Learning Sexism Detection Using Multi-Agent Perspectivist Preference Optimization

2026-08-04

AI Summary

Existing approaches to gender bias detection often rely on majority voting for unified annotation, thereby overlooking legitimate disagreements among different demographic groups regarding bias perception and consequently erasing valuable perspective diversity. This work proposes a multi-agent framework that identifies distinct annotator groups through clustering of labeling behaviors, fine-tunes dedicated large language models for each group, and employs a collaborative training strategy integrating both individual- and team-level preference optimization. For the first time, this approach combines behavior-based clustering for personalized modeling with a team-level reward mechanism. Evaluated across four language–model configurations, the method demonstrates the necessity of cluster-specific training and shows that team-level signals effectively calibrate agent behavior, accurately reproducing both group-specific and aggregate label distributions.

Abstract

When people label text for sexism, they often disagree, and not because some of them are wrong: they genuinely perceive sexism differently. Most NLP systems discard this disagreement by collapsing it into a majority vote. We propose the Multi-Agent Perspectivist Preference Optimization (MAP-PO) framework to keep these different perspectives. On the EXIST 2024 dataset of labeled English and Spanish tweets, we first cluster annotators by their labeling behavior rather than their demographic attributes. We then fine-tune one Large Language Model agent per cluster to reproduce that cluster's annotation behavior, and coordinate the agents with preference optimization that combines individual and team-level rewards. We evaluate MAP-PO in four settings defined by two languages and two backbone language models, asking whether each agent reproduces the annotations of its own cluster and whether the agents together reproduce the majority label. Two findings hold in all four settings. First, without fine-tuning the agents behave almost identically, so cluster-specific training is necessary. Second, we show that training each agent only on the labels of its own cluster pushes the agents far beyond the clusters they should represent, while adding a shared team-level training signal consistently keeps each agent calibrated to its cluster.

Problem

• sexism detection
• annotation disagreement
• perspectivist modeling
• multi-agent learning
• labeling behavior

Innovation

• Multi-Agent Learning
• Perspectivist Annotation
• Preference Optimization
• Label Disagreement Modeling
• Cluster-Based Fine-Tuning