Pith. sign in

REVIEW 3 major objections 66 references

LLM-Driven Evolutionary Generation of Multi-Objective Bayesian Optimization Algorithms

T0 review · 3 major / 0 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read LLM-driven evolution can invent multi-objective Bayesian optimizers that match top accuracy at a fraction of the runtime.

desk verdict Solid LLaMEA extension that produces real, cheap MOBO code; the 0.971 synthetic headline rests on a development-found, post-hoc-fixed design rather than the systematic selection loop. read the letter →

arxiv 2607.08791 v1 pith:XIOBL2B6 submitted 2026-07-06 cs.NE cs.AI

Georgios Laskaris , Reuben Brasher , Niki van Stein , Elena Raponi , Thomas Bäck , Florian Neukart This is my paper
classification cs.NEcs.AI
keywords multi-objectiveBayesianoptimizationLLM-drivenalgorithmgenerationevolutionarystrategieshyperparameternormalizedhypervolumeParetofrontapproximationwall-clockefficiency
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

Designing multi-objective Bayesian optimizers means juggling interdependent choices—surrogates, scalarizations, acquisition functions, candidate generators—whose best combination is problem-dependent and usually requires expert trial-and-error. This paper shows that an evolutionary loop can treat a large language model as the mutation and crossover operator, generating complete, runnable algorithm code and then tuning each candidate’s hyperparameters inside the same loop. From roughly nine hundred generated programs the search produces designs that, on twelve standard synthetic problems, reach higher mean normalized hypervolume than a state-of-the-art qParEGO baseline while running about sixty times faster; on three unseen real-world engineering problems one of those designs is again the most accurate method and still several times cheaper. The practical claim is that automatic evolutionary search can discover Pareto-efficient algorithm architectures that human designers rarely reach by hand.

What carries the argument

LLaMEA-MOBO: an evolutionary strategy in which a large language model acts as both mutation and crossover operator, each generated algorithm is immediately configured by multi-fidelity hyperparameter optimization, and fitness is mean normalized hypervolume on a multi-problem suite.

What would settle it

Re-run the same generated algorithms and the qParEGO baseline on a larger set of constrained, noisy, or higher-dimensional real-world problems with budgets well above 400 evaluations; if the generated designs lose both their accuracy edge and their runtime advantage, the central transfer claim fails.

Watch

Extended reading notes

Core claim

Across nine evolutionary runs the framework discovers complete multi-objective Bayesian optimization algorithms whose strongest members attain the highest mean normalized hypervolume on twelve synthetic benchmarks (0.971 versus 0.869 for qParEGO) at roughly sixty-fold lower wall-clock cost, remain competitive or superior on three held-out engineering problems, and occupy the efficient frontier of the accuracy–runtime trade-off that manual design has not filled.

Load-bearing premise

That average normalized hypervolume after a fixed 400-evaluation budget on a handful of synthetic training problems plus three unconstrained engineering cases is a sufficient signal for claiming generally useful multi-objective Bayesian optimizers.

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

3 major / 0 minor

Summary. The paper extends LLaMEA to multi-objective Bayesian optimization (LLaMEA-MOBO), using an LLM (Gemini-2.5-flash) as mutation/crossover operators inside (1+1), (4+16), and (8,16) evolution strategies, with SMAC multi-fidelity HPO inside the evolutionary loop. Across nine runs (~900 candidates) it produces complete MOBO implementations scored by mean normalized hypervolume. On twelve synthetic problems the development-found MOEAD-EI Hybrid reports the best mean normalized HV (0.971 vs 0.869 for BoFire qParEGO) at ~60× lower wall-clock cost and is significantly better on 7/12 problems (never worse). On three held-out RE engineering problems, systematically selected Improved-Scalarized-EI is best (0.985 vs 0.971) and significantly better on 2/3 problems at ~3.4× lower cost. Classical EA baselines and Random Search are included; time–accuracy plots and Friedman/Nemenyi plus per-problem Welch/Wilcoxon tests support a Pareto-efficient trade-off claim.

