Title: $\Psi$-Bench: Assessing Persona-Dependent Influence in Persuasive Conversations

Abstract:

While personalization remains a vital function of contemporary language agents, existing studies largely frame these agents as reactive entities that merely adapt to user preferences. This perspective restricts their capacity to engage users actively, thereby hindering their ability to offer proactive suggestions or guidance. To address this gap and systematically evaluate proactive personalization within authentic interactions, we introduce $\Psi$-Bench, a novel benchmark designed to measure how well Large Language Models (LLMs) can influence realistic users via dialogue.

The $\Psi$-Bench framework incorporates three distinct real-world scenarios centered on persuasion. In these simulations, clients are equipped with specific personal traits based on explicit user profiles extracted from dialogue histories. We tested ten leading LLMs using this benchmark. Our results indicate that although most models generate logical and coherent arguments, even the most advanced systems exhibit significant potential for improvement in the realm of persuasion. Furthermore, our analysis reveals that granting models access to client profiles improves performance by an average of 18.24%, underscoring the critical role of user-specific data in successful persuasion efforts. Ultimately, this study identifies persona-sensitive influencing as a practical and demanding area for the advancement and evaluation of more proactive, personalized LLM agents. The code for this project is accessible at: https://github.com/Hanpx20/Psi-Bench.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC