pith. sign in

arxiv: 2606.24166 · v1 · pith:XEAJIFVKnew · submitted 2026-06-23 · 💻 cs.NE

Distributed Quality-Diversity Search for Toxicity in Large Language Models

Onkar Shelar , Travis Desell This is my paper

Pith reviewed 2026-06-25 22:10 UTC · model grok-4.3

classification 💻 cs.NE
keywords toxicity searchevolutionary algorithmsquality-diversitylarge language modelsred teamingspeciationMPI parallelizationniche maintenance
0
0 comments X

The pith

ToxSearch-S reaches competitive peak toxicity in LLM prompts with a less toxic search trajectory than prior methods.

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

The paper presents ToxSearch-S as a speciated version of evolutionary prompt search for eliciting toxic responses from large language models. It uses incremental embedding-driven niche maintenance to partition the search space and runs the process on an MPI master-worker architecture that parallelizes evaluation. Under fixed budgets, the method matches the highest toxicity scores of earlier approaches while recording lower toxicity in the best-so-far sequence, which the authors interpret as reduced cumulative search pressure. It also produces more localized behavioral clusters and delivers clear wall-clock speedups from parallel workers without changing final peak performance.

Core claim

ToxSearch-S, a speciated extension of toxicity-focused evolutionary prompt search with incremental, embedding-driven niche maintenance, attains peak toxicity competitive with both ToxSearch and RainbowPlus while following a measurably less toxic best-so-far trajectory under a common budget, indicating lower cumulative search pressure. Diversity is non-uni-dimensional: RainbowPlus yields greater embedding-level spread, whereas ToxSearch-S partitions high-toxicity prompts into more localized behavioral pockets, reflected by a higher DBSCAN cluster count. MPI distribution delivers substantial wall-clock gains, approximately 1.8 times with two workers and 3.2 times with four, while leaving Best@

What carries the argument

Incremental speciation via embedding-driven niche maintenance with DBSCAN clustering, which maintains separate behavioral niches in the population while the MPI master-worker setup centralizes bookkeeping on rank zero and distributes prompt generation and evaluation.

If this is right

  • Red-teaming can reach high-toxicity prompts while recording lower average toxicity across the search history.
  • Four-worker MPI runs increase final species count and the number of toxicity-bearing species without raising global peak toxicity.
  • Embedding-level spread and cluster count measure different aspects of diversity, with speciation favoring localized pockets.
  • Parallel workers compress wall-clock time by up to 3.2 times while preserving the same best outcomes as sequential runs.

Where Pith is reading between the lines

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

  • The lower cumulative toxicity trajectory could reduce the volume of harmful content generated during the red-teaming process itself.
  • The speciation mechanism may transfer to other quality-diversity tasks in AI safety where managing exposure risk during search matters.
  • Larger species cardinality with more workers suggests the method could scale to broader prompt spaces if cluster quality holds.

Load-bearing premise

The embedding space and DBSCAN clustering used for niche maintenance are assumed to produce behaviorally meaningful partitions that do not systematically miss important toxicity failure modes or bias the evolutionary search.

What would settle it

If repeated runs show that the DBSCAN-derived clusters do not align with distinct toxicity categories identified by human review, or if ToxSearch-S peak toxicity falls statistically below the baselines under identical budgets, the central performance claims would be refuted.

Figures

Figures reproduced from arXiv: 2606.24166 by Onkar Shelar, Travis Desell.

