvers? The current experiment only changes instance scale within the mix_2 family. Please clarify whether portfolio discovery and selector training must be repeated and whether any discovered configurations can be reused.

3 comments

We thank the reviewer for the positive assessment and constructive comments. We appreciate the opportunity to clarify our contributions and experimental findings, and respond to each point below.

Response to Weaknesses:

W1: Limited methodological novelty. LEAP’s methodological contribution centers on coupling problem-informed initialization with portfolio-aware search-space refinement throughout iterative configuration discovery. In related MIP configuration work, Hydra-MIP discovers complementary configurations through repeated solver evaluations, whose cost limits exploration of large parameter spaces under finite budgets. Lawless et al. (2025) provide effective cold-start separator configurations, but their family-level selection remains constrained by the generated candidates and cannot fully exploit differing configuration preferences across instances. LEAP addresses these limitations by using problem-family knowledge and parameter semantics to propose strategically distinct starting configurations, then using portfolio performance and search history to focus or expand subsequent numerical search. This provides a stronger starting portfolio and directs further search toward complementary gains beyond the initial proposals.

W2: Limited transfer evaluation. Thank you for highlighting the scope of transfer evaluation. Please see our response to Question 3.

Response to Questions:

Q1: Offline costs and evaluation counts. Thank you for this helpful suggestion. We provide a stage-by-stage cost breakdown across all 12 datasets and both solvers, covering portfolio discovery, label generation, and selector training, with test evaluation and online inference reported separately. Discovery uses five 2-hour SMAC rounds. Outside this search budget, evaluating ten initial configurations and one configuration after each of five rounds schedules (15n) configuration–instance evaluation tasks for a discovery set of (n) instances.

Note that the solver evaluations for initialization, portfolio updates, and label generation are naturally parallelizable across instances. For a given configuration, evaluating one instance requires no feedback from evaluations on other instances. We therefore execute these stages using 20 parallel single-thread CPU workers. SMAC search and GNN training are accounted for separately. Offline total = 10h + Init. + Updates + Train labels + Val. labels + GNN, calculated before rounding. All stage costs are in hours; inference is in seconds per instance. Test evaluation and online inference are excluded from the offline total.

SCIP

DatasetInit. (h)Updates (h)Train labels (h)Val. labels (h)GNN (h)Offline total (h)Inference (s/instance)
bnd_20.7510.2680.4870.1380.07611.7200.064
bnd_31.2530.5400.5950.1620.07012.6200.067
rhs_20.2090.0660.0590.0160.15210.5020.141
rhs_30.8570.2621.1680.2900.27812.8540.916
rhs_40.2390.0870.0770.0200.15210.5750.146
obj_10.7070.2750.3290.1240.07211.5080.061
obj_21.0610.4240.7790.2010.16612.6300.236
mix_10.3340.1110.4990.1090.08111.1350.134
mix_20.5390.1970.4510.1130.06311.3620.427
mix_30.2480.0530.3470.0900.07910.8170.134
nn0.6070.1701.1241.4420.57113.9130.088
load2.9340.8263.3073.6120.72321.4020.235

Gurobi

DatasetInit. (h)Updates (h)Train labels (h)Val. labels (h)GNN (h)Offline total (h)Inference (s/instance)
bnd_20.7980.3610.1550.0370.07511.4260.064
bnd_30.8270.3730.1660.0430.07011.4790.067
rhs_20.0350.0090.0080.0020.15210.2060.140
rhs_30.0440.0110.1330.0140.27010.4730.909
rhs_40.0300.0070.0070.0020.15210.1980.143
obj_10.1490.0640.0460.0240.07110.3550.061
obj_20.0430.0110.0290.0050.16510.2530.263
mix_10.0280.0060.1260.0350.08010.2760.136
mix_20.0260.0070.1280.0260.06210.2480.419
mix_30.0290.0110.0140.0050.07810.1370.136
nn0.0480.0120.0960.1420.56510.8640.088
load0.0600.0040.0340.0260.71410.8380.233

The same stage-level breakdown is included in the revised Appendix A.3. GNN training reads cached labels without additional solver calls; deployment requires neither further LLM calls nor configuration search or retraining.

Q2: Selector performance and regret. Thank you for suggesting this controlled comparison. On the same final LEAP portfolios, LEAP achieves lower or comparable mean test solving time than a validation-selected fixed configuration across all 24 solver–dataset settings. For each repetition, Fixed (val-selected) chooses the configuration with the lowest mean validation solving time from the final portfolio and applies it to all test instances; LEAP uses the GNN to select a configuration for each instance. Both use the same portfolio and test set. LEAP-PI reports the portfolio’s minimum measured solving time for each test instance, averaged over the test set, representing its performance under ideal selection.

We define selection regret as mean test solving time minus LEAP-PI, measuring how much of the available portfolio potential remains unrealized. LEAP improves mean solving time over Fixed by 39.8% on SCIP mix_3 and 85.9% on Gurobi obj_2; in the latter case, regret decreases from 9.2s to 0.1s. LEAP achieves lower or comparable mean solving times than Fixed across all settings, with substantial gains in several cases, while some portfolio potential remains unrealized by the deployed selector. On some instances, the best configuration in the portfolio solves quickly, while some other configurations can take up to 500s. Selecting one of these slower configurations therefore incurs a large per-instance regret, so even a small number of such errors can substantially increase the average regret. This highlights opportunities for future work to improve selection and better exploit the discovered portfolios.