Significance. If the results hold under a cleaner selection and reporting protocol, the work is a solid empirical contribution to automated algorithm design for expensive multi-objective optimization: it shows that LLM-driven evolutionary search can synthesize complete MOBO pipelines (surrogate + acquisition + candidate generation) that match or beat a maintained SOTA BO baseline at substantially lower wall-clock cost, with partial transfer to real engineering problems. Strengths include an independent BoFire/BoTorch qParEGO baseline, external fitness (normalized HV), multi-strategy ES comparison, SMAC-in-the-loop HPO, public code for the four highlighted algorithms, and honest per-problem significance rather than only suite averages. The time–accuracy frontier (especially the RF variants at 2–5 s/run) is practically relevant for deployment under tight compute budgets.

major comments (3)
  1. Abstract and §V.B.2 lead with the synthetic champion result (0.971 mean normalized HV, 60× speedup, better on 7/12 problems) as if it were the product of the advertised nine-run selection pipeline. §V.A and §VI.A state that MOEAD-EI Hybrid was “discovered during the development of our experimentation,” retained for exceptional performance, and is not among the three Phase-1 systematically ranked algorithms (Improved-Scalarized-EI 0.855, RF-LCB-PBI 0.795, RF-ParEGO-Batch 0.756). The systematically selected synthetic leader (Improved-Scalarized-EI) is behind qParEGO on Phase 2 (0.811 vs 0.869). The abstract and strongest claim must be rewritten so that (i) development-found vs systematically selected algorithms are labeled, (ii) the primary synthetic accuracy claim is either restricted to the systematic set or clearly caveated, and (iii) the real-world win for Improved-Scalarized-EI is not
  2. §VI.B documents two post-hoc human interventions that affect the headline algorithms: (1) a one-line population-size cap so MOEAD-EI Hybrid does not index out of bounds when N_pop > n_init on lower-dimensional problems; (2) re-running Improved-Scalarized-EI across five seeds after discovering it was generated as a deterministic (fixed-seed) heuristic with zero run-to-run variance. These repairs are load-bearing for the reported means, stds, and Welch tests. The paper should report, side-by-side, the unrepaired automated outputs (or failure modes) versus the repaired versions, state which results depend on the fixes, and add the promised generation-time checks (stochasticity / seed handling; population consistency) so that the pure LLaMEA-MOBO+SMAC loop is what is being claimed.
  3. §IV.B–C and §VI.B: generalization is asserted from a generation fitness suite that uses only a subset of the synthetic problems (ZDT1–4,6 and DTLZ1–2,4) plus three unconstrained RE problems in Phase 3. The paper already notes that the RE suite is small and underpowered for rank tests. Given that the central transfer claim (“gains transfer beyond the synthetic regime”) rests on RE21/34/37 only, either expand the held-out real-world set or substantially soften the transfer language in the abstract and conclusion, and report per-problem effect sizes with the same honesty used for the synthetic Welch tests rather than leading with mean normalized HV over three problems.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical generation-and-benchmark paper whose performance claims rest on external hypervolume measurements against an independent baseline, not on any self-referential derivation.

full rationale

The paper is an automated-algorithm-design study. Its load-bearing claims are experimental: LLM-driven ES + SMAC produces complete MOBO implementations whose mean normalized hypervolume (and wall-clock time) on a fixed 400-evaluation budget is competitive with or better than BoFire/BoTorch qParEGO on twelve synthetic problems and three held-out RE engineering problems. Fitness during search is the external, reference-point-normalized hypervolume of Eq. (1) averaged over a training subset of the synthetic suite; final claims use absolute HV values, Welch t-tests, Wilcoxon signed-rank, and Friedman/Nemenyi tests against an independent baseline implementation. Within-phase max-normalization is used only for ranking which generated candidates to promote; it does not force the absolute numbers or the statistical comparisons. Self-citations are to the authors’ prior LLaMEA / LLaMEA-BO / LLaMEA-HPO framework papers that supply the generation loop; those works do not contain the MOBO results or the performance numbers reported here. No equation, uniqueness theorem, or fitted constant is redefined as a “prediction.” Post-hoc mechanical fixes (population-size cap, multi-seed re-run) affect experimental validity but do not create a circular reduction of claim to input. Consequently the derivation chain is empty of circular steps.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim is empirical and rests on standard multi-objective optimization practice plus the LLaMEA generation machinery. Free parameters are the usual experimental knobs (budget, ES settings, SMAC budget, LLM choice). No new physical entities are postulated; the invented entities are the generated algorithm designs themselves, which are fully specified by code.