Figure 1
Figure 1. Figure 1: Best-so-far toxicity versus evaluated genomes at a common budget (B=1000); the legend reports max@1000 for each method. schedule (G=50) with steady-state controls α=30 and β=3, yielding approxi￾mately 103 accumulated genomes without a hard integrated cap. ToxSearch-S used θsim=0.25, θmerge=0.25, Cmin=5, Cspecies=25, Creserves=500, and Tspecies=7, terminating at B integrated evaluations. Full hyperparameter… view at source ↗
Figure 2
Figure 2. Figure 2: Run-level quality summaries shown as raincloud plots (half-violin density + per-run points + median line). Left panel: Best@B. Right panel: AUC@B. ToxSearch ToxSearch-S RainbowPlus 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 # clusters (higher: more diverse) DBSCAN clusters (top-50) ToxSearch ToxSearch-S RainbowPlus 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Mean cosine distance (higher: more diverse) Mean pairwise distance (top… view at source ↗
Figure 3
Figure 3. Figure 3: Run-level diversity summaries on top-K prompts (K=50) shown as raincloud plots (half-violin density + per-run points + median line). Left panel: DBSCAN clus￾ter count on top-K embeddings. Right panel: semantic spread (mean pairwise cosine distance) on top-K embeddings. stone curves ( [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance versus evaluated genomes (n=7 runs per mode; solid mean, min– max band): left-cumulative wall-clock time; right-throughput. results indicate that, at least up to four workers, MPI converts additional paral￾lel resources into substantial reductions in wall-clock time without changing the underlying search procedure. Under a limited evaluation budget B, increasing parallelism yields large, system… view at source ↗
Figure 5
Figure 5. Figure 5: Outcome trajectories versus evaluated genomes (n=7 runs per mode; median with IQR; solid lines). Top row: inter- and intra-species diversity; bottom row: best-so￾far toxicity and species count. pairwise two-sided Mann–Whitney U tests with Holm adjustment and Cliff’s δ with percentile-bootstrap 95% confidence intervals. Because n=7 per mode provides limited power, non-rejection is not interpreted as evidenc… view at source ↗
Figure 6
Figure 6. Figure 6: MPI communication flow [PITH_FULL_IMAGE:figures/full_fig_p025_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Gen-0 calibration diagnostic (n=100 seeds). Top: pairwise densemble with θsim=θmerge=0.25 marked. Middle: (dgenotype-norm, dphenotype) coloured by densemble. Bottom: ECDFs of tight triangle slack s1 and tight 2-inframetric slack s2 over all [PITH_FULL_IMAGE:figures/full_fig_p028_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: 2D Landmark-MDS embedding map of deduplicated prompt embeddings (com￾puted with all-MiniLM-L6-v2); point size scales with toxicity. Landmark MDS fits MDS on a representative subset and maps all points into the learned 2D space [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Best-so-far toxicity versus cumulative wall-clock time. Dashed vertical lines mark the first wall-time milestone at which each mode’s median has reached its final best-so-far value [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Top ten (run, species) pairs ranked by pool-wide peak toxicity Tmax after pool￾ing all 21 runs. Each horizontal boxplot summarizes the per-genome toxicity distribu￾tion within that species in that run (median and IQR), with jittered points showing individual genomes. Rows are ordered from highest to lowest peak Tmax. The y-axis shows truncated c-TF-IDF labels for readability. Ranking is based only on toxi… view at source ↗
read the original abstract

Large Language Models remain vulnerable to adversarial prompts that elicit harmful responses, and scaling red-teaming to cover a broad range of failure modes is constrained by the cost of text generation and evaluation. We present \emph{ToxSearch-S}, a speciated extension of toxicity-focused evolutionary prompt search with incremental, embedding-driven niche maintenance, together with an MPI master-worker realization that centralizes population and species bookkeeping on rank~0 while offloading prompt evolution and evaluation to $n_w$ parallel workers. Under a common budget, ToxSearch-S attains peak toxicity competitive with both ToxSearch and RainbowPlus while following a measurably less toxic best-so-far trajectory, indicating lower cumulative search pressure. Diversity is non-uni-dimensional: RainbowPlus yields greater embedding-level spread, whereas ToxSearch-S partitions high-toxicity prompts into more localized behavioral pockets, reflected by a higher DBSCAN cluster count. MPI distribution delivers substantial wall-clock gains, approximately $1.8\times$ with two workers and $3.2\times$ with four, while leaving Best@B statistically indistinguishable from sequential execution. Four-worker runs also produce significantly larger final species cardinality and more toxicity-bearing species, without a reliable gain in global peak toxicity. These results position incremental speciation as a practical quality-diversity mechanism for AI Safety and MPI as an effective means of compressing time-to-result while preserving measured search outcomes.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 2 minor

