arxiv: 2605.03339 · v1 · submitted 2026-05-05 · 💻 cs.AI

Recognition: unknown

Automated Large-scale CVRP Solver Design via LLM-assisted Flexible MCTS

Tong Guo , Caishun Chen , Yew Soon Ong

Authors on Pith no claims yet

Pith reviewed 2026-05-07 16:41 UTC · model grok-4.3

classification 💻 cs.AI

keywords large-scale CVRPautomated algorithm designlarge language modelsMonte Carlo tree searchdecomposition methodsvehicle routingoptimization solversLLM-assisted search

0 comments

The pith

LLMs guided by flexible Monte Carlo tree search autonomously compose decomposition-based solvers for large-scale CVRP that outperform existing state-of-the-art methods.

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

The paper presents LaF-MCTS, a framework that directs large language models to design solvers for capacitated vehicle routing problems involving hundreds or thousands of nodes. It organizes the design task into a three-tier hierarchy so the model can build decomposition rules and sub-solvers step by step within context limits. Semantic pruning discards duplicate or similar code proposals while branch regrowth restores variety to keep the search productive. On standard CVRPLib benchmarks the resulting solvers exceed the performance of several leading CVRP methods. The approach targets the practical bottleneck that creating scalable decomposition logic has remained a manual, expertise-heavy task.

Core claim

LaF-MCTS uses a three-tier decision hierarchy to let LLMs incrementally specify decomposition policies and configure sub-solvers for LSCVRP instances. Semantic pruning removes semantically and structurally redundant code while branch regrowth regenerates alternatives to maintain diversity. Extensive tests on CVRPLib show that the autonomously composed solvers surpass various state-of-the-art CVRP solvers.

What carries the argument

LaF-MCTS, an LLM-guided Monte Carlo Tree Search that employs a three-tier decision hierarchy for incremental decomposition policy and sub-solver design, augmented by semantic pruning to eliminate redundant codes and branch regrowth to sustain search diversity.

If this is right

Decomposition-enhanced CVRP solvers can be produced without manual expert design of the decomposition logic.
The same automated process scales to instances with hundreds to thousands of nodes where direct solvers currently fail.
LLM context limits no longer block generation of sophisticated search strategies once the hierarchy and pruning mechanisms are in place.
Performance gains appear on standard CVRPLib sets against multiple published baselines.

Where Pith is reading between the lines

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

The same hierarchy-and-pruning pattern could be applied to other large-scale combinatorial problems such as scheduling or network design.
Replacing the underlying LLM with a stronger model would likely raise the quality ceiling of the generated solvers without changing the framework.
The method opens a route to hybrid design loops in which a human supplies only high-level constraints and the system fills in the detailed decomposition code.

Load-bearing premise

The three-tier hierarchy combined with semantic pruning and branch regrowth lets LLMs produce effective decomposition policies and sub-solvers for large CVRP instances even though context windows are limited.

What would settle it

Generate a solver with LaF-MCTS on a fresh large-scale CVRP instance outside CVRPLib and measure whether its solution quality or runtime on that instance exceeds the best published solvers such as LKH or modern decomposition heuristics run on the same data.

Figures

Figures reproduced from arXiv: 2605.03339 by Caishun Chen, Tong Guo, Yew Soon Ong.

**Figure 1.** Figure 1: The diagram of LaF-MCTS. The three-tier decision hierarchy leverages LLMs to generate component codes for the framework, view at source ↗

**Figure 2.** Figure 2: The dilemma of code redundancy detection. (a) Syntactically different yet equivalent codes are well-handled by semantic embed view at source ↗

**Figure 3.** Figure 3: Convergence curves of HGS, HGS+BS and the CVRP solver automatically designed from LaF-MCTS on representative instances. view at source ↗

