Pith. sign in

REVIEW 5 major objections 5 minor 44 references

EALG: Evolutionary Adversarial Generation of Language Model-Guided Generators for Combinatorial Optimization

T0 review · 5 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 EALG's adversarial co-evolution loop generates harder combinatorial instances and synthesizes solvers that generalize better than current heuristic-search baselines.

desk verdict A genuinely new co-evolution idea, but the empirical section is an appendix placeholder: no experimental specification, a confounded headline comparison, and no code or data. read the letter →

arxiv 2506.02594 v1 pith:KXQYUHLS submitted 2025-06-03 cs.AI

Ruibo Duan , Yuxin Liu , Xinyao Dong , Chenglin Fan This is my paper
classification cs.AI
keywords combinatorialoptimizationlargelanguagemodelsadversarialco-evolutioninstancegenerationheuristicsynthesistravelingsalesmanproblemorienteeringevolutionaryalgorithms
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

EALG is a framework that uses a large language model to run an automated arms race between two programs: an instance generator that produces harder Traveling Salesman and Orienteering Problem instances, and a heuristic solver that is mutated and improved against those instances. The paper's central claim is that this co-evolution loop produces instance distributions that expose solver weaknesses far more sharply than standard random benchmarks, and heuristic solvers that transfer to unseen, human-designed problems from TSPLIB. If true, this would make benchmark construction and solver design a single automated process, removing human effort from both and giving a principled way to stress-test optimization algorithms. The reported numbers are that four established heuristic search methods degrade to roughly 7 to 13 percent optimality gaps on EALG-generated TSP instances while the EALG solver stays below them, and the EALG solver also achieves the lowest gap on all nine tested TSPLIB instances.

What carries the argument

The engine is a symbolic adversarial co-evolution loop formalized as a minimax game over the space of instance generators $\mathcal{G}$ and heuristic solvers $\mathcal{H}$: $\min_{h\sim \mathcal{L}_{\mathrm{heur}}} \max_{g\sim \mathcal{L}_{\mathrm{inst}}} \mathbb{E}_{z\sim\mathcal{Z}}[H(g(z);h)]$, where hardness $H$ is measured by the relative optimality gap between a heuristic's cost and a reference cost. Mutations are LLM program rewrites, and evolutionary reflection turns observed failures into natural-language feedback embedded in future prompts. This lets the whole process run without gradient-based optimization and without hand-crafted instance templates, which is what allows both the generators and the solvers to adapt to each other's improvements.

What would settle it

Run an exact TSP solver (for example, Concorde) or a single fixed strong reference on the EALG-generated instances and recompute the optimality gap from the paper's formula; if the baseline solvers' gaps collapse toward their standard-dataset levels, the hardness claim is an artifact of a weak reference rather than intrinsic instance difficulty. Likewise, evaluate the EALG-evolved solvers on held-out TSPLIB instances with different seeds and see whether their margin over the baselines persists.

Watch

Extended reading notes

Core claim

The central discovery is a two-loop evolutionary game played in the space of executable programs. A heuristic solver $h$ is sampled from an LLM conditioned on prompts and accumulated reflections; an instance generator $g$ is mutated to maximize the relative optimality gap $\mathrm{Gap}(I;h) = \mathbb{E}[f(h(I))]/\mathbb{E}[f^*(I)] - 1$ over instances $I$ produced by $g$; then $h$ is mutated again to reduce that gap. Working entirely through code rewrites and behavioral evaluation, with no gradients, this mutual escalation yields instance generators that produce spatially structured, clustered, spiral-like patterns rather than uniform random points, and solvers that incorporate spatial information and adaptive exploration. The paper reports that on TSP400 through TSP1000, EALG-generated instances raise the optimality gaps of FunSearch, EoH, KGLS, and ReEvo to roughly 7 to 13 percent while the EALG-evolved solver achieves about 5.5 to 8.2 percent, and that on nine TSPLIB instances the EALG solver has the lowest gap in every case, which the authors interpret as evidence of generalization rather than overfitting to synthetic instances.

