Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

RedAHD: Reduction-Based End-to-End Automatic Heuristic Design with Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read LLMs can now invent their own reductions to solve NP-hard problems

desk verdict A genuinely end-to-end LLM-EPS framework with useful new components, but the 'reduction' premise is weaker than it looks and the reproducibility details are lacking. read the letter →

arxiv 2505.20242 v1 pith:5SE2N6NE submitted 2025-05-26 cs.LG

Nguyen Thach , Aida Riahifar , Nathan Huynh , Hau Chan This is my paper · ORCID
classification cs.LG
keywords automaticheuristicdesignlargelanguagemodelscombinatorialoptimizationproblemreductionLLM-basedevolutionaryprogramsearchend-to-endmulti-problem
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces RedAHD, a framework that makes LLM-based automatic heuristic design fully end-to-end. Its central proposal is that an LLM can perform the reduction step of algorithm design: transform the combinatorial optimization problem (COP) at hand into a similar, better-understood problem, and implement the two functions that map instances and solutions between them. Existing evolutionary LLM heuristic-search methods can then design a heuristic for the transformed problem, and its solutions are mapped back and scored on the original problem. If this works, practitioners no longer have to hand-build a generalized algorithmic framework such as ant colony optimization or guided local search for each new COP. The paper reports competitive or improved results over state-of-the-art LLM-based heuristic designers across six COPs, with minimal human involvement.

What carries the argument

The load-bearing object is the language reduction: an approximate reduction from COP $A$ to COP $B$ in which both the instance map $f$ and the solution map $g$ are LLM-generated code rather than human-specified framework components. It replaces the generalized algorithmic framework (GAF), and it carries the argument because it converts an arbitrary COP into a form that off-the-shelf LLM-EPS can solve directly; the score that drives evolution is still $A$'s objective after mapping solutions back through $g$. The second mechanism is multi-problem LLM-EPS, which samples reductions by score and allows any heuristic in the population to serve as a parent for any target problem, and the third is reduction refinement, which re-prompts the LLM to update $f$, $g$, and the code template after a fixed number of generations without score improvement.

What would settle it

For each COP, take the final reductions selected by RedAHD and compare the best heuristic found through a non-identity reduction with a heuristic found by the same LLM-EPS method when the reduction is forced to be the identity map. If forced-identity search performs equally well on the same budget, then the reduction step is not what carries RedAHD's performance. A direct inspection also works: the MKP example in Figure S7 maps solutions back with a passthrough, so replacing the transformation by the identity would change almost nothing.

Watch

Extended reading notes

Core claim

The central claim is that the manually designed generalized algorithmic framework (GAF), which every previous LLM-based evolutionary program search method required, can be replaced by an LLM-generated reduction. Formally, for a COP $A$ and a target COP $B$, the LLM proposes functions $f$ (mapping instances of $A$ to instances of $B$) and $g$ (mapping solutions of $B$ back to feasible solutions of $A$), and a heuristic $h'$ is evolved directly for $B$; the fitness of $h'$ is evaluated as $q(x, g(h'(f(x))))$ on the original problem's objective $q$. The paper calls this pair a language reduction and treats it as approximate: it does not enforce any approximation-ratio guarantee. RedAHD maintains several such reductions at once, evolves heuristics jointly across the corresponding target problems so that ideas cross from one to another, and auto-refines a reduction when its associated heuristics stagnate. Reported experiments on TSP, CVRP, KP, MKP, BPP, and OBPP show RedAHD matching or beating LLM-EPS methods that operate inside ACO or iterative-construction frameworks.

Load-bearing premise

The paper treats a reduction as valid as long as the solutions it produces are feasible, without requiring the mapped-back solutions to be anywhere near optimal; if the LLM's reductions are just re-descriptions of the original problem, the framework's claimed benefit collapses.

