Eliciting Personality Traits in Large Language Models

This paper examines the personality traits that large language models display, particularly in contexts like recruitment where they are increasingly being used. Unlike past work that directly gave models personality tests, the authors designed prompts inspired by interview questions and trait-activation cues to see if models express personality traits through their outputs. They tested a wide range of models, from smaller ones like Falcon and Bloom to larger ones like GPT and LLaMA-2, including both base and fine-tuned versions. Their findings show that most LLMs generally exhibit high openness and low extraversion. Smaller models tend to behave more similarly in their traits, while larger models demonstrate a broader range, with higher agreeableness, emotional stability, and conscientiousness. Fine-tuned models also show small but noticeable shifts in personality depending on the dataset. Overall, the study highlights how model size and fine-tuning shape personality expression and calls attention to the ethical implications of using LLMs in decision-making contexts like hiring.