acceptodds
Under review as a conference paper at ICLR 2027

Towards LLM-Based Algorithm Design for Heterogeneous Datasets: Decoupled Configuration Policy and Parameterized Algorithm

Abstract

Automated Algorithm Design (AAD) based on Large Language Models (LLMs) has shown promising results, but existing methods often struggle to generalize across heterogeneous problem instances. Most of these methods either learn a single static algorithm to handle all cases or generate a finite set of algorithms to cover different types of instances. While these strategies can improve performance on seen data, they remain limited to a finite set of learned behaviors and may fail when encountering unseen or mixed cases that do not clearly belong to any known type. In this work, we aim to overcome this limitation by moving beyond selecting from a finite set of algorithms. We propose the decoupled configuration-policy and parameterized-algorithm (DCPPA) structure, in which the algorithm encodes behavioral differences as parameters, and the policy maps observable instance or state features to these parameters. DCPPA represents algorithmic behavior in a continuous, parameterized space, and enables the algorithm to capture gradual variations among diverse instances and smoothly adjust its behavior for cases not well covered by the training data. To construct DCPPA-structured algorithms, we develop TS-DCPPA, a two-stage pipeline. The first stage discovers complementary source algorithms and fuses them into a unified structure. With this structure fixed, the second stage employs a warm-start mechanism and learns a continuous parameter-tuning policy. Experiments on real-world problem instances demonstrate that TS-DCPPA achieves greater robustness and cross-distribution generalization than existing methods.

Primary area:
optimization
Keywords:
Automated algorithm design, large language model

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

TS-DCPPA addresses combinatorial optimization heuristic design for heterogeneous instances through a two-stage pipeline. It first discovers complementary algorithms and fuses them into a parameterized algorithm, then fixes that algorithm and learns a state-dependent configuration program. Deployment requires no LLM calls. Experiments cover TSP, OVMP, and FJSP.

Critical strengths

The two-stage design—fusing complementary algorithms before learning a state-dependent parameter policy—is sensible and requires no LLM calls at deployment. TS-DCPPA performs best across the reported ID and OOD settings, with especially strong TSP OOD results. The study includes ablations, repeated runs, paired tests, and a partial replication with another LLM backend.

Critical weaknesses
  1. The “no fusion” ablation changes both the algorithm representation and the policy space, so it does not isolate dynamic parameter generation. A cleaner comparison would fix the fused algorithm and compare optimized fixed parameters, dynamic source-preset selection, and free parameter generation. The uneven warm-start benefit and the shared TSP scoring terms also motivate a simple coefficient-based parameterization. Without these controls, the gains cannot be attributed specifically to continuous adaptation.

  2. The experiments omit relevant configuration and adaptive baselines. TIDE would test whether fixed-structure tuning recovers similar gains, while DASH and DyCA are closer adaptive alternatives. Comparisons under compatible interfaces and budgets, or a concrete explanation of infeasibility, are needed. The OVMP policy relies heavily on discrete threshold branches before continuous modulation, so it does not isolate whether smooth adaptation drives the OOD gain.

  3. The compute comparison is not fully controlled: candidate timeouts differ, and the reported evaluation budget excludes fusion, review, and revision calls. The paper should report total LLM calls, wall-clock search cost, final-algorithm runtime, and run-to-run variation. Reproducibility is also limited because the OVMP OOD evaluation relies entirely on the non-public BD dataset. The authors should report how often fusion reached the soft limit and was accepted without structural approval.

Questions for authors
  1. With the fused algorithm and compute fixed, how do optimized fixed parameters, dynamic preset selection, and dynamic parameter generation compare?

  2. Under comparable interfaces and budgets, how does TS-DCPPA compare with DASH and DyCA? On TSP, does automatic fusion outperform a simple coefficient parameterization of the same source algorithms?

  3. What are the total search cost, final-algorithm runtime, and run-to-run variation? How often was fusion accepted without structural approval, and can the OVMP OOD result be verified on public data?

