Large language models (LLMs) have demonstrated strong potential for impersonating real individuals, yet achieving faithful role‑playing remains difficult. Existing in‑context learning (ICL) approaches fail to capture how a person reacts across different situations, and LLM‑only evaluation struggles with obscure targets.
To tackle these issues, we introduce the Situation‑Internal state‑Behavior Persona (SIB) method. SIB integrates three components—situational context, internal state, and behavioral strategy—into a persona that adapts its response generation according to the current situation.
We also devise an evaluation protocol that supplies LLM evaluators with reference materials about the impersonated individual (e.g., background, past statements), reducing subjective bias.
Experiments on a newly built dataset of social‑media reply generation show that SIB outperforms state‑of‑the‑art ICL baselines, and the protocol achieves a moderate correlation with human judgments. Additional tests on fictional‑character benchmarks confirm that the approach generalizes beyond real‑person scenarios.
These findings suggest that incorporating behavioral information into personas broadly improves the fidelity of role‑playing for both real individuals and fictional characters on social media.
Review