Load-bearing premise

The reported gaps all rely on reference solution costs $f^*(I)$ in the gap formula, and the paper does not state how those reference costs are obtained, so if the reference is a weak heuristic or is applied inconsistently, the measured hardness amplification and the ranking of methods could change.

Editorial extensions

If this is right

  • Benchmark generation becomes a controllable stress test: the same pipeline can produce instance distributions with a chosen difficulty level by selecting the target gap.
  • Solver comparison becomes more discriminative, because EALG-generated instances amplify performance differences among competing heuristics.
  • Solver synthesis and benchmark construction are unified into one automated pipeline, removing the need for manually designed datasets and hand-tuned heuristics.
  • The loop transfers across underlying solver paradigms (guided local search and ant colony optimization) and across routing problems (TSP and orienteering), suggesting broader applicability than a single benchmark.
  • If the generalization to TSPLIB holds, co-evolved solvers are not just artifacts of the synthetic training distribution and can be deployed on practical, human-designed instances.

Reading between the lines

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

  • A natural next step would be to publish the EALG-generated instances as a standalone hard benchmark with reference costs fixed by an exact solver, making the hardness measure reproducible and comparable across future papers.
  • The minimax formulation is problem-agnostic in principle, so extending the loop to scheduling, packing, or graph partitioning would test whether the arms race transfers beyond routing problems.
  • The gap-based objective measures difficulty only relative to the current solver population, so the reported hardness numbers are time-stamped: as baselines improve, the same generators may stop being hard, which is exactly the dynamic the co-evolution loop is designed to exploit.
  • A concrete test of the framework's limit is to run the loop for many more generations and check whether generated instance structures keep diversifying or settle into a small set of recurring adversarial patterns.
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

5 major / 5 minor

Summary. The manuscript introduces EALG, a framework that co-evolves LLM-generated instance generators and heuristic solvers for the Traveling Salesman Problem and the Orienteering Problem. Sections 3-6 formalize the generator and solver spaces, define a minimax hardness objective, and describe an adversarial mutation/reflection loop. Section 7 reports empirical claims that EALG-generated instances increase the optimality gaps of existing solvers, that the EALG-synthesized solver outperforms FunSearch, EoH, KGLS, and ReEvo on those instances, and that the solver generalizes to nine TSPLIB instances. The manuscript explicitly defers all experimental details and implementation specifics to a future version.

Significance. The proposed idea of co-evolving instance generators and solvers addresses a real limitation of static benchmarks and, if substantiated, would be a useful addition to LLM-based algorithm design. The paper is fully open to that possibility: a working framework that hardens solvers on distribution shift would be of interest. However, the current manuscript does not substantiate the central claims. All quantitative evidence rests on unspecified protocols: the reference solver behind Eq. (2) is not defined, no run counts or error bars are provided, and Section 7 explicitly says details are 'provided in the future.' Moreover, the main solver comparison in Tables 1-2 is between a solver co-evolved on the test distribution and baselines seeing that distribution for the first time, which is a circular evaluation design. The TSPLIB results are the only held-out evidence, but they lack implementation details. The paper contains no code, data, or executable artifacts, and its central empirical claims are therefore not verifiable from the submitted version.