Editorial extensions

If this is right

  • A new COP can be attacked by giving RedAHD only its description; the human no longer chooses or codes a search framework for it.
  • The same heuristic-design engine can serve many COPs at once, because reductions park them all in target problems the search already understands.
  • Heuristic ideas transfer across problems: mutation and crossover can borrow algorithmic motifs from one reduction's heuristics when designing for another.
  • Stronger designer LLMs or stronger LLM-EPS engines should improve RedAHD without any change to the reduction loop, as the paper demonstrates with a reasoning model and two alternative search methods.
  • Even poor initial reductions are not fatal: stagnation-triggered refinement can repair them during search.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable implication of accepting feasibility-only validity is that the framework could be gamed by near-identity reductions: if $B$ is just $A$ with renamed variables, the 'reduction' adds no new algorithmic leverage, so RedAHD's gains would have to come from multi-problem search alone. I would test this by comparing top-scoring LRs against identity mappings on the same budget.
  • The multi-problem crossover idea suggests a broader principle: heuristics for one COP can serve as mutations for another whenever a (possibly loose) reduction connects them. One could exploit this intentionally by seeding a fixed library of solved benchmark COPs as permanent reduction targets.
  • Because reductions are approximate, RedAHD's output heuristics inherit no worst-case guarantee; for applications requiring provable bounds, the framework's reductions would need an approximation-preserving check that the current validity test does not provide.
  • The stagnation-based refinement loop implicitly defines a meta-heuristic over reductions; a natural extension would be to let the LLM propose completely new reductions during search rather than only refine existing ones.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes RedAHD, an end-to-end automatic heuristic design framework that uses LLMs to generate 'language reductions' (LRs): for a target combinatorial optimization problem A, the LLM produces a description of a related problem B together with functions f (mapping instances of A to instances of B) and g (mapping solutions of B back to A). Existing LLM-EPS methods such as EoH are then run on B, and the resulting heuristics are evaluated indirectly on A through g. RedAHD maintains multiple LRs, scores them by the objective of A, refines them when their scores stagnate, and allows crossover/mutation across heuristics of different LRs. Experiments on TSP, CVRP, KP, MKP, BPP, and OBPP report competitive or improved objective values relative to LLM-EPS baselines operating under IC or ACO frameworks, with ablations showing that both the multi-LR setup (M=3 vs. M=1) and the reduction-refinement mechanism contribute to performance.

Significance. If the reduction premise holds, RedAHD addresses a genuine limitation of current LLM-EPS methods: the need for a human-specified generalized algorithmic framework (GAF). The framework is coherent, the iterative refinement and multi-problem evolution are sensible, and the TSPLib results are numerically striking. The ablations in Tables 6 and S12 give explicit evidence that the two new components (multi-problem search and reduction refinement) matter. The paper also verifies solution feasibility with per-instance checks, which is a concrete safeguard against invalid heuristics. However, the significance hinges on whether the LLM-generated pairs (f,g) are genuine reductions to 'better-understood' problems rather than identity-like re-descriptions; the current evidence for that premise is weak, and the comparative experiments lack variance estimates and rerun baselines, so the central empirical claim is not yet fully established.

