acceptodds
Under review as a conference paper at ICLR 2027

From Tuning Trials to Multi-Level Experience: LLM-Guided MIP Solver Configuration

Abstract

Mixed-integer programming (MIP) solvers play a vital role in solving large-scale combinatorial optimization problems, which often originate from important real-world applications. It is well-known that the parameter configuration of the MIP solvers has a significant impact on the solving efficiency, while users often rely on static defaults or a single dataset-wide configuration for all instances. However, different instances often favor substantially different configurations, as real-world MIP instances usually originate from diverse sources and they are highly heterogeneous. To tackle this problem, we propose **H**ierarchical **E**xperience **G**raph for LLM-Guided Parameter **Config**uration (**HEG-Config**), which uses an LLM to generate an instance-specific configuration by exploiting tuning experience accumulated across instances. The key challenges to develop HEG-Config are twofold: (1) LLMs lack specialized tuning experience, and (2) tuning itself must balance generalization and instance-specific adaptation. By introducing the novel **hierarchical experience graph**, which can effectively distill historical tuning trials into multi-level experience from instance-specific rules to general tuning strategies, HEG-Config can well address the aforementioned challenges simultaneously. Then, during inference, by retrieving relevant experience at different levels, HEG-Config enables the LLM to integrate general and instance‑specific tuning experience to generate effective configurations. Experiments demonstrate that HEG-Config achieves better aggregate performance than the baselines, consistently improving solving efficiency over the default configuration and reducing the primal-dual integral within the given time limit by 21%–31%.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Learning to Optimize, Solver Configuration, Mixed-integer programming, Memory-Augmented LLMs

Reviews

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

This paper introduces HEG-Config, a framework using LLMs to generate instance-specific parameter configurations for MIP solvers. Offline tuning trajectories are distilled into feature-conditioned experience cards and organized into a hierarchical experience graph ranging from instance-level rules to general strategies. Then a hierarchy-aware retrieval module provides multi-level context to an LLM to generate customized solver configurations for new instances. Across MIPLIB 2017 and D-MIPLIB benchmarks, HEG-Config reduces the NPDI by 21%–31% relative to solver defaults and outperforms several traditional selection/configuration baselines.

Critical strengths

Tackling instance-specific parameter tuning via LLM-driven generation rather than selecting from a static parameter addresses an important challenge in combinatorial optimization. Across heterogeneous benchmarks and solver backends, the method delivers consistent reductions in NPDI and optimality gap.

Critical weaknesses

The proposed method incurs substantial computational overhead, and the omission of online inference latency undermines the practical value of the reported gains. As shown in Table 14, HEG-Config requires significantly higher runtime. For easy problem instances whose solving times are often below 60 seconds, a 15.6-second inference delay is a heavy penalty that can easily offset the downstream solving speedup. However, the evaluation metrics (Gap, NPDI) disregard inference latency entirely.

The baseline evaluation is incomplete due to the lack of comparisons with existing LLM-based algorithm configuration methods (as mentioned in Appendix A).

Questions for authors

Appendix G.3 (Figure 3) indicates that performance degrades when transitioning from a 30B to a 7B model. Does scaling to even larger models yield further gains, and why did the authors not consider larger reasoning models (e.g., DeepSeek reasoning series)?

4 comments
AuthorsOpenReview ↗

Response to Reviewer GTak

Dear Reviewer GTak,

Thank you for your insightful and valuable comments. We sincerely hope our rebuttal could adequately address your concerns. If so, we would deeply appreciate it if you could consider raising your score. If not, please let us know your further concerns, and we will continue actively responding to your comments.

W1. Inference latency and practical runtime

We agree that configuration latency must be included when assessing practical benefit, especially for short solves. The revised Section 4.4, Table 5 reports each method's inference time and “+Infer.”, the sum of solver runtime and inference time. HEG-Config takes 15.60 s for inference and still has the lowest mean total online runtime among the compared methods in all three benchmark groups:

Benchmark groupDefault SCIPHEG-Config solver timeInferenceHEG-Config totalRelative Reduction
MIPLIB623.61s486.37s15.60s501.97s19.5%
Easy D-MIPLIB133.73s65.79s15.60s81.39s39.1%
Medium D-MIPLIB620.22s433.71s15.60s449.31s27.6%

Relative to default SCIP, these correspond to reductions of approximately 19.5%, 39.1%, and 27.6%, respectively, after adding inference. While we acknowledge that fixed inference latency can outweigh solver-time savings on trivial instances, SC-easy is the only evaluated dataset where this occurs (default SCIP takes 22.36s, whereas HEG-Config takes 13.81 + 15.60 = 29.41s). Across all other evaluated datasets, solver-time savings substantially exceed the inference overhead.

