Harnessing In-Context Learning for Personality Elicitation in Large Language Models

With the rapid advancements in deep learning and natural language processing, large language models (LLMs) have demonstrated remarkable performance across a variety of tasks. While previous studies have explored the existence of personality traits within LLMs, the potential to systematically control and induce specific personalities remains under-researched. This paper introduces a novel approach using In-Context Learning (ICL) to induce personality traits in LLMs, specifically grounded in the Big Five personality theory. We present a curated example set designed to evoke each of the five key personality dimensions: openness, conscientiousness, extraversion, agreeableness, and neuroticism. Using the Machine Personality Inventory (MPI), we quantitatively measure LLMs' personality traits and analyze the effects of example complexity, relevance, and diversity on the stability and effectiveness of personality induction. Experimental results indicate that LLMs can exhibit personality traits similar to humans and that the ICL method significantly enhances the prominence of these traits. Additionally, LLMs exhibit strong generalization capabilities in scenario-based tests, suggesting their potential for personalized AI applications. This study advances the understanding of personality control in LLMs and highlights the implications for human-AI interaction.