6 comments

Thank you for your detailed and constructive review. We appreciate your positive assessment of our sensible two-stage design, TS-DCPPA's strong ID and OOD performance, and our thorough evaluation across ablations, repeated runs, paired statistical tests, and a second LLM backend. Below, we respond to your comments and questions point by point. We have revised the manuscript accordingly and highlighted the changes in blue.

Warm start and Stage-2 policy controls (W1/Q1)

Thank you for these constructive comments. We agree that discussing how Stage-2 policy search behaves under different warm-start schedules and parameter spaces is valuable.

Discussion on warm start. Warm start spends an initial share of the Stage-2 budget learning to switch among the source presets before free parameter generation. Varying this share on the same fused algorithm ϕ\phi shows that warm start is beneficial across all three tasks, while the best fixed ratio is task-dependent (test Gap %, five-run mean, lower is better; ID = training distributions, OOD = unseen distributions):

Stage-2 warm startTSP IDTSP OODOVMP IDOVMP OODFJSP IDFJSP OOD
w/o warm start18.5720.721.471.7811.499.18
20% (original)18.4920.991.031.6511.078.89
40%18.3420.671.051.4711.6110.09
60%18.4921.200.991.3511.119.62
80%18.4521.141.051.4411.519.49

Relative to w/o warm start, the best fixed ratio lowers the gap in every column (40% on TSP, 60% on OVMP, and 20% on FJSP), and all four nonzero ratios lower OVMP's ID gap by 28–33%. However, the sensitivity to the fixed ratio varies across tasks. This observation motivates an adaptive extension, TS-DCPPA (adaptive), which replaces the fixed ratio with Thompson sampling to choose between preset-guided and free parameter generation based on their recent search gains.

Discussion on fixed parameters vs. source presets vs. non-source parameters. Next, we examine what happens on the same ϕ\phi and Stage-2 budget if the policy is restricted to a single learned parameter vector (fixed parameters) or to the discrete source presets (per-instance preset or per-step preset), versus generating non-source parameters freely under TS-DCPPA (original) and TS-DCPPA (adaptive):

Policy on the same ϕ\phiTSP IDTSP OODOVMP IDOVMP OODFJSP IDFJSP OOD
Fixed parameters18.3020.971.221.4111.469.47
Per-instance preset18.5921.801.391.6811.209.61
Per-step preset18.7121.540.981.4211.319.57
TS-DCPPA (original)18.4920.991.031.6511.078.89
TS-DCPPA (adaptive)18.2220.070.961.3411.179.15

Generating non-source parameters at each step achieves the lowest gap in all six columns, and TS-DCPPA (adaptive) alone beats all three restricted controls in every column. This confirms that dynamic non-source parameter generation brings gains beyond both a static ϕ\phi (vs. fixed parameters) and discrete preset switching (vs. per-instance and per-step presets), and shows that better Stage-2 search allocation can further unlock the benefit of non-source parameters, which is a promising direction for future study. We have added these discussions on Stage-2 policy search to the appendix.

Baselines and parameterization (W2/Q2)

We evaluate TIDE, DASH, and DyCA under the same budget of 1,000 candidate evaluations. Like EoH-S, DyCA discovers a set of complementary algorithms in training without a test-time selector, so it naturally serves as an alternative Stage-1 searcher in TS-DCPPA (500 evaluations in Stage 1 + 500 in Stage 2). The table reports five-run mean test Gap (%; lower is better):

MethodTSP IDTSP OODOVMP IDOVMP OODFJSP IDFJSP OOD
TIDE23.0130.461.792.4111.7711.55
DASH24.5926.393.154.6313.4511.04
TS-DCPPA (DyCA in Stage 1)19.8920.711.591.5211.6010.84
TS-DCPPA (EoH-S in Stage 1)18.4920.991.031.6511.078.89

Both TS-DCPPA variants beat TIDE and DASH in all six columns. Moreover, plugging DyCA into Stage 1 remains superior to all baselines and even achieves the lowest OOD gaps on TSP (20.71) and OVMP (1.52), confirming that TS-DCPPA's gains stem from the two-stage framework itself rather than a specific Stage-1 method.

