Pith. sign in

REVIEW 3 major objections 8 minor 48 references

Specializing different repair heuristics to different local regions inside one routing solution beats scoring one program on whole instances.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 16:08 UTC pith:E3LXRNTS

load-bearing objection Solid bilevel AHD package with real feasibility engineering, but the stress-test is right: headline gains are mostly the checked-LNS scaffold, not the specialization story the abstract sells. the 3 major comments →

arxiv 2607.23676 v1 pith:E3LXRNTS submitted 2026-07-26 cs.AI

SpecAHD: Localize to Specialize for Automated Heuristic Design in Large-Scale Routing Problems

Kezhao Lai , Yutao Lai , Hai-Lin Liu This is my paper
classification cs.AI
keywords automated heuristic designlarge language modelslarge neighborhood searchvehicle routingbilevel optimizationsubmodular repertoire selectionwithin-instance specialization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Large routing solutions are usually improved by opening small pieces and rebuilding them, but those pieces can look very different from one another. Standard LLM-based heuristic design still rewards a single construction rule for average performance across whole instances or fixed solver slots, so the rule must compromise. SpecAHD instead runs two coupled searches: one learns which bounded regions of an incumbent to open, and the other evolves a small set of executable repair programs tailored to the tasks those regions create. Checked full-solution outcomes both score the region chooser and select complementary specialists via a coverage-aware objective that is monotone submodular for a fixed task table, so greedy selection has a proven approximation factor. The same checked responses fill an archive that, at test time, routes each new local structure to a specialist. On four routing problems the method cuts held-out cost by as much as about 58% versus the strongest competing automated-design baseline and beats the best per-instance baseline on most public benchmarks.

Core claim

Within-instance specialization—jointly learning where to expose bounded repair regions and which complementary executable heuristics should repair each local structure—substantially outperforms automated heuristic design that scores one program on complete instances or fixed components, cutting held-out routing cost by up to 57.7% against the strongest competing baseline.

What carries the argument

A coupled bilevel loop: an upper program induces validated repair tasks; a lower objective J_g that mixes average quality with taskwise max-response coverage is monotone submodular on a fixed dense response table, so greedy repertoire selection has a (1−1/e) guarantee; checked region–heuristic scores also populate a Repair Response Archive for nearest-neighbor routing at inference.

Load-bearing premise

Hand-built local structure descriptors plus nearest-neighbor lookup in the training archive are enough to pick the right specialist when test regions differ from what was seen in training.

What would settle it

Freeze the trained upper program and repertoire, then on held-out public instances replace archive routing with the single global-best lower heuristic (or random choice); if costs do not rise relative to SpecAHD routing, the claimed value of within-instance specialization and response-based routing fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Automated heuristic design for large routing should optimize region exposure and a specialist set together, not a single global construction rule.
  • Task-level checked responses can replace aggregate instance scores when selecting small complementary program sets, with a greedy (1−1/e) guarantee on the fixed table.
  • A frozen response archive can serve as a nonparametric router from local structure to repair code without training a separate selector.
  • Checked sequential merge with rollback lets generated repair code fail safely while still preserving complete-solution feasibility.
  • The same bilevel pattern can be instantiated on CVRP, TSP, VRPTW, and SDVRP through problem-specific validators and checkers only.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If descriptors are the fragile link, learned or problem-agnostic fingerprints of local solution state could extend the method beyond hand-specified routing features.
  • The same localize-then-specialize loop may transfer to other large combinatorial settings where LNS-style destroy/repair already splits the work (scheduling, packing, network design).
  • Successive-repair interactions flagged as future work suggest the archive may eventually need multi-region or order-aware records, not only single-region responses.
  • Budget shared between upper and lower LLM search implies that methods which over-allocate to one level may leave complementarity on the table.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 8 minor