Summary. The manuscript introduces ToxSearch-S, a speciated quality-diversity extension of evolutionary prompt search for eliciting toxicity in LLMs. It uses incremental embedding-driven niche maintenance with DBSCAN clustering and an MPI master-worker parallel implementation. Under a fixed evaluation budget, ToxSearch-S is reported to reach peak toxicity competitive with ToxSearch and RainbowPlus baselines while exhibiting a less toxic best-so-far trajectory (indicating lower cumulative search pressure), higher DBSCAN cluster counts reflecting more localized behavioral pockets, and wall-clock speedups of approximately 1.8× (two workers) and 3.2× (four workers) with no change in Best@B outcomes.

Significance. If the embedding + DBSCAN speciation produces partitions that align with distinct toxicity failure modes, the work provides a practical QD mechanism for red-teaming that balances peak performance against reduced cumulative exposure, together with a scalable distributed realization. The common-budget empirical comparisons and reported MPI speedups are concrete strengths; the parallelization results appear reproducible from the described master-worker design.

major comments (1)
  1. [Abstract] Abstract and Results (diversity claims): The central interpretation that a higher DBSCAN cluster count indicates 'more localized behavioral pockets' and supports lower cumulative search pressure assumes the embedding space and DBSCAN produce behaviorally meaningful partitions of toxicity failure modes. No validation is described (e.g., cluster content inspection against known toxicity categories or ablation on embedding choice), which is load-bearing for attributing the flatter trajectory and competitiveness with RainbowPlus to the speciation mechanism rather than arbitrary metric-space behavior.
minor comments (2)
  1. [Abstract] The abstract states 'Best@B statistically indistinguishable' and 'significantly larger final species cardinality' but does not name the statistical test, significance threshold, or correction method; these details belong in the experimental protocol section.
  2. [Methods] Dataset descriptions, exact prompt templates, and full hyperparameter tables for the evolutionary operators and DBSCAN (eps, min_samples) are referenced but not reproduced in the provided abstract; ensure they appear explicitly in the methods for reproducibility.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and for recognizing the reproducibility of the MPI results and the practical value of the QD approach for red-teaming. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract and Results (diversity claims): The central interpretation that a higher DBSCAN cluster count indicates 'more localized behavioral pockets' and supports lower cumulative search pressure assumes the embedding space and DBSCAN produce behaviorally meaningful partitions of toxicity failure modes. No validation is described (e.g., cluster content inspection against known toxicity categories or ablation on embedding choice), which is load-bearing for attributing the flatter trajectory and competitiveness with RainbowPlus to the speciation mechanism rather than arbitrary metric-space behavior.

    Authors: We agree that the manuscript provides no direct validation (cluster inspection against toxicity taxonomies or embedding ablations) that the DBSCAN partitions correspond to distinct, semantically meaningful toxicity failure modes. This limits the strength of any causal claim that the speciation mechanism itself produces the observed lower search pressure. The lower best-so-far toxicity trajectory is an empirical measurement independent of cluster semantics; the higher DBSCAN count is reported simply as the outcome of applying incremental embedding-driven niche maintenance. In the revised manuscript we will (i) add an explicit limitations paragraph stating that cluster validity was not verified against external toxicity categories and (ii) qualify the diversity claim to emphasize that ToxSearch-S yields more embedding-space clusters while RainbowPlus yields greater overall spread, without asserting that the clusters map to known behavioral modes. No new experiments are planned for this revision. revision: partial

Circularity Check

0 steps flagged

No circularity; empirical method comparisons are independent of inputs

full rationale

The paper presents an empirical study of ToxSearch-S, a speciated evolutionary search method using embeddings and DBSCAN for niche maintenance, evaluated against baselines (ToxSearch, RainbowPlus) under fixed computational budgets. All reported outcomes—peak toxicity, best-so-far trajectories, cluster counts, and MPI speedups—are direct measurements from experiments rather than derivations, predictions fitted to the same data, or results justified solely by self-citations. No equations, uniqueness theorems, or ansatzes are invoked that reduce to the method's own definitions or prior author work by construction. The central claims rest on observable differences in search trajectories and diversity metrics, which are falsifiable against external baselines and do not contain self-referential loops.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no information on fitted parameters, background axioms, or new postulated entities.