major comments (5)
  1. [Section 7] The paper states at the start of Section 7: 'Experimental details and implementation specifics are provided in the future.' This is a load-bearing omission because every central claim is empirical. None of Tables 1-3 or Figures 2-3 can be reproduced or meaningfully interpreted without specifying: the exact reference solver used to compute f* in Eq. (2), the number of independent runs, error bars, time budgets, seeds, the LLM and prompts used, evolutionary hyperparameters, and the rule for selecting the final solver and generator. Without these, the reported gap values and comparisons are unverifiable.
  2. [Section 7.2 / Table 1 / Eq. (4)] The evaluation of solver superiority is circular. In Eq. (4) the solver h is explicitly optimized to minimize hardness on instances drawn from the co-evolved generator g, and Section 7.2 then compares this distribution-fitted solver against FunSearch, EoH, KGLS, and ReEvo on instances produced by that same generator without giving those baselines any adaptation phase. A lower gap for EALG is therefore expected by construction and does not establish general solver strength or 'distribution-aware' superiority. The claim would require either giving baselines a matched co-evolution/adaptation phase on the same distribution or evaluating all methods on independent held-out distributions with matched computational budgets.
  3. [Eq. (2)] The hardness metric Gap(I;h) = E[f(h(I))]/E[f*(I)] - 1 requires reference costs f*(I_i), but the manuscript never states how these are obtained: exact TSP solver, high-quality heuristic, time limit, or tolerance. Because Tables 1 and 2 and Figure 2 are all comparisons of gaps, a weak or inconsistent reference solver could change the ordering of methods and the magnitude of the reported hardness amplification. This measurement assumption must be specified before any of the quantitative conclusions can be assessed.
  4. [Section 7.3 / Figure 3] Figure 3 compares raw objective values on standard versus EALG-generated instances without normalizing by reference cost or controlling coordinate/reward scaling. TSP tour length scales linearly with coordinate range, and OP reward depends on the reward model, so the observed 'degradation' could reflect coordinate scaling or reward magnitude rather than intrinsic combinatorial hardness. The paper needs normalized gaps or equivalent scale-invariant measures to support the claim that EALG instances are harder.
  5. [Section 7.5 / Table 3] Table 3 is the only evidence that avoids the distribution-leakage issue, but it is not adequately specified. The manuscript does not describe how FunSearch and ReEvo heuristics were instantiated under the shared GLS solver, how many runs were averaged, what termination criteria or time budgets were used, or whether the EALG solver received more evaluations during evolution. The claim that EALG 'consistently achieves the lowest gap on every tested instance' is unverifiable from the provided information.
minor comments (5)
  1. [Eq. (2)] The notation 'En i=1' in Eq. (2) is nonstandard and ambiguous; it should be replaced with either an explicit sum with indices or a defined empirical expectation over the n generated instances.
  2. [Section 7.4 / Table 2] Table 2 reports objective differences that are often below 2%, but no standard errors, number of runs, or OP parameter settings (e.g., time budget, node coordinates, reward values) are given, so it is impossible to tell whether the differences are statistically meaningful.
  3. [Figure 2] Figure 2 does not state whether the plotted curves are means over a fixed set of instances or single runs, and no error bars are shown; adding this information is necessary for interpreting the gap comparisons.
  4. [References [26] and [27]] References [26] and [27] appear to cite the same Nature paper by Romera-Paredes et al. in different forms; the duplicate citation should be merged or clearly disambiguated.
  5. [Section 7 opening] The phrase 'provided in the future' should be replaced with a concrete experimental-setup section; as written, it signals that the core specifications are absent from the submitted manuscript.

Circularity Check

2 steps flagged · score 6.0 of 10

The EALG solver's low gap on EALG-generated hard instances is the selection objective of the co-evolution loop, so Tables 1 and 2 compare a distribution-fitted solver against cold-start baselines; the solver-superiority claim is partially circular, although the TSPLIB results provide external, non-circular evidence.

  1. fitted input called prediction [Section 5, Eq. (4) and Section 7.2, Table 1]
    "min h∼Lheur max g∼Linst Ez∼Z [H(g(z); h)] / "EALG (ours) consistently generalizes to harder instances while baseline performance degrades.""

    The hard instances in Table 1 come from the generator g that Eq. (4) maximizes against the solver family from which EALG selects h, and the EALG solver is the minimizer of that same objective. Its low gap on those instances is therefore the optimized training signal, not an independent prediction on a held-out distribution. FunSearch, EoH, KGLS, and ReEvo were not part of this minimax loop and encounter the EALG distribution cold-start, so the comparison reports a distribution-fitted solver against unadapted baselines. The reported EALG superiority is substantially built into the evolutionary selection process rather than established as general combinatorial skill.

  2. fitted input called prediction [Section 7.4, Table 2]
    "heuristics(ReEvo)[39] ⊕hard instance(EALG) vs. heuristics(EALG) ⊕hard instance(EALG) / "those evolved within the EALG framework consistently outperform across all tested problem sizes.""

