Diversifying Toxicity Search in Large Language Models Through Speciation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Shelar, Onkar, Desell, Travis
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908981630337024
author Shelar, Onkar
Desell, Travis
author_facet Shelar, Onkar
Desell, Travis
contents Evolutionary prompt search is a practical black-box approach for red teaming large language models, however existing methods often collapse onto a small family of high-performing prompts, limiting coverage of distinct failure modes. We present a speciated quality-diversity extension of \textit{ToxSearch} that maintains multiple high-toxicity prompt niches in parallel rather than optimizing a single best prompt. \textit{ToxSearch-S} introduces unsupervised prompt speciation via a search methodology that maintains capacity-limited species with exemplar leaders, a reserve pool for emerging niches, and species-aware parent selection that trades off within-niche exploitation and cross-niche exploration. Preliminary results show \textit{ToxSearch-S} reaching higher peak toxicity ($\approx 0.73$ vs.\ $\approx 0.47$) with a heavier tail (top-10 median $0.66$ vs.\ $0.45$) than the baseline. Speciation also yields broader semantic coverage under a topics-as-species analysis (higher effective topic diversity and larger unique topic coverage). Finally, species formed are well-separated in embedding space (mean separation ratio $\approx 1.93$) and exhibit distinct toxicity distributions, indicating that speciation partitions the adversarial space into behaviorally differentiated niches rather than superficial lexical variants.
format Preprint
id arxiv_https___arxiv_org_abs_2601_20981
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Diversifying Toxicity Search in Large Language Models Through Speciation
Shelar, Onkar
Desell, Travis
Neural and Evolutionary Computing
Populations and Evolution
Evolutionary prompt search is a practical black-box approach for red teaming large language models, however existing methods often collapse onto a small family of high-performing prompts, limiting coverage of distinct failure modes. We present a speciated quality-diversity extension of \textit{ToxSearch} that maintains multiple high-toxicity prompt niches in parallel rather than optimizing a single best prompt. \textit{ToxSearch-S} introduces unsupervised prompt speciation via a search methodology that maintains capacity-limited species with exemplar leaders, a reserve pool for emerging niches, and species-aware parent selection that trades off within-niche exploitation and cross-niche exploration. Preliminary results show \textit{ToxSearch-S} reaching higher peak toxicity ($\approx 0.73$ vs.\ $\approx 0.47$) with a heavier tail (top-10 median $0.66$ vs.\ $0.45$) than the baseline. Speciation also yields broader semantic coverage under a topics-as-species analysis (higher effective topic diversity and larger unique topic coverage). Finally, species formed are well-separated in embedding space (mean separation ratio $\approx 1.93$) and exhibit distinct toxicity distributions, indicating that speciation partitions the adversarial space into behaviorally differentiated niches rather than superficial lexical variants.
title Diversifying Toxicity Search in Large Language Models Through Speciation
topic Neural and Evolutionary Computing
Populations and Evolution
url https://arxiv.org/abs/2601.20981