major comments (4)
  1. [Section 4.1 / Definition 2 / Appendix C.6] The acceptance criterion for an LR is solution feasibility alone: Section 4.1 states that as long as the returned solutions are valid, the reduction is deemed valid, and Definition 2 explicitly allows approximate reductions without a performance-ratio guarantee. This criterion cannot support the central claim that RedAHD transforms A into a 'better-understood' COP B: a reduction that maps A to an unrelated or trivially different B whose solutions are feasible but arbitrarily poor for A would be accepted. The MKP example in Figure S7 illustrates the risk: convert_input_A_to_B returns the original values, weights, and constraints with an appended ratios array, and convert_solution_B_to_A is the identity on the selected item indices; B is essentially MKP with a greedy value/weight-ratio hint, not a distinct better-understood problem. The paper should either report and analyze the selected LRs for all six COPs (e.g., measure how often they are identity-like re-descriptions and show that g maps good B-solutions to good A-solutions in a non-trivial sense) or substantially weaken the reduction claim and demonstrate the advantage over direct multi-prompt or diverse-population LLM-EPS without reductions. As written, the M=1 vs. M=3 ablation shows that diversity helps, but it does not distinguish genuine reduction from prompt-induced diversity.
  2. [Tables 3, 5, 8, and S12] The main comparative tables report each RedAHD result as a single average over three runs, without standard deviations, and the LLM-EPS baselines are copied from prior work (as stated in the table captions and Section 5.1) rather than rerun under the same evaluation protocol and computational budget. Several head-to-head differences are small (e.g., TSP n=50 in Table 5: RedAHD 5.819 vs. MCTS-AHD 5.801; CVRP n=50 in Table 5: RedAHD 9.826 vs. EoH 9.359), so without variance estimates or paired significance tests the claimed competitive or improved performance is not statistically established. Please report run-level statistics and, if feasible, rerun or re-evaluate the baselines under identical conditions.
  3. [Table 4 and Section 5.2] The TSPLib results are obtained by first selecting the best-performing heuristic among the three RedAHD runs and then averaging three executions of that selected heuristic with different random starting nodes. This protocol introduces selection bias over runs, and the reported plus/minus intervals reflect only the three random executions of the chosen heuristic, not the run-to-run variation of RedAHD itself. The comparison with baselines should use a consistent selection rule, or the full distribution across all runs and heuristics should be reported.
  4. [Table 7 / Section 5.3] The claim that RedAHD with o3-mini achieves a CVRP objective of 13.516 on the OOD setting, 'even better than' OR-Tools, is an important headline result, but no information is given about OR-Tools' configuration, time budget, or whether it was run with a comparable per-instance time limit. Please state the OR-Tools settings and the computational budgets for both methods, since the per-heuristic evaluation cap is reported as 60 seconds in Appendix C.1.
minor comments (5)
  1. [Appendix C.4 / Figure S7] The notation 'Problem B1' in the MKP example is confusing because the LR score and selection process in Section 4.1 refer to LRs by subscript j; consider using the same indexing throughout the supplementary examples.
  2. [Definition 2] The term 'approximate reduction' is used here to mean a reduction without a performance-ratio guarantee, which differs from the standard approximation-preserving reduction terminology in the cited reference [9]; a brief note on this terminological choice would avoid confusion.
  3. [Section 2] There is a stray parenthesis in the sentence 'Please refer to the latest survey by [27]) for a comprehensive review' that should be removed.
  4. [Appendix C.1] The statement that the code 'will be uploaded to GitHub upon publication' means the experiments are not currently reproducible from the manuscript; I encourage releasing the code, generated LRs, and evaluation scripts with the revision, especially because the reduction examples are central to assessing the method's claims.
  5. [Section 4.2 / Figure 3] The claimed benefit of cross-LR heuristic transfer is illustrated by a single anecdotal TSP example; a quantitative example showing that offspring heuristics inherited ideas across different B_j would strengthen the narrative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RedAHD is an empirical LLM search method evaluated on external benchmarks; the feasibility-only acceptance of reductions is a limitation, not a circular derivation.

full rationale

RedAHD makes no first-principles derivation claim; it is an empirical search method. The reductions (f,g) are LLM-generated code, not fitted parameters, and the score s_j is computed from the original COP objective q on fixed benchmark instances (Sec. 4.1: Q = (1/D) \sum_k q(x^(k), g_j(h_i^(j)(f_j(x^(k)))))). The reported results are objective values of final heuristics on unseen in-distribution and out-of-distribution test sets in Tables 3-5, not quantities used to fit the method. The selection and refinement of LRs using the same heuristic fitness is a feedback loop within an evolutionary search, not a derivation of the result from the result, because the fitness values come from external problem instances and the problem's own objective. Feasibility-only acceptance of LRs (Sec. 4.1, Definition 2, Appendix C.6) weakens the 'better-understood B' premise, and the MKP example in Figure S7 is a near-identity re-description of MKP, but these are correctness/scope limitations rather than circularity: no equation reduces to itself and no load-bearing claim is justified only by a self-citation, since the cited LLM-EPS baselines are by other author groups.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no data-fitted constants. The hand-chosen hyperparameters M, Minit, l, T and the 60-second evaluation cap are not ablated, so results may be sensitive to them. The framework additionally assumes that LLM sampling will produce valid reductions, that validity on training data implies useful reductions, and that imported baseline numbers are comparable.