**Figure 4.** Figure 4: Convergence curves of EoH/ReEvo variants and LaF view at source ↗

**Figure 5.** Figure 5: Parameter analysis. Left: The impact of backbone LLMs view at source ↗

read the original abstract

Solving large-scale CVRP (LSCVRP) with hundreds to thousands of nodes remains difficult for even state-of-the-art solvers. Divide-and-conquer can scale by decomposing the instance into size-reduced subproblems, but designing decomposition logic and configuring sub-solvers is highly expertise- and labor-intensive. Large Language Models (LLMs) have emerged as promising tools for automated algorithm design. However, existing LLM-driven approaches struggle with LSCVRP primarily due to the difficulty in generating sophisticated search strategies within a limited context window. To bridge this gap, we propose the LLM-assisted Flexible Monte Carlo Tree Search (LaF-MCTS), a novel framework that automates the design of high-performance LSCVRP solvers. We develop a three-tier decision hierarchy to enable incremental design of decomposition policies and sub-solvers for LSCVRP. To enable efficient search within the algorithmic hypothesis space, we introduce semantic pruning to eliminate semantically and structurally redundant codes, and branch regrowth to regenerate codes and preserve diversity. Extensive experiments on CVRPLib demonstrate that LaF-MCTS autonomously composes and optimizes decomposition-enhanced solvers that surpasses various state-of-the-art CVRP solvers.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LaF-MCTS gives LLMs a three-tier MCTS setup with pruning and regrowth to generate CVRP decomposition solvers, but the abstract's performance claims lack numbers and the evaluation risks overfitting to CVRPLib.

read the letter

The paper's main contribution is LaF-MCTS, which wraps an LLM in a flexible Monte Carlo Tree Search that uses a three-tier hierarchy to build decomposition policies and sub-solvers incrementally for large-scale CVRP. Semantic pruning drops redundant code and branch regrowth keeps the search from collapsing, both aimed at working around context limits when generating optimization code. This is a concrete extension of prior LLM algorithm design efforts to a logistics problem where manual divide-and-conquer design is expensive. The framing is straightforward: existing solvers struggle at scale, and automating the decomposition logic could cut down on expert effort. The hierarchy and search controls look like reasonable engineering choices for keeping the generated solvers both sophisticated and varied. The approach sits at the intersection of automated algorithm design and operations research, so readers working on LLM-guided solvers or scalable routing might find the framework description useful as a starting point. The soft spots are in the evidence. The abstract states that the method surpasses state-of-the-art solvers on CVRPLib after extensive experiments, yet supplies no metrics, baseline list, or protocol. That makes it hard to judge effect size or reproducibility. The stress-test concern also lands: because candidate solvers are scored inside the MCTS loop on CVRPLib instances, any final reported win could reflect tuning to those exact cases rather than a general policy. Without a held-out partition or explicit separation between design-time and test-time instances, the generalization claim stays unproven. I would bring this to a reading group focused on AI for combinatorial optimization as a maybe, mainly to discuss the hierarchy and pruning mechanics. I would not cite it yet because the current writeup does not give enough data to rely on. A serious editor should send it to peer review; the framework has enough structure to merit referee input on the experimental design and held-out testing, even if heavy revision is likely.

Referee Report

2 major / 2 minor

Summary. The paper proposes LaF-MCTS, an LLM-assisted flexible Monte Carlo Tree Search framework for automating the design of decomposition-enhanced solvers for large-scale CVRP (LSCVRP). It introduces a three-tier decision hierarchy for incremental policy and sub-solver design, along with semantic pruning and branch regrowth to manage context limits, and reports that extensive experiments on CVRPLib show the resulting solvers surpass various state-of-the-art CVRP solvers.

Significance. If the performance claims hold with proper generalization controls, the work would be significant for automated algorithm design in combinatorial optimization: it demonstrates a structured way to use LLMs for composing sophisticated divide-and-conquer strategies that scale to hundreds or thousands of nodes, potentially lowering the expertise barrier for high-performance LSCVRP solvers.

major comments (2)

[Section 5] Experimental evaluation (Section 5): The manuscript does not describe a held-out test partition of CVRPLib instances that is disjoint from those used inside the MCTS evaluation loop. Because candidate solvers are scored by running them on CVRPLib instances during search, any selected solver may have been implicitly tuned to the exact instances (or a subset) later used for final reporting; without an explicit train/test separation, the reported superiority over SOTA solvers cannot be distinguished from instance-specific adaptation.
[Section 3] LaF-MCTS framework (Section 3): The three-tier decision hierarchy is presented as enabling LLMs to generate effective decomposition policies despite limited context windows, yet no concrete examples of generated code, policy fragments, or ablation results on context-window usage are supplied. This makes it impossible to verify whether semantic pruning and branch regrowth actually produce the claimed sophisticated sub-solvers or merely prune to simpler heuristics.

minor comments (2)

[Abstract] The abstract states superior performance on CVRPLib but supplies no quantitative metrics, baseline names, or statistical test details; these should be summarized in the abstract for clarity.
[Section 3] Notation for the three-tier hierarchy (e.g., definitions of the decision levels) is introduced without a compact diagram or pseudocode; adding one would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and detailed comments. We address each major point below and commit to revisions that strengthen the experimental validity and illustrative clarity of the work.

read point-by-point responses

Referee: [Section 5] Experimental evaluation (Section 5): The manuscript does not describe a held-out test partition of CVRPLib instances that is disjoint from those used inside the MCTS evaluation loop. Because candidate solvers are scored by running them on CVRPLib instances during search, any selected solver may have been implicitly tuned to the exact instances (or a subset) later used for final reporting; without an explicit train/test separation, the reported superiority over SOTA solvers cannot be distinguished from instance-specific adaptation.