    The EALG heuristics were evolved by minimizing the same hardness objective on the EALG hard instance distribution, while the ReEvo heuristics were not exposed to that distribution during evolution. Table 2 therefore compares an in-distribution solver with an out-of-distribution solver. The consistent outperformance of EALG on these instances is the expected consequence of selection pressure on exactly those instances, not a demonstration that EALG produces generally stronger OP heuristics.

full rationale

The central empirical claim that EALG's synthesized solvers 'generalize effectively across a broad spectrum of combinatorial tasks' is partially circular. In the co-evolution loop formalized by Eq. (4), the generator is mutated to maximize H(g(z); h) for the current EALG solver, and the solver is mutated to minimize H on the generated distribution. Tables 1 and 2 then report the EALG solver's performance on instances produced by that same generator, against baselines that never participated in the loop. The EALG solver's lower gap and higher objective values are the quantities it was explicitly selected for, so those comparisons do not independently establish generalization. The TSPLIB results in Table 3 are external and do provide non-circular evidence, which prevents a higher circularity score. However, Section 7 states 'Experimental details and implementation specifics are provided in the future,' so the parity of baseline instantiation, time budgets, seeds, and prompts cannot be verified, weakening the external check. Additionally, Figure 3 uses raw objective values rather than the scale-invariant Gap of Eq. (2), so coordinate or reward scaling alone could explain part of the reported 'degradation'; this is a measurement confound rather than a circularity. No load-bearing self-citation chain or imported uniqueness theorem was found. Overall, the paper's hardness-generation contribution has independent content, but the solver-superiority claims on the co-evolved distributions reduce substantially to the training objective.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central results depend on unstated measurement choices: the reference solution behind Eq. (2), the evolutionary and LLM hyperparameters, and the rule for selecting the final solver. The framework also assumes that LLM mutation and reflection reliably improve programs, that Eq. (4) faithfully models the implemented loop, and that the optimality gap is a faithful difficulty measure independent of the reference solver. No formal entities are introduced.

free parameters (3)
  • Reference solver for f* in Eq. (2) = Not reported
    The optimality gap depends on the cost f*(I), but no solver, tolerance, or time limit is reported.
  • LLM and evolutionary hyperparameters = Not reported
    Number of generations, population size, mutation rate, model, temperature, and compute budget are not given.
  • Final solver selection rule = Not reported
    The paper does not say whether the reported solver is the best, last, or an ensemble from the evolution, which strongly affects gap values.
assumptions (3)
  • domain assumption LLM-sampled generators and heuristics form a search space rich enough that mutation and reflection improve program quality.
    The entire co-evolution depends on this; no coverage or diversity evidence is provided.
  • ad hoc to paper The minimax objective in Eq. (4) captures the behavior of the implemented co-evolutionary loop.
    The equation is presented as a formalization but is not derived from the algorithm and has no convergence analysis.
  • domain assumption Relative optimality gap in Eq. (2) is a scale-invariant, faithful measure of instance difficulty independent of the reference solver.
    No validation that gap correlates with solver difficulty or is stable across reference solvers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EALG: Evolutionary Adversarial Generation of Language Model-Guided Generators for Combinatorial Optimization." pith.science (2026). https://pith.science/paper/KXQYUHLS

@misc{pith2026250602594,
  author       = {Pith},
  title        = {Pith review of: EALG: Evolutionary Adversarial Generation of Language Model-Guided Generators for Combinatorial Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KXQYUHLS}},
  note         = {Machine review of arXiv:2506.02594}
}
read the original abstract

Generating challenging instances is crucial for the evaluation and advancement of combinatorial optimization solvers. In this work, we introduce EALG (Evolutionary Adversarial Generation of Language Model-Guided Generators), a novel framework that automates the co-evolution of optimization problem instances and their corresponding heuristic solvers using large language models (LLMs). EALG leverages a mutation-based adversarial approach that dynamically evolves instance generation procedures to create increasingly difficult problems, while simultaneously synthesizing adaptive heuristic algorithms through interactions with LLMs guided by algorithmic structure. Unlike existing approaches that focus solely on static benchmark creation or manual solver design, EALG provides a seamless pipeline from instance generation to solver synthesis. Experimental results demonstrate that EALG generates significantly harder instances than current benchmarks, and its synthesized solvers generalize effectively across a broad spectrum of combinatorial tasks. This work explores a new paradigm for combinatorial optimization that integrates instance generation with solver design, resulting in state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2506.02594 by the authors.

