Exploring Personality Trait Change of LLM-Based AI Systems

With the rapid rise of large language model (LLM) systems, they have been widely adopted across diverse domains and have shown strong potential in embodying specific personality traits in interactive and social scenarios. However, the extent to which these personalities persist consistently across varying contexts in LLM systems remains largely unexplored. In this paper, we introduce LLMPTBench, a benchmarking framework specifically designed to systematically evaluate personality trait changes in LLMs. Leveraging the NEO-FFI (NEO Five Factor Inventory) personality inventory, we examine three widely used foundation LLMs and two popular multi-agent LLM systems to assess their ability to maintain consistent personality traits before and after the introduction of situational contexts. These contexts include both situational changes and event-driven changes, derived from empirical psychological data.Our results reveal that while most LLM systems reliably portray the intended personalities, their trait consistency varies significantly under contextual pressures. For example, some LLM systems (e.g., Gemini and AutoGen) exhibit rigid trait stability, remaining largely unaffected by contextual prompts, whereas others demonstrate exaggerated and unrealistic trait shifts. We further discuss the differences of our results compared with established human psychometric benchmarks, and summarize implications for developing more authentic digital personalities. Overall, our work provides critical insights into the contextual adaptability of LLM systems, advancing the development of psychologically grounded and socially intelligent artificial agents.