free parameters (5)
  • evaluation budget B = 400
    Fixed at 400 expensive evaluations per run; standard but arbitrary for the claimed regime.
  • ES population and offspring sizes = µ,λ as listed; p_crossover=0.6
    (1+1), (4+16), (8,16) configurations chosen by the authors; crossover probability 0.6.
  • SMAC multi-fidelity HPO budget and instance subset
    Each generated algorithm’s hyperparameters are tuned by SMAC on a subset of the suite; exact SMAC iteration counts are not fully enumerated.
  • LLM backbone = Gemini-2.5-flash
    Gemini-2.5-flash used as sole mutation/crossover operator; results may be sensitive to model choice.
  • normalized-HV reference points and per-problem max normalizers = see Table I
    Reference points for synthetic problems are fixed; RE points set to 1.1×95th percentile of 10k random samples; Phase ranking uses within-phase max HV.
assumptions (4)
  • domain assumption Normalized hypervolume is a valid scalar fitness for ranking multi-objective optimizers under a fixed evaluation budget.
    Used throughout Sections III–V as the sole fitness and comparison metric.
  • domain assumption A 400-evaluation budget and the chosen ZDT/DTLZ/WFG/RE suites are representative enough to claim practical MOBO utility.
    Stated experimental protocol in Section IV; limitations acknowledged in Section VI.
  • ad hoc to paper LLM-generated code that compiles and passes a random-configuration smoke test is a legitimate candidate algorithm.
    Generation pipeline in Section III; one mechanical population-size fix was still required for MOEAD-EI Hybrid.
  • domain assumption BoFire qParEGO is a faithful state-of-the-art Bayesian multi-objective baseline.
    Adopted in Sections II and IV as the primary SOTA comparator.
invented entities (2)
  • LLaMEA-MOBO generation loop independent evidence
    purpose: Extend LLaMEA to synthesize complete multi-objective Bayesian optimizers with SMAC in the loop.
    Core methodological contribution; fully specified by the pipeline in Section III and Figure 1.
  • MOEAD-EI Hybrid / Improved-Scalarized-EI / RF-LCB-PBI / RF-ParEGO-Batch independent evidence
    purpose: Concrete generated algorithms that realize the claimed accuracy–cost trade-offs.
    Described in Section VI and Appendix algorithms 1–4; source linked on GitHub.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-Driven Evolutionary Generation of Multi-Objective Bayesian Optimization Algorithms." pith.science (2026). https://pith.science/paper/XIOBL2B6

@misc{pith2026260708791,
  author       = {Pith},
  title        = {Pith review of: LLM-Driven Evolutionary Generation of Multi-Objective Bayesian Optimization Algorithms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XIOBL2B6}},
  note         = {Machine review of arXiv:2607.08791}
}
read the original abstract

Designing effective multi-objective Bayesian optimization (MOBO) algorithms requires balancing many interdependent design choices whose optimal configuration is problem-dependent and typically demands deep expertise. We extend the LLaMEA framework to MOBO, using large language models as mutation and crossover operators within evolutionary strategies to generate complete algorithm implementations, with SMAC hyperparameter optimization integrated into the evolutionary loop. Across nine evolutionary runs we generated approximately 900 algorithms and benchmarked them on twelve synthetic problems (ZDT, DTLZ, WFG) and three real-world engineering problems (RE), using a BoFire qParEGO implementation as a state-of-the-art Bayesian-optimization baseline. On the synthetic suite the strongest generated algorithm attains the highest mean normalized hypervolume (0.971, vs. 0.869 for qParEGO) while requiring roughly 60x less wall-clock time; a Friedman test with post-hoc analysis places the two in a single top-performing group, and per-problem tests find the generated algorithm significantly better than qParEGO on 7 of the 12 problems and never worse, matching state-of-the-art accuracy at an order-of-magnitude lower cost. On the three unseen real-world engineering problems a generated algorithm attains the best mean normalized hypervolume (0.985, vs. 0.971 for qParEGO)--significantly better than qParEGO on two of the three problems--at roughly 3.4x lower wall-clock cost, confirming that the gains transfer beyond the synthetic regime. LLM-driven evolutionary search can thus discover algorithm designs that achieve Pareto-efficient trade-offs difficult to reach through manual design.