SCIP

DatasetFixed (val-selected)LEAPLEAP-PIFixed regretLEAP regret
bnd_2233.4195.6179.054.416.6
bnd_3412.8394.3384.428.49.9
rhs_253.252.250.72.41.5
rhs_467.159.555.811.33.7
obj_1405.9399.6374.931.124.8
rhs_3328.3290.7147.8180.4142.9
obj_2387.7364.3316.171.648.2
mix_1128.1125.339.788.485.6
mix_2283.1264.4177.8105.286.6
mix_3142.886.058.384.527.7
nn53.252.930.522.722.4
load243.8217.7130.2113.687.5

Gurobi

DatasetFixed (val-selected)LEAPLEAP-PIFixed regretLEAP regret
bnd_2270.6269.3267.03.62.3
bnd_3258.5258.5258.50.00.0
rhs_26.56.15.90.60.2
rhs_45.25.25.20.00.0
obj_1114.1103.7100.213.93.5
rhs_342.422.95.736.717.2
obj_210.61.51.49.20.1
mix_112.712.010.22.51.8
mix_224.022.53.520.519.0
mix_312.46.63.88.62.8
nn6.14.62.14.02.5
load0.80.80.60.20.2

Q3: Transfer scope and configuration reuse. Thank you for asking us to clarify the transfer setting and the reuse of discovered configurations. LEAP targets scenarios that require repeatedly solving related MILP instances, enabling reuse of a discovered portfolio and trained selector within a given solver and problem family. For example, supply-chain planning repeatedly solves instances of the same underlying model, but planning horizons, customer counts, and batch requirements may change between planning runs. Reusing the portfolio and selector across these changes avoids repeating discovery and training for each incoming instance and amortizes offline costs over repeated solves.

Consistent with this scope, our transfer experiment changes more than instance scale. Beyond extending the planning horizon from 6 to 13/21 weeks, we jointly change discrete decision periods (3 → 6/10), customer counts (40 → 35/50/65), and batch factors (20 → 1/2/3). These changes reflect the planning, customer-coverage, and batch-requirement variations described above. Both the portfolio and selector remain frozen, without target-domain configuration search, LLM refinement, training, or model selection. On the 13- and 21-week groups, the deployed selector retains mean solving-time improvements over Default of 7.1% and 5.4%, respectively; the corresponding PI improvements are 26.1% and 17.7% (Table 3). These results support direct reuse under the tested changes within the same family.

We discuss the reviewer’s two broader transfer scenarios separately below.

Across solvers: we perform solver-specific discovery and training. SCIP and Gurobi expose different parameter names, semantics, and admissible values, so configurations cannot be directly copied between them. The selector’s outputs also correspond to particular configurations and their costs on the chosen solver. Our experiments therefore construct portfolios and train selectors separately for each solver. Results on both solvers demonstrate the applicability of LEAP’s framework; they do not establish cross-solver transfer of a portfolio or selector. We are interested in exploring whether semantic correspondences between parameters with similar algorithmic roles can support transferring configuration strategies across solvers in future work.

Across problem families: we construct a portfolio and train a selector for each family; configuration reuse is possible but untested. Under the same solver and a compatible parameter space, previously discovered configurations can be applied to another family or used as initial candidates for subsequent discovery. Their effectiveness and the original selector’s predictive accuracy in that family remain untested. Transfer between such unrelated applications is not a prerequisite for the repeated-solving scenario we target: our evaluated protocol uses budget-limited discovery for each family and spreads offline costs over many subsequent solves.

Comments on the paper

1 comment

Summary of Rebuttal and Additional Evidence

We thank the reviewers for their positive assessment and constructive comments, and the Area Chair for overseeing the review process. We summarize the main additions below.

  1. Benefits of LLM guidance. We added LHS-Init, using ten non-LLM initial candidates, and Staged-Refine, using a numerical block-search rule. Each replaces one LLM component while retaining the other. LEAP achieves lower mean test solving time with GNN selection than both controls across all 14 evaluated solver–dataset settings.

  2. Discovery-budget comparison and cost reporting. We reran LEAP with initialization evaluations, SMAC search, and inter-round evaluations jointly charged to a 10-hour wall-clock discovery budget. LEAP achieves leading mean test solving performance in 12/14 settings, while baselines retain their main-experiment results and search allocations. Appendix A.3 provides stage-by-stage costs across all 12 datasets and both solvers, together with CPU evaluation costs from the supplementary experiments.

  3. Complete results with standard deviations. Appendix E.1 now provides the complete test solving results with standard deviations, covering both baseline comparisons and component ablations. We also added a controlled selector comparison on identical final portfolios: LEAP achieves lower mean test solving time than the validation-selected fixed configuration in 21 of 24 settings, with comparable performance in the remaining three.

Detailed protocols, results, and clarifications are provided in our responses to the individual reviewers. We respectfully ask that this additional evidence be considered in the final assessment. Thank you again for your time and consideration.

Best, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.