pith-pipeline@v0.9.1-grok · 5770 in / 1078 out tokens · 20437 ms · 2026-06-25T22:10:16.329125+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 22 canonical work pages

  1. [1]

    Google perspective api (Jan 2026),https://perspectiveapi.com

    2026

  2. [2]

    Openai moderation api (Jan 2026),https://platform.openai.com/docs/ api-reference/moderations

    2026

  3. [3]

    Alba, E., Tomassini, M.: Parallelism and evolutionary algorithms. Trans. Evol. Comp6(5), 443–462 (Oct 2002).https://doi.org/10.1109/TEVC.2002.800880, https://doi.org/10.1109/TEVC.2002.800880

    work page doi:10.1109/tevc.2002.800880 2002

  4. [4]

    In: Proceed- ings of the 9th annual conference on Genetic and evolutionary computation

    Ando, S.: Heuristic speciation for evolving neural network ensemble. In: Proceed- ings of the 9th annual conference on Genetic and evolutionary computation. pp. 1766–1773 (2007)

    2007

  5. [5]

    In: Ku, L.W., Martins, A., Srikumar, V

    Bhardwaj, R., Do, D.A., Poria, S.: Language models are Homer simpson! safety re-alignment of fine-tuned language models through task arithmetic. In: Ku, L.W., Martins, A., Srikumar, V. (eds.) Proceedings of the 62nd Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 14138–14149. Association for Computational Lin...

    work page doi:10.18653/v1/2024.acl-long.762 2024

  6. [6]

    ArXivabs/2308.09662(2023),https://api

    Bhardwaj, R., Poria, S.: Red-teaming large language models using chain of utterances for safety-alignment. ArXivabs/2308.09662(2023),https://api. semanticscholar.org/CorpusID:261030829

    arXiv 2023

  7. [7]

    Population diversity and inheritance in genetic programming for symbolic regression

    Burlacu, B., Yang, K., Affenzeller, M.: Population diversity and inheritance in genetic programming for symbolic regression. Natural Computing23(01 2023). https://doi.org/10.1007/s11047-022-09934-x

    work page doi:10.1007/s11047-022-09934-x 2023

  8. [8]

    semanticscholar.org/CorpusID:14264381

    Cantú-Paz, E.: A survey of parallel genetic algorithms (2000),https://api. semanticscholar.org/CorpusID:14264381

    2000

  9. [9]

    Cantú-Paz, E., Goldberg, D.E.: On the scalability of parallel genetic algorithms. Evol. Comput.7(4), 429–449 (Dec 1999).https://doi.org/10.1162/evco.1999. 7.4.429,https://doi.org/10.1162/evco.1999.7.4.429

    work page doi:10.1162/evco.1999 1999

  10. [10]

    Cantú-Paz, E., Goldberg, D.E.: Efficient parallel genetic algorithms: the- ory and practice. Computer Methods in Applied Mechanics and Engineer- ing186(2), 221–238 (2000).https://doi.org/https://doi.org/10.1016/ S0045-7825(99)00385-0,https://www.sciencedirect.com/science/article/ pii/S0045782599003850

    2000

  11. [11]

    In: Proceedings of the 62nd Annual Meeting of the Asso- ciation for Computational Linguistics (Volume 1: Long Papers)

    Cao, B., Cao, Y., Lin, L., Chen, J.: Defending against alignment-breaking attacks via robustly aligned llm. In: Proceedings of the 62nd Annual Meeting of the Asso- ciation for Computational Linguistics (Volume 1: Long Papers). pp. 10542–10560 (2024)

    2024

  12. [12]

    Artificial Intelli- gence Review58(11), 335 (2025)

    Chauhan, D., Shivani, Jung, D., Yadav, A.: Advancements in multimodal differ- ential evolution: a comprehensive review and future perspectives. Artificial Intelli- gence Review58(11), 335 (2025)

    2025

  13. [13]

    arXiv preprint arXiv:2501.01741 (2025)