Figures

Figures reproduced from arXiv: 2607.08791 by the authors.

Figure 1
Figure 1. FIG. 1: Overview of the LLaMEA-MOBO framework. An [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2: Convergence of the nine LLaMEA-generated algorithms (Phase 1) on representative problems from each [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3: Convergence of all nine algorithms—the four LLaMEA-generated (the top-3 plus MOEAD-EI Hybrid) and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: FIG. 4: Convergence on the three real-world RE [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5: Time–accuracy trade-off on (a) the twelve synthetic and (b) the three real-world (RE) problems. Horizontal [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6: Critical-difference diagram (Nemenyi, [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7: Best-so-far search fitness (mean normalized HV [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8: Convergence on all twelve synthetic problems (Phase 1) for the nine LLaMEA-generated algorithms. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9: Convergence on all twelve synthetic problems (Phase 2) for the nine benchmarked algorithms (four [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: FIG. 10: Non-dominated fronts on the twelve synthetic benchmark problems for the four LLaMEA-generated [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: FIG. 11: Non-dominated fronts on the real-world RE problems for the four generated algorithms and qParEGO. [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 4 linked inside Pith

  1. [1]

    From each run we select the best-performing generated algorithm for benchmark- ing

    This yields 100 evaluated algorithms for (1 + 1)-ES and (4 + 16)-ES (4 initial parents + 6 generations×16 offspring) and 104 for (8,16)-ES (8 initial parents + 6 generations×16 offspring). From each run we select the best-performing generated algorithm for benchmark- ing. Additionally, we compare against a pool of base- lines spanning a state-of-the-art B...

  2. [2]

    Re- sults are summarized in terms of HV convergence curves and final normalized HV

    Phase 1: Comparison Among Generated Algorithms In Phase 1, the nine best-performing generated algo- rithms are evaluated on the twelve synthetic benchmark problems listed in Table I, using a budget of 400 function evaluations and 5 independent repeats per problem. Re- sults are summarized in terms of HV convergence curves and final normalized HV. Improved...

  3. [3]

    MOEAD-EI Hybrid achieves the highest mean nor- malized HV of 0.971, beating qParEGO on 8 of the 12 problems and outperforming every classical baseline; 7 FIG

    Phase 2: Comparison Against Baselines and State-of-the-Art on Synthetic Problems In Phase 2, the best three generated algorithms — together with the development-found MOEAD-EI Hy- brid — are benchmarked alongside five established al- gorithms — Multi-objective Random Search, NSGA- II [32], NSGA-III [33], IOC-SAMO-COBRA [31], and qParEGO [15] — on the same...

  4. [4]

    Here the ordering changes

    Phase 3: Generalization to Real-World Problems Phase 3 evaluates all nine algorithms on three uncon- strained real-world engineering optimization problems from the RE benchmark suite [40] (Table I), which were unseen during the LLaMEA generation phase and during Phase 1 and Phase 2 benchmarking. Here the ordering changes. Improved-Scalarized-EI at- tains ...

  5. [5]

    Wall-clock time is measured as the to- tal execution time for one complete optimization run of 400 function evaluations on a single problem instance, 8 FIG

    Time–Accuracy Trade-off A central contribution of this work is the identifica- tion of LLaMEA-generated algorithms that match or ex- ceed state-of-the-art accuracy at a fraction of the com- putational cost. Wall-clock time is measured as the to- tal execution time for one complete optimization run of 400 function evaluations on a single problem instance, ...

