Reproduce “Towards Strategic Persuasion with Language Models” on OpenResearch

About this paper

Zirui Cheng, Jiaxuan You

Researchers from the University of Illinois Urbana-Champaign established a theory-driven framework using Bayesian persuasion to evaluate and enhance the strategic persuasive capabilities of large language models (LLMs). Their work demonstrated that frontier LLMs exhibit sophisticated strategic information disclosure, and reinforcement learning effectively improved persuasion gains in smaller models, generalizing across different LLM receivers.

Key insights

• Frontier LLMs exhibit sophisticated strategic behaviors in information disclosure, adapting their signaling strategies over time in dynamic settings, consistent with theoretical predictions from dynamic Bayesian persuasion.
• Reinforcement learning can significantly enhance the strategic persuasive capabilities of even smaller LLMs, and these learned improvements generalize effectively to different, unseen LLM receiver architectures.
• Persuasion is most effective when the receiver's prior beliefs are at intermediate levels, aligning with Bayesian persuasion theory, as these are the contexts where information has the greatest potential to shift decisions.

Summary by alphaXiv.