HistActor: Summon Your Designated Historical Persona

ACL ARR 2025 February Submission6406 Authors

Abstract: History encompasses the profound heritage of human civilization, wherein historical figures, as architects of cultural legacies, possess event-specific knowledge and ideological frameworks that facilitate historical reconstruction. Recent years have witnessed remarkable advancements in large language models (LLMs), which have demonstrated exceptional learning capabilities across diverse domains. However, current LLMs still exhibit limitations in historical figure role-play applications, the model with the substantial parameter count presents deployment challenges in local environments, while its parameter-efficient counterpart typically underperforms in role-specific factual knowledge representation. To address these limitations, we propose HistActor, a role-play framework incorporating data generation, model training, and performance optimization mechanisms. Furthermore, we introduce RoleFactPsyBench, a multidimensional evaluation benchmark that simultaneously assesses factual accuracy and psychological verisimilitude in role-play scenarios, with replicability across diverse historical figures. Taking the Su Shi persona simulation model as a case study, Our HistActor framework achieves performance comparable to large-scale models while maintaining parameter efficiency in small-scale architectures, thereby providing an effective solution for historical character simulation tasks.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: evaluation and metrics,factuality,knowledge augmented,interactive storytelling,applications

Contribution Types: NLP engineering experiment

Languages Studied: English, Chinese

Submission Number: 6406