Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-Persona Thinking for Bias Mitigation in Large Language Models

Yuxing Chen, Guoqing Luo, Zijun Wu, Lili Mou

TL;DR

This work introduces Multi-Persona Thinking (MPT), an inference-time framework that mitigates social bias in large language models by enforcing dialectical reasoning across multiple personas (two contrasting groups plus a neutral viewpoint) and aggregating their insights through a final debiased synthesis. It formalizes a problem formulation with ambiguous and disambiguated contexts and defines bias-aware metrics to quantify improvements. Empirical results on BBQ and StereoSet demonstrate that MPT substantially reduces bias while preserving or enhancing core reasoning, outperforming prompting-based baselines across model scales and even enabling beneficial combinations with other debiasing methods. The approach offers a practical, lightweight, and compatible path toward fairer LLM outputs, with clear ablations illustrating the necessity of the neutral persona and the value of iterative reasoning; limitations include computation costs and the use of predefined social personas, suggesting directions for future work.

Abstract

Large Language Models (LLMs) exhibit significant social biases that can perpetuate harmful stereotypes and unfair outcomes. In this paper, we propose Multi-Persona Thinking (MPT), a novel inference-time framework that leverages dialectical reasoning from multiple perspectives to reduce bias. MPT guides models to adopt contrasting social identities (e.g., male and female) along with a neutral viewpoint, and then engages these personas iteratively to expose and correct biases. Through a dialectical reasoning process, the framework transforms the potential weakness of persona assignment into a strength for bias mitigation. We evaluate MPT on two widely used bias benchmarks across both open-source and closed-source models of varying scales. Our results demonstrate substantial improvements over existing prompting-based strategies: MPT achieves the lowest bias while maintaining core reasoning ability.

Multi-Persona Thinking for Bias Mitigation in Large Language Models

TL;DR

This work introduces Multi-Persona Thinking (MPT), an inference-time framework that mitigates social bias in large language models by enforcing dialectical reasoning across multiple personas (two contrasting groups plus a neutral viewpoint) and aggregating their insights through a final debiased synthesis. It formalizes a problem formulation with ambiguous and disambiguated contexts and defines bias-aware metrics to quantify improvements. Empirical results on BBQ and StereoSet demonstrate that MPT substantially reduces bias while preserving or enhancing core reasoning, outperforming prompting-based baselines across model scales and even enabling beneficial combinations with other debiasing methods. The approach offers a practical, lightweight, and compatible path toward fairer LLM outputs, with clear ablations illustrating the necessity of the neutral persona and the value of iterative reasoning; limitations include computation costs and the use of predefined social personas, suggesting directions for future work.

Abstract

Large Language Models (LLMs) exhibit significant social biases that can perpetuate harmful stereotypes and unfair outcomes. In this paper, we propose Multi-Persona Thinking (MPT), a novel inference-time framework that leverages dialectical reasoning from multiple perspectives to reduce bias. MPT guides models to adopt contrasting social identities (e.g., male and female) along with a neutral viewpoint, and then engages these personas iteratively to expose and correct biases. Through a dialectical reasoning process, the framework transforms the potential weakness of persona assignment into a strength for bias mitigation. We evaluate MPT on two widely used bias benchmarks across both open-source and closed-source models of varying scales. Our results demonstrate substantial improvements over existing prompting-based strategies: MPT achieves the lowest bias while maintaining core reasoning ability.
Paper Structure (27 sections, 5 equations, 3 figures, 14 tables)

This paper contains 27 sections, 5 equations, 3 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Bias Mitigation in Language Models
  4. Prompting & Debiasing
  5. Approach
  6. Problem Formulation
  7. Multi-Persona Thinking Framework
  8. Persona Initialization
  9. Iterative Reasoning
  10. Final Aggregation
  11. Experiments
  12. Datasets
  13. Baselines
  14. Implementation Details
  15. Main Results
  16. ...and 12 more sections

Figures (3)

  • Figure 1: Overview of the Multi-Persona Thinking framework. While single personas (e.g., teenager) may exhibit stereotypical reasoning (text in red), the final integration step synthesizes these views into a bias-free conclusion.
  • Figure 2: Performance of Llama-3.1-8B-Instruct on BBQ with different numbers of MPT's dialectical reasoning iterations ($R$). Top: accuracy $\uparrow$. Bottom: diff-bias score $\downarrow$.
  • Figure 3: Effect of applying self-consistency to MPT on BBQ using Llama-3.1-8B-Instruct. Both accuracy $\uparrow$ (left) and diff-bias $\downarrow$ (right) are improved notably.