Figure 1
Figure 1. Illustration of the EALG framework and generated instance examples. 5 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Gap (%) comparison across four methods on standard (dashed) and hard (solid) datasets. Lower values indicate better performance, while higher values reflect greater degradation under hard instance distributions. 7.2 Hardness Amplification and Solver Superiority on EALG Instances [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Objective values on TSP_ACO (left) and OP_ACO (right) using ReEvo heuristics on standard versus EALG-generated hard instances. In TSP, higher objective values indicate worse solutions, while in OP, lower values imply reduced solver performance. Across both tasks, EALG￾generated instances consistently lead to more adverse outcomes, indicating a substantial increase in optimization difficulty. 7.4 Solver–Instance Syne… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of EALG-generated instance distributions across TSP_GLS, TSP_ACO, and OP_ACO using t-SNE, UMAP, and standard instance layouts. Top row shows optimization tasks; left column shows visualization methods. The standard instance layout consists of problem inst…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 32 canonical work pages

  1. [1]

    Knowledge-guided local search for the vehicle routing problem.Computers & Operations Research, 105:32–46, 2019

    Florian Arnold and Kenneth Sörensen. Knowledge-guided local search for the vehicle routing problem.Computers & Operations Research, 105:32–46, 2019

    work page 2019

  2. [2]

    Metaheuristics in combinatorial optimization: Overview and conceptual comparison.ACM Computing Surveys, 35(3):268–308, 2003

    Christian Blum and Andrea Roli. Metaheuristics in combinatorial optimization: Overview and conceptual comparison.ACM Computing Surveys, 35(3):268–308, 2003

    work page 2003

  3. [3]

    Leveraging large language models to develop heuristics for emerging optimization problems.CoRR, abs/2503.03350, 2025

    Thomas Bömer, Nico Koltermann, Max Disselnmeyer, Laura Dörr, and Anne Meyer. Leveraging large language models to develop heuristics for emerging optimization problems.CoRR, abs/2503.03350, 2025

    arXiv 2025

  4. [4]

    Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, and et al

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, and et al. Language models are few-shot learners. InNeurIPS 2020, 2020

    work page 2020

  5. [5]

    Evaluating large language models trained on code.CoRR, abs/2107.03374, 2021

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, and et al. Evaluating large language models trained on code.CoRR, abs/2107.03374, 2021

    arXiv 2021

  6. [6]

    Dorigo and L.M

    M. Dorigo and L.M. Gambardella. Ant colony system: A cooperative learning approach to the traveling salesman problem.IEEE Transactions on Evolutionary Computation, 1(1):53–66, 1997

    work page 1997

  7. [7]

    M. R. Garey and David S. Johnson.Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman, 1979

    work page 1979

  8. [8]

    Metaheuristics in combinatorial optimization.Annals of Operations Research, 140(1):189–213, 2005

    Michel Gendreau and Jean-Yves Potvin. Metaheuristics in combinatorial optimization.Annals of Operations Research, 140(1):189–213, 2005

    work page 2005

Show all 44 references
  1. [9]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. InNeurIPS 2014, pages 2672–2680, 2014

    2014

  2. [10]

    Multi-objective combinatorial generative adversarial optimization and its application in crowdsensing

    Yi-Nan Guo, Jianjiao Ji, Ying Tan, and Shi Cheng. Multi-objective combinatorial generative adversarial optimization and its application in crowdsensing. InICSI 2020, volume 12145 of Lecture Notes in Computer Science, pages 423–434. Springer, 2020

    2020

  3. [11]

    Evolutionary multiob- jective optimization driven by generative adversarial networks (gans).IEEE Transactions on Cybernetics, 51(6):3129–3142, 2021