  6. [6]

    Schneider, W

    P. Schneider, W. P. Walters, A. T. Plowright, N. Sieroka, J. Listgarten, R. A. Goodnow Jr, J. Fisher, J. M. Jansen, J. S. Duca, T. S. Rush,et al., Rethinking drug design in the artificial intelligence era, Nature reviews drug discov- ery19, 353 (2020)

    2020

  7. [7]

    J. M. Stokes, K. Yang, K. Swanson, W. Jin, A. Cubillos- Ruiz, N. M. Donghia, C. R. MacNair, S. French, L. A. Carfrae, Z. Bloom-Ackermann,et al., A deep learning approach to antibiotic discovery, Cell180, 688 (2020)

    2020

  8. [8]

    K. Deb, Multi-objective optimisation using evolutionary algorithms: an introduction, inMulti-objective evolution- ary optimisation for product design and manufacturing (Springer, 2011) pp. 3–34. 13

    2011

Show all 66 references
  1. [9]

    R. T. Marler and J. S. Arora, Survey of multi-objective optimization methods for engineering, Structural and multidisciplinary optimization26, 369 (2004)

    2004

  2. [10]

    Shahriari, K

    B. Shahriari, K. Swersky, Z. Wang, R. P. Adams, and N. De Freitas, Taking the human out of the loop: A re- view of Bayesian optimization, Proceedings of the IEEE 104, 148 (2015)

    2015

  3. [11]

    D. R. Jones, M. Schonlau, and W. J. Welch, Efficient global optimization of expensive black-box functions, Journal of Global optimization13, 455 (1998)

    1998

  4. [12]

    Romera-Paredes, M

    B. Romera-Paredes, M. Barekatain, A. Novikov, M. Ba- log, M. P. Kumar, E. Dupont, F. J. R. Ruiz, J. S. Ellen- berg, P. Wang, O. Fawzi,et al., Mathematical discoveries from program search with large language models, Nature 625, 468 (2024)

    2024

  5. [13]

    F. Liu, X. Tong, M. Yuan, X. Lin, F. Luo, Z. Wang, Z. Lu, and Q. Zhang, Evolution of heuristics: Towards efficient automatic algorithm design using large language model, inInternational Conference on Machine Learning (ICML), Proceedings of Machine Learning Research, Vol. 235 (...

    2024

  6. [14]

    H. Ye, J. Wang, Z. Cao, F. Berto, C. Hua, H. Kim, J. Park, and G. Song, ReEvo: Large language models as hyper-heuristics with reflective evolution, inAdvances in Neural Information Processing Systems, Vol. 37 (2024)

    2024

  7. [15]

    van Stein and T

    N. van Stein and T. B¨ ack, LLaMEA: A large language model evolutionary algorithm for automatically generat- ing metaheuristics, Transactions on Evolutionary Com- putation29, 331–345 (2025)

    2025

  8. [16]

    Novikov, N

    A. Novikov, N. V˜ u, M. Eisenberger, E. Dupont, P.-S. Huang, A. Z. Wagner, S. Shirobokov, B. Kozlovskii, F. J. Ruiz, A. Mehrabian,et al., Alphaevolve: A coding agent for scientific and algorithmic discovery, arXiv preprint arXiv:2506.13131 (2025)

    2025 arXiv

  9. [17]

    W. Li, N. van Stein, T. B¨ ack, and E. Raponi, LLaMEA- BO: A large language model evolutionary algorithm for automatically generating Bayesian optimization al- gorithms, arXiv preprint arXiv:2505.21034 (2025)

    2025 arXiv

  10. [18]

    van Stein, D

    N. van Stein, D. Vermetten, and T. B¨ ack, In-the-loop hyper-parameter optimization for LLM-based automated design of heuristics, ACM Transactions on Evolutionary Learning (2024)

    2024

  11. [19]

    J. Knowles, ParEGO: A hybrid algorithm with on-line landscape approximation for expensive multiobjective optimization problems, IEEE Transactions on Evolution- ary Computation10, 50 (2006)

