Rethinking Role-Playing Evaluation: Anonymous Benchmarking and a Systematic Study of Personality Effects
This addresses evaluation fairness and scalability for LLM-based role-playing agents, though it's incremental on existing methods.
The paper tackles bias in evaluating Role-Playing Agents (RPAs) by proposing anonymous benchmarking, which significantly degrades performance, confirming models rely on character name memory. It also shows personality augmentation improves performance, with self-generated traits matching human-annotated ones.
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.