Accepted at Customizable NLP @ EMNLP 2024; the public preprint is available on arXiv.
Rather than identifying a person's traits, this paper asks the reverse question: can an LLM consistently simulate a specified personality when generating responses? The framework uses the Big Five model.
The study constructs a dataset of texts with predefined Big Five characteristics and introduces an analytical framework for evaluating how well models maintain a target personality profile.