Summary. The paper proposes SpecAHD, a bilevel LLM-based automated heuristic design (AHD) framework for large-scale routing: an upper-level evolved program exposes bounded repair regions in an incumbent solution y0 = B(x), and a lower-level search evolves a small repertoire of constructive repair heuristics selected via a coverage-aware objective. Task-level checked repair outcomes populate a Repair Response Archive, which a deterministic k-NN router over hand-designed structural descriptors queries at test time to pick a specialist per region. The authors prove (Prop. 1) that checked sequential repair with rollback preserves feasibility under an exact checker, and (Prop. 2) that the repertoire objective Jg is monotone submodular for fixed upper program and candidate pool, giving a (1−1/e) greedy guarantee; they explicitly disclaim any convergence claim for the alternating search. Experiments span CVRP, TSP, VRPTW, SDVRP with two LLM backbones, ten AHD baselines under a shared 1,000-program budget, held-out synthetic sets, and public benchmarks (CVRP-XL, TSPLIB, Solomon/Homberger, SDVRP SET-1–4), reporting up to 57.7% improvement over the strongest AHD baseline and envelope wins on most public instances. Ablations cover generation-time memory components, and a routing study (Table 6) compares the archive router against global-best, random, and oracle selection on held-out regions.

Significance. If the results hold, the paper makes a real contribution: moving AHD specialization from across-instance to within-instance is a well-motivated idea, the checked-repair interface with a feasibility guarantee is practically valuable for LLM-generated code, and the evaluation is unusually broad (four problem classes, two backbones, controlled budget, public benchmarks). The reproducibility posture is a genuine strength: frozen checkpoints, run-level audit trails, fixed seeds, disjoint train/validation/public splits, no archive write-back at test time, and honest reporting of timeout/partial-run accounting in the appendix. The theory is modest but correctly scoped — the authors state plainly that the (1−1/e) guarantee covers only greedy selection over a fixed dense response table, not the alternating search, which is the right level of claim. Table 6 is also commendably honest in showing that archive routing beats the global-best specialist by only 0.659% (TSP) and 0.188% (CVRP) against a ~2–2.5% oracle ceiling. The weakness is that the framing attributes the headline gains to the specialization mechanism, while the paper's own evidence indicates that mechanism contributes at most ~2.

