A Model of Multi-turn Human Persuadability Using Probabilistic Belief Tracing
Researchers have developed a new framework called PERSUASIONTRACE to study how large language models (LLMs) influence human beliefs over multiple conversational turns. This framework includes a platform for conducting multi-turn persuasion studies and a protocol for evaluating the fidelity of simulated persuaders to real human belief dynamics. The study found that human participants grouped into two distinct clusters regarding belief updates and showed susceptibility to rhetorical strategies, while LLMs demonstrated persuasiveness across various topics and modalities. Notably, a new Bayesian-network simulated target within the framework achieved human-like belief dynamics, outperforming standard LLM simulators. AI
IMPACT Provides a new methodology for evaluating the persuasive capabilities of LLMs and understanding their impact on human cognition.