Authors: We acknowledge the validity of this concern. The current manuscript performs both the LaF-MCTS search and the final reporting on CVRPLib instances without an explicit disjoint held-out partition, which leaves open the possibility of implicit adaptation. In the revised manuscript we will introduce a clear train/test split of the CVRPLib benchmark: a search subset used inside the MCTS loop for scoring candidate solvers, and a completely disjoint held-out test subset on which all final performance numbers (including comparisons against SOTA solvers) will be reported. We will also add a short discussion of how the split was constructed to ensure no overlap. revision: yes
Referee: [Section 3] LaF-MCTS framework (Section 3): The three-tier decision hierarchy is presented as enabling LLMs to generate effective decomposition policies despite limited context windows, yet no concrete examples of generated code, policy fragments, or ablation results on context-window usage are supplied. This makes it impossible to verify whether semantic pruning and branch regrowth actually produce the claimed sophisticated sub-solvers or merely prune to simpler heuristics.

Authors: We agree that the lack of concrete examples reduces the transparency of the framework. In the revised version we will augment Section 3 with (i) representative code snippets and policy fragments produced at each of the three decision tiers, (ii) before-and-after examples illustrating the effect of semantic pruning and branch regrowth, and (iii) a new ablation study that measures solver complexity and solution quality with and without these context-management mechanisms. These additions will allow readers to directly assess whether the generated sub-solvers are indeed sophisticated divide-and-conquer strategies. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical algorithmic framework is self-contained

full rationale

The paper describes an LLM-assisted MCTS framework (LaF-MCTS) with a three-tier decision hierarchy, semantic pruning, and branch regrowth to automate decomposition-enhanced CVRP solver design. Its central claim rests on empirical experiments showing the generated solvers surpass SOTA methods on CVRPLib instances. No equations, derivations, fitted parameters renamed as predictions, or self-citations appear in the provided text. The algorithmic description does not reduce to its inputs by construction, nor does it invoke uniqueness theorems or ansatzes from prior self-work. The evaluation uses an external benchmark library, making the result independent of any internal fitting loop.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on assumptions about LLM code-generation reliability and the effectiveness of the introduced search controls; no explicit free parameters or new physical entities are described in the abstract.

axioms (1)

domain assumption LLMs can produce valid and high-performing decomposition policies and sub-solver code when guided by the three-tier hierarchy and search mechanisms
This is the core premise enabling the automated design process for LSCVRP.

pith-pipeline@v0.9.0 · 5505 in / 1311 out tokens · 71308 ms · 2026-05-07T16:41:44.515862+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

40 extracted references · 9 canonical work pages · 2 internal anchors

[1]

Routefinder: Towards founda- tion models for vehicle routing problems

[Bertoet al., ] Federico Berto, Chuanbo Hua, Nayeli Gast Zepeda, Andr ´e Hottung, Niels Wouda, Leon Lan, Kevin Tierney, and Jinkyoo Park. Routefinder: Towards founda- tion models for vehicle routing problems. InICML 2024 Workshop on Foundation Models in the Wild. [Caceres-Cruzet al., 2014 ] Jose Caceres-Cruz, Pol Arias, Daniel Guimarans, Daniel Riera, and...

2024
[2]

City vehicle routing problem (city vrp): A review

Heng, Puay Siew Tan, and Nengsheng Allan Zhang. City vehicle routing problem (city vrp): A review. IEEE Transactions on Intelligent Transportation Systems, 16(4):1654–1666, 2015. [Kusneret al., 2015 ] Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to docu- ment distances. InInternational conference on machine learning, p...

