Not Just Imitation: Enhancing Role-Playing Through Retrieve Fine-Tuned Collaborative Large Language Model

In this paper, we focus on enhancing the credibility and personalization of large language models in role-playing scenarios, aiming to facilitate natural and believable humancomputer interaction. Current dialogue systems often fall short in accurately portraying character styles, leading to issues of role illusion. To address this, we introduce an innovative method, the Retrieve Fine-Tuned Collaborative Large Language Model (RFTC-LLM), which employs a two-stage strategy combining fine-tuning with retrieval-augmented generation (RAG). By employing instruction-based fine-tuning, the model is directed to learn the character’s expression style and preferences, and the use of a role-specific knowledge graph aids in constraining responses to ensure relevance and avoid misleading answers. This approach effectively resolves the confusion surrounding role-specific knowledge, thereby boosting the credibility and authenticity of the dialogue system. Experimental results demonstrate the superiority of our method across multiple role-playing dialogue datasets, particularly in significantly reducing instances of role illusion.