major comments (3)
  1. [Experiments; Table 6 / Appendix E; §Problem Formulation] The central attribution claim is not supported by the reported controls. The headline gains (up to 57.7% vs AHD baselines; envelope wins on most public instances) are attributed to the coupled specialization mechanism, but the paper's own Table 6 shows the region-conditioned routing mechanism contributes +0.659% (TSP) and +0.188% (CVRP) over the single global-best repertoire member, with a hindsight-oracle ceiling of only +2.552%/+1.998% — i.e., the entire specialization mechanism caps out near 2.5% on these tasks while headline gains are one to two orders of magnitude larger. Meanwhile every SpecAHD evaluation starts from a classical-solver incumbent y0 = B(x) (PASA + PyVRP for SDVRP; B is never specified for TSP/CVRP/VRPTW) and applies validated ruin-and-recreate, whereas all ten baselines are scored as constructive heuristics without that scaffold. The large-TSPLIB rows (usa13509, brd
  2. [Experiments, Tables 2–3; Appendix C (run selection and aggregation)] Tables 2–3 aggregate 'completed runs' and the resulting rows exhibit non-monotonicities that need explanation: EoH-S CVRP drops from 57.98k at size 100 to 44.71k at size 200; SpecAHD-G VRPTW drops from 90.81k at 400 to 88.28k at 600; SpecAHD-D SDVRP drops from 7.43k at 200 to 4.84k at 250. Aggregate cost over instance pools of increasing size should not decrease unless the per-size instance pools differ in difficulty or completion filtering differs across cells. The protocol states costs average completed runs while partial runs and timeouts are retained only 'in the record'; if timeout rates differ across methods and sizes, per-cell means over completed runs are biased in favor of timeout-prone methods, and the claim that a SpecAHD backbone achieves the lowest cost in all 19 size groups could be an artifact of differential completion. Please report per-cell completion counts for Tables
  3. [Table 5; Appendix D (public results by named instance)] The public-benchmark summary counts only rows with a valid SpecAHD result: Appendix D states 'All printed rows contain a valid SpecAHD result', and the printed rows are 79/100 CVRP, 30/31 TSP, 331/356 VRPTW, 95/95 SDVRP. Excluding the 21 CVRP, 1 TSP, and 25 VRPTW instances where SpecAHD produced no valid result — rather than counting them as losses — inflates the reported win rates (71/79, 28/30, 315/331, 68/95) that support the abstract's 'outperforms the per-instance baseline envelope on most public instances'. These rows should either be included as losses or the summary should state the exclusion explicitly in the main text. Relatedly, on rows where the strongest baselines time out (e.g., large TSPLIB), the 'baseline envelope' is effectively uncontested; the win count conflates beating the best available baseline with being the only method with a usable pipeline. A breakdown of wins
minor comments (8)
  1. [Abstract; Experiments] The 'up to 57.7%' figure in the abstract is not traceable to any reported cell in the provided tables. Please identify the instance/size group where it is attained and report a distributional summary (median or per-problem mean improvement over the strongest baseline) alongside the maximum.
  2. [Experiments; Table 5; Appendix C Table 2] The text states 'Across the 19 validation size groups ... (Table 2)' but the 19 groups span Tables 2 and 3 (CVRP 5 + TSP 4 + VRPTW 5 + SDVRP 5). Also, Table 5 reports 30 TSP instances while Appendix Table 2 lists 31 TSPLIB instances; please reconcile.
  3. [Ablation Summary, Table 4] Table 4's metrics are not defined: the lower-level-search columns mix units (TSP in %, CVRP in k) and the frozen-repertoire columns are all percentages with the Full row at −0.337% on TSP. What is the reference for these percentages, and why is the full system's frozen-repertoire change negative on TSP? A one-line definition per column would suffice.
  4. [Ablation Summary; Appendix F] The ablation variants (single embedding, w/o semantic filter, w/o skill guidance) ablate generation-time memory accessories inherited from the EoH lineage, not the paper's claimed contributions (bilevel coupling, coverage-aware repertoire, archive routing). Readers looking for 'localize to specialize' ablations will find only Table 6. At minimum, the framing of Table 4 should clarify what is and is not being ablated.
  5. [§Conditioning Repair on Measured Responses; Appendix B] The main text states 'Appendix B lists each problem-specific descriptor together with its repair interface', but Appendix B (Table 1) gives only the repair interfaces; the coordinates of the structural descriptors Φp are never listed. Since Φp and k=5 are the router's only inputs, the descriptor definitions should be given explicitly (even if the code is released).
  6. [Conclusion and Limitations] The conclusion claims SpecAHD 'generalizes beyond its training distributions', but the VRPTW upper-level training pool is explicitly built from Solomon/Homberger C/R/RC families and type-1/type-2 regimes, and the public VRPTW test is Solomon/Homberger — same family. The claim is supported for size scaling but overstated for distribution shift; please qualify it.
  7. [Tables 2–3] The backbone variance is occasionally extreme and deserves a sentence of discussion: e.g., SDVRP size 250, SpecAHD-G 12.17k vs SpecAHD-D 4.84k; VRPTW size 600, 88.28k vs 142.69k. Reporting both variants is good practice, but the instability of the search across backbones bears on how robust the method is in deployment.
  8. [References; Figure 1] Reference formatting: 'Garey, Johnson et al. 1990' is an unusual citation for Maximum Coverage (the standard reference is Garey & Johnson 1979, and the greedy set-cover analysis of Hochbaum or Feige would be more apt); 'et al.' on a two-author book is also odd. Figure 1's text is dense and hard to read at column width.