free parameters (5)
  • M (number of surviving language reductions) = 3
    Hand-chosen in Section 5.1 with no sensitivity analysis; the main results depend on this value.
  • Minit (number of candidate reductions generated) = 10
    Hand-chosen in Section 5.1; no ablation varies the number of candidate reductions.
  • l (number of top heuristics used to score each reduction) = 3
    Hand-chosen in Section 5.1; affects LR selection and stagnation tracking.
  • T (stagnation threshold in generations before refinement) = 3
    Hand-chosen in Sections 4.3 and 5.1; no sensitivity analysis is provided.
  • per-heuristic evaluation time cap = 60 seconds
    Set in Appendix C.1; this budget affects whether stronger local search such as 2-opt can run on large TSPLib instances.
assumptions (4)
  • ad hoc to paper The LLM can generate at least one valid and useful reduction among Minit candidates for each COP.
    The entire framework depends on stochastic LLM outputs from the prompt templates in Section 4.1; no formal guarantee or diversity analysis is provided.
  • domain assumption Solution validity on the finite evaluation dataset suffices to certify a reduction as useful.
    Section 4.1 accepts any reduction whose composed solutions are valid on evaluated instances; Definition 2 deliberately drops approximation-ratio guarantees, so unseen instances may reveal invalid or unhelpful reductions.
  • ad hoc to paper Stagnation of the LR score for T generations is a reliable trigger for refining reductions.
    Section 4.3 sets T=3; only one ablation (Table 6) supports the refinement step, not the specific threshold.
  • domain assumption Baseline results from MCTS-AHD and related papers are directly comparable despite differences in test sets and computational budgets.
    Numbers are imported from prior papers in Section 5.2; Table 2 itself notes different test-set sizes for IC versus ACO settings, so cross-table comparisons inherit this hazard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RedAHD: Reduction-Based End-to-End Automatic Heuristic Design with Large Language Models." pith.science (2026). https://pith.science/paper/5SE2N6NE

@misc{pith2026250520242,
  author       = {Pith},
  title        = {Pith review of: RedAHD: Reduction-Based End-to-End Automatic Heuristic Design with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5SE2N6NE}},
  note         = {Machine review of arXiv:2505.20242}
}
read the original abstract

Solving NP-hard combinatorial optimization problems (COPs) (e.g., traveling salesman problems (TSPs) and capacitated vehicle routing problems (CVRPs)) in practice traditionally involves handcrafting heuristics or specifying a search space for finding effective heuristics. The main challenges from these approaches, however, are the sheer amount of domain knowledge and implementation efforts required from human experts. Recently, significant progress has been made to address these challenges, particularly by using large language models (LLMs) to design heuristics within some predetermined generalized algorithmic framework (GAF, e.g., ant colony optimization and guided local search) for building key functions/components (e.g., a priori information on how promising it is to include each edge in a solution for TSP and CVRP). Although existing methods leveraging this idea have shown to yield impressive optimization performance, they are not fully end-to-end and still require considerable manual interventions. In this paper, we propose a novel end-to-end framework, named RedAHD, that enables these LLM-based heuristic design methods to operate without the need of GAFs. More specifically, RedAHD employs LLMs to automate the process of reduction, i.e., transforming the COP at hand into similar COPs that are better-understood, from which LLM-based heuristic design methods can design effective heuristics for directly solving the transformed COPs and, in turn, indirectly solving the original COP. Our experimental results, evaluated on six COPs, show that RedAHD is capable of designing heuristics with competitive or improved results over the state-of-the-art methods with minimal human involvement.

Figures

Figures reproduced from arXiv: 2505.20242 by the authors.

Figure 1
Figure 1. Timeline of LLM-EPS methods developed thus far. for iterative refinement of heuristics, also known as LLM-based evolutionary program search (LLM￾EPS) [28, 11], has attracted increasing attention. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of RedAHD. First, the designer LLM generates a set of LRs, encoded as two reduction functions (one for mapping instances and the other for mapping solutions of A to B). The LRs are then used to generate a set of heuristics that are iteratively refined using existing LLM-EPS methods, in which offspring heuristics of an LR may be generated using algorithmic ideas from heuristics of any other LRs. When the… view at source ↗
Figure 3
Figure 3. A demonstration of multi-problem LLM-EPS for TSP, in which the parent heuristic (blue) during EoH mutation [26] is not intended to solve the COP at hand (“Problem B3”). As a result, the offspring heuristic for B3 (green) is generated with the novel idea of 2-opt edge swap and hence yields better performance. Once the set of LRs R and the resulting set of heuristics P are initialized, the evolutionary search procedur… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    cs.AI 2026-07 conditional novelty 6.5 of 10

    A coupled bilevel LLM search that specializes repair heuristics to local regions within one routing solution cuts held-out cost by up to 57.7% versus competing AHD methods.

