Pith. sign in

REVIEW 2 major objections 2 minor 32 references

A Fresh Look at Lamarckian Evolution and the Baldwin Effect

T0 review · 2 major / 2 minor · reviewed 2026-06-29 · grok-4.3

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

Pith's one-line read Baldwinian evolution is asymptotically faster than Lamarckian which is faster than Darwinian for block lengths greater than two.

desk verdict The empirical comparison on GraphBench is the solid part here, but the theoretical claim that Baldwinian stays fastest once local search cost is included rests on an underspecified cost model. read the letter →

arxiv 2605.28703 v1 pith:Q6GYB65S submitted 2026-05-27 cs.NE cs.AIcs.DSmath.OC

Inès Benito , Johannes F. Lutzeyer , Benjamin Doerr This is my paper · ORCID
classification cs.NEcs.AIcs.DSmath.OC
keywords evolutionaryalgorithmsLamarckianevolutionBaldwineffectruntimeanalysismaximumindependentsetcutlocalsearchgraphoptimization
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

The paper compares three styles of evolutionary algorithms on both graph problems and a theoretical benchmark. It demonstrates that versions using local search outperform plain Darwinian evolution in experiments on Maximum Independent Set and Maximum Cut across six datasets, often beating deep learning baselines. On the theoretical side the authors extend the Deceptive Leading Block problem to arbitrary block lengths and prove runtime bounds showing Baldwinian evolution requires fewer fitness evaluations than Lamarckian, which in turn beats Darwinian, once blocks exceed length two. When the cost of local search is folded into the fitness count the ordering can shift, yet Baldwinian stays fastest from small blocks onward in the cost models examined. These results matter because they quantify when and why hybrid local-global search improves efficiency on hard combinatorial tasks.

What carries the argument

The Deceptive Leading Block benchmark extended to arbitrary block lengths, together with upper and lower bounds on expected runtime derived from modern theoretical runtime analysis.

What would settle it

An empirical runtime measurement on the extended Deceptive Leading Block problem with block length four in which the average number of fitness evaluations required by Darwinian evolution is smaller than that required by Baldwinian evolution.

Watch

Extended reading notes

Core claim

On the Deceptive Leading Block benchmark extended to arbitrary block length, for block lengths greater than two Baldwinian evolution is asymptotically faster than Lamarckian which is asymptotically faster than Darwinian evolution. When the cost of the local search procedure is included in the fitness evaluation count, Baldwinian evolution remains the fastest from small block lengths onwards. Empirically, on graphs drawn from six datasets, Baldwinian and Lamarckian variants consistently outperform Darwinian evolution, exceed recent deep learning baselines in most cases, and approach the performance of specialised heuristics and exact solvers.

Load-bearing premise

The cost of the local search procedure used inside the Baldwinian and Lamarckian variants can be assigned to fitness evaluations without changing the runtime ordering obtained from the number of evaluations alone.

Editorial extensions

If this is right

  • Baldwinian and Lamarckian evolution outperform Darwinian evolution on the great majority of Maximum Independent Set and Maximum Cut instances tested.
  • All three evolutionary algorithm variants exceed recent deep learning baselines in most cases examined.
  • A single set of generalist parameters works well across Darwinian, Lamarckian and Baldwinian evolution on the studied graph problems.
  • Baldwinian evolution retains the best asymptotic runtime even after local search cost is included, for the cost models considered.

Reading between the lines

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

  • The Baldwinian advantage may arise because learned local improvements guide the population without locking the genotype into a local optimum that must be inherited.
  • The reported generalist parameters could serve as a practical starting point for applying these hybrids to other combinatorial problems where local search is cheap relative to global search.
  • The runtime ordering may change on landscapes whose deception structure differs from the leading-block pattern, suggesting targeted experiments on other benchmark families.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper empirically evaluates Baldwinian, Lamarckian, and Darwinian evolutionary algorithms on Maximum Independent Set and Maximum Cut instances from six GraphBench datasets, reporting consistent outperformance of the local-search-augmented variants over plain Darwinian evolution and often over deep learning baselines, while also supplying a recommended generalist parameter set. Theoretically, it extends the Deceptive Leading Blocks function to arbitrary block length k and derives asymptotic runtime bounds showing that, for k>2, Baldwinian evolution is fastest, followed by Lamarckian, then Darwinian when counting only fitness evaluations; when local-search costs are included the ordering is stated to depend on implementation details yet Baldwinian remains fastest from small k onward.