Circularity Check

0 steps flagged

No significant circularity: submodular guarantee, bilevel coupling, and held-out gains are independently defined and measured, not forced by construction or self-citation.

full rationale

The paper’s load-bearing formal claim is that, for a fixed upper program g and dense response table S_g, the repertoire objective J_g (Eq. 3) is nonnegative, normalized, monotone, and submodular, so greedy selection attains the standard (1−1/e) bound (Prop. 2). That property follows from the definition of J_g as a modular average-quality term plus a pointwise max-coverage term; it is a correct application of Nemhauser–Wolsey–Fisher, not a tautology that equates the training objective to the reported test metric. Upper-level utility U_x is computed from complete-solution costs under a frozen (H, A) pipeline with checked merge (Eq. 1) and no archive write-back at test time; training, checkpoint, and public splits are disjoint. Self-citations to EoH/LLM4AD supply the mutation interface and sandbox, not a uniqueness theorem that forces the result. The empirical headline (up to 57.7% vs AHD baselines; envelope wins on public instances) is an external comparison, not a fitted parameter renamed as prediction. Attribution questions (scaffold vs specialization) are correctness/experimental-design issues, not circular derivation. No step reduces a claimed prediction to its own inputs by construction.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 3 invented entities

Load-bearing content is mostly standard combinatorial optimization plus LLM program search. The mathematical guarantee rests on classical submodular maximization. Empirical claims rest on problem-specific validators/checkers, hand-built descriptors, finite alternating budgets, and LLM sampling. Free parameters are the usual method knobs; invented entities are methodological (archive, bilevel coupling), not physical postulates.

free parameters (7)
  • β (quality vs complementarity weight in Jg) = ∈[0,1], exact default not singled out as fitted in main text
    Trades average heuristic quality against taskwise envelope coverage in repertoire selection; chosen by authors, not derived.
  • q (repertoire cardinality) = m≤q; experiments use small fixed repertoire (e.g., ten programs in routing study)
    Hard cap on retained lower heuristics; controls specialization capacity.
  • K (region budget) and validator size/count limits
    Bound how many/how large repair regions the upper program may expose.
  • λ, α1, α2, α3 (upper penalty weights) = ≥0, author-set
    Penalize empty validated output, validator rejection rate, and merge rollbacks in upper utility.
  • k (archive neighbors) and descriptor Φp coordinates = k=5 in routing validation
    Hand-designed local features and k-NN weighting drive test-time specialist choice; central to transfer.
  • Search budgets (TU, TL, BU, BL) and shared 1,000-program cap = 1,000 programs total per method/backbone
    Finite alternating budgets define the practical algorithm; results depend on this compute envelope.
  • LLM backbone and sampling stack = GLM-4-Flash and DeepSeek-V4-Flash
    Generated code quality is backbone-dependent; two flash models reported with uneven winners by problem/size.
axioms (5)
  • domain assumption Exact problem-specific checker Fp recognizes feasible complete solutions Yp(x); baseline y0 is feasible.
    Proposition 1 and the merge operator Γp depend on exact feasibility recognition including capacity/time-window effects beyond one support.
  • standard math For fixed upper program g, task set Rg, and evaluated pool Pg, Jg is nonnegative, normalized, monotone, and submodular, so greedy attains (1−1/e) of the ideal budget-q repertoire (Nemhauser–Wolsey–Fisher).
    Proposition 2; hardness via Maximum Coverage when β=0 and binary scores.
  • ad hoc to paper Alternating finite-budget upper/lower EoH-style search is an adequate practical surrogate for the ideal nested coupled objective g⋆.
    Paper states exact nested optimization is too costly and that the guarantee does not imply bilevel convergence; empirical claims rest on this surrogate.
  • ad hoc to paper Structural descriptors Φp and cosine k-NN similarity on standardized features identify repair tasks that share specialist suitability.
    Archive routing (Eq. 4) and test-time specialization depend on this similarity model; descriptors are hand-designed per problem.
  • domain assumption LLM-generated executable programs under sandbox timeouts, with zero credit for failures, are a valid search space for constructive repair and region exposure.
    Standard AHD premise shared with FunSearch/EoH; feasibility is recovered by checked merge rather than by correct code alone.
invented entities (3)
  • SpecAHD coupled bilevel AHD framework independent evidence
    purpose: Jointly search region-exposure programs and complementary repair programs with mutual evaluation coupling.
    Core methodological object; evaluated empirically, not an external physical entity.
  • Repair Response Archive A with response-based router ρA,H independent evidence
    purpose: Store checked (region, descriptor, heuristic, score) tuples and select specialists at inference by local structure.
    Distinguishes SpecAHD from archives that only keep code diversity; evidence is the held-out routing study and public transfer protocol.
  • Repertoire objective Jg (average quality + taskwise max coverage) independent evidence
    purpose: Select a small complementary set of lower heuristics for induced repair tasks.
    Variant of set-function design used in EoH-S-style work, specialized to within-instance tasks; submodularity gives selection guarantee only for fixed tables.