    2006

  12. [20]

    Daulton, M

    S. Daulton, M. Balandat, and E. Bakshy, Differentiable expected hypervolume improvement for parallel multi- objective Bayesian optimization, inAdvances in Neural Information Processing Systems, Vol. 33 (2020) pp. 9851– 9864

    2020

  13. [21]

    J. P. D¨ urholt, T. S. Asche, J. Kleinekorte, G. Mancino- Ball, B. Schiller, S. Sung, J. Keupp, A. Osburg, T. Boyne, R. Misener, R. Eldred, C. Kappatou, R. M. Lee, D. Linzner, W. S. Costa, D. Walz, N. Wulkow, and B. Shafei, BoFire: Bayesian optimization framework in- tended fo...

    2025

  14. [22]

    Balandat, B

    M. Balandat, B. Karrer, D. R. Jiang, S. Daulton, B. Letham, A. G. Wilson, and E. Bakshy, BoTorch: A framework for efficient Monte-Carlo Bayesian optimiza- tion, inAdvances in Neural Information Processing Sys- tems, Vol. 33 (2020) pp. 21524–21538

    2020

  15. [23]

    M. T. M. Emmerich, K. C. Giannakoglou, and B. Nau- joks, Single- and multiobjective evolutionary optimiza- tion assisted by Gaussian random field metamodels, inIEEE Transactions on Evolutionary Computation, Vol. 10 (IEEE, 2006) pp. 421–439

    2006

  16. [24]

    Belakaria, A

    S. Belakaria, A. Deshwal, and J. R. Doppa, Max-value entropy search for multi-objective Bayesian optimization, Advances in neural information processing systems32 (2019)

    2019

  17. [25]

    Hern´ andez-Lobato, J

    D. Hern´ andez-Lobato, J. Hernandez-Lobato, A. Shah, and R. Adams, Predictive entropy search for multi- objective Bayesian optimization, inInternational Con- ference on Machine Learning(PMLR, 2016) pp. 1492– 1501

    2016

  18. [26]

    Zhang, W

    Q. Zhang, W. Liu, E. Tsang, and B. Virginas, Expensive multiobjective optimization by MOEA/D with Gaussian process model, IEEE Transactions on Evolutionary Com- putation14, 456 (2010)

    2010

  19. [27]

    Grochow, New applications of the polynomial method: the cap set conjecture and beyond, Bulletin of the Amer- ican Mathematical Society56, 29 (2019)

    J. Grochow, New applications of the polynomial method: the cap set conjecture and beyond, Bulletin of the Amer- ican Mathematical Society56, 29 (2019)

    2019

  20. [28]

    Sharma, Openevolve: an open-source evolutionary coding agent (2025)

    A. Sharma, Openevolve: an open-source evolutionary coding agent (2025)

    2025

  21. [29]

    R. T. Lange, Y. Imajuku, and E. Cetin, Shinkaevolve: Towards open-ended and sample-efficient program evolu- tion, arXiv preprint arXiv:2509.19349 (2025)

    2025 arXiv

  22. [30]

    Z. Xie, F. Liu, Z. Wang, and Q. Zhang, Llm-driven neigh- borhood search for efficient heuristic design., inCEC (2025) pp. 1–8

    2025

  23. [31]

    Zheng, Z

    Z. Zheng, Z. Xie, Z. Wang, and B. Hooi, Monte carlo tree search for comprehensive exploration in LLM-based automatic heuristic design, inInternational Conference on Machine Learning (ICML), Proceedings of Machine Learning Research (PMLR, 2025)

    2025

  24. [32]

    A. Chen, D. Dohan, and D. So, EvoPrompting: Lan- guage models for code-level neural architecture search, inAdvances in Neural Information Processing Systems, Vol. 36 (2023)

    2023

  25. [33]

    Lindauer, K

    M. Lindauer, K. Eggensperger, M. Feurer, A. Biedenkapp, D. Deng, C. Benjamins, T. Ruhkopf, R. Sass, and F. Hutter, SMAC3: A versatile Bayesian optimization package for hyperparameter optimization, Journal of Machine Learning Research23, 1 (2022)