2015
[2]

Guided local search with an adaptive neighbourhood size heuristic for large scale ve- hicle routing problems

[Costaet al., 2022 ] Joao Guilherme Cavalcanti Costa, Yi Mei, and Mengjie Zhang. Guided local search with an adaptive neighbourhood size heuristic for large scale ve- hicle routing problems. InProceedings of the Genetic and Evolutionary Computation Conference, pages 213–221,

2022
[3]

Competition-level code generation with alphacode

Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, et al. Competition-level code generation with alphacode. Science, 378(6624):1092–1097, 2022. [Liet al., 2024 ] Han Li, Fei Liu, Zhi Zheng, Yu Zhang, and Zhenkun Wang. Cada: Cross-problem routing solver with constraint-aware dual-attention, 2024. [Liuet al., 2023 ] Fei Liu, Chengyu Lu, Lin Gui, Qingfu

2022
[3]

The truck dispatching problem.Manage

[Dantzig and Ramser, 1959] George B Dantzig and John H Ramser. The truck dispatching problem.Manage. Sci., 6(1):80–91,

1959
[4]

Large language models as end-to-end combinatorial optimization solvers.arXiv preprint arXiv:2509.16865, 2025

[Jianget al., 2025a ] Xia Jiang, Yaoxin Wu, Minshuo Li, Zhiguang Cao, and Yingqian Zhang. Large language models as end-to-end combinatorial optimization solvers. arXiv preprint arXiv:2509.16865,

work page arXiv
[5]

Formulations and exact algorithms for the vehicle routing problem with time windows.Computers & Operations Research, 35(7):2307–2330,

[Kallehauge, 2008] Brian Kallehauge. Formulations and exact algorithms for the vehicle routing problem with time windows.Computers & Operations Research, 35(7):2307–2330,

2008
[6]

City vehicle routing problem (city vrp): A review

[Kimet al., 2015 ] Gitae Kim, Yew-Soon Ong, Chen Kim Heng, Puay Siew Tan, and Nengsheng Allan Zhang. City vehicle routing problem (city vrp): A review. IEEE Transactions on Intelligent Transportation Systems, 16(4):1654–1666,

2015
[7]

Large language mod- els as evolutionary optimizers

Qu, Ke Tang, and Yew-Soon Ong. Large language mod- els as evolutionary optimizers. In2024 IEEE Congress on Evolutionary Computation (CEC), pages 1–8. IEEE, 2024. [Liuet al., 2025 ] Aixin Liu, Aoxue Mei, Bangcai Lin, Bing

2024
[7]

From word embeddings to docu- ment distances

[Kusneret al., 2015 ] Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to docu- ment distances. InInternational conference on machine learning, pages 957–966. PMLR,

2015
[8]

Learning to delegate for large-scale vehicle routing

[Liet al., 2021 ] Sirui Li, Zhongxia Yan, and Cathy Wu. Learning to delegate for large-scale vehicle routing. Advances in Neural Information Processing Systems, 34:26198–26211,

2021
[9]

Mathematical discoveries from program search with large language models.Nature, 625(7995):468–475, 2024

Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, et al. Mathematical discoveries from program search with large language models.Nature, 625(7995):468–475, 2024. [Santiniet al., 2023 ] Alberto Santini, Michael Schneider, Thibaut Vidal, and Daniele Vigo. Decomposition strate- gies for vehicle routing heuristics.INFORMS Journal on Computing, 35(3):543–55...

2024
[9]

Competition-level code generation with alphacode

[Liet al., 2022 ] Yujia Li, David Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, R ´emi Leblond, Tom Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, et al. Competition-level code generation with alphacode. Science, 378(6624):1092–1097,

2022
[10]

New benchmark instances for the capacitated ve- hicle routing problem.European Journal of Operational Research, 257(3):845–858, 2017

Pessoa, Marcus Poggi, Thibaut Vidal, and Anand Subra- manian. New benchmark instances for the capacitated ve- hicle routing problem.European Journal of Operational Research, 257(3):845–858, 2017. [Vayeret al., 2020 ] Titouan Vayer, Laetitia Chapel, R ´emi

2017
[10]

Cada: Cross-problem routing solver with constraint-aware dual-attention,

[Liet al., 2024 ] Han Li, Fei Liu, Zhi Zheng, Yu Zhang, and Zhenkun Wang. Cada: Cross-problem routing solver with constraint-aware dual-attention,

2024
[11]

Heuristics for vehicle routing problem: A survey and recent advances

[Liuet al., 2023 ] Fei Liu, Chengyu Lu, Lin Gui, Qingfu Zhang, Xialiang Tong, and Mingxuan Yuan. Heuristics for vehicle routing problem: A survey and recent advances. arXiv preprint arXiv:2303.04147,

work page arXiv 2023
[12]

