Introducing PMIYC: A Framework for Evaluating Persuasion Effectiveness and Susceptibility
Date:
In this presentation we introduce PMIYC, an automated framework for evaluating persuasion effectiveness and susceptibility in large language models through multi-agent interactions. We discuss how Persuader agents engage in multi-turn conversations with Persuadee agents, allowing us to measure LLMs’ persuasive effectiveness and their susceptibility to persuasion. We present our findings on various models, showing significant differences in persuasive capabilities and resistance to misinformation. This work contributes to understanding the dynamics of persuasion in AI systems and aims to enhance the safety and ethical alignment of language models.
Watch the presentation here.