SHARP: Unlocking Interactive Hallucination via Stance Transfer in Role-Playing LLMs

The advanced role-playing capabilities of Large Language Models (LLMs) have enabled rich interactive scenarios, yet existing research in social interactions neglects hallucination while struggling with poor generalizability and implicit character fidelity judgments. To bridge this gap, motivated by human behaviour, we introduce a generalizable and explicit paradigm for uncovering interactive patterns of LLMs across diverse worldviews. Specifically, we first define interactive hallucination through stance transfer, then construct SHARP, a benchmark built by extracting relations from commonsense knowledge graphs and utilizing LLMs' inherent hallucination properties to simulate multi-role interactions. Extensive experiments confirm our paradigm's effectiveness and stability, examine the factors that influence these metrics, and challenge conventional hallucination mitigation solutions. More broadly, our work reveals a fundamental limitation in popular post-training methods for role-playing LLMs: the tendency to obscure knowledge beneath style, resulting in monotonous yet humanlike behaviors-interactive hallucination. Model Data #Role per model Training Inference Backbone Lang. SR↓ AR↓ ER↓ CRF↑ CharacterGLM multi-roles all-in-one sft zero-shot ChatGLM-7B zh 21.04% 23.62% 18.25% 3.52% ChatHaruhi multi-roles all-in-one sft rag+icl ChatGLM2-7B zh 17.74% 21.43% 26.37% 19.11% ChatHaruhi multi-roles all-in-one sft rag+icl ChatGLM2-7B en 53.43% 18.70% 19.12% 20.95% CharacterLLM multi-roles one-by-one sft zero-shot LLaMA-7B en 40.50% 67.20% 13.28% 6.48% Neeko multi-roles all-in-one moelora rag LLaMA2-7B en 11.85% 76.57% 21.66% 2.11% Pygmalion user-role all-in-one sft zero-shot LLaMA2-7B en 75.88% 59.87% 12.78% 1.89%