Multi-task learning for routing problem with cross-problem zero-shot generalization

[Liuet al., 2024a ] Fei Liu, Xi Lin, Zhenkun Wang, Qingfu Zhang, Tong Xialiang, and Mingxuan Yuan. Multi-task learning for routing problem with cross-problem zero-shot generalization. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1898–1908,

1908
[13]

Evolution of heuristics: Towards efficient automatic algorithm design using large language model,

[Liuet al., 2024b ] Fei Liu, Xialiang Tong, Mingxuan Yuan, Xi Lin, Fu Luo, Zhenkun Wang, Zhichao Lu, and Qingfu Zhang. Evolution of heuristics: Towards efficient auto- matic algorithm design using large language model.arXiv preprint arXiv:2401.02051,

work page arXiv
[14]

Llm4ad: A platform for algorithm design with large language model

[Liuet al., 2024c ] Fei Liu, Rui Zhang, Zhuoliang Xie, Rui Sun, Kai Li, Xi Lin, Zhenkun Wang, Zhichao Lu, and Qingfu Zhang. Llm4ad: A platform for algo- rithm design with large language model.arXiv preprint arXiv:2412.17287,

work page arXiv
[15]

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

[Liuet al., 2025 ] Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, et al. Deepseek-v3. 2: Pushing the frontier of open large language models.arXiv preprint arXiv:2512.02556,

work page internal anchor Pith review arXiv 2025
[16]

Vrpdiv: a divide and conquer framework for large ve- hicle routing problems.ACM Transactions on Spatial Algorithms and Systems (TSAS), 7(4):1–41,

[Mariescu-Istodoret al., 2021 ] Radu Mariescu-Istodor, Alexandru Cristian, Mihai Negrea, and Peiwei Cao. Vrpdiv: a divide and conquer framework for large ve- hicle routing problems.ACM Transactions on Spatial Algorithms and Systems (TSAS), 7(4):1–41,

2021
[17]

Decomposition-based memetic algorithm for multiobjec- tive capacitated arc routing problem.IEEE Transactions on Evolutionary Computation, 15(2):151–165,

[Meiet al., 2011 ] Yi Mei, Ke Tang, and Xin Yao. Decomposition-based memetic algorithm for multiobjec- tive capacitated arc routing problem.IEEE Transactions on Evolutionary Computation, 15(2):151–165,

2011
[18]

Vehicle routing problems over time: a survey

[Mor and Speranza, 2022] Andrea Mor and Maria Grazia Speranza. Vehicle routing problems over time: a survey. Annals of Operations Research, 314(1):255–275,

2022
[19]

Multi- task vehicle routing solver via mixture of specialized experts under state-decomposable mdp.arXiv preprint arXiv:2510.21453,

[Panet al., 2025 ] Yuxin Pan, Zhiguang Cao, Chengyang Gu, Liu Liu, Peilin Zhao, Yize Chen, and Fangzhen Lin. Multi- task vehicle routing solver via mixture of specialized experts under state-decomposable mdp.arXiv preprint arXiv:2510.21453,

work page arXiv 2025
[20]

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

[Perron and Furnon, ] Laurent Perron and Vincent Furnon. Or-tools. [Reimers and Gurevych, 2019] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks.arXiv preprint arXiv:1908.10084,

work page internal anchor Pith review arXiv 2019
[21]

Mathematical discoveries from program search with large language models.Nature, 625(7995):468–475,

[Romera-Paredeset al., 2024 ] Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco JR Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, et al. Mathematical discoveries from program search with large language models.Nature, 625(7995):468–475,

2024
[22]

Decomposition strate- gies for vehicle routing heuristics.INFORMS Journal on Computing, 35(3):543–559,

[Santiniet al., 2023 ] Alberto Santini, Michael Schneider, Thibaut Vidal, and Daniele Vigo. Decomposition strate- gies for vehicle routing heuristics.INFORMS Journal on Computing, 35(3):543–559,

2023
[23]

Optimal transport for structured data with application on graphs

[Titouanet al., 2019 ] Vayer Titouan, Nicolas Courty, Ro- main Tavenard, and R ´emi Flamary. Optimal transport for structured data with application on graphs. InInter- national Conference on Machine Learning, pages 6275–

2019
[24]

New benchmark instances for the capacitated ve- hicle routing problem.European Journal of Operational Research, 257(3):845–858,

