Enhancing the Role-playing Capabilities of Reasoning-optimized Large Language Models

Currently, advancements in reasoning techniques are constantly pushing the performance boundaries of LLMs. This confluence motivates a pivotal research question: can advanced reasoning capabilities enhance the role-playing performance of LLMs? To systematically investigate this question, this paper conducts a large-scale empirical study across 6 mainstream benchmarks and 26 representative models, evaluating three distinct role-playing paradigms based on zero-shot, Chain-of-Thought, and reasoning model, respectively. Our core findings reveal two counter-intuitive conclusions. First, introducing reasoning capabilities, whether explicit or internalized, often impairs role-playing performance and disrupts its scaling laws. Secondly, current models exhibit significant shortcomings in advanced role-playing tasks, and their performance in Chinese is generally superior to English. Based on these findings, this paper proposes two promising future research directions, "Role-aware Chain-of-Thought" and "Reinforcement Learning for Role-Playing", and further refines the latter into a "Fine-grained Character Reward-driven Reinforcement Learning" framework. By validating these directions through experiments, this work provides new perspectives and empirical evidences to enhance the adaptability, consistency, and effectiveness of next-generation role-playing agents.