pith-pipeline@v1.2.0-grok45-kimik3 · 50997 in / 4106 out tokens · 88347 ms · 2026-07-30T16:08:40.554756+00:00 · methodology

0 comments
read the original abstract

LLM-based automated heuristic design (AHD) typically scores executable programs on complete instances or within fixed solver components. In large-scale routing problems, localized reconstruction reduces the size of each optimization task, but repair regions within the same incumbent can exhibit substantially different structures. One construction rule must therefore compromise across them. In this paper, we propose SpecAHD, a coupled bilevel framework for within-instance specialization. An upper-level search learns where to expose bounded repair regions, while a lower-level search evolves a complementary repertoire of executable heuristics for the induced repair tasks. The upper-level program determines the repair tasks seen by the lower level, while checked repair outcomes determine how upper-level programs are evaluated. The lower-level objective favors heuristics that perform well on average or solve tasks that the current repertoire handles poorly. For the repair tasks induced by a fixed upper-level program and a fixed lower-level candidate pool, this objective is monotone submodular, allowing greedy repertoire selection with a (1-1/e) approximation guarantee. Across four routing problems and multiple LLM backbones, SpecAHD reduces held-out objective cost by up to 57.7% against the strongest competing AHD baseline and outperforms the per-instance baseline envelope on most public instances.

Figures

Figures reproduced from arXiv: 2607.23676 by Hai-Lin Liu, Kezhao Lai, Yutao Lai.

Figure 1
Figure 1. Figure 1: Overview of SpecAHD. Training alternates upper-level region design and lower-level repair search. Checked task-level [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Running-best training scores after the first 50 eval [PITH_FULL_IMAGE:figures/full_fig_p024_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Wall-clock validation time for the lower-level [PITH_FULL_IMAGE:figures/full_fig_p026_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 9 linked inside Pith

  1. [1]

    2014 , publisher=

    Vehicle routing: problems, methods, and applications , author=. 2014 , publisher=

    2014

  2. [2]

    1999 , publisher=

    Designing and managing the supply chain: Concepts, strategies, and cases , author=. 1999 , publisher=

    1999

  3. [3]

    Integration , volume=

    Recent directions in netlist partitioning: A survey , author=. Integration , volume=. 1995 , publisher=

    1995

  4. [4]

    Current medicinal chemistry , volume=

    Virtual screening strategies in drug discovery: a critical review , author=. Current medicinal chemistry , volume=. 2013 , publisher=

    2013

  5. [5]

    Computers and intractability , pages=

    A Guide to the Theory of NP-Completeness , author=. Computers and intractability , pages=. 1990 , publisher=

    1990

  6. [6]

    Nature , volume=

    Mathematical discoveries from program search with large language models , author=. Nature , volume=. 2024 , publisher=

    2024

  7. [7]

    arXiv preprint arXiv:2401.02051 , year=

    Evolution of heuristics: Towards efficient automatic algorithm design using large language model , author=. arXiv preprint arXiv:2401.02051 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2501.08603 , year=