    Cheng He, Shihua Huang, Ran Cheng, Kay Chen Tan, and Yaochu Jin. Evolutionary multiob- jective optimization driven by generative adversarial networks (gans).IEEE Transactions on Cybernetics, 51(6):3129–3142, 2021

    2021

  4. [12]

    Self-guiding exploration for combinatorial problems, 2024

    Zangir Iklassov, Yali Du, Farkhad Akimov, and Martin Takac. Self-guiding exploration for combinatorial problems, 2024

    2024

  5. [13]

    Large language models for combinatorial optimization of design structure matrix.CoRR, abs/2411.12571, 2024

    Shuo Jiang, Min Xie, and Jianxi Luo. Large language models for combinatorial optimization of design structure matrix.CoRR, abs/2411.12571, 2024

    2024

  6. [14]

    Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi

    Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew P. Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In2017 IEEE Conf...

    2017

  7. [15]

    From distribution learning in training to gradient search in testing for combinatorial optimization

    Yang Li, Jinpei Guo, Runzhong Wang, and Junchi Yan. From distribution learning in training to gradient search in testing for combinatorial optimization. InNeurIPS 2023, 2023

    2023

  8. [16]

    Yujia Li, David H. Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, Rémi Leblond, Tom Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, Thomas Hubert, Peter Choy, Cyprien de Masson d’Autume, Igor Babuschkin, Xinyun Chen, Po-Sen Huang, Johannes Welbl, Sven Gowal, ...

    2022 arXiv

  9. [17]

    Heuristics for vehicle routing problem: A survey and recent advances.CoRR, abs/2303.04147, 2023

    Fei Liu, Chengyu Lu, Lin Gui, Qingfu Zhang, Xialiang Tong, and Mingxuan Yuan. Heuristics for vehicle routing problem: A survey and recent advances.CoRR, abs/2303.04147, 2023

    2023 arXiv

  10. [18]

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

    Fei Liu, Xialiang Tong, 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. InICML 2024, 2024

    2024

  11. [19]

    Algorithm evolution using large language model.CoRR, abs/2311.15249, 2023

    Fei Liu, Xialiang Tong, Mingxuan Yuan, and Qingfu Zhang. Algorithm evolution using large language model.CoRR, abs/2311.15249, 2023. arXiv:2311.15249

    2023 arXiv

  12. [20]

    Large language models as evolutionary optimizers

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

    2024

  13. [21]

    Neural combinatorial optimization with heavy decoder: Toward large scale generalization

    Fu Luo, Xi Lin, Fei Liu, Qingfu Zhang, and Zhenkun Wang. Neural combinatorial optimization with heavy decoder: Toward large scale generalization. InNeurIPS 2023, 2023

    2023

  14. [22]

    McInnes, J

    L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction.ArXiv e-prints, February 2018

    2018

  15. [23]

    Planning of heuristics: Strategic planning on large language models with monte carlo tree search for automating heuristic optimization.CoRR, abs/2502.11422, 2025

    Chaoxu Mu, Xufeng Zhang, and Hui Wang. Planning of heuristics: Strategic planning on large language models with monte carlo tree search for automating heuristic optimization.CoRR, abs/2502.11422, 2025. arXiv:2502.11422

    2025 arXiv

  16. [24]

    A branch-and-cut algorithm for the resolution of large-scale symmetric traveling salesman problems.SIAM Review, 33(1):60–100, 1991

    Manfred Padberg and Giovanni Rinaldi. A branch-and-cut algorithm for the resolution of large-scale symmetric traveling salesman problems.SIAM Review, 33(1):60–100, 1991

    1991

  17. [25]

    Tsplib-atravelingsalesmanproblemlibrary.INFORMS Journal on Computing, 3(4):376–384, 1991

    GerhardReinelt. Tsplib-atravelingsalesmanproblemlibrary.INFORMS Journal on Computing, 3(4):376–384, 1991

    1991

  18. [26]

    Pawan Kumar, Emilien Dupont, Francisco J

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with large l...