W2. Comparisons with existing LLM-based configuration methods

We first clarify that LLAMBO and SLLMBO follow the dataset-level AC setting: they use LLMs within a search procedure to optimize one configuration for a relatively homogeneous dataset. This differs from our main task of adapting configurations to heterogeneous instances and transferring to unseen problem families; tuning on each target family would change the original evaluation setting.

To accommodate their dataset-level objective, we extend the original homogeneous-mix experiment in Appendix G.2, Table 12. It covers four families—CA-easy, SC-easy, CFLP-medium, and MIRP-medium—each relatively homogeneous and represented in both training and evaluation. This makes a shared configuration within each family a suitable comparison. LLAMBO and SLLMBO are optimized separately for each family, while HEG-Config and the original instance-specific baselines learn from the combined training data of all four families. Their configurations are then evaluated on the same test instances:

Dataset—NPDILLAMBO (per-dataset)SLLMBO (per-dataset)HEG-Config (mixed training)
CA0.830.550.43
SC0.870.330.28
CFLP0.320.300.30
MIRP0.840.950.79
Mean0.710.530.45

HEG-Config outperforms LLAMBO on all four datasets and SLLMBO on three, with a tie on CFLP. Its mean NPDI of 0.45 is approximately 36.6% lower than LLAMBO's 0.71 and 15.1% lower than SLLMBO's 0.53, demonstrating its advantage over both LLM-based configuration methods.

Q. Would a larger or more capable model improve performance?

Yes. In our expanded evaluation in Figure 3, integrating DeepSeek-V4.1-Flash into HEG-Config yields lower NPDI than Qwen-30B across all five datasets—for example, improving from 0.72 to 0.54 on CA-easy and from 0.69 to 0.43 on SC-easy. This confirms that our framework scales effectively with stronger underlying LLMs beyond the Qwen models used in the main experiments.

AuthorsOpenReview ↗

We are looking forward to your feedback.

Dear Reviewer GTak,

We are writing as the authors of the paper titled "From Tuning Trials to Multi-Level Experience: LLM-Guided MIP Solver Configuration" (ID: 20814). We sincerely thank you for your time and efforts during the rebuttal process. We are looking forward to your feedback to understand if our responses have adequately addressed your concerns. If so, we would deeply appreciate it if you could consider raising your score. If not, please let us know your further concerns, and we will continue actively responding to your comments. We sincerely thank you once more for your insightful comments and kind support.

Best,

Authors

Thanks for the rebuttal, I have raised my score to 3.

AuthorsOpenReview ↗

Thanks for your kind support

Dear Reviewer GTak,

Thanks for your kind support and for helping us improve the paper. We sincerely appreciate your valuable suggestions.

Best regards,

Authors

Reviewer_vdLEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This work proposes an LLM-based approach to mixed mixed integer programming (MIP) solver configuration. The approach is a RAG-based hierarchical graph approach, which progressively summarizes, using LLMs, solver configurations based on a Bayesian optimization approach called Tree-structured Parzen Estimator (TPE). It benchmarks against classical instance-specific algorithm configuration approaches for MIP, showing a 20-30% improvement on the primal-dual integral (PDI) on a strong benchmark (MIPLIB 2017).

Critical strengths
  1. The problem studied is valuable.
  2. The evaluation metrics and benchmark selection are solid.
  3. This paper may be the first LLM-based approach to MIP solver configuration. However, see Weakness #1.
Critical weaknesses

Overall, the benchmarking is insufficient to support the design of the proposed method:

  1. There is a rich literature on MIP separator configuration, which designs algorithms for a prominent subset of MIP solver configuration. This literature includes LLM-based and other ML-based approaches. This literature is not acknowledged, benchmarked, or adapted.
  2. The proposed method is not sufficiently justified, especially as the first method of this type. Why this hierarchical design? Why not a simpler approach, like direct prompting.
  3. A qwen3-30b model is used, which is not representative of frontier model capabilities. Direct prompting of frontier models would help gauge the efficacy of the proposed method.
  4. Classical algorithm configuration and LLM for algorithm configuration methods are discussed in the related work section, but it is not sufficiently discussed why they are not benchmarked against the proposed method.
Questions for authors
  1. What is the specific claimed contribution of the paper and is the supporting evidence for it strong?
  2. Address or clarify the weaknesses
5 comments
AuthorsOpenReview ↗

Response to Reviewer vdLE — Part 1

Dear Reviewer vdLE,

Thank you for your insightful and valuable comments. We sincerely hope our rebuttal could adequately address your concerns. If so, we would deeply appreciate it if you could consider raising your score. If not, please let us know your further concerns, and we will continue actively responding to your comments.

