Rethinking Role-Playing Evaluation: Anonymous Benchmarking and a Systematic Study of Personality Effects

TL;DR

This paper introduces an anonymous benchmarking approach for role-playing agents, revealing the impact of character names on performance, and demonstrates that self-generated personalities can effectively enhance role fidelity.

Contribution

It proposes a fairer, anonymous evaluation method and systematically compares human-annotated versus self-generated personalities for improving RPAs.

Findings

Anonymization significantly reduces role-playing performance.

Self-generated personalities perform comparably to human-annotated ones.

Personality augmentation improves RPA robustness and fidelity.

Abstract

Large language models (LLMs) have demonstrated significant potential in developing Role-Playing Agents (RPAs). However, current research primarily evaluates RPAs using famous fictional characters, allowing models to rely on memory associated with character names. This dependency creates a bias that limits the generalization of RPAs to unseen personas. To address this issue, we propose an anonymous evaluation method. Experiments across multiple benchmarks reveal that anonymization significantly degrades role-playing performance, confirming that name exposure carries implicit information. Furthermore, we investigate personality augmentation to enhance role fidelity under anonymous setting. We systematically compare the efficacy of personality traits derived from human annotations versus those self-generated by the model. Our results demonstrate that incorporating personality information consistently improves RPA performance. Crucially, self-generated personalities achieve performance comparable to human-annotated ones. This work establishes a fairer evaluation protocol and validates a scalable, personality-enhanced framework for constructing robust RPAs.