    Corbo, S., Bancale, L., De Gennaro, V., Lestingi, L., Scotti, V., Camilli, M.: How toxic can you get? search-based toxicity testing for large language models. arXiv preprint arXiv:2501.01741 (2025)

    arXiv 2025

  14. [14]

    In: Proceedings of the Genetic and Evolution- ary Computation Conference

    Cully, A.: Multi-emitter map-elites: improving quality, diversity and data efficiency with heterogeneous sets of emitters. In: Proceedings of the Genetic and Evolution- ary Computation Conference. p. 84–92. GECCO ’21, Association for Comput- ing Machinery, New York, NY, USA (2021).https://doi.org/10.1145/3449639. 3459326,https://doi.org/10.1145/3449639.34...

    work page doi:10.1145/3449639 2021

  15. [15]

    arXiv preprint arXiv:2504.15047 (2025)

    Dang, Q.A., Ngo, C., Hy, T.S.: Rainbowplus: Enhancing adversarial prompt gen- eration via evolutionary quality-diversity search. arXiv preprint arXiv:2504.15047 (2025)

    arXiv 2025

  16. [16]

    Evolutionary Computation 21(2), 261–291 (May 2013).https://doi.org/10.1162/EVCO_a_00076

    Depolli, M., Trobec, R., Filipič, B.: Asynchronous master-slave parallelization of differential evolution for multi-objective optimization. Evolutionary Computation 21(2), 261–291 (May 2013).https://doi.org/10.1162/EVCO_a_00076

    work page doi:10.1162/evco_a_00076 2013

  17. [17]

    In: IEEE Congress on Evolutionary Computation

    Desell, T., Anderson, D.P., Magdon-Ismail, M., Newberg, H., Szymanski, B.K., Varela, C.A.: An analysis of massively distributed evolutionary algorithms. In: IEEE Congress on Evolutionary Computation. pp. 1–8 (2010).https://doi.org/ 10.1109/CEC.2010.5586073

    work page doi:10.1109/cec.2010.5586073 2010

  18. [18]

    In: 2008 IEEE International Symposium on Parallel and Distributed Processing

    Desell, T., Szymanski, B., Varela, C.: Asynchronous genetic search for scientific modeling on large-scale heterogeneous environments. In: 2008 IEEE International Symposium on Parallel and Distributed Processing. pp. 1–12 (April 2008).https: //doi.org/10.1109/IPDPS.2008.4536169

    work page doi:10.1109/ipdps.2008.4536169 2008

  19. [19]

    In: Proceed- ings of the 10th Annual Conference on Genetic and Evolutionary Computation

    Desell, T., Szymanski, B., Varela, C.: An asynchronous hybrid genetic-simplex search for modeling the milky way galaxy using volunteer computing. In: Proceed- ings of the 10th Annual Conference on Genetic and Evolutionary Computation. p. 921–928. GECCO ’08, Association for Computing Machinery, New York, NY, USA(2008).https://doi.org/10.1145/1389095.138927...

    work page doi:10.1145/1389095.1389273 2008

  20. [20]

    Ethayarajh, K.: How contextual are contextualized word representations? compar- ing the geometry of bert, elmo, and gpt-2 embeddings. In: Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th inter- national joint conference on natural language processing (EMNLP-IJCNLP). pp. 55–65 (2019)

    2019

  21. [21]

    SIAM Journal on discrete mathematics17(1), 134–160 (2003)

    Fagin, R., Kumar, R., Sivakumar, D.: Comparing top k lists. SIAM Journal on discrete mathematics17(1), 134–160 (2003)

    2003

  22. [22]

    International Journal of Computer Vision30(3), 219–231 (1998)

    Fagin, R., Stockmeyer, L.: Relaxing the triangle inequality in pattern matching. International Journal of Computer Vision30(3), 219–231 (1998)

    1998

  23. [23]

    arXiv preprint arXiv:2309.16797 (2023)

    Fernando, C., Banarse, D., Michalewski, H., Osindero, S., Rocktäschel, T.: Prompt- breeder: Self-referential self-improvement via prompt evolution. arXiv preprint arXiv:2309.16797 (2023)

    Pith/arXiv arXiv 2023

  24. [24]

