Empathetic Cascading Networks: A Multi-Stage Prompting Technique for Reducing Social Biases in Large Language Models
Wangjiaxuan Xin
TL;DR
ECN tackles the challenge of producing empathetic and unbiased responses from large language models by introducing a four-stage prompting framework that mimics human empathy: Perspective Adoption, Emotional Resonance, Reflective Understanding, and Integrative Synthesis. It demonstrates that multi-stage prompts can outperform standard and basic empathy baselines in terms of Empathy Quotient ($EQ$) while maintaining competitive Regard and acceptable Perplexity across GPT-3.5-turbo and GPT-4 on a curated 150-item Personae dataset. The approach is deployment-friendly, requiring no retraining, and shows potential for more inclusive and trustworthy human-AI interactions in diverse populations. These findings highlight ECN as a scalable method to reduce social biases in conversational AI and to improve emotionally resonant user experiences.
Abstract
This report presents the Empathetic Cascading Networks (ECN) framework, a multi-stage prompting method designed to enhance the empathetic and inclusive capabilities of large language models. ECN employs four stages: Perspective Adoption, Emotional Resonance, Reflective Understanding, and Integrative Synthesis, to guide models toward generating emotionally resonant and contextually aware responses. Experimental results demonstrate that ECN achieves the highest Empathy Quotient (EQ) scores across GPT-3.5-turbo and GPT-4, while maintaining competitive Regard and Perplexity metrics. These findings emphasize ECN's potential for applications requiring empathy and inclusivity in conversational AI.