Automatic fusion vs. simple coefficient parameterization. We use automatic LLM fusion to construct ϕ\phi rather than prescribing a fixed coefficient template for three reasons:

  • No manual term alignment. Instantiating a coefficient template for the TSP source programs requires manual identification and alignment of their scoring terms. LLM fusion automates this process directly from the source code.
  • Beyond simple parameterization. LLM fusion identifies shared computations across the source programs and exposes differences among them as parameters. For example, regret_sharpness in the fused TSP algorithm controls the distance threshold that determines which candidate nodes receive a regret adjustment, rather than simply adjusting the overall weight of the regret term.
  • Applicability across tasks. Extending a coefficient template to source programs with different control flows or composite decisions, such as (operation, machine) assignments in FJSP, requires additional task-specific design. In contrast, LLM fusion constructs a task-specific parameterization directly from the source programs without requiring a manually designed template.

Threshold branches on OVMP. DCPPA favors continuous parameters for ϕ\phi but also allows discrete or mixed parameters (such as integer window sizes and feature toggles on OVMP), and it does not restrict how the programmatic policy π\pi maps states to parameters. Using threshold branches on state variables to identify workload regimes before continuously modulating parameters within each regime is a natural way for π\pi to combine regime switching with fine-grained adjustment. Moreover, the comparison on the same ϕ\phi (W1/Q1) shows that generating non-source parameters freely beats both fixed parameters and discrete preset selection.

Cost, variation, and reproducibility (W3/Q3)

Thank you for these concrete reporting suggestions.

(1) Search cost. TS-DCPPA matches the 1,000 candidate evaluations of all Table-1 baselines, and each candidate evaluation uses one LLM call. The wall-clock time is shown as follows.

Method / stageCandidate evaluationsTSP (min)OVMP (min)FJSP (min)
EoH1,00032.481.677.1
ReEvo1,00095.099.5156.0
PartEvo1,00065.669.735.6
HiFo1,00034.259.525.6
RefineEvo1,000157.6176.3140.4
InstSpecHH1,00019.922.517.4
TS-DCPPA1,00053.8102.353.1
TS-DCPPA: Stage 150019.055.324.9
TS-DCPPA: Fusion-2.010.94.6
TS-DCPPA: Stage 250032.736.023.6

(2) Fusion acceptance. The following table summarizes five fusion results per task. Note that acceptance without structural approval is allowed after the 20-round soft limit only if the behavioral checks pass.

TaskMean LLM callsMean roundsBehavioral checks passedAccepted without structural approval
TSP11.46.25/55/50/50/5
OVMP17.89.05/55/51/51/5
FJSP27.612.05/55/51/51/5

(3) Run-to-run variation. Below is Table 1 with the SD over five runs.

MethodTSP IDTSP OODOVMP IDOVMP OODFJSP IDFJSP OOD
Seed heuristic26.3328.132.303.8425.4127.49
EoH19.55 ± 0.4221.44 ± 0.381.67 ± 0.402.75 ± 0.5512.77 ± 1.1112.88 ± 4.14
ReEvo22.42 ± 4.2624.56 ± 4.502.55 ± 1.373.57 ± 0.4713.02 ± 2.1111.69 ± 2.26
PartEvo20.56 ± 1.0221.50 ± 1.631.94 ± 0.552.67 ± 0.8714.10 ± 2.0014.63 ± 2.59
HiFo20.12 ± 0.4321.27 ± 0.521.48 ± 0.712.73 ± 1.2011.71 ± 1.9211.27 ± 2.00
RefineEvo20.82 ± 2.8770.10 ± 108.792.08 ± 0.822.31 ± 0.6111.71 ± 1.1911.54 ± 4.17
InstSpecHH26.44 ± 2.8130.20 ± 3.953.80 ± 1.604.78 ± 1.9516.77 ± 2.5913.83 ± 3.40
TS-DCPPA (Ours)18.49 ± 0.4820.99 ± 0.661.03 ± 0.251.65 ± 0.1611.07 ± 0.418.89 ± 0.32

(4) Candidate time limits and final-algorithm runtime. Each TS-DCPPA candidate is evaluated on all 20 training instances under a time limit. Because the time each stage needs is hard to know before the search, we leave a margin in each stage's limit. These limits are safety caps against non-terminating or very slow candidates, not actual runtimes.