    Multiview Symbolic Regression , year =

    Flageat, M., Lim, B., Cully, A.: Enhancing map-elites with multiple parallel evolution strategies. In: Proceedings of the Genetic and Evolutionary Computa- tion Conference. p. 1082–1090. GECCO ’24, Association for Computing Machin- ery, New York, NY, USA (2024).https://doi.org/10.1145/3638529.3654089, https://doi.org/10.1145/3638529.3654089

    work page doi:10.1145/3638529.3654089 2024

  25. [25]

    arXiv preprint arXiv:2309.08532 (2023)

    Guo, Q., Wang, R., Guo, J., Li, B., Song, K., Tan, X., Liu, G., Bian, J., Yang, Y.: Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. arXiv preprint arXiv:2309.08532 (2023)

    Pith/arXiv arXiv 2023

  26. [26]

    Gustafson, S., Burke, E.K.: The speciating island model: An alternative parallel evolutionary algorithm. Journal of Parallel and Distributed Com- puting66(8), 1025–1036 (2006).https://doi.org/https://doi.org/10.1016/ j.jpdc.2006.04.017,https://www.sciencedirect.com/science/article/pii/ S0743731506001067, special Issue: Parallel Bioinspired Algorithms

    2006

  27. [27]

    arXiv preprint arXiv:2406.11654 (2024)

    Han, V.T.Y., Bhardwaj, R., Poria, S.: Ruby teaming: Improving quality diversity search with memory for automated red teaming. arXiv preprint arXiv:2406.11654 (2024)

    arXiv 2024

  28. [28]

    In: Proceed- ings of the Genetic and Evolutionary Computation Conference Companion

    Karns, J., Desell, T.: Improving the scalability of distributed neuroevolution using modular congruence class generated innovation numbers. In: Proceed- ings of the Genetic and Evolutionary Computation Conference Companion. p. Distributed Quality-Diversity Search for Toxicity in Large Language Models 17 1299–1307. GECCO ’21, Association for Computing Mach...

    work page doi:10.1145/3449726.3463202 2021

  29. [29]

    In: Leung, C.S., Lee, M., Chan, J.H.(eds.)NeuralInformationProcessing.pp.630–637.SpringerBerlinHeidelberg, Berlin, Heidelberg (2009)

    Kim, K.J., Cho, S.B.: Evaluation of distance measures for speciated evolutionary neural networks in pattern classification problems. In: Leung, C.S., Lee, M., Chan, J.H.(eds.)NeuralInformationProcessing.pp.630–637.SpringerBerlinHeidelberg, Berlin, Heidelberg (2009)

    2009

  30. [30]

    Evolutionary Computation19(2), 189–223 (2011) https://doi

    Lehman, J., Stanley, K.O.: Abandoning objectives: Evolution through the search for novelty alone. Evolutionary Computation19(2), 189–223 (June 2011).https: //doi.org/10.1162/EVCO_a_00025

    work page doi:10.1162/evco_a_00025 2011

  31. [31]

    Machine Intelligence Research19(1), 3–23 (2022).https: //doi.org/https://doi.org/10.1007/s11633-022-1317-4

    Li, J.Y., Zhan, Z.H., Zhang, J.: Evolutionary computation for expensive opti- mization: A survey. Machine Intelligence Research19(1), 3–23 (2022).https: //doi.org/https://doi.org/10.1007/s11633-022-1317-4

    work page doi:10.1007/s11633-022-1317-4 2022

  32. [32]

    In: Proceedings of the Genetic and Evolutionary Computation Conference Companion

    Lim, B., Allard, M., Grillotti, L., Cully, A.: Qdax: on the benefits of massive par- allelization for quality-diversity. In: Proceedings of the Genetic and Evolutionary Computation Conference Companion. p. 128–131. GECCO ’22, Association for Computing Machinery, New York, NY, USA (2022).https://doi.org/10.1145/ 3520304.3528927,https://doi.org/10.1145/3520...

    work page doi:10.1145/3520304.3528927 2022

  33. [33]

    arXiv preprint arXiv:2410.05295 (2024)

