Do AI Personas Grow? Analyzing and Benchmarking Personality Evolution in LLM Agents After Life Events
Ming Wang, Peidong Wang, Xiaocui Yang, Daling Wang, Shi Feng, Fiona Fui-Hoon Nah, Ee-Peng Lim
cs.CL, cs.AI, cs.SI
2026-08-07
Across 14 LLMs playing 100 personas through 11 life events, Big Five personality does shift, but often in the wrong direction, at about a tenth of human magnitude, with all personas collapsed onto one shared trajectory.
A growing set of applications (emotional companionship, social simulation, role-playing) need personality-conditioned LLM agents (PC-Agents) that stay coherent over long interactions. A key question is whether their personality should evolve like a human's after life events. Human psychology has longitudinal evidence: conscientiousness rises after job entry and promotion, neuroticism rises after chronic illness and unemployment, conscientiousness falls after retirement. An agent that ignores these dynamics stays locally consistent but turns into a static script wearing a person's name. Prior work showed LLM personality can shift under perturbation, but how systematically, across which traits, events, and models, and whether direction and magnitude resemble humans, was largely unchecked.
The authors chose 11 major life events (job entry, promotion, unemployment, retirement, marriage, divorce, childbirth, chronic illness, and others) and built 100 demographically controlled personas from a 2 gender by 5 region by 10 personality-type factorial design. Each persona takes the standard Big Five inventory (BFI-44), goes through a structured event reflection, then retakes the same inventory; the difference is the personality change. This runs across 11 models, about 12,000 event-conditioned trajectories. A validation suite (no-event retest, independent paraphrase, scenario-based choices, delayed remeasurement after unrelated dialogue) confirms the signal is real.
Analysis runs along four axes: does change exist, do direction and magnitude match humans, does it split by gender and culture, and do different personas within one event differ individually. From this they build BFI-Adapt, a benchmark that scores whether event-induced shifts follow expected human directions, and rank 14 models.
All four axes agree: models move, but weakly and monotonically.
On direction, across the 27 event-trait pairs with a definite human expectation, only about 56% follow the human direction. Retirement is a universal failure: humans decline in conscientiousness after retirement, but every model predicts a rise, with a median correct rate of just 11.5%. Occupational events align best (graduation 42 to 76%, work entry 49 to 96%). Agreeableness is the weakest dimension: models default to making personas more agreeable after major events, even when human evidence says otherwise.
| Phenomenon | Result |
| Directional hit rate (27 pairs) | about 56% (nearly half reversed) |
| Retirement, conscientiousness | all models reverse, 11.5% median correct |
| Magnitude inside human range | only 11.0 to 16.4% |
| Between-persona SD | median 0.19 (human 0.5 to 0.8) |
Magnitude is worse. Projected onto the human direction, only 11.0 to 16.4% of changes fall inside the human effect-size band (about 0.035 to 0.14 Likert units); the median absolute change is about 0.02, roughly ten times smaller than in humans. The rest reverse, under-shift, or overshoot. The best-calibrated model, Gemini-3-flash, still lands only 16.4% in range.
Both shape axes collapse. Gender and cultural region barely moderate anything; no gender or continent comparison survives multiple-testing correction among the definite-direction pairs. Individually, the between-persona change SD sits at a median of 0.19 versus 0.5 to 0.8 in human samples, a three- to four-fold compression. Different personas put through the same event trace nearly the same trajectory. BFI-Adapt ranking: Gemini-3-flash (0.348), GLM-4.6 (0.322), Qwen3-235B (0.287) lead; the field spans 7.9x.
This turns the question of whether LLM personalities develop like humans from an impression into a quantified diagnosis. The verdict is sober: current PC-Agents simulate the mean of human personality dynamics, not its shape. They shift, but with ambiguous direction, systematically muted magnitude, and erased individual differences. For builders of companion agents and social simulators, believable long-horizon personality evolution is not here yet; you need event specificity and individual heterogeneity baked into training or architecture, not prompted in.
The pre-post design captures only immediate post-event reaction, not the acute-to-adaptation trajectory of multi-year human panels; the authors add a three-turn delayed retest, still far short of human horizons. The human direction priors come from a single source (Specht 2017), and the human sigma and effect-size bands are coarse meta-analytic ranges, capping how precisely models can be judged against them. BFI is self-report, and model self-report converges only weakly with actual behavior (scenario-decision correlation 0.003 to 0.105), so the measured change and the choices made need not be the same thing.