To verify that the final TS-DCPPA programs finish within the baseline time limits, we evaluate the final programs on all 20 training instances. The table below reports the maximum observed runtime (seconds; InstSpecHH excludes LLM latency for algorithm selection), showing that TS-DCPPA finishes well within the baseline limits:

MethodTSP (s)OVMP (s)FJSP (s)
Baseline time limit306030
EoH4.48118.8331.415
ReEvo28.22215.7760.456
PartEvo5.77840.5071.435
HiFo5.68419.2060.658
RefineEvo12.64934.4110.393
InstSpecHH21.36825.96419.578
TS-DCPPA8.81819.3930.244

(5) Public-data validation. We additionally evaluate TS-DCPPA and all six Table-1 baselines on public CERIT-SC traces, an OVMP source unseen during training. Using the five runs trained on Huawei and Azure without retraining, we test 60 instances (30 each at 50 and 80 PMs, matching our Table-1 OOD split) and normalize by first fit to avoid costly exact offline optimization over thousands of placement decisions (mean Gap %; lower is better, negative means more accepted requests than first fit):

CERIT-SC (OOD)EoHReEvoPartEvoHiFoRefineEvoInstSpecHHTS-DCPPA
Gap (%)-2.18-0.51-1.64-2.29-2.490.18-3.01

TS-DCPPA achieves the lowest mean Gap, confirming its OOD advantage on public data. Last but not least, we will release the code upon acceptance.

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

This paper addresses the insufficient generalization capability of LLM-driven Automated Algorithm Design (LLM-based AAD) on heterogeneous problem instances by proposing the DCPPA framework, which decouples the algorithm into a parameterized algorithm ϕ\phi and a configuration policy π\pi. The configuration policy dynamically generates parameters based on the current instance or solving state, thereby adjusting the decision behavior of the underlying algorithm. The authors further design a two-stage implementation method, TS-DCPPA: the first stage uses EoH-S to search for complementary algorithms and fuses them into a unified parameterized algorithm via an LLM; the second stage fixes this algorithm structure and uses EoH to search for a configuration policy, first learning to select among existing parameter presets and then shifting to free parameter generation. Overall, this paper has clear practical motivation and fairly sufficient experimental results. Its main contribution lies in using an LLM to automatically construct a configurable algorithm space and then learning an instance-dependent configuration policy, rather than proposing an entirely new principle of parameter adaptation or algorithm configuration. My main current concern is how to determine whether the performance improvement actually stems from the continuous instance-adaptive configuration, or mainly from the stronger heuristic algorithm constructed through the fusion of complementary algorithms.

Critical strengths
  1. The proposed DCPPA framework, through Equation (1), represents algorithmic behavior as the composition of a configuration policy and a parameterized decision rule, allowing weights, thresholds, and other parameters to be adjusted within the same algorithm structure. This design offers a larger behavioral expressive space than relying solely on static heuristics, and it also provides a concrete approach for turning LLM-generated programs into configurable algorithm families.

  2. The two-stage design has clear engineering motivation. It reduces the complexity of simultaneously searching for the algorithm structure and the configuration policy, while also avoiding the need to modify the underlying algorithm code every time the policy changes. In addition, the authors introduce behavioral consistency validation, requiring the fused algorithm to approximately reproduce the original algorithm's behavior under the source presets. This is more reliable than relying entirely on the LLM's subjective judgment of fusion quality.

Critical weaknesses
  1. Does continuous parameter adaptation genuinely yield benefits beyond discrete algorithm selection? The authors treat the continuous parameter space as the primary advantage of DCPPA over a finite algorithm portfolio, and argue that this representation can handle intermediate states under unknown distributions. However, the current experiments do not sufficiently isolate the contribution of this factor. The paper provides a "w/o fusion" ablation, but this experiment simultaneously alters both the underlying algorithm structure and the configuration space, and thus cannot directly demonstrate that continuous parameter generation is superior to finite preset selection.