    Liu, X., Li, P., Suh, E., Vorobeychik, Y., Mao, Z., Jha, S., McDaniel, P., Sun, H., Li, B., Xiao, C.: Autodan-turbo: A lifelong agent for strategy self-exploration to jailbreak llms. arXiv preprint arXiv:2410.05295 (2024)

    arXiv 2024

  34. [34]

    arXiv preprint arXiv:2310.04451 (2023)

    Liu, X., Xu, N., Chen, M., Xiao, C.: Autodan: Generating stealthy jailbreak prompts on aligned large language models. arXiv preprint arXiv:2310.04451 (2023)

    Pith/arXiv arXiv 2023

  35. [35]

    arXiv preprint arXiv:2005.07376 (2020)

    Lyu, Z., Karns, J., ElSaid, A., Desell, T.: Improving neuroevolution using island extinction and repopulation. arXiv preprint arXiv:2005.07376 (2020)

    arXiv 2005

  36. [36]

    arXiv preprint arXiv:1504.04909 (2015)

    Mouret, J.B., Clune, J.: Illuminating search spaces by mapping elites. arXiv preprint arXiv:1504.04909 (2015)

    Pith/arXiv arXiv 2015

  37. [37]

    In: 1999 Third International Conference on Knowledge-Based Intelligent Information Engineering Systems

    Nowostawski, M., Poli, R.: Parallel genetic algorithm taxonomy. In: 1999 Third International Conference on Knowledge-Based Intelligent Information Engineering Systems. Proceedings (Cat. No.99TH8410). pp. 88–92 (Aug 1999).https://doi. org/10.1109/KES.1999.820127

    work page doi:10.1109/kes.1999.820127 1999

  38. [38]

    Pons Mir, M.: Follow the new leader: similarity-based clustering algorithms. B.S. thesis, Universitat Politècnica de Catalunya (2024),https://upcommons.upc.edu/ entities/publication/ac7edf57-fae7-4907-a4b3-68a1799185e9

    2024

  39. [39]

    Frontiers in Robotics and AI3(2016) https://doi

    Pugh, J.K., Soros, L.B., Stanley, K.O.: Quality diversity: A new frontier for evolu- tionary computation. Frontiers in Robotics and AIV olume 3 - 2016(2016). https://doi.org/10.3389/frobt.2016.00040,https://www.frontiersin.org/ journals/robotics-and-ai/articles/10.3389/frobt.2016.00040

    work page doi:10.3389/frobt.2016.00040 2016

  40. [40]

    In: Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP)

    Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert- networks. In: Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). pp. 3982–3992 (2019)

    2019

  41. [41]

    Rochester Institute of Technology: Research Computing Services (2019).https:// doi.org/10.34788/0S3G-QD15,https://doi.org/10.34788/0S3G-QD15, accessed 2026-01-23

    work page doi:10.34788/0s3g-qd15 2019

  42. [42]

    Advances in Neural Infor- mation Processing Systems37, 69747–69786 (2024)

    Samvelyan, M., Raparthy, S.C., Lupu, A., Hambro, E., Markosyan, A.H., Bhatt, M., Mao, Y., Jiang, M., Parker-Holder, J., Foerster, J., et al.: Rainbow teaming: Open-ended generation of diverse adversarial prompts. Advances in Neural Infor- mation Processing Systems37, 69747–69786 (2024)

    2024

  43. [43]

    In: International Confer- ence on Similarity Search and Applications

    Schubert, E.: A triangle inequality for cosine similarity. In: International Confer- ence on Similarity Search and Applications. pp. 32–44. Springer (2021) 18 O. Shelar et al

    2021

  44. [44]

    Expert Systems40(5), e13100 (2023).https://doi.org/https://doi.org/10.1111/exsy.13100, https://onlinelibrary.wiley.com/doi/abs/10.1111/exsy.13100

    Scott, E.O., Coletti, M., Schuman, C.D., Kay, B., Kulkarni, S.R., Parsa, M., Gunaratne, C., De Jong, K.A.: Avoiding excess computa- tion in asynchronous evolutionary algorithms. Expert Systems40(5), e13100 (2023).https://doi.org/https://doi.org/10.1111/exsy.13100, https://onlinelibrary.wiley.com/doi/abs/10.1111/exsy.13100

    work page doi:10.1111/exsy.13100 2023

  45. [45]

