Thought-Driven Role-Playing for Consistent Character Simulation

Role-Playing involves making Large Language Models (LLMs) generate responses in the manner of a specific character. Existing works primarily focus on maintaining the consistency of a character’s personality, information, and tone through fine-tuning or specialized prompts. However, these models often struggle to fully embody the character’s thought, making it difficult for them to generate responses aligned with the character’s way of thinking. This limitation negatively impacts the user experience. To address this issue, we propose the Thought-Process-Driven Role-Playing model (TRP), which mimics the character’s logical reasoning process and generates thoughts before responding. We enhance the training data for each set of dialogues by incorporating logical reasoning based on the character’s profile and the contextual content of the conversations. This approach enables the model to learn and replicate the character’s thought. Additionally, we incorporate a small set of questions beyond the character’s knowledge scope to train the model to appropriately decline answering. To verify the effectiveness of our model, we prepare new evaluation datasets and metrics. Experimental results demonstrate that the TRP model achieves superior role-playing performance in tone, information, and thoughts.