  2. The behavioral fidelity of the fused algorithm is insufficient to establish that the continuous parameter space has effective generalization capability. Even if the fused algorithm can accurately reproduce the original algorithms under each source preset, this does not guarantee that it exhibits smooth, valid behavioral variation between the presets, because the final decisions are typically discrete, and continuously varying parameters may still produce abrupt action sequences.

Questions for authors
  1. The proposed configuration-policy formulation is closely related to instance-specific algorithm configuration and learned parameter prediction. Could the authors clarify the methodological novelty beyond automatically constructing the parameterized algorithm and compare against an existing instance-dependent configurator applied to the same fused algorithm?

  2. The fusion procedure validates the recovery of source algorithms at predefined parameter settings, but it remains unclear whether intermediate configurations yield meaningful or competitive behaviors. Could the authors analyze solution quality along parameter interpolation paths and report how frequently the learned policy generates configurations outside the source presets?

  3. The warm-start mechanism provides inconsistent improvements, and removing it slightly improves TSP-OOD performance. Could the authors analyze the sensitivity to the warm-start budget fraction and explain under which conditions this phase is beneficial?

2 comments

Thank you for your careful review and constructive questions. We appreciate your recognition of our clear practical motivation and experimental results, the expressiveness of DCPPA in turning LLM-generated programs into configurable algorithm families, and our well-motivated two-stage design with behavioral-consistency validation. Below, we respond to your comments and questions point by point. We have revised the manuscript accordingly and highlighted the changes in blue.

Warm start and Stage-2 policy controls (W1/Q3)

Thank you for these constructive comments. We agree that discussing how Stage-2 policy search behaves under different warm-start schedules and parameter spaces is valuable.

Discussion on warm start. Warm start spends an initial share of the Stage-2 budget learning to switch among the source presets before free parameter generation. Varying this share on the same fused algorithm ϕ\phi shows that warm start is beneficial across all three tasks, while the best fixed ratio is task-dependent (test Gap %, five-run mean, lower is better; ID = training distributions, OOD = unseen distributions):

Stage-2 warm startTSP IDTSP OODOVMP IDOVMP OODFJSP IDFJSP OOD
w/o warm start18.5720.721.471.7811.499.18
20% (original)18.4920.991.031.6511.078.89
40%18.3420.671.051.4711.6110.09
60%18.4921.200.991.3511.119.62
80%18.4521.141.051.4411.519.49

Relative to w/o warm start, the best fixed ratio lowers the gap in every column (40% on TSP, 60% on OVMP, and 20% on FJSP), and all four nonzero ratios lower OVMP's ID gap by 28–33%. However, the sensitivity to the fixed ratio varies across tasks. This observation motivates an adaptive extension, TS-DCPPA (adaptive), which replaces the fixed ratio with Thompson sampling to choose between preset-guided and free parameter generation based on their recent search gains.

Discussion on fixed parameters vs. source presets vs. non-source parameters. Next, we examine what happens on the same ϕ\phi and Stage-2 budget if the policy is restricted to a single learned parameter vector (fixed parameters) or to the discrete source presets (per-instance preset or per-step preset), versus generating non-source parameters freely under TS-DCPPA (original) and TS-DCPPA (adaptive):

Policy on the same ϕ\phiTSP IDTSP OODOVMP IDOVMP OODFJSP IDFJSP OOD
Fixed parameters18.3020.971.221.4111.469.47
Per-instance preset18.5921.801.391.6811.209.61
Per-step preset18.7121.540.981.4211.319.57
TS-DCPPA (original)18.4920.991.031.6511.078.89
TS-DCPPA (adaptive)18.2220.070.961.3411.179.15

Generating non-source parameters at each step achieves the lowest gap in all six columns, and TS-DCPPA (adaptive) alone beats all three restricted controls in every column. This confirms that dynamic non-source parameter generation brings gains beyond both a static ϕ\phi (vs. fixed parameters) and discrete preset switching (vs. per-instance and per-step presets), and shows that better Stage-2 search allocation can further unlock the benefit of non-source parameters, which is a promising direction for future study. We have added these discussions on Stage-2 policy search to the appendix.

