Study finds LLM personas show compressed, unrealistic personality growth after life events

A new arXiv paper from Ming Wang and colleagues investigates whether LLM agents with personality conditioning evolve credibly after simulated life events. The study tests PC-Agents across 11 major life events—from marriage to job loss—using the Big Five psychometric framework. Results show that while agents do exhibit measurable trait shifts, the magnitude is often below documented human effect sizes, and the dispersion across personas is compressed three-to-four-fold relative to human samples. Gender and cultural region prompts had little moderating effect, indicating that current models fail to capture demographic nuance in personality response.

To support ongoing research, the team introduces BFI-Adapt, a reusable benchmark that scores the directional fidelity of event-induced personality change. They validated the benchmark across 14 models, confirming that shifts exceed retest noise, persist across paraphrased prompts and unrelated dialogue, and remain stable. The findings have practical implications for emotional support bots, social simulations, and role-playing agents: they may appear psychologically consistent on average but lack the individual variability that makes human personality dynamics believable. The authors conclude that PC-Agents simulate the mean of human personality dynamics but not its shape, calling for new methods to model realistic, context-dependent evolution.