W1. Relation to learning-based separator configuration

We thank the reviewer for pointing out learning-based separator configuration, and we have expanded Appendix A to discuss L2Sep [1], DynSep [2], LLM-based cold-start separator configuration [3], and LLM-guided cutting-plane management [4].

The distinction lies in the configuration decisions being optimized. The cited methods target decisions within the cutting-plane component: which separators to activate, how to configure individual separators, or which generated cuts to retain. Some also adapt these decisions during solving. Our task is to jointly configure existing parameters across six solver components before solving: presolving, separating, branching, node selection, primal heuristics, and LP solving. HEG-Config outputs a static configuration for 25 parameters through the solver's parameter interface (Appendix C.1, Table 8). Even within the separating component, our four parameters only set global limits on separation rounds and cut counts, at the root and other nodes. HEG-Config does not choose individual separators or select individual cuts during solving; these decisions remain with the solver's built-in procedures under the supplied settings.

Within this setting, our contribution is to distill historical tuning trajectories into reusable rules and strategies, organize them hierarchically, and combine relevant experience to generate a configuration for each new instance. Accordingly, the baselines in Tables 1 and 12 address configuration across solver components, matching the scope of this task.

[1] Learning to Configure Separators in Branch-and-Cut

[2] Dynamic Configuration for Cutting Plane Separators via Reinforcement Learning on Incremental Graph

[3] LLMs for Cold-Start Cutting Plane Separator Configuration

[4] LLM-guided Cutting-plane Management for Mixed-integer Linear Programming

W2. Why a hierarchy rather than direct prompting or Vanilla RAG?

The hierarchy addresses a specific transfer problem: individual tuning trajectories provide concrete, local evidence, whereas a new instance may require guidance shared across several training instances. Base cards preserve the former, higher-level summaries capture the latter, and the hierarchy integrates both levels of experience.

To evaluate this design against simpler alternatives, we added Direct Prompting and Vanilla RAG (retrieving experience via standard RAG) in Appendix G.7 (Table 18). Across all five datasets, retrieving experience improves upon Direct Prompting, and HEG-Config yields substantial further gains over Vanilla RAG:

Method—NPDIMIPLIBCA-easySC-easyCFLP-mediumMIRP-medium
Direct Prompting1.041.300.880.871.08
Vanilla RAG0.961.100.830.621.05
HEG-Config0.790.720.690.360.82

These findings align with our original ablations: removing either the hierarchy or neighbor expansion degrades NPDI across all five datasets (Table 2), and our granularity analysis further confirms the benefit of combining multi-level experience (Table 6).

W3. Comparison with stronger models

We expanded Appendix G.3, Figure 3 to include DeepSeek-V4.1-Flash alongside Qwen-7B and Qwen-30B. When integrated into HEG-Config, DeepSeek-V4.1-Flash consistently achieves lower NPDI than Qwen-30B across all five datasets—for instance, improving from 0.72 to 0.54 on CA-easy and from 0.69 to 0.43 on SC-easy. This demonstrates that our framework enjoys strong scalability and readily benefits from more capable underlying LLMs.

AuthorsOpenReview ↗

Response to Reviewer vdLE — Part 2

W4. Classical and LLM-based algorithm-configuration baselines

We first clarify that TPE, LLAMBO, and SLLMBO follow the dataset-level AC setting: they search for one shared parameter configuration for a homogeneous dataset. Our main experiment instead learns from heterogeneous MIPLIB instances and evaluates instance-specific configuration on MIPLIB and cross-dataset transfer to D-MIPLIB. Dataset-level tuning on each target family does not directly match this main evaluation protocol, since it uses family-specific tuning data rather than testing transfer from the heterogeneous training set.

