ZICQ
中 Log in / Sign up
Newsroom Research & Papers #Multi-Intelligence #Preferences Optimization #Sex Discrimination Testing #Mark differences #LLMs & Foundation Models

Multi-Agent Perspectivist Preference Optimization (MAP-PO): Preserving Annotator Disagreement for Robust Sexism Detectio

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 9 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:A new arXiv paper introduces the Multi-Agent Perspectivist Preference Optimization (MAP-PO) framework to address the issue of annotator disagreement being collapsed by majority voting in NLP. On the EXIST 2024 dataset, annotators are clustered by labeling behavior rather than demographics, and each cluster is assigned an LLM agent fine-tuned to reproduce its annotation behavior. Agents are coordinated via preference optimization combining individual and team-level rewards. Experiments across two


Background

In NLP, annotator disagreement is often treated as noise, and majority voting is the standard approach. However, for subjective tasks like sexism detection, disagreement often reflects genuine differences in perspective. Discarding this disagreement loses information and can introduce bias.

Core Method: MAP-PO Framework

The proposed MAP-PO framework consists of three key steps:

  1. Behavioral Clustering: Cluster annotators based on their labeling behavior (not demographics) to ensure similar annotation patterns within each cluster.
  2. Agent Fine-tuning: Fine-tune an LLM agent for each cluster to reproduce its annotation behavior.
  3. Preference Optimization: Coordinate agents via preference optimization that combines individual and team-level rewards, balancing individual characteristics and collective approximation of the majority label.

Experimental Design and Findings

Experiments were conducted on the EXIST 2024 dataset (English and Spanish tweets) with two backbone models (e.g., BERT and RoBERTa). Key findings:

  • Untuned agents converge: All agents behave nearly identically without fine-tuning, highlighting the necessity of cluster-specific training.
  • Individual training overfits: Training on individual cluster labels alone pushes agents beyond their intended clusters, overfitting to training data.
  • Team signal maintains calibration: Adding a shared team-level training signal consistently keeps each agent calibrated to its cluster while improving overall performance.

Technical Highlights and Innovations

  • Perspective Preservation: Unlike traditional methods, MAP-PO explicitly models and preserves different annotator perspectives, enhancing model expressiveness for subjective tasks.
  • Behavioral Clustering: Clustering based on behavior rather than demographics avoids introducing unnecessary bias and aligns with real annotation scenarios.
  • Multi-Agent Coordination: Preference optimization balances individual and team objectives, offering new insights for multi-agent system design.

Industry Impact and Developer Recommendations

This framework can be applied to any task with annotator disagreement, such as hate speech detection and sentiment analysis. Developers can adopt the idea of preserving disagreement during data annotation and training multi-agent systems to improve model robustness. Future work may explore more sophisticated clustering methods and team reward designs.


Source: arXiv:2608.04056


Source: ArXiv NLP/LLM (cs.CL) (2026-08-06)

— END —

Tags: #Multi-Intelligence #Preferences Optimization #Sex Discrimination Testing #Mark differences #LLMs & Foundation Models

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…