Methodological novelty and configurator comparison (Q1)

The novelty of DCPPA lies in the framework rather than in a specific configurator. DCPPA decouples an algorithm into a parameterized algorithm ϕ\phi and a configuration policy π\pi, and TS-DCPPA constructs both automatically with an LLM. Existing instance-specific configurators can therefore serve as alternative Stage-2 implementations for learning π\pi on the same fused algorithm ϕ\phi.

To illustrate this, we plug two representative configurators, OPTICAT (TELO 2026) and ISAC++ (AIJ 2016), into Stage 2 on the same fused ϕ\phi, parameter space, and training/test split. Both set parameters once per instance from handcrafted instance features (following our InstSpecHH baseline). Under generous Stage-2 budgets, OPTICAT trains for 30 iterations and ISAC++ tunes for two hours per run, using 76,200 and 26,000–35,000 instance evaluations, respectively (2.6–7.6 times the 500×20=10,000500 \times 20 = 10\text{,}000 instance evaluations of our Stage 2). The table reports five-run mean test Gap (%; lower is better):

MethodTSP IDTSP OODOVMP IDOVMP OODFJSP IDFJSP OOD
TS-DCPPA (OPTICAT in Stage 2)19.0519.982.032.2213.1911.87
TS-DCPPA (ISAC++ in Stage 2)18.1821.121.652.0511.5210.29
TS-DCPPA (EoH in Stage 2)18.4920.991.031.6511.078.89

Original TS-DCPPA (EoH in Stage 2) outperforms both configurators in five of the six columns with a fraction of their evaluation budget. Both OPTICAT and ISAC++ rely on predefined handcrafted instance features, which may not fully capture the problem structure across distributions, and they only predict one static configuration per instance. By contrast, evolving a programmatic policy π\pi in Stage 2 extracts informative quantities directly from the raw inputs and solver state without handcrafted features, and adapts parameters at every decision step.

Parameter space and behavior (W2/Q2)

We agree that recovering the source presets does not guarantee smooth or effective behavior between them, since decisions are discrete and a small parameter change can switch the chosen action. This is why TS-DCPPA learns a policy in Stage 2 to select configurations state by state, rather than assuming that every intermediate configuration is good. DCPPA does not require parameters to be continuous or behavior to be smooth; parameters may be continuous, discrete, or mixed. Its aim is to keep every source behavior reachable, which the fidelity check ensures, while letting the policy reach behaviors that no source algorithm produces. The two analyses below support both parts.

Interpolation. For each task, we fix the fused algorithm and evaluate 500 configurations, comprising the five source presets and 495 points on uniform grids along all ten preset pairs (integer and categorical parameters are rounded or switched at the midpoint). Each configuration is fixed throughout an instance and run on all ID/OOD test instances; all 307,000 runs complete successfully. Even without any state information, some single intermediate configurations beat all five source presets on both ID and OOD instances in every task:

TSPOVMPFJSP
Intermediate configs better than all five presets on both ID and OOD9 (1.8%)17 (3.4%)31 (6.3%)
Gap reduction of the best such config vs. the best preset (ID / OOD)0.6% / 0.4%6.9% / 7.9%3.3% / 6.8%

Learned behavior. We replay the five final TS-DCPPA policies per task on all test instances (4.1 million decisions). At least 99.96% of the generated parameter vectors differ from every source preset. To check behavior rather than parameters, at each decision we also run the fused algorithm with each of the five source presets on the same state and test whether any of them selects the policy's action. The table reports the share of decisions where none does (mean over five policies):

TSPOVMPFJSP
ID2.8%24.8%4.4%
OOD4.5%19.2%4.9%

Most decisions thus agree with at least one source algorithm, so the source behavior is retained, while the remaining decisions are actions that no source algorithm would take in that state. Such actions cannot be obtained even by switching among the sources at every step; they are the extra behavior that distinguishes TS-DCPPA from per-step preset selection in W1/Q3.