To enable a fair comparison on shared test sets, we incorporated these baselines into our homogeneous-mix experiment (Appendix G.2, Table 12), which spans four problem families (CA-easy, SC-easy, CFLP-medium, and MIRP-medium). In this setting, HEG-Config and the instance-specific baselines are trained jointly on the mixed training set, whereas TPE, LLAMBO, and SLLMBO are tuned separately on each individual family to produce a dedicated family-level configuration:

Dataset—NPDIKNN (mixed)DGCAC (mixed)TPE (per-dataset)LLAMBO (per-dataset)SLLMBO (per-dataset)HEG-Config (mixed)
CA0.600.440.360.830.550.43
SC0.350.430.560.870.330.28
CFLP0.300.300.280.320.300.30
MIRP0.791.230.930.840.950.79
Mean0.510.600.530.710.530.45

HEG-Config achieves the lowest mean NPDI across all compared methods (0.45), outperforming TPE (0.53), LLAMBO (0.71), and SLLMBO (0.53). These results demonstrate the benefit of instance-specific configuration even in this setting, where each dataset-level baseline is tuned specifically for the target problem family. The full table also includes all six Hydra variants and MILPTune.

Q1. Specific contribution and supporting evidence

Our core contribution is a hierarchical experience mechanism for instance-specific configuration generation. Specifically, HEG-Config distills historical tuning trajectories into both concrete, instance-level parameter rules and higher-level cross-instance strategies, then retrieves and synthesizes relevant multi-level experience to generate tailored configurations for new instances—enabling parameter combinations beyond those directly observed during tuning.

This contribution is supported by two lines of empirical evidence:

  • Performance advantage over baselines (Tables 1, 12, and 18): Main benchmark evaluations and expanded comparisons against instance-specific, dataset-level, and LLM/RAG baselines demonstrate consistent NPDI reductions across heterogeneous and homogeneous settings.
  • Mechanism effectiveness and reliability (Tables 2, 6, and 20): Component ablations, hierarchy granularity analyses, and repeated-generation audits confirm the necessity of multi-level experience integration as well as the stability and validity of the synthesized configurations.

Thank you for the detailed responses. I have some follow-up questions and comments:

Can the authors clarify whether any direct prompting / vanilla RAG baselines are conducted with frontier LLMs? I understand if token costs limit the evaluation of the proposed method with frontier models, but it would be interesting to benchmark how frontier LLMs perform 'out-of-the-box' (or close to it) on the task in question, to understand where the contributed work is relative to the state of the art.

Thank you for clarifying that separators are not configured by the proposed approach. I still suggest adding a benchmark comparison to the separator configuration line of work. This seems like a notable omission for work on 'MIP solver configuration' given the well-documented effectiveness of separator configuration for MIP. Adding a comparison would allow the reader to better interpret the results, such as whether superior results can be achieved even without separator configuration or that it would be an important type of configuration to incorporate in the future.

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

The paper proposes HEG-Config, an LLM-based system for instance-specific SCIP parameter configuration. Offline: for each of 125 training MIPLIB instances, TPE produces a 50-trial tuning trajectory, which an LLM distills into "experience cards" containing strict feature-conditioned quantitative rules ϕ(f)→θ\phi(f)\rightarrow\theta. The cards form the base level of a hierarchical experience graph (built on LightRAG) whose higher-level nodes are LLM-generated summaries of related lower-level nodes (aggregation threshold τ=5\tau=5; four levels). Online: relevant base cards are retrieved via RAG for a new instance, expanded upward through the hierarchy with Top-Ψ(k)\Psi(k) neighbor expansion, and the LLM (Qwen3-30B-A3B) generates a 25-parameter configuration from the multi-level context. On 50 held-out MIPLIB instances and D-MIPLIB cross-dataset splits, evaluated by dataset-level normalized primal-dual integral (NPDI), HEG-Config achieves the best aggregate performance (0.79 on MIPLIB, ~21% better than the default; 0.69/0.77 mean on easy/medium D-MIPLIB, ~31%/23%), against Hydra (six variants), MILPTune, and KNN. Ablations support each component: removing the hierarchy, neighbor expansion, or strict rule-format prompts degrades performance; single-level retrieval fails; and gains persist with smaller LLMs, different time budgets, and Gurobi.

Critical strengths
  1. A genuine conceptual step over configuration selection. The paper correctly identifies that selection-based methods (Hydra, MILPTune, KNN) are confined to configurations explored during tuning, and demonstrates a recombination mechanism — parameter-level rules retrieved at multiple granularities and synthesized by an LLM — that beats selection on a heterogeneous benchmark (MIPLIB 2017) and transfers zero-shot to D-MIPLIB. The ablations (levels, τ\tau, prompt type, training-set scaling, model size, solver, budget) are unusually thorough, and each supports the design.