Significance. If the empirical protocol is fully reproducible and the local-search cost model is made explicit, the work would supply both practical guidance (generalist parameters) and the first rigorous asymptotic comparison of the three evolutionary regimes on an extended benchmark, strengthening the case for hybrid local-search EAs. The use of modern runtime-analysis tools to obtain matching upper and lower bounds is a methodological strength.

major comments (2)
  1. [Theoretical analysis] Theoretical analysis (paragraph immediately after the fitness-evaluation-only bounds): the claim that 'Baldwinian evolution staying fastest from small block lengths onwards' when local-search cost is included rests on an unstated cost model; no explicit function is supplied that defines the per-step cost of local search (e.g., as a function of block length k, neighborhood size, or whether Lamarckian inheritance alters the cost), so it is impossible to verify whether the reported ordering survives plausible implementations such as exhaustive enumeration inside each block.
  2. [Empirical evaluation] Empirical evaluation section: the abstract asserts 'consistent outperformance' and 'in the great majority of cases' superiority over deep-learning baselines, yet the manuscript provides neither the precise data-exclusion rules, the statistical test used for 'consistent', nor the full experimental protocol (random seeds, termination criteria, how ties are broken), rendering the central empirical claim difficult to assess for robustness.
minor comments (2)
  1. [Theoretical analysis] Notation: the extended Deceptive Leading Blocks function is introduced without an explicit equation number or definition of the fitness function for arbitrary k; adding a displayed equation would improve readability.
  2. Table/figure captions: several result tables lack explicit indication of whether reported runtimes include or exclude local-search overhead, which is directly relevant to the cost-model discussion.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major point below and will revise the paper accordingly to improve clarity and reproducibility.

read point-by-point responses

  1. Referee: [Theoretical analysis] Theoretical analysis (paragraph immediately after the fitness-evaluation-only bounds): the claim that 'Baldwinian evolution staying fastest from small block lengths onwards' when local-search cost is included rests on an unstated cost model; no explicit function is supplied that defines the per-step cost of local search (e.g., as a function of block length k, neighborhood size, or whether Lamarckian inheritance alters the cost), so it is impossible to verify whether the reported ordering survives plausible implementations such as exhaustive enumeration inside each block.

    Authors: We agree that an explicit cost model was not provided. In the revised manuscript we will introduce a concrete cost function for local search (e.g., linear in block length k for exhaustive enumeration within each block, with an additional constant factor for Lamarckian inheritance of the improved genotype). We will then restate the asymptotic comparison under this model and show that the claimed ordering (Baldwinian fastest from small k onward) holds for the standard implementation we have in mind. The fitness-evaluation-only bounds remain unchanged. revision: yes

  2. Referee: [Empirical evaluation] Empirical evaluation section: the abstract asserts 'consistent outperformance' and 'in the great majority of cases' superiority over deep-learning baselines, yet the manuscript provides neither the precise data-exclusion rules, the statistical test used for 'consistent', nor the full experimental protocol (random seeds, termination criteria, how ties are broken), rendering the central empirical claim difficult to assess for robustness.

    Authors: We acknowledge the lack of protocol detail. The revision will add a dedicated experimental-setup subsection that specifies: (i) the exact statistical test and significance threshold used to support 'consistent outperformance', (ii) random seeds and number of independent runs, (iii) termination criteria, (iv) tie-breaking rule, and (v) any data-exclusion criteria. We will also make the full result tables and code repository reference explicit so that the claims can be verified. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: runtime bounds derived directly from model definitions

full rationale

The paper extends the Deceptive Leading Blocks function and applies standard theoretical runtime analysis to prove asymptotic upper and lower bounds on expected runtime for the three evolution types. These proofs follow from the algorithm descriptions and problem structure without any parameter fitting, self-referential definitions, or load-bearing self-citations. The statement on local-search cost is presented as implementation-dependent rather than derived from a fitted model or renamed input. Empirical results on GraphBench datasets are reported separately and do not feed back into the theoretical claims.

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

Abstract provides no explicit free parameters, axioms, or invented entities; all modeling choices remain implicit.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Fresh Look at Lamarckian Evolution and the Baldwin Effect." pith.science (2026). https://pith.science/paper/Q6GYB65S

@misc{pith2026260528703,
  author       = {Pith},
  title        = {Pith review of: A Fresh Look at Lamarckian Evolution and the Baldwin Effect},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q6GYB65S}},
  note         = {Machine review of arXiv:2605.28703}
}
read the original abstract

Baldwinian and Lamarckian evolution have existed for a long time in evolutionary algorithms (EAs) without ever dominating the academic literature or practical applications. In this work, we use modern empirical and theoretical methods to revisit Lamarckian and Baldwinian evolution and rigorously compare them with the generic Darwinian evolution. On the empirical side, we run a comprehensive suite of experiments on graphs from six different datasets from the recent GraphBench benchmark on Maximum Independent Set and Maximum Cut problems. Our results show that Baldwinian and Lamarckian evolution consistently outperform Darwinian evolution, confirming the great potential of local search augmented evolutionary algorithms. Notably, in the great majority of cases, all EAs outperform recent deep learning baselines and approach the performance of highly specialised heuristic and exact solvers. We furthermore report a high-performing set of generalist parameters for all studied evolution types that we hope will be of use to practitioners in future. On the theoretical side, we extend the existing Deceptive Leading Block benchmark to arbitrary block length and use tools from modern theoretical runtime analysis to prove upper and lower bounds on the expected runtime. For block lengths greater than two, Baldwinian evolution is asymptotically faster than Lamarckian which is asymptotically faster than Darwinian evolution. When accounting for the cost of the local search procedure in fitness evaluations, the ordering depends on the implementation with Baldwinian evolution staying fastest from small block lengths onwards, explaining its strong empirical performance.

