Identify which Attributive Identity attributes drive RPA performance

Determine which specific fine-grained character attributes comprising the Attributive Identity layer—such as personality traits, moral values, interpersonal relationship styles, motivations, and abilities—most significantly influence the role-playing performance of large language model–based role-playing agents across evaluation settings.

Background

The paper introduces Character Identity as a two-layer framework for role-playing agents: Parametric Identity (character-specific knowledge encoded during pre-training) and Attributive Identity (fine-grained properties including personality traits, moral values, and interpersonal styles). While many benchmarks and studies have focused on coarse personality constructs (e.g., MBTI or Big Five), the authors emphasize that it remains unresolved which specific attributes within Attributive Identity most strongly impact role-playing fidelity.

This open question motivates the paper’s dataset and analyses that compare famous versus synthetic characters and quantify sensitivity to attribute valence. The authors frame it as a central gap in current RPA evaluation and design, where understanding attribute-level drivers would enable more precise characterization and improvement of RPA performance.