  2. Robust, realistic evaluation target. Using NPDI on heterogeneous MIPLIB rather than homogeneous mixtures (plus a homogeneous-mix control) tests the regime where per-instance configuration actually matters, and HEG-Config stays below 1.0 — i.e., better than the default — on essentially every dataset, where several baselines frequently exceed 1.0.

Critical weaknesses
  1. Statistical rigor is missing. All headline numbers (21–31% NPDI improvement) come from single runs with stochastic LLM generation and stochastic solver behavior; no seeds, confidence intervals, or paired significance tests are reported. Margins over the second-best method (0.79 vs. 0.91 for MILPTune on MIPLIB; 0.69 vs. 0.85 for Hydra-KNN on easy D-MIPLIB) could plausibly be within noise given 50 test instances and a ratio-of-sums aggregate. The relative-gap metric, with a 10,000 penalty per infeasible instance, makes ablation tables swing on one- or two-instance events (e.g., MIPLIB gap 16.60 vs. 3964.85).

  2. Baseline and budget comparability gaps. All baselines are selection methods from 2018–2023; there is no comparison against spending part of the same offline budget on online per-instance tuning at test time (e.g., short TPE/SMAC runs on the test instance), nor against recent learned configurators. Offline costs (125×50 TPE trials ≈ thousands of CPU-hours, LLM calls, graph maintenance) are not reported comparatively, so it is unclear how many test-instance solves amortize the offline investment.

Questions for authors
  1. (Assessment) Please provide multi-seed results (repeated LLM generation and repeated solves) with paired per-instance comparisons and significance tests for NPDI. How much of the aggregate gain is driven by the worst few instances?
  2. (Clarification) How were distilled cards validated against the underlying tuning trajectories (rule factuality / hallucination rate), and what fraction of generated configurations are invalid or partially ignored by SCIP?
  3. (Assessment) How does HEG-Config compare against spending, e.g., 200–500 s of online TPE on each test instance at solve time, given your offline budget? This comparison would most affect my assessment.
  4. (Clarification) What is the marginal cost and marginal gain of each added training instance in Fig. 2 (tuning time + LLM calls + graph construction vs. NPDI)?
  5. (Clarification) Why 25 parameters, and does the strict-rule approach remain reliable as the parameter set scales toward SCIP's full space with conditional dependencies?
5 comments
AuthorsOpenReview ↗

Response to Reviewer kwtB — Part 1

Dear Reviewer kwtB,

Thank you for your insightful and valuable comments. We sincerely hope our rebuttal could adequately address your concerns. If so, we would deeply appreciate it if you could consider raising your score. If not, please let us know your further concerns, and we will continue actively responding to your comments.

W1 & Q1. Statistical reliability and the influence of difficult instances

We thank the reviewer for emphasizing the importance of statistical reliability. To account for solver stochasticity and generation variance, we have added repeated evaluations, paired significance tests, and an analysis of hard-instance influence in Appendix H.

Repeated solver runs (Appendix H.1, Table 19). We evaluate each method's original configurations over three solver runs. Across all three benchmark groups, HEG-Config consistently achieves the lowest aggregate NPDI while maintaining low variance across runs:

Method—NPDIMIPLIBEasy D-MIPLIBMedium D-MIPLIB
Hydra (best per group)1.07±0.011.03±0.391.05±0.12
MILPTune0.97±0.040.82±0.301.10±0.36
KNN0.91±0.020.92±0.111.04±0.24
DGCAC1.06±0.110.98±0.301.03±0.20
HEG-Config0.79±0.010.69±0.150.77±0.19

Repeated configuration generation (Appendix H.1, Table 20). We evaluate three independently generated configuration sets with three solver runs each for HEG-Config and its ablated variants. Overall, HEG-Config achieves the lowest total Gap and NPDI across all five datasets with stable performance across generations. Additionally, we observe that on MIPLIB, the three generations of w/o Hierarchical yield total Gaps of 3938.16 ± 19.05, 120.17 ± 0.55, and 13507.63 ± 4.48 (where one instance fails to find a feasible solution, while the remaining instances still contribute an average total Gap of 3507.63). In contrast, HEG-Config and w/o Expand remain substantially more stable, with mean total Gaps tightly bounded within 16.78–18.30 and 33.35–55.51, respectively. This stark contrast underscores that the hierarchical experience structure is critical for stabilizing configuration generation.