Figures

Figures reproduced from arXiv: 2605.28703 by the authors.

Figure 1
Figure 1. Mean fitness score (↑) over fitness evaluations per EA evolution across datasets for MIS and MC. Curves use the per-(dataset, problem) best parameter configuration. Darwinian (solid blue), Baldwinian (dashed green), Lamarckian (dash-dot red), L-B (dotted orange) [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗
Figure 2
Figure 2. Mean pairwise Hamming distance as a measure of population diversity [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 5 canonical work pages

  1. [1]

    (eds.): Theory of Randomized Search Heuristics

    Auger, A., Doerr, B. (eds.): Theory of Randomized Search Heuristics. World Sci- entific Publishing (2011)

    2011

  2. [2]

    The American Naturalist30, 441–451 (1896)

    Baldwin, J.M.: A new factor in evolution. The American Naturalist30, 441–451 (1896)

  3. [3]

    In: Parallel Problem Solving from Nature, PPSN

    Böttcher, S., Doerr, B., Neumann, F.: Optimal fixed and adaptive mutation rates for the LeadingOnes problem. In: Parallel Problem Solving from Nature, PPSN

  4. [4]

    pp. 1–10. Springer (2010)

    2010

  5. [5]

    In: Genetic and Evolutionary Computation Conference, GECCO 2019

    Doerr, B.: A tight runtime analysis for the cGA on jump functions: EDAs can cross fitness valleys at no extra cost. In: Genetic and Evolutionary Computation Conference, GECCO 2019. pp. 1488–1496. ACM (2019)

    2019

  6. [6]

    In: Doerr, B., Neumann, F

    Doerr,B.:Probabilistictoolsfortheanalysisofrandomizedoptimizationheuristics. In: Doerr, B., Neumann, F. (eds.) Theory of Evolutionary Computation: Recent Developments in Discrete Optimization, pp. 1–87. Springer (2020), also available at https://arxiv.org/abs/1801.06733

    work page arXiv 2020

  7. [7]

    Al- gorithmica83, 3059–3107 (2021)

    Doerr, B.: The runtime of the compact genetic algorithm on Jump functions. Al- gorithmica83, 3059–3107 (2021)

    2021

  8. [8]

    IEEE Transactions on Evolutionary Computation29, 1864–1872 (2025)

    Doerr,B.,Kelley,A.J.:TheruntimeofRandomizedLocalSearchonthegeneralized Needle problem. IEEE Transactions on Evolutionary Computation29, 1864–1872 (2025). https://doi.org/10.1109/TEVC.2024.3453776

    work page doi:10.1109/tevc.2024.3453776 2025

Show all 32 references
  1. [9]

    Algorithmica 86, 367–395 (2024)

    Doerr, B., Kötzing, T.: Lower bounds from fitness levels made easy. Algorithmica 86, 367–395 (2024)

    2024

  2. [10]

    IEEE Transactions on Evolutionary Computation24, 1025–1034 (2020)

    Doerr, B., Krejca, M.S.: Significance-based estimation-of-distribution algorithms. IEEE Transactions on Evolutionary Computation24, 1025–1034 (2020)

    2020

  3. [11]

    Evolutionary Computation29, 543–563 (2021)

    Doerr, B., Krejca, M.S.: The univariate marginal distribution algorithm copes well with deception and epistasis. Evolutionary Computation29, 543–563 (2021)

    2021

  4. [12]

    (eds.): Theory of Evolutionary Computation—Recent Developments in Discrete Optimization

    Doerr, B., Neumann, F. (eds.): Theory of Evolutionary Computation—Recent Developments in Discrete Optimization. Springer (2020), also available at http://www.lix.polytechnique.fr/Labo/Benjamin.Doerr/doerr_neumann_book.html

    2020

  5. [13]

    Springer (2015)

    Eiben, A.E., Smith, J.E.: Introduction to Evolutionary Computing, Second Edi- tion. Springer (2015)

    2015

  6. [14]

    Gurobi Optimization, LLC: Gurobi Optimizer Reference Manual (2024)

    2024

  7. [15]

    Complex Systems 1(1987)

    Hinton, G.E., Nowlan, S.J.: How learning can guide evolution. Complex Systems 1(1987)

