CharacterGPT: A Persona Reconstruction Framework for Role-Playing Agents

TL;DR

CharacterGPT presents a framework to reconstruct and maintain consistent character personas for role-playing agents by continually updating a structured persona document through Chapter-wise Trait Extraction (CPT). It defines eight traits, classifies them into internal and external categories, and uses a two-stage initialization+training process to reflect narrative progression. Evaluations with Big Five personality assessments and human story-generation tasks show that CPT-enhanced personas improve coherence, controllability, and role-specific knowledge compared with unstructured document inputs. The approach enables interaction with characters at specific narrative moments and reduces information loss and computational overhead associated with standard retrieval. The work demonstrates potential for more reliable NPCs in games and enhanced RP experiences.

Abstract

The recent introduction of the Assistants API highlights its potential for large language models (LLMs) in role-playing agents (RPA). However, maintaining consistent character personas remains a significant challenge due to variability in information extraction, which frequently omits critical elements such as backstory or interpersonal relationships. To address this limitation, we introduce CharacterGPT, a framework designed to dynamically reconstruct character personas through Character Persona Training (CPT). This approach incrementally updates personas by extracting traits from chapter-wise novel summaries, reflecting the progression of the narrative. Our framework is evaluated through Big Five personality evaluations and creative tasks, in which characters generate original narratives, demonstrating the efficacy of CharacterGPT in preserving persona consistency. The code and results are available at https://github.com/Jeiyoon/charactergpt

Figures (11)

• Figure 1: Comparison of response accuracy between persona-based GPT-4 assistants utilizing unstructured versus structured character traits as input. When provided with unstructured traits, the assistant demonstrates limited success in generating accurate responses. In contrast, the use of structured traits significantly improves the correctness of the assistant’s responses.
• Figure 2: An example of CharacterGPT (Anya Forger). (Top) Character Persona Training process. (Bottom) CharacterGPT generating responses that align with the character’s persona.
• Figure 3: Total sum of # Wins for each character in ChatGPT and GPT-4 settings ($\Sigma$ # Wins). The larger value, the better.
• Figure 4: Total sum of $|d|$ for each character in ChatGPT and GPT-4 settings ($\Sigma \Sigma |d|$). The smaller value, the better.
• Figure 5: Case study of the character (Megumin) implemented at three different points in time.