Evaluation of LLMs-based Hidden States as Author Representations for Psychological Human-Centered NLP Tasks