Deliberative-Role-Playing: A Reasoning-Enhanced Role-Playing Method for Large Language Model

With the advancements in generative capabilities and instruction-following abilities of Large Language Models (LLMs), role-playing tasks using LLMs have garnered significant attention. However, current research on Role-Playing Language Models (RPLMs) primarily focuses on imitating role-specific speaking styles and replicating knowledge, lacking efforts to enable LLMs to deeply analyze the characters they portray. This paper introduces Deliberative-RP (Deliberative-Role-Playing), an reasoning-enhanced role-playing method for large language model. Deliberative-RP enabling LLMs to first understand their target roles before engaging in role-playing, which realizes rationalized role-playing behaviors for LLMs. The evaluations were conducted on WikiRoleEval benchmark and used LLM-as-a-Judge. The LLMs demonstrated better across evaluation metrics by using Deliberative-RP.