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

The paper presents TS-DCPPA, a two-stage automated algorithm design pipeline which uses LLMs to improving generalization performance on heterogeneous datasets. The approach divides algorithm design into two stages. First, the LLM explores complementary algorithms to fuse them into a parametrized algorithm. Second, the LLM designs a policy to control the parameters to tailor performance to the specific instance. The approach shows promising results on various popular problems like the Traveling Salesman Problem, Online VM placement and flexible job scheduling.

Critical strengths

⁠The study provides a interesting approach to the problem. Many design decisions are well-supported by the experiments and ablations. The approach shows genuine performance improvements on some problems.

Critical weaknesses

Current experiments do not isolate the benefits of the learned policy. The study would benefit from an experiment comparing the discrete algorithms (or presets from the fused algorithm) from Stage 1 with the full pipeline. This is significant, as the DCPPA approach is performant even on homogeneous data (Appendix D.2), where supposedly the policy should not vary significantly. A possible ablation would be the fused algorithm run with a single, fixed theta (learned or preset) to isolate the benefits of the policy on heterogeneous data.

Two-stage design is not fully supported by the ablation: The authors provide intuitive justifications for the two-stage design. The ablation run, however, differs from the proposed TS-DCPPA approach in more than staging, as EoH-DCPPA does not have complementary algorithm selection, fusion and other steps. Therefore, performance improvements by the two-stage design are not isolated.

Questions for authors

Pl refer to the weaknesses.

Can you report the results of the fused algorithm run with a single, fixed theta?

Can you run EoH-DCPPA with the complementary algorithms from Stage 1 as its starting point, so that the performance improvements by the two-stage design are isolated?

With reference to Appendix A.3, can you report results with matched time limits or explain why the current budgets are comparable?

2 comments

Thank you for your thoughtful review and valuable suggestions. We appreciate your positive feedback that our two-stage approach is interesting, that our design decisions are well supported by experiments and ablations, and that TS-DCPPA delivers genuine performance improvements. Below, we respond to your comments and questions point by point. We have revised the manuscript accordingly and highlighted the changes in blue.

Stage-2 policy controls and homogeneous data (W1/Q1)

Thank you for this constructive suggestion. Following the suggestions from you and the other reviewers, we have expanded our discussion on Stage-2 policy search on the same fused algorithm ϕ\phi.

Specifically, we examine your question on running ϕ\phi with a single learned parameter vector (fixed parameters), alongside policies restricted to the source presets (per-instance preset and per-step preset) and policies that freely generate non-source parameters. In addition to TS-DCPPA (original) (20% fixed warm start), we include TS-DCPPA (adaptive), an extension that replaces the fixed ratio with Thompson sampling to choose between preset-guided and free parameter generation based on their recent search gains (see our responses to Reviewer owwW, W1/Q3, and Reviewer UeSb, W1/Q1). The table reports five-run mean test Gap (%; lower is better; ID = training distributions, OOD = unseen distributions):

Policy on the same ϕ\phiTSP IDTSP OODOVMP IDOVMP OODFJSP IDFJSP OOD
Fixed parameters18.3020.971.221.4111.469.47
Per-instance preset18.5921.801.391.6811.209.61
Per-step preset18.7121.540.981.4211.319.57
TS-DCPPA (original)18.4920.991.031.6511.078.89
TS-DCPPA (adaptive)18.2220.070.961.3411.179.15

For the Fixed parameters control you asked about, step-level non-source parameter generation achieves lower gaps across all six columns, and TS-DCPPA (adaptive) beats Fixed parameters as well as both preset-only controls in every column. These results show that dynamic non-source parameters contribute beyond both a single fixed θ\theta and discrete preset selection, and suggest that refining Stage-2 search allocation to better exploit non-source parameters is a worthwhile direction for future work. We have added this discussion on Stage-2 policy search to the appendix.