Reference graph

Works this paper leans on

48 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

    A computation investigation of the impact of convex hull subtour on the nearest neighbour heuristic

    Emmanuel O Asani, Aderemi E Okeyinka, and Ayodele Ariyo Adebiyi. A computation investigation of the impact of convex hull subtour on the nearest neighbour heuristic. In 2023 International Conference on Science, Engineering and Business for Sustainable Development Goals (SEB-SDG), volume 1, pages 1–7. IEEE, 2023

    work page 2023

  2. [2]

    Handbook of evolutionary computa- tion

    Thomas Bäck, David B Fogel, and Zbigniew Michalewicz. Handbook of evolutionary computa- tion. Release, 97(1):B1, 1997

    1997

  3. [3]

    Machine learning for combinatorial optimization: a methodological tour d’horizon

    Yoshua Bengio, Andrea Lodi, and Antoine Prouvost. Machine learning for combinatorial optimization: a methodological tour d’horizon. European Journal of Operational Research, 290(2):405–421, 2021

    2021

  4. [4]

    The approximation ratio of the greedy algorithm for the metric traveling salesman problem

    Judith Brecklinghaus and Stefan Hougardy. The approximation ratio of the greedy algorithm for the metric traveling salesman problem. Operations Research Letters, 43(3):259–261, 2015

    work page 2015

  5. [5]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  6. [6]

    Hyper-heuristics: A survey of the state of the art

    Edmund K Burke, Michel Gendreau, Matthew Hyde, Graham Kendall, Gabriela Ochoa, Ender Özcan, and Rong Qu. Hyper-heuristics: A survey of the state of the art. Journal of the Operational Research Society, 64(12):1695–1724, 2013

    work page 2013

  7. [7]

    Efficient meta neural heuristic for multi-objective combinatorial optimization

    Jinbiao Chen, Jiahai Wang, Zizhen Zhang, Zhiguang Cao, Te Ye, and Siyuan Chen. Efficient meta neural heuristic for multi-objective combinatorial optimization. Advances in Neural Information Processing Systems, 36:56825–56837, 2023

    work page 2023

  8. [8]

    Worst-case analysis of a new heuristic for the travelling salesman problem

    Nicos Christofides. Worst-case analysis of a new heuristic for the travelling salesman problem. In Operations Research Forum, volume 3, page 20. Springer, 2022

    work page 2022

Show all 48 references
  1. [9]

    A short guide to approximation preserving reductions

    Pierluigi Crescenzi. A short guide to approximation preserving reductions. In Proceedings of Computational Complexity. Twelfth Annual IEEE Conference, pages 262–273. IEEE, 1997

    1997

  2. [10]

    The truck dispatching problem

    George B Dantzig and John H Ramser. The truck dispatching problem. Management science, 6(1):80–91, 1959

    1959

  3. [11]

    Hsevo: Elevating automatic heuristic design with diversity-driven harmony search and genetic algorithm using llms

    Pham Vu Tuan Dat, Long Doan, and Huynh Thi Thanh Binh. Hsevo: Elevating automatic heuristic design with diversity-driven harmony search and genetic algorithm using llms. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 26931–26938, 2025

    2025

  4. [12]

    Heuristic and meta-heuristic algorithms and their relevance to the real world: a survey

    Sachin Desale, Akhtar Rasool, Sushil Andhale, and Priti Rane. Heuristic and meta-heuristic algorithms and their relevance to the real world: a survey. Int. J. Comput. Eng. Res. Trends, 351(5):2349–7084, 2015

    2015

  5. [13]

    Ant colony optimization

    Marco Dorigo, Mauro Birattari, and Thomas Stutzle. Ant colony optimization. IEEE computa- tional intelligence magazine, 1(4):28–39, 2007

    2007

  6. [14]

    Recent advances in selection hyper-heuristics

    John H Drake, Ahmed Kheiri, Ender Özcan, and Edmund K Burke. Recent advances in selection hyper-heuristics. European Journal of Operational Research, 285(2):405–428, 2020

    2020

  7. [15]

    A gp hyper-heuristic approach for generating tsp heuristics

    Gabriel Duflo, Emmanuel Kieffer, Matthias R Brust, Grégoire Danoy, and Pascal Bouvry. A gp hyper-heuristic approach for generating tsp heuristics. In 2019 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pages 521–529. IEEE, 2019. 10

    2019

  8. [16]

    From evolutionary computation to the evolution of things

    Agoston E Eiben and Jim Smith. From evolutionary computation to the evolution of things. Nature, 521(7553):476–482, 2015

    2015

  9. [17]

    Or-tools routing library

    Vincent Furnon and Laurent Perron. Or-tools routing library

  10. [18]

    Connecting large language models with evolutionary algorithms yields powerful prompt optimizers

    Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang Bian, and Yujiu Yang. Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. In The Twelfth International Conference on Learning Representations, 2024

    2024

  11. [19]

    Evolving code with a large language model

    Erik Hemberg, Stephen Moskal, and Una-May O’Reilly. Evolving code with a large language model. Genetic Programming and Evolvable Machines, 25(2):21, 2024

    2024

  12. [20]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qiang- long Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Informat...