Method—NPDIMIPLIBCA-easySC-easyCFLP-mediumMIRP-medium
HEG-Config0.80±0.010.69±0.050.70±0.050.35±0.030.83±0.01
w/o Hierarchical0.94±0.020.90±0.050.85±0.030.73±0.031.05±0.01
w/o Expand0.89±0.010.89±0.060.85±0.030.58±0.081.02±0.02
Method—GapMIPLIBCA-easySC-easyCFLP-mediumMIRP-medium
HEG-Config17.63±0.650.00±0.000.00±0.000.00±0.0018.70±1.25
w/o Hierarchical5855.32±5631.030.00±0.000.00±0.000.01±0.0010021.08±1.35
w/o Expand41.69±9.850.00±0.000.00±0.000.01±0.0010018.51±0.84

Paired per-instance comparisons (Appendix H.2, Table 22). We perform paired PDI tests with Holm–Bonferroni correction across 15 comparisons. As shown below, HEG-Config significantly outperforms default SCIP across all five datasets after correction, confirming consistent, statistically robust improvements at the individual instance level.

DatasetWin/Tie/LossΔNPDIPointwise 95% CIAdjusted p
MIPLIB41/0/9+0.204[+0.055,+0.354]0.025
CA-easy96/0/4+0.311[+0.253,+0.369]<0.001
SC-easy99/0/1+0.305[+0.230,+0.379]<0.001
CFLP-medium100/0/0+0.647[+0.612,+0.683]<0.001
MIRP-medium19/0/1+0.175[+0.120,+0.230]<0.001

Influence of the hardest instances (Appendix H.4 & H.3). To verify that gains are not driven by a few extreme outliers, we conduct two analyses:

  1. Instance removal (Figure 4). Removing up to the top five hardest instances (ranked by default PDI or Gap) still leaves HEG-Config with a mean NPDI below 1 and better than both ablations across all five datasets.
  2. Top-instance contribution (Table 23). We measure the share of total PDI reduction contributed by the top-k instances ranked by default PDI. Specifically, the top 5 instances account for 59.0% and 39.8% of the reduction on MIPLIB and MIRP-medium (reflecting their smaller test sets of 50 and 20 instances, where top 5 represent 10–25% of the total). On the other three datasets, improvements predominantly come from instances outside the top 5 (reaching only 1.0% to 14.2% for top 5), confirming that HEG-Config's gains are broadly distributed.
AuthorsOpenReview ↗

Response to Reviewer kwtB — Part 2

W2 & Q3. Baselines, online TPE, and computational cost

We thank the reviewer for the constructive comments regarding baseline comparisons and computational overhead. Below, we address online TPE, additional instance-specific baselines, and offline construction costs in detail.

Online TPE. Our setting targets unseen instances solved once, where no repeated solves exist to amortize online search costs. Under our 1,000-second evaluation horizon, a single solver run can already take hundreds of seconds, making multi-trial online tuning impractical within a 200–500-second budget. To still provide a strong TPE comparison, we added dataset-level TPE (Appendix G.2, Table 12), tuning one reusable configuration per dataset. HEG-Config achieves a lower mean NPDI than dataset-level TPE (0.45 vs. 0.53), demonstrating the clear benefit of instance-specific configuration over a static dataset-wide setting.

Instance-specific baselines (Table 1). To further strengthen our comparison against instance-specific configurators, we incorporated DGCAC, a state-of-the-art method for instance-level parameter tuning. HEG-Config consistently outperforms DGCAC in aggregate NPDI across all three benchmark groups: 0.79 vs. 0.90 on MIPLIB, 0.69 vs. 1.00 on easy D-MIPLIB, and 0.77 vs. 0.96 on medium D-MIPLIB. While we aim to cover representative baselines for this single-shot instance-specific setting, we welcome any suggestions from the reviewer on additional methods for further discussion in the final version.