    2022

  26. [34]

    L´ opez-Ib´ a˜ nez, J

    M. L´ opez-Ib´ a˜ nez, J. Dubois-Lacoste, L. P. C´ aceres, M. Birattari, and T. St¨ utzle, The irace package: Iterated racing for automatic algorithm configuration, Operations Research Perspectives3, 43 (2016)

    2016

  27. [35]

    Beyer and D

    H.-G. Beyer and D. V. Arnold, Theory of evolution strategies—a tutorial, Theoretical aspects of evolution- ary computing , 109 (2001)

    2001

  28. [36]

    De Winter, B

    R. De Winter, B. Milatz, J. Blank, N. Van Stein, T. B¨ ack, and K. Deb, Parallel multi-objective opti- mization for expensive and inexpensive objectives and constraints, Swarm and Evolutionary Computation86, 101508 (2024)

    2024

  29. [37]

    K. Deb, S. Agrawal, A. Pratap, and T. Meyarivan, A fast elitist non-dominated sorting genetic algorithm for multi- objective optimization: NSGA-II, inInternational con- ference on parallel problem solving from nature(Springer,

  30. [38]

    Wietheger and B

    S. Wietheger and B. Doerr, A mathematical runtime analysis of the non-dominated sorting genetic algorithm iii (NSGA-III), inProceedings of the Genetic and Evolu- tionary Computation Conference Companion(2024) pp. 14 63–64

    2024

  31. [39]

    Zitzler, K

    E. Zitzler, K. Deb, and L. Thiele, Comparison of multiob- jective evolutionary algorithms: Empirical results, Evo- lutionary computation8, 173 (2000)

    2000

  32. [40]

    K. Deb, L. Thiele, M. Laumanns, and E. Zitzler, Scalable test problems for evolutionary multiobjective optimiza- tion, inEvolutionary multiobjective optimization: the- oretical advances and applications(Springer, 2005) pp. 105–145

    2005

  33. [41]

    Huband, P

    S. Huband, P. Hingston, L. Barone, and L. While, A review of multiobjective test problems and a scalable test problem toolkit, IEEE Transactions on Evolution- ary Computation10, 477 (2006)

    2006

  34. [42]

    Zitzler and L

    E. Zitzler and L. Thiele, Multiobjective evolutionary al- gorithms: a comparative case study and the strength pareto approach, IEEE transactions on Evolutionary Computation3, 257 (1999)

    1999

  35. [43]

    Zitzler, L

    E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and V. G. Da Fonseca, Performance assessment of multiob- jective optimizers: An analysis and review, IEEE Trans- actions on evolutionary computation7, 117 (2003)

    2003

  36. [44]

    Blank and K

    J. Blank and K. Deb, pymoo: Multi-objective optimiza- tion in python, IEEE Access8, 89497 (2020)

    2020

  37. [45]

    Tanabe and H

    R. Tanabe and H. Ishibuchi, An easy-to-use real-world multi-objective optimization problem suite, Applied Soft Computing89, 106078 (2020)

    2020

  38. [46]

    Bajer, Z

    L. Bajer, Z. Pitra, J. Repick` y, and M. Holeˇ na, Gaussian process surrogate models for the cma evolution strategy, Evolutionary computation27, 665 (2019)

    2019

  39. [47]

    P. I. Frazier, A tutorial on Bayesian optimization, arXiv preprint arXiv:1807.02811 (2018)

    2018 arXiv

  40. [48]

    Q. Lu, K. D. Polyzos, B. Li, and G. B. Giannakis, Surro- gate modeling for Bayesian optimization beyond a single gaussian process, IEEE Transactions on Pattern Analysis and Machine Intelligence45, 11283 (2023)

    2023

  41. [49]

    Bhattacharjee, K

    B. Bhattacharjee, K. Unni, and M. Pratap, Bayesian- optimized extreme gradient boosting models for classi- fication problems: an experimental analysis of product return case, Journal of Systems and Information Tech- nology26, 495 (2024)