    1987

  8. [16]

    Evolutionary Computation5, 31–60 (1997)

    Houck, C.R., Joines, J.A., Kay, M.G., Wilson, J.R.: Empirical investigation of the benefits of partial lamarckianism. Evolutionary Computation5, 31–60 (1997)

    1997

  9. [17]

    Springer (2013)

    Jansen, T.: Analyzing Evolutionary Algorithms – The Computer Science Perspec- tive. Springer (2013)

    2013

  10. [18]

    In: Algorithm Engineering and Experiments, ALENEX 2019

    Lamm, S., Schulz, C., Strash, D., Williger, R., Zhang, H.: Exactly solving the max- imum weight independent set problem on large real-world graphs. In: Algorithm Engineering and Experiments, ALENEX 2019. pp. 144–158. SIAM (2019)

    2019

  11. [19]

    In: Foundations of Genetic Algorithms, FOGA 2019

    Lehre,P.K.,Nguyen,P.T.H.:Onthelimitationsoftheunivariatemarginaldistribu- tion algorithm to deception and where bivariate EDAs might help. In: Foundations of Genetic Algorithms, FOGA 2019. pp. 154–168. ACM (2019)

    2019

  12. [20]

    Algorithmica64, 623–642 (2012)

    Lehre, P.K., Witt, C.: Black-box search by unbiased variation. Algorithmica64, 623–642 (2012)

    2012

  13. [21]

    Caltech concurrent computation program, C3P Report (1989) 16 I

    Moscato, P.: On evolution, search, optimization, genetic algorithms and martial arts: Towards memetic algorithms. Caltech concurrent computation program, C3P Report (1989) 16 I. Benito et al

    1989

  14. [22]

    Springer (2010)

    Neumann, F., Witt, C.: Bioinspired Computation in Combinatorial Optimization – Algorithms and Their Computational Complexity. Springer (2010)

    2010

  15. [23]

    Artificial Intelligence287, 103345 (2020)

    Nguyen, P.T.H., Sudholt, D.: Memetic algorithms outperform evolutionary algo- rithms in multimodal optimisation. Artificial Intelligence287, 103345 (2020)

    2020

  16. [24]

    CoRR abs/2512.04475(2025)

    Stoll, T., Qian, C., Finkelshtein, B., Parviz, A., Weber, D., Frasca, F., Shavit, H., Siraudin, A., Mielke, A., Anastacio, M., Müller, E., Bechler-Speicher, M., Bronstein, M.M., Galkin, M., Hoos, H.H., Niepert, M., Perozzi, B., Tönshoff, J., Morris, C.: Graphbench: Next-genera...

    2025 arXiv

  17. [25]

    Algorithmica59, 343–368 (2011)

    Sudholt, D.: Hybridizing evolutionary algorithms with variable-depth search to overcome local optima. Algorithmica59, 343–368 (2011)

    2011

  18. [26]

    In: Auger, A., Doerr, B

    Sudholt, D.: Memetic evolutionary algorithms. In: Auger, A., Doerr, B. (eds.) The- ory of Randomized Search Heuristics: Foundations and Recent Developments, pp. 141–169. World Scientific (2011)

    2011

  19. [27]

    CoRRabs/cs/0603004(2006)

    Valdivieso, P.Á.C., Arenas, M.G., Castellano, F.J.G., Guervós, J.J.M., Prieto, A., Rivas, V.M., Romero, G.: Lamarckian evolution and the baldwin effect in evolu- tionary neural networks. CoRRabs/cs/0603004(2006)

    2006 arXiv

  20. [28]

    Information and Computation296, 105125 (2024)

    Wang, S., Zheng, W., Doerr, B.: Choosing the right algorithm with hints from complexity theory. Information and Computation296, 105125 (2024)

    2024

  21. [29]

    In: Automata, Languages and Programming, ICALP 2005

    Wegener, I.: Simulated annealing beats Metropolis in combinatorial optimization. In: Automata, Languages and Programming, ICALP 2005. pp. 589–601. Springer (2005)

    2005

  22. [30]

    In: Parallel Problem Solving from Nature, PPSN

    Whitley, L.D., Gordon, V.S., Mathias, K.E.: Lamarckian evolution, the baldwin effect and function optimization. In: Parallel Problem Solving from Nature, PPSN

  23. [31]

    pp. 6–15. Springer (1994)

    1994

  24. [32]

    Springer (2019) Baldwinian vs

    Zhou, Z.H., Yu, Y., Qian, C.: Evolutionary Learning: Advances in Theories and Algorithms. Springer (2019) Baldwinian vs. Lamarckian Evolution 17 A Problem Formulations In this section, we formally define the two combinatorial optimisation problems studied in this paper. Recall...

    2019

Pith tools

Reviewed June 29, 2026 · model on record in the stance chip above.