Offline costs (Appendix G.6, Tables 15–17). Substantial offline computation is a shared requirement across learning-based configurators rather than unique to HEG-Config. Specifically, KNN, MILPTune, and HEG-Config share the exact same 24.11h data-collection phase (50 TPE trials per training instance)—with HEG-Config distilling the full tuning trajectories rather than retaining only the single best configuration—while DGCAC requires 26.21h for joint data collection and training. For HEG-Config's additional 6.28h experience-construction phase (Table 16), 6.18h is spent on LightRAG indexing, whereas hierarchical construction itself takes only 372s. Notably, replacing LightRAG with a lightweight naive RAG backend (HEG-Config-NR, Table 17) reduces construction time to 1,435.1s (≈24 min\approx 24\text{ min}) while keeping the five-dataset mean NPDI nearly unchanged (0.70 vs. 0.68).

From an amortization perspective, given HEG-Config's mean online time saving of 121.64s per instance on MIPLIB (623.61s →\rightarrow 501.97s, including inference; Table 5), the combined 30.39h setup time is recovered after approximately 900 sequential solves under our experimental setup (and substantially fewer if historical tuning logs are already available).

MethodData collectionTraining/experience construction
Hydra, six variants6.46h7–27s
MILPTune24.11h163s
KNN24.11h0s
DGCACIncluded in training26.21h
HEG-Config24.11h6.28h
HEG-Config-NR24.11h1,435.1s

Q2. Factuality of experience cards and validity of generated configurations

We thank the reviewer for raising this important point regarding rule factuality and configuration safety. We address both aspects below:

Factuality of experience cards (Table 3). Experience cards are LLM-distilled summaries conditioned on instance features and measured tuning trajectories under a structured template (Appendix E). Because rules extrapolated from finite trials cannot be formally proven universally valid—and we do not separately measure a rule-level hallucination rate—we treat these cards as empirical configuration guidance rather than formal guarantees. Consequently, their reliability is validated end-to-end through held-out solver performance and ablations: as shown in Table 3, structured rule-based distillation consistently achieves lower NPDI than free-form summaries across all five datasets, confirming the practical utility of this representation.

Validity of generated configurations (Appendix I, Table 24). To ensure that malformed or unsupported settings never reach the solver, all generated outputs undergo strict programmatic whitelist and value-domain verification before execution. In a comprehensive audit of 6,620 generation records across our main experiments and analyses, only 10 records (≈\approx 0.15%) contained a violation (5 invalid values for branching/scorefunc and 5 parameters outside the whitelist). Our validation filter automatically discards any invalid entries prior to solver invocation, guaranteeing that 100% of applied configurations are syntactically and semantically valid for SCIP.

AuthorsOpenReview ↗

Response to Reviewer kwtB — Part 3

Q4. Marginal cost and gain from additional training instances

We thank the reviewer for asking about the scaling trade-off between training-set size and performance gains.

Cost per training instance. Each training instance requires 50 solver evaluations followed by experience construction. Under our 50-process parallel setup, the measured data-collection time averages 24.11 h/125≈694 s24.11\text{ h} / 125 \approx 694\text{ s} per instance, while LLM-based experience construction and retrieval indexing add 6.28 h/125≈181 s6.28\text{ h} / 125 \approx 181\text{ s} (including approximately 3s for hierarchical construction; Tables 15–16). This yields a combined amortized cost of approximately 875s (14.6 minutes) per training instance.

Performance gain per added instance. Based on the results in Figure 2, expanding the training set from 20 to 125 instances reduces NPDI across all five datasets—by 0.21 (MIPLIB), 0.48 (CA-easy), 0.05 (SC-easy), 0.44 (CFLP-medium), and 0.26 (MIRP-medium). Across these five datasets, the arithmetic mean NPDI decreases from 0.964 to 0.676. Averaged over the 105 added instances, each additional training instance contributes a mean NPDI reduction of 0.00274, with consistent gains observed across all five datasets.

Q5. Why 25 parameters, and what about the full conditional space?

We thank the reviewer for raising this question regarding parameter subset selection and scalability to the full conditional space.

Rationale and generalizability of the 25-parameter subset (Appendices C.1–C.2 & G.5). The 25 SCIP parameters form a tractable subset selected to span six core solver components: presolving, separating, branching, node selection, primal heuristics, and LP solving (Table 8). Importantly, 25 is not an intrinsic requirement of HEG-Config. As demonstrated by our Gurobi experiments—which use a distinct set of 15 parameters without exhaustive manual tuning—HEG-Config consistently achieves an NPDI below 1 across all five datasets (Tables 9 and 14), supporting its applicability beyond SCIP.

