cs.NENov 18, 2026

Incremental Speciation in Evolutionary Red Teaming of Large Language Models

Authors: Onkar Shelar, Travis Desell

Organizations: Rochester Institute of Technology, Rochester NY 14623, USA

Abstract

Speciation maintains distinct subpopulations in evolutionary search through similarity-based grouping and species-aware parent selection. For text-based individuals, niches depend on both the compatibility representation and the assignment operator that turns pairwise relations into species. We examine this design space in black-box evolutionary red teaming of large language models, comparing embedding and natural language inference (NLI) dissimilarities under representative-based and density-based speciation. Offline calibration on a 3,139-prompt archive shows that MiniLM embedding dissimilarity (dEd_E) and DeBERTa-v3-small NLI non-entailment (d¬Ed_{\neg E}) produce different species structures at the same nominal threshold. Even with pairwise edge density matched to MiniLM, NLI density connectivity forms a few concentrated, mixed-topic species, with substantially more off-topic membership. These structural differences persist online under a matched evaluation budget. The endpoint top-50 median fitness favors incremental DBSCAN over leader-follower, but bootstrap intervals from 10 runs include zero, providing no reliable evidence of an advantage. Semantic diversity among top prompts declines in every configuration, with no consistent method advantage. The clearest difference concerns high-fitness population occupancy: MiniLM incremental DBSCAN maintains mature niches across seed topics, whereas NLI incremental DBSCAN attains the same highest observed median fitness while concentrating high-fitness prompts in variants of a few seed topics. Thus, representation-dependent connectivity predicts population structure more consistently than endpoint fitness. Species count alone does not establish quality-diversity, and more species do not necessarily preserve semantically diverse high-fitness prompts.

Figures & tables

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Jun 23, 2026

Distributed Quality-Diversity Search for Toxicity in Large Language Models

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 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 nwn_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 1.8×1.8\times with two workers and 3.2×3.2\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.
May 26, 2026

DEI: Diversity in Evolutionary Inference for Quality-Diversity Search

We present DEI: Diversity in Evolutionary Inference, a distributed Quality-Diversity (QD) search framework that assigns heterogeneous large language models (LLMs) as mutation operators across peer nodes communicating with non-blocking collective operations. Unlike homogeneous parallel search, which replicates a single model's inductive biases across all workers, DEI treats each LLM's distinct creative prior as a complementary source of behavioral novelty. Extending the Digital Red Queen framework with DEI, nodes share local optimal solutions at the end of each round to seed the next round's population. This creates cross-model adversarial pressure that drives robustness beyond intra-model self-play. Evaluated on the Core War domain, a competitive programming benchmark in which Redcode warrior programs battle inside a simulated machine, a four-node heterogeneous ensemble (GPT-5.4-mini, Claude Sonnet 4.6, GPT-5.2, and Claude Haiku 4.5) achieves 124 percent higher merged-archive QD-Score (45.90 vs. 20.46) and 28 percent higher coverage (80.6 percent vs. 63.0 percent of cells) than a single-node baseline at equal total LLM-call budget. The heterogeneous ensemble also outperforms an equally-budgeted homogeneous ensemble on QD-Score, coverage, and held-out solution generality across all four model families. These results provide the first empirical evidence that model diversity, not merely parallelism, is the key driver of gain in distributed LLM-based QD search.
May 31, 2026

Consistent and Distinctive: LLM Benchmark Efficiency via Maximum Independent Set Prompt Selection on Similarity Graphs

Evaluating large language models (LLMs) across comprehensive benchmarks is expensive and time-consuming. We propose a graph-based prompt selection framework that models each benchmark as a similarity graph -- nodes are prompts connected if their embedding-space distance falls above a configurable threshold -- and applies Maximum Independent Set (MIS) algorithms to select a maximally diverse, non-redundant subset. We evaluate four MIS solvers (CPLEX, GREEDY, Online-MIS, ReduMIS) across six embedding models, three distance measures, six percentile thresholds, and four benchmarks (GPQA, IFEval, MMLU-Pro, Omni-MATH) covering 66 LLMs. Our central hypothesis -- that repeated selection under different random seeds yields consistent LLM rankings that may also differ from the full-benchmark baseline -- is strongly confirmed: Kendall's W≥0.90W \geq 0.90 in 99.2% of stochastic configurations (mean W=0.997±0.008W = 0.997 \pm 0.008), while at higher percentile thresholds selected subsets achieve 25--48% prompt reduction on average. Ranking divergence from the full benchmark (ρ<0.95ρ< 0.95) occurs in only 15.95% of configurations, concentrated at low thresholds (p10p_{10}--p20p_{20}) and benchmarks (GPQA, IFEval), identifying overly dense graphs as the primary failure mode.