    Monte carlo tree search for comprehensive exploration in llm-based automatic heuristic design , author=. arXiv preprint arXiv:2501.08603 , year=

    Pith/arXiv arXiv

  9. [9]

    Advances in neural information processing systems , volume=

    Reevo: Large language models as hyper-heuristics with reflective evolution , author=. Advances in neural information processing systems , volume=

  10. [10]

    ACM Transactions on Evolutionary Learning , year=

    In-the-loop hyper-parameter optimization for llm-based automated design of heuristics , author=. ACM Transactions on Evolutionary Learning , year=

  11. [11]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Efficient heuristics generation for solving combinatorial optimization problems using large language models , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2 , pages=

  12. [12]

    arXiv preprint arXiv:2505.12285 , year=

    Calm: Co-evolution of algorithms and language model for automatic heuristic design , author=. arXiv preprint arXiv:2505.12285 , year=

    Pith/arXiv arXiv

  13. [13]

    Workshop on Differentiable Learning of Combinatorial Algorithms , year=

    Generalizable heuristic generation through large language models with meta-optimization , author=. Workshop on Differentiable Learning of Combinatorial Algorithms , year=

  14. [14]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages=

    Eoh-s: Evolution of heuristic set using llms for automated heuristic design , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=

  15. [15]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages=

    Hsevo: Elevating automatic heuristic design with diversity-driven harmony search and genetic algorithm using llms , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=

  16. [16]

    arXiv preprint arXiv:2506.00490 , year=

    Llm-driven instance-specific heuristic generation and selection , author=. arXiv preprint arXiv:2506.00490 , year=

    arXiv

  17. [17]

    arXiv preprint arXiv:2203.13392 , year=

    Automated algorithm selection: from feature-based to feature-free approaches , author=. arXiv preprint arXiv:2203.13392 , year=

    Pith/arXiv arXiv

  18. [18]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Automating Thought of Search: A Journey Towards Soundness and Completeness (Student Abstract) , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  19. [19]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages=

    Co-bench: Benchmarking language model agents in algorithm search for combinatorial optimization , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=

  20. [20]

    arXiv preprint arXiv:2505.20242 , year=

    RedAHD: Reduction-Based End-to-End Automatic Heuristic Design with Large Language Models , author=. arXiv preprint arXiv:2505.20242 , year=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2602.08253 , year=

    G-LNS: Generative large neighborhood search for LLM-based automatic heuristic design , author=. arXiv preprint arXiv:2602.08253 , year=

    arXiv

  22. [22]

    Forty-second International Conference on Machine Learning , year=

    Large language model-driven large neighborhood search for large-scale milp problems , author=. Forty-second International Conference on Machine Learning , year=

  23. [23]

    arXiv preprint arXiv:2603.06996 , year=

    Large language model-driven full-component evolution of adaptive large neighborhood search , author=. arXiv preprint arXiv:2603.06996 , year=

    Pith/arXiv arXiv

  24. [24]

    Mathematical programming , volume=

    An analysis of approximations for maximizing submodular set functions—I , author=. Mathematical programming , volume=. 1978 , publisher=

    1978

  25. [25]

    IEEE transactions on information theory , volume=

    Nearest neighbor pattern classification , author=. IEEE transactions on information theory , volume=. 1967 , publisher=

    1967

  26. [26]

    arXiv preprint arXiv:1504.04909 , year=

    Illuminating search spaces by mapping elites , author=. arXiv preprint arXiv:1504.04909 , year=

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2307.09326 , year=

    Bop-elites, a bayesian optimisation approach to quality diversity search with black-box descriptor functions , author=. arXiv preprint arXiv:2307.09326 , year=

    Pith/arXiv arXiv

  28. [28]

    International conference on principles and practice of constraint programming , pages=

    Using constraint programming and local search methods to solve vehicle routing problems , author=. International conference on principles and practice of constraint programming , pages=. 1998 , organization=

    1998

  29. [29]