    2025

  13. [21]

    Position: Llms can’t plan, but can help planning in llm-modulo frameworks

    Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Mudit Verma, Kaya Stechly, Siddhant Bhambri, Lucas Paul Saldyt, and Anil B Murthy. Position: Llms can’t plan, but can help planning in llm-modulo frameworks. In Forty-first International Conference on Machine Learning, 2024

    2024

  14. [22]

    Pomo: Policy optimization with multiple optima for reinforcement learning

    Yeong-Dae Kwon, Jinho Choo, Byoungjip Kim, Iljoo Yoon, Youngjune Gwon, and Seungjai Min. Pomo: Policy optimization with multiple optima for reinforcement learning. Advances in Neural Information Processing Systems, 33:21188–21198, 2020

    2020

  15. [23]

    Foundations of genetic programming

    William B Langdon and Riccardo Poli. Foundations of genetic programming. Springer Science & Business Media, 2013

    2013

  16. [24]

    Large language models as evolution strategies

    Robert Lange, Yingtao Tian, and Yujin Tang. Large language models as evolution strategies. In Proceedings of the Genetic and Evolutionary Computation Conference Companion, pages 579–582, 2024

    2024

  17. [25]

    Evolution through large models

    Joel Lehman, Jonathan Gordon, Shawn Jain, Kamal Ndousse, Cathy Yeh, and Kenneth O Stanley. Evolution through large models. In Handbook of evolutionary machine learning, pages 331–366. Springer, 2023

    2023

  18. [26]

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

    Fei Liu, Tong Xialiang, Mingxuan Yuan, Xi Lin, Fu Luo, Zhenkun Wang, Zhichao Lu, and Qingfu Zhang. Evolution of heuristics: Towards efficient automatic algorithm design using large language model. In Forty-first International Conference on Machine Learning, 2024

    2024

  19. [27]

    A systematic survey on large language models for algorithm design

    Fei Liu, Yiming Yao, Ping Guo, Zhiyuan Yang, Xi Lin, Xialiang Tong, Mingxuan Yuan, Zhichao Lu, Zhenkun Wang, and Qingfu Zhang. A systematic survey on large language models for algorithm design. arXiv preprint arXiv:2410.14716, 2024

  20. [28]

    Large language models as evolutionary optimizers

    Shengcai Liu, Caishun Chen, Xinghua Qu, Ke Tang, and Yew-Soon Ong. Large language models as evolutionary optimizers. In 2024 IEEE Congress on Evolutionary Computation (CEC), pages 1–8. IEEE, 2024

    2024

  21. [29]

    Metabox: A benchmark platform for meta-black-box optimization with reinforcement learning

    Zeyuan Ma, Hongshu Guo, Jiacheng Chen, Zhenrui Li, Guojun Peng, Yue-Jiao Gong, Yining Ma, and Zhiguang Cao. Metabox: A benchmark platform for meta-black-box optimization with reinforcement learning. Advances in Neural Information Processing Systems, 36:10775–10795, 2023

    2023

  22. [30]

    Dissociating language and thought in large language models

    Kyle Mahowald, Anna A Ivanova, Idan A Blank, Nancy Kanwisher, Joshua B Tenenbaum, and Evelina Fedorenko. Dissociating language and thought in large language models. Trends in cognitive sciences, 2024

    2024

  23. [31]

    Traveling salesman problem: an overview of applications, formulations, and solution approaches

    Rajesh Matai, Surya Prakash Singh, and Murari Lal Mittal. Traveling salesman problem: an overview of applications, formulations, and solution approaches. Traveling salesman problem, theory and applications, 1(1):1–25, 2010

    2010

  24. [32]

    Explainable artificial intelli- gence by genetic programming: A survey

    Yi Mei, Qi Chen, Andrew Lensen, Bing Xue, and Mengjie Zhang. Explainable artificial intelli- gence by genetic programming: A survey. IEEE Transactions on Evolutionary Computation, 27(3):621–641, 2022. 11