Why TS-DCPPA also excels on homogeneous data. "Homogeneous" in Appendix D.2 means instances share the same generation distribution, but they still span three problem scales (n∈{100,1000,10000}n \in \{100, 1000, 10000\} for TSP; 20,50,8020, 50, 80 PMs for OVMP) and undergo large within-instance state transitions during solving (e.g., from abundant unvisited nodes or server capacity early on to tight bottlenecks later). Because π\pi adapts parameters at every decision step on a fused ϕ\phi that integrates complementary heuristics, it naturally improves solving even within a single distribution. This is consistent with step-level generation outperforming both Fixed parameters and Per-instance preset in the table above.

Initializing EoH-DCPPA with Stage-1 algorithms (W2/Q2)

Thank you for this insightful suggestion. We view initializing EoH-DCPPA with the Stage-1 complementary algorithms as a natural extension of Stage 2, allowing it to evolve both the parameterized algorithm and the configuration policy rather than only the policy. This aligns closely with the future direction discussed in the conclusion section, where we noted that jointly optimizing the parameterized algorithm and policy from strong initial candidates could further improve performance.

Following your suggestion, we initialize EoH-DCPPA with the five Stage-1 source algorithms and run 500 further evaluations after Stage 1 (test Gap %, five-run mean, lower is better):

MethodTSP IDTSP OODOVMP IDOVMP OODFJSP IDFJSP OOD
EoH-DCPPA, from scratch21.3723.901.562.1313.5214.40
EoH-DCPPA, initialized with Stage-1 algorithms17.9120.331.211.7811.228.80
TS-DCPPA (original)18.4920.991.031.6511.078.89
TS-DCPPA (adaptive)18.2220.070.961.3411.179.15

Two findings emerge from this exploration:

  • Stage-1 complementary algorithms are a strong starting point. Initializing with Stage-1 algorithms reduces the gap of from-scratch EoH-DCPPA by 15–39% across all six columns (e.g., 23.90→20.3323.90 \rightarrow 20.33 on TSP-OOD), confirming the value of Stage-1 complementary discovery.
  • Policy-only search and joint refinement are competitive alternatives for Stage 2. While decoupled policy search (TS-DCPPA) leads in four of six columns (TS-DCPPA (adaptive) on TSP-OOD and OVMP, TS-DCPPA (original) on FJSP-ID), the Stage-1-initialized extension is competitive and edges ahead on TSP-ID and FJSP-OOD. Exploring how to combine decoupled policy search with lightweight structural refinement in Stage 2 is an interesting avenue for future work. We have added a discussion of this extension to the appendix.

Evaluation budget and candidate time limits (Q3)

TS-DCPPA and each baseline use the same budget of 1,000 candidate evaluations per run. TS-DCPPA uses 500 evaluations in Stage 1 to discover complementary source algorithms and 500 in Stage 2 to learn the configuration policy.

As for the time limits, they cap the evaluation time of each candidate on all 20 training instances. Because the time each stage of TS-DCPPA needs is hard to know before the search, we leave a margin in each stage's limit. These limits are safety caps against non-terminating or very slow candidates, not actual runtimes.

To verify that the final TS-DCPPA programs finish within the baseline time limits, we evaluate the final programs on all 20 training instances. The table below reports the maximum observed runtime (seconds; InstSpecHH excludes LLM latency for algorithm selection), showing that TS-DCPPA finishes well within the baseline limits:

MethodTSP (s)OVMP (s)FJSP (s)
Baseline time limit306030
EoH4.48118.8331.415
ReEvo28.22215.7760.456
PartEvo5.77840.5071.435
HiFo5.68419.2060.658
RefineEvo12.64934.4110.393
InstSpecHH21.36825.96419.578
TS-DCPPA8.81819.3930.244

Comments on the paper

2 comments

Dear Reviewers,

As the discussion phase draws to a close, please acknowledge the author's response and let them know whether it has addressed some or all of your concerns.

Thank you for the time and effort you and the other reviewers have put into this discussion.

Best,

Your AC

Dear Reviewers,

We thank you again for your thoughtful reviews. With the discussion period concluding in one day, we would like to kindly check if our point-by-point responses, additional experiments, and updated PDF have resolved your concerns. If any questions remain or further clarification would be helpful, we are readily available to respond promptly.

Thank you for your time and consideration.

Best regards,

Authors of Submission 32633

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.