Do personality-conditioned agents change like humans do?

The paper asks whether personality-conditioned LLM agents (PC-Agents) "grow" the way people do when major life events happen. It measures Big Five profiles before and after 11 life events and reads the trajectories against longitudinal human personality psychology. The agents do move: trait shifts are measurable. But the abstract reports that they occur "at similar rates for event–trait pairs with and without documented human change directions," so the shifts do not track where human research says change should occur. Where the direction does match, magnitudes "usually fall below human effect-size ranges." Gender and cultural-region prompts "show little moderating effect," and persona-level dispersion is "compressed three- to four-fold relative to human samples."

The conclusion states the pattern compactly. Current models can move, but their movement is "weakly event-specific, poorly calibrated in magnitude, and compressed across demographic and individual variation." The agents therefore approximate "a generic pattern of change" more readily than the event-specific and person-specific structure of human development. The paper backs the claim with robustness checks: event-conditioned trajectories exceed retest noise, keep their event–trait structure under independent paraphrases, show model-dependent convergence with scenario-based decisions, and remain detectable after unrelated dialogue. So the shifts are real and stable enough to measure, and the weakness lies in their content rather than in noise. The paper builds a benchmark, BFI-Adapt, from these trajectories.

This extends Does conditioning LLMs on personal profiles improve prediction? along a different axis. That note shows that conditioning on who someone is does not sharpen predictions of what they will do. This paper suggests a temporal counterpart: persona conditioning does not make an agent's change over time person-specific either, since dispersion across personas is compressed. It also qualifies Why do static persona descriptions produce repetitive dialogue?. Dynamic modeling is needed, but movement alone does not supply psychological plausibility, because agents can shift without shifting in the right places. Systems like the one in Can personas evolve in real time to match what users actually want? let a persona evolve, and this paper offers a yardstick for what evolving should mean, though the excerpt does not evaluate such architectures. One untested link is Why do open language models converge on one personality type?. A shared model-level default could plausibly produce a generic pattern of change, but the excerpt does not say why the change is generic.

The excerpt is silent on the number and identity of models, the personas and prompts used, how events were presented to the agents, and any effect sizes beyond the directional statement. It gives no cause for the compression or the weak event specificity, and it does not test whether training, memory, or longer interaction would fix them. Its consequence is at the level the evidence supports: for applications that lean on lifelong personas, such as emotional support, social simulation, and role-play, a persona that changes after an event is not yet a persona that changes as a person would. Whether a simulated character has "grown" therefore needs checking against event-specific, person-specific human baselines and not just against the presence of change.