    2024

  42. [50]

    Y. Lai, P. Zheng, C. Ji, C. Qiu, T. Wang, S. Zhang, Z. Wang, and Y. Du, Coffeeboost: Gradient boosting native conformal inference for Bayesian optimization, in Proceedings of the AAAI Conference on Artificial Intel- ligence, Vol. 39 (2025) pp. 18017–18025

    2025

  43. [51]

    X. Wang, Y. Jin, S. Schmitt, and M. Olhofer, Recent ad- vances in Bayesian optimization, ACM computing sur- veys55, 1 (2023)

    2023

  44. [52]

    H. Li, Y. Jin, and T. Chai, Evolutionary multi-objective Bayesian optimization based on multisource online trans- fer learning, IEEE Transactions on Emerging Topics in Computational Intelligence8, 488 (2023)

    2023

  45. [53]

    van Stein, A

    N. van Stein, A. V. Kononova, H. Yin, and T. B¨ ack, BLADE: Benchmark suite for LLM-driven automated de- sign and evolution of iterative optimisation heuristics, in Proceedings of the Genetic and Evolutionary Computa- tion Conference Companion(2025) pp. 2336–2344

    2025

  46. [54]

    B. L. Welch, The generalization of ‘student’s’problem when several different population varlances are involved, Biometrika34, 28 (1947)

    1947

  47. [55]

    Wilcoxon, Individual comparisons by ranking meth- ods, Biometrics bulletin1, 80 (1945)

    F. Wilcoxon, Individual comparisons by ranking meth- ods, Biometrics bulletin1, 80 (1945)

    1945

  48. [56]

    Friedman, The use of ranks to avoid the assumption of normality implicit in the analysis of variance, Journal of the american statistical association32, 675 (1937)

    M. Friedman, The use of ranks to avoid the assumption of normality implicit in the analysis of variance, Journal of the american statistical association32, 675 (1937)

    1937

  49. [57]

    P. B. Nemenyi,Distribution-free multiple comparisons. (Princeton University, 1963)

    1963

  50. [58]

    Demˇ sar, Statistical comparisons of classifiers over mul- tiple data sets, Journal of Machine learning research7, 1 (2006)

    J. Demˇ sar, Statistical comparisons of classifiers over mul- tiple data sets, Journal of Machine learning research7, 1 (2006)

    2006

  51. [59]

    C. K. Williams and C. E. Rasmussen,Gaussian processes for machine learning, Vol. 2 (MIT press Cambridge, MA, 2006)

    2006

  52. [60]

    Miettinen,Nonlinear multiobjective optimization, Vol

    K. Miettinen,Nonlinear multiobjective optimization, Vol. 12 (Springer Science & Business Media, 1999)

    1999

  53. [61]

    A. E. Eiben and J. E. Smith,Introduction to evolutionary computing(Springer, 2015)

    2015

  54. [62]

    Zhang and H

    Q. Zhang and H. Li, MOEA/D: A multiobjective evolu- tionary algorithm based on decomposition, IEEE Trans- actions on evolutionary computation11, 712 (2007)

    2007

  55. [63]

    R. E. Steuer and E.-U. Choo, An interactive weighted Tchebycheff procedure for multiple objective program- ming, Mathematical programming26, 326 (1983)

    1983

  56. [64]

    Breiman, Random forests, Machine learning45, 5 (2001)

    L. Breiman, Random forests, Machine learning45, 5 (2001)

    2001

  57. [65]

    Hutter, H

    F. Hutter, H. H. Hoos, and K. Leyton-Brown, Sequential model-based optimization for general algorithm configu- ration, inInternational conference on learning and intel- ligent optimization(Springer, 2011) pp. 507–523

    2011

  58. [66]

    Srinivas, A

    N. Srinivas, A. Krause, S. M. Kakade, and M. W. Seeger, Information-theoretic regret bounds for Gaussian process optimization in the bandit setting, IEEE Transactions on Information Theory58, 3250 (2012). APPENDIX A Selected LLaMEA-Generated Algorithms The full source code of ...

    2012

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.