New AI Framework Multi-Persona Thinking Reduces Social Bias in Language Models

A research paper introduces Multi-Persona Thinking (MPT), an inference-time framework designed to mitigate social biases in Large Language Models (LLMs). The approach guides models to reason from multiple contrasting social identities, such as male and female perspectives, alongside neutral viewpoints. These different perspectives interact through iterative reasoning processes to identify and correct biased judgments. The framework transforms persona assignment from a potential weakness into a bias mitigation mechanism. Researchers evaluated MPT on two widely used bias benchmarks using both open-source and closed-source models. Results demonstrate that MPT achieves lower bias levels than existing prompting-based methods while maintaining core reasoning capabilities. The paper addresses concerns about harmful stereotypes and unfair outcomes that can emerge from biased LLM outputs. The research was published on arXiv, a repository for scientific papers in fields including computer science and computational linguistics. The framework represents a technical advancement in addressing ethical challenges in artificial intelligence development.