    In: Proceedings of the Companion Publication of the 2015 Annual Conference on Genetic and Evolutionary Computation

    Scott, E.O., De Jong, K.A.: Evaluation-time bias in asynchronous evolution- ary algorithms. In: Proceedings of the Companion Publication of the 2015 Annual Conference on Genetic and Evolutionary Computation. p. 1209–1212. GECCO Companion ’15, Association for Computing Machinery, New York, NY, USA(2015).https://doi.org/10.1145/2739482.2768482,https://doi.o...

    work page doi:10.1145/2739482.2768482 2015

  46. [46]

    arXiv preprint arXiv:2511.12487 (2025)

    Shelar, O., Desell, T.: Evolving prompts for toxicity search in large language mod- els. arXiv preprint arXiv:2511.12487 (2025)

    Pith/arXiv arXiv 2025

  47. [47]

    arXiv preprint arXiv:2601.20981 (2026)

    Shelar, O., Desell, T.: Diversifying toxicity search in large language models through speciation. arXiv preprint arXiv:2601.20981 (2026)

    Pith/arXiv arXiv 2026

  48. [48]

    arXiv preprint arXiv:2310.00892 (2023)

    Srivastava, A., Ahuja, R., Mukku, R.: No offense taken: Eliciting offensiveness from language models. arXiv preprint arXiv:2310.00892 (2023)

    arXiv 2023

  49. [49]

    Evolutionary Computation10(2), 99–127 (06 2002).https://doi.org/ 10.1162/106365602320169811,https://doi.org/10.1162/106365602320169811

    Stanley, K.O., Miikkulainen, R.: Evolving neural networks through augmenting topologies. Evolutionary Computation10(2), 99–127 (06 2002).https://doi.org/ 10.1162/106365602320169811,https://doi.org/10.1162/106365602320169811

    work page doi:10.1162/106365602320169811 2002

  50. [50]

    Oxford Univer- sity Press (2009)

    Sutherland, W.A.: Introduction to metric and topological spaces. Oxford Univer- sity Press (2009)

    2009

  51. [51]

    RSC Adv.2, 5337–5348 (2012).https://doi.org/10

    Tan, S.T., Chew, W.: Applications of the improved leader-follower cluster analysis (ilfca) algorithm on large array (la) and very large array (vla) hyperspectral mid- infrared imaging datasets. RSC Adv.2, 5337–5348 (2012).https://doi.org/10. 1039/C2RA20495A,http://dx.doi.org/10.1039/C2RA20495A

    work page doi:10.1039/c2ra20495a 2012

  52. [52]

    arXiv preprint arXiv:2506.07121 (2025)

    Wang, R.J., Xue, K., Qin, Z., Li, Z., Tang, S., Li, H.T., Liu, S., Qian, C.: Quality- diversity red-teaming: Automated generation of high-quality and diverse attackers for large language models. arXiv preprint arXiv:2506.07121 (2025)

    arXiv 2025

  53. [53]

    IEEE Computational Intelligence Magazine20(3), 41– 62 (2025).https://doi.org/10.1109/MCI.2025.3563425

    Wei, F.F., Chen, W.N., Zhao, T.F., Tan, K.C., Zhang, J.: A survey on distributed evolutionary computation. IEEE Computational Intelligence Magazine20(3), 41– 62 (2025).https://doi.org/10.1109/MCI.2025.3563425

    work page doi:10.1109/mci.2025.3563425 2025

  54. [54]

    In- formation Sciences700, 121842 (2025).https://doi.org/https://doi.org/10

    Zhou, X., Li, N., Fan, L., Li, H., Cheng, B., Wang, M.: Adaptive niching differen- tial evolution algorithm with landscape analysis for multimodal optimization. In- formation Sciences700, 121842 (2025).https://doi.org/https://doi.org/10. 1016/j.ins.2024.121842,https://www.sciencedirect.com/science/article/ pii/S0020025524017560 Distributed Quality-Diversi...

    arXiv 2025