Scaling to the full conditional parameter space. Extending directly to SCIP's entire parameter space (comprising thousands of parameters) remains an open challenge that we have not yet validated. At that scale, prompt-based experience representation would face severe context-length and search-coverage bottlenecks, while complex conditional dependencies between parameters would require structured validation beyond individual value-domain checks. Addressing the full space would likely require automated parameter-group screening, explicit dependency graph representation, and dependency-aware post-generation validation. Accordingly, our current empirical claims focus on curated, multi-component subspaces rather than unrestricted full-space configuration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear Area Chair,

We sincerely thank you for your careful work during the review and rebuttal process. We are also grateful to all reviewers for their constructive and insightful comments. While some reviewers assigned relatively low scores, in our assessment, no fatal flaws or fundamental limitations were identified. Most concerns focused on the completeness of baseline comparisons, statistical reliability, and computational costs. We have provided detailed responses to all questions and implemented all feasible suggestions. At this stage, we do not see any unresolved technical issues remaining.

For your convenience, we summarize below how we have addressed the main concerns raised by each reviewer.

Response to Reviewer kwtB (Rating: 3)

Reviewer kwtB appreciated the advance beyond configuration selection and the heterogeneous evaluation, while raising concerns about statistical reliability, baseline comparability, card and configuration reliability, computational costs, and scalability. We have addressed these points as follows:

  • Statistical reliability (W1, Q1). We added repeated solver and generation evaluations, paired tests with multiple-testing correction, and hard-instance analyses (Appendix H), confirming stable gains that persist after removing the hardest instances.
  • Baselines and costs (W2, Q3). We added comparisons with instance-specific DGCAC and dataset-level TPE, explained online TPE's limitations within the proposed 200–500s test-time budget, and reported offline costs and amortization (Appendix G.6).
  • Card reliability and configuration validity (Q2). We clarified how solver evaluations and ablations support card effectiveness (Table 3), and audited generation records, confirming that invalid parameter entries are rare and filtered before execution (Appendix I).
  • Training-set size (Q4). We reported amortized offline costs per training instance and used Figure 2 to quantify performance gains from additional training data, clarifying the trade-off between offline cost and solver performance.
  • Parameter space (Q5). We explained the parameter subset selection, cited the original Gurobi experiments to support applicability across solvers and parameter sets, and discussed extensions to the full conditional parameter space.

Response to Reviewer vdLE (Rating: 1)

Reviewer vdLE recognized the value of the problem and benchmark setup, but requested further justification of the hierarchical design and broader comparisons, including stronger LLMs and separator configuration methods. We have addressed these points as follows:

  • Separator configuration (W1 & follow-up). We clarified the configuration scope (Appendix A) and added Lawless et al.'s LLM-based separator baseline, showing that HEG-Config achieves lower NPDI across all compared datasets.
  • Hierarchical design (W2). We clarified the role of multi-level experience and added Direct Prompting and Vanilla RAG baselines (Appendix G.7). HEG-Config outperforms both across all five datasets, complementing the original hierarchy ablations.
  • Stronger LLM baselines (W3 & follow-up). We added Direct Prompting and Vanilla RAG evaluations with DeepSeek-V4.1-Flash (follow-up response), showing that HEG-Config consistently outperforms both using the same backbone.
  • Classical and LLM-based AC (W4, Q1). We clarified dataset-level versus instance-specific evaluation protocols and added TPE, LLAMBO, and SLLMBO comparisons in the homogeneous-mix setting (Appendix G.2), where HEG-Config achieves the lowest mean NPDI.

Response to Reviewer GTak (Rating: 2 →\to 3)

Reviewer GTak appreciated LLM-driven instance-specific configuration and the consistent gains across benchmarks and solvers, while raising concerns about inference overhead, missing LLM-based baselines, and scaling to stronger models. We have addressed these points as follows:

  • Inference latency (W1). We added a runtime comparison including inference (Table 5), showing lower mean total runtime than default SCIP across all benchmark groups. SC-easy is the only evaluated dataset without a net runtime reduction.
  • LLM-based baselines (W2). We added LLAMBO and SLLMBO comparisons in the homogeneous-mix setting (Appendix G.2), showing that HEG-Config achieves lower mean NPDI than both baselines.
  • Stronger models (Q). We expanded Figure 3 to include DeepSeek-V4.1-Flash, which yields lower NPDI than Qwen-30B within HEG-Config across all five datasets, demonstrating further gains from a stronger backbone.

We sincerely hope this summary is helpful and reduces your workload in assessing the rebuttals. Thank you again for your careful work during the review and rebuttal process.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.