    2022

  25. [33]

    Language model crossover: Variation through few-shot prompting

    Elliot Meyerson, Mark J Nelson, Herbie Bradley, Adam Gaier, Arash Moradi, Amy K Hoover, and Joel Lehman. Language model crossover: Variation through few-shot prompting. ACM Transactions on Evolutionary Learning, 4(4):1–40, 2024

    2024

  26. [34]

    Open issues in genetic programming

    Michael O’Neill, Leonardo Vanneschi, Steven Gustafson, and Wolfgang Banzhaf. Open issues in genetic programming. Genetic Programming and Evolvable Machines, 11:339–363, 2010

    2010

  27. [35]

    Hyper-heuristics: theory and applications

    Nelishia Pillay and Rong Qu. Hyper-heuristics: theory and applications. Springer, 2018

    2018

  28. [36]

    Tsplib—a traveling salesman problem library

    Gerhard Reinelt. Tsplib—a traveling salesman problem library. ORSA journal on computing, 3(4):376–384, 1991

    1991

  29. [37]

    Mathematical discoveries from program search with large language models

    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...

    2024

  30. [38]

    An analysis of several heuristics for the traveling salesman problem

    Daniel J Rosenkrantz, Richard E Stearns, and Philip M Lewis, II. An analysis of several heuristics for the traveling salesman problem. SIAM journal on computing, 6(3):563–581, 1977

    1977

  31. [39]

    Guided local search

    Christos V oudouris, Edward PK Tsang, and Abdullah Alsheddy. Guided local search. In Handbook of metaheuristics, pages 321–361. Springer, 2010

    2010

  32. [40]

    Pokorny, Xiao Huang, and Xinrun Wang

    Chang Yang, Ruiyu Wang, Junzhe Jiang, Qi Jiang, Qinggang Zhang, Yanchen Deng, Shuxin Li, Shuyue Hu, Bo Li, Florian T. Pokorny, Xiao Huang, and Xinrun Wang. Nondeterministic polynomial-time problem challenge: An ever-scaling reasoning benchmark for llms, 2025

    2025

  33. [41]

    A survey on reinforcement learning for combinatorial optimization

    Yunhao Yang and Andrew Whinston. A survey on reinforcement learning for combinatorial optimization. In 2023 IEEE World Conference on Applied Intelligence and Computing (AIC), pages 131–136. IEEE, 2023

    2023

  34. [42]

    Multi-objective evolution of heuristic using large language model

    Shunyu Yao, Fei Liu, Xi Lin, Zhichao Lu, Zhenkun Wang, and Qingfu Zhang. Multi-objective evolution of heuristic using large language model. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 27144–27152, 2025

    2025

  35. [43]

    Reevo: Large language models as hyper-heuristics with reflective evolution

    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. In Advances in Neural Information Processing Systems, 2024. https://github. com/ai4co/reevo

    2024

  36. [44]

    Deepaco: Neural-enhanced ant systems for combinatorial optimization

    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

    2023

  37. [45]

    Automated design of metaheuristic algorithms: A survey

    Qi Zhao, Qiqi Duan, Bai Yan, Shi Cheng, and Yuhui Shi. Automated design of metaheuristic algorithms: A survey. arXiv preprint arXiv:2303.06532, 2023

    2023 arXiv

  38. [46]

    A review on edge large language models: Design, execution, and applications

    Yue Zheng, Yuhao Chen, Bin Qian, Xiufang Shi, Yuanchao Shu, and Jiming Chen. A review on edge large language models: Design, execution, and applications. ACM Computing Surveys, 57(8):1–35, 2025

    2025

  39. [47]

    Monte carlo tree search for com- prehensive exploration in llm-based automatic heuristic design

    Zhi Zheng, Zhuoliang Xie, Zhenkun Wang, and Bryan Hooi. Monte carlo tree search for com- prehensive exploration in llm-based automatic heuristic design. In Forty-Second International Conference on Machine Learning, 2025

    2025

  40. [48]

    heuristics

    Zhi Zheng, Changliang Zhou, Tong Xialiang, Mingxuan Yuan, and Zhenkun Wang. Udc: A unified neural divide-and-conquer framework for large-scale combinatorial optimization problems. arXiv preprint arXiv:2407.00312, 2024. 12 A Considered COPs In this appendix, we introduce the co...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.