[Uchoaet al., 2017 ] Eduardo Uchoa, Diego Pecin, Artur Pessoa, Marcus Poggi, Thibaut Vidal, and Anand Subra- manian. New benchmark instances for the capacitated ve- hicle routing problem.European Journal of Operational Research, 257(3):845–858,

2017
[25]

Fused gromov-wasserstein distance for structured objects.Algo- rithms, 13(9):212,

[Vayeret al., 2020 ] Titouan Vayer, Laetitia Chapel, R ´emi Flamary, Romain Tavenard, and Nicolas Courty. Fused gromov-wasserstein distance for structured objects.Algo- rithms, 13(9):212,

2020
[26]

A unified solution framework for multi-attribute vehicle routing problems

[Vidalet al., 2014 ] Thibaut Vidal, Teodor Gabriel Crainic, Michel Gendreau, and Christian Prins. A unified solution framework for multi-attribute vehicle routing problems. European Journal of Operational Research, 234(3):658– 673,

2014
[27]

Hybrid genetic search for the cvrp: Open-source implementation and swap* neighbor- hood.Computers & Operations Research, 140:105643,

[Vidal, 2022] Thibaut Vidal. Hybrid genetic search for the cvrp: Open-source implementation and swap* neighbor- hood.Computers & Operations Research, 140:105643,

2022
[28]

Pyvrp: A high-performance vrp solver package.IN- FORMS Journal on Computing, 36(4):943–955,

[Woudaet al., 2024 ] Niels A Wouda, Leon Lan, and Wouter Kool. Pyvrp: A high-performance vrp solver package.IN- FORMS Journal on Computing, 36(4):943–955,

2024
[29]

Improving word mover’s distance by leveraging self-attention matrix

[Yamagiwaet al., 2023 ] Hiroaki Yamagiwa, Sho Yokoi, and Hidetoshi Shimodaira. Improving word mover’s distance by leveraging self-attention matrix. InFindings of the As- sociation for Computational Linguistics: EMNLP 2023, pages 11160–11183,

2023
[30]

Deepaco: Neural-enhanced ant systems for combinatorial optimization.Advances in neural information processing systems, 36:43706–43728,

[Yeet al., 2023 ] Haoran Ye, Jiarui Wang, Zhiguang Cao, Helan Liang, and Yong Li. Deepaco: Neural-enhanced ant systems for combinatorial optimization.Advances in neural information processing systems, 36:43706–43728,

2023
[31]

Reevo: Large language models as hyper-heuristics with reflective evolution.Advances in neural information processing systems, 37:43571–43608,

[Yeet al., 2024 ] Haoran Ye, Jiarui Wang, Zhiguang Cao, Federico Berto, Chuanbo Hua, Haeyeon Kim, Jinkyoo Park, and Guojie Song. Reevo: Large language models as hyper-heuristics with reflective evolution.Advances in neural information processing systems, 37:43571–43608,

2024
[32]

An agentic framework with llms for solving complex vehicle routing problems

[Zhanget al., 2025 ] Ni Zhang, Zhiguang Cao, Jianan Zhou, Cong Zhang, and Yew-Soon Ong. An agentic framework with llms for solving complex vehicle routing problems. arXiv preprint arXiv:2510.16701,

work page arXiv 2025
[33]

Udc: A uni- fied neural divide-and-conquer framework for large-scale combinatorial optimization problems.Advances in Neural Information Processing Systems, 37:6081–6125,

[Zhenget al., 2024 ] Zhi Zheng, Changliang Zhou, Tong Xi- aliang, Mingxuan Yuan, and Zhenkun Wang. Udc: A uni- fied neural divide-and-conquer framework for large-scale combinatorial optimization problems.Advances in Neural Information Processing Systems, 37:6081–6125,

2024
[34]

Monte carlo tree search for comprehensive explo- ration in llm-based automatic heuristic design,

[Zhenget al., 2025a ] Zhi Zheng, Zhuoliang Xie, Zhenkun Wang, and Bryan Hooi. Monte carlo tree search for com- prehensive exploration in llm-based automatic heuristic design.arXiv preprint arXiv:2501.08603,

work page arXiv
[35]

Mvmoe: Multi-task vehicle routing solver with mixture-of-experts

[Zhouet al., 2024 ] Jianan Zhou, Zhiguang Cao, Yaoxin Wu, Wen Song, Yining Ma, Jie Zhang, and Chi Xu. Mvmoe: Multi-task vehicle routing solver with mixture-of-experts. In41st International Conference on Machine Learning, ICML 2024, pages 61804–61824. PMLR, 2024

2024