Robust Persona-Aware Toxicity Detection with Prompt Optimization and Learned Ensembling

TL;DR

Toxicity perception is inherently subjective, merit-based pluralistic evaluation across demographic personas. The authors systematically compare persona-conditioned prompting methods, introduce an automated prompt optimization via TextGrad, and propose a lightweight SVM meta-ensemble over four prompt outputs. Results show no single prompting method dominates across all model–persona pairs; however, the SVM ensemble consistently achieves the strongest performance and outperforms traditional majority voting. The work advances robust pluralistic evaluation in subjective NLP tasks and demonstrates the practical value of learned ensembling for toxicity detection.

Abstract

Toxicity detection is inherently subjective, shaped by the diverse perspectives and social priors of different demographic groups. While ``pluralistic'' modeling as used in economics and the social sciences aims to capture perspective differences across contexts, current Large Language Model (LLM) prompting techniques have different results across different personas and base models. In this work, we conduct a systematic evaluation of persona-aware toxicity detection, showing that no single prompting method, including our proposed automated prompt optimization strategy, uniformly dominates across all model-persona pairs. To exploit complementary errors, we explore ensembling four prompting variants and propose a lightweight meta-ensemble: an SVM over the 4-bit vector of prompt predictions. Our results demonstrate that the proposed SVM ensemble consistently outperforms individual prompting methods and traditional majority-voting techniques, achieving the strongest overall performance across diverse personas. This work provides one of the first systematic comparisons of persona-conditioned prompting for toxicity detection and offers a robust method for pluralistic evaluation in subjective NLP tasks.

Figures (8)

• Figure 1: Prompting comparisons with default prompting as the baseline.
• Figure 2: Prompting comparisons with persona prompt as the baseline.
• Figure 3: Prompting comparisons with persona optimized prompt as the baseline.
• Figure 4: Ensembling comparisons with Value Profile prompting as the baseline.
• Figure 5: Ensembling comparisons with Accuracy-based Weighted Majority as the baseline.