    Journal of Computational Physics , volume=

    Record breaking optimization results using the ruin and recreate principle , author=. Journal of Computational Physics , volume=. 2000 , publisher=

    2000

  30. [30]

    Transportation science , volume=

    An adaptive large neighborhood search heuristic for the pickup and delivery problem with time windows , author=. Transportation science , volume=. 2006 , publisher=

    2006

  31. [31]

    Advances in Neural Information Processing Systems , volume=

    Partition to evolve: Niching-enhanced evolution with llms for automated algorithm discovery , author=. Advances in Neural Information Processing Systems , volume=

  32. [32]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages=

    Multi-objective evolution of heuristic using large language model , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=

  33. [33]

    International Conference on Parallel Problem Solving from Nature , pages=

    Understanding the importance of evolutionary search in automated heuristic design with large language models , author=. International Conference on Parallel Problem Solving from Nature , pages=. 2024 , organization=

    2024

  34. [34]

    , author=

    LLM-Driven Neighborhood Search for Efficient Heuristic Design. , author=. CEC , pages=

  35. [35]

    arXiv preprint arXiv:2508.05433 , year=

    Multimodal LLM-assisted Evolutionary Search for Programmatic Control Policies , author=. arXiv preprint arXiv:2508.05433 , year=

    arXiv

  36. [36]

    arXiv preprint arXiv:2412.17287 , year=

    Llm4ad: A platform for algorithm design with large language model , author=. arXiv preprint arXiv:2412.17287 , year=

    arXiv

  37. [37]

    arXiv preprint arXiv:2601.11467 , year=

    The XL Instances for the Capacitated Vehicle Routing Problem , author=. arXiv preprint arXiv:2601.11467 , year=

    Pith/arXiv arXiv

  38. [38]

    ORSA journal on computing , volume=

    TSPLIB—A traveling salesman problem library , author=. ORSA journal on computing , volume=. 1991 , publisher=

    1991

  39. [39]

    Operations research , volume=

    Algorithms for the vehicle routing and scheduling problems with time window constraints , author=. Operations research , volume=. 1987 , publisher=

    1987

  40. [40]

    INFOR: Information Systems and Operational Research , volume=

    Two evolutionary metaheuristics for the vehicle routing problem with time windows , author=. INFOR: Information Systems and Operational Research , volume=. 1999 , publisher=

    1999

  41. [41]

    Networks: An International Journal , volume=

    The split delivery vehicle routing problem: Applications, algorithms, test problems, and computational results , author=. Networks: An International Journal , volume=. 2007 , publisher=

    2007

  42. [42]

    Operations research , volume=

    A lower bound for the split delivery vehicle routing problem , author=. Operations research , volume=. 2000 , publisher=

    2000

  43. [43]

    Transportation Science , volume=

    An optimization-based heuristic for the split delivery vehicle routing problem , author=. Transportation Science , volume=. 2008 , publisher=

    2008

  44. [44]

    and Golden, Bruce , journal=

    Torkzaban, Nariman and Gholami, Anousheh and Baras, John S. and Golden, Bruce , journal=. 2023 , doi=

    2023

  45. [45]

    and Lan, Leon and Kool, Wouter , journal=

    Wouda, Niels A. and Lan, Leon and Kool, Wouter , journal=. 2024 , doi=

    2024

  46. [46]

    Advances in Neural Information Processing Systems , volume=

    Learning to Delegate for Large-Scale Vehicle Routing , author=. Advances in Neural Information Processing Systems , volume=. 2021 , url=

    2021

  47. [47]

    ECAI 2020 , pages =

    Neural Large Neighborhood Search for the Capacitated Vehicle Routing Problem , author =. ECAI 2020 , pages =. 2020 , publisher =

    2020

  48. [48]

    European Journal of Operational Research , volume =

    New Benchmark Instances for the Capacitated Vehicle Routing Problem , author =. European Journal of Operational Research , volume =. 2017 , publisher =

    2017