    2023

  19. [27]

    Pawan Kumar, Emilien Dupont, Francisco J

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with larg...

    2024

  20. [28]

    Improving existing optimization algorithms with llms.CoRR, abs/2502.08298, 2025

    Camilo Chacón Sartori and Christian Blum. Improving existing optimization algorithms with llms.CoRR, abs/2502.08298, 2025. 13

    2025 arXiv

  21. [29]

    Skiscim and Bruce L

    Christopher C. Skiscim and Bruce L. Golden. Optimization by simulated annealing: A prelimi- nary computational study for the tsp. InWinter Simulation Conference (WSC), pages 523–535. ACM, 1983

    1983

  22. [30]

    van Hemert

    Kate Smith-Miles and Jano I. van Hemert. Discovering the suitability of optimisation algorithms by learning from evolved instances.Annals of Mathematics and Artificial Intelligence, 61(2):87– 104, 2011

    2011

  23. [31]

    Difusco: Graph-based diffusion solvers for combinatorial optimization

    Zhiqing Sun and Yiming Yang. Difusco: Graph-based diffusion solvers for combinatorial optimization. InNeurIPS 2023, 2023

    2023

  24. [32]

    Unleashing the potential of large language models as prompt optimizers: Analogical analysis with gradient-based model optimizers

    Xinyu Tang, Xiaolei Wang, Wayne Xin Zhao, Siyuan Lu, Yaliang Li, and Ji-Rong Wen. Unleashing the potential of large language models as prompt optimizers: Analogical analysis with gradient-based model optimizers. InAAAI 2025, pages 25264–25272, 2025

    2025

  25. [33]

    Visualizing data using t-sne.Journal of machine learning research, 9(11), 2008

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of machine learning research, 9(11), 2008

    2008

  26. [34]

    Qi Wang, Yongsheng Hao, and Jiawei Zhang. Generative inverse reinforcement learning for learning 2-opt heuristics without extrinsic rewards in routing problems.Journal of King Saud University - Computer and Information Sciences, 35(9):101787, 2023

    2023

  27. [35]

    As-llm: When algorithm selection meets large language model.CoRR, abs/2311.13184, 2023

    Xingyu Wu, Yan Zhong, Jibin Wu, and Kay Chen Tan. As-llm: When algorithm selection meets large language model.CoRR, abs/2311.13184, 2023

    2023 arXiv

  28. [36]

    Generative adversarial training for neural combinatorial optimization models, 2022

    Liang Xin, Wen Song, Zhiguang Cao, and Jie Zhang. Generative adversarial training for neural combinatorial optimization models, 2022

    2022

  29. [37]

    Le, Denny Zhou, and Xinyun Chen

    Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. InICLR 2024, 2024

    2024

  30. [38]

    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. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI-25), pages 27144–27152. AAAI Press, 2025

    2025

  31. [39]

    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. InNeurIPS 2024, 2024

    2024

  32. [40]

    Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. Star: Bootstrapping reasoning with reasoning. InNeurIPS 2022, 2022

    2022

  33. [41]

    Bridging visualization and optimization: Multimodal large language models on graph-structured combinatorial optimization.CoRR, abs/2501.11968, 2025

    Jie Zhao, Kang Hao Cheong, and Witold Pedrycz. Bridging visualization and optimization: Multimodal large language models on graph-structured combinatorial optimization.CoRR, abs/2501.11968, 2025

    2025 arXiv

  34. [42]

    Can large language models be trusted as black-box evolutionary optimizers for combinatorial problems?CoRR, abs/2501.15081, 2025

    Jie Zhao, Tao Wen, and Kang Hao Cheong. Can large language models be trusted as black-box evolutionary optimizers for combinatorial problems?CoRR, abs/2501.15081, 2025

    2025

  35. [43]

    Large language models as commonsense knowledge for large-scale task planning

    Zirui Zhao, Wee Sun Lee, and David Hsu. Large language models as commonsense knowledge for large-scale task planning. InNeurIPS 2023, 2023

    2023

  36. [44]

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. InICCV 2017, pages 2242–2251, 2017. 14

    2017

Pith tools

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