Do LLM personas actually embody their assigned cultural values?

The paper builds a World Values Survey (WVS)-grounded simulation in which culturally diverse agents with different communication styles hold longitudinal, value-laden discussions. Across approximately 4,000 conversations, 1,200 personas, 15 topics, and three models (GPT-4o, Gemini-2.5-Flash, and Gemma-4-E4B), it reports that "more than 50% of personas fail to express their assigned WVS profiles from the outset, while 2-7% drift after repeated conversations." The title states the ordering plainly: the failure happens before the drift. The discussion adds the practical warning that "conversational plausibility can mask failures of representational fidelity."

The paper frames validity as two requirements, that agents can "reliably instantiate and sustain" culturally grounded value profiles during interaction. Most of the measured failure sits in the first requirement, instantiation, and only a small share in the second, sustaining. The ablation supports this reading. Removing demographic details improves faithfulness for some models but "does not change the broader trend," since simulated value distributions "still systematically deviate" from the assigned profiles. The realism comparison points the same way: against human discussions, simulated dialogues trade stylistic consistency against semantic diversity differently, "often producing content-wise varied but stylistically repetitive exchanges." Fluent, varied surface text is what makes the value failure hard to see. From this the authors conclude that richer demographic prompting is not a sufficient safeguard and that such simulations should be treated as model-mediated instruments requiring validation against the populations they claim to represent.

Against the nearest notes, this extends How do we generate realistic personas at population scale? from demographic joint distributions to values: even when a persona is specified, the value profile may not come through. It also sharpens Why do LLM persona prompts produce inconsistent outputs across runs?. That note describes noise swamping persona signal, while this paper describes value distributions that deviate systematically. It qualifies optimism such as Can AI agents learn people better from interviews than surveys? and Can AI personas reliably replicate human experiment results?: success on individual responses or robust main effects does not show that an agent holds a stable value system inside a conversation.

The excerpt is silent on several points a reader would need. It does not say how faithfulness was scored, how the over-50% figure divides across the three models, which models improved when demographics were removed, or what a human baseline for drift would be. It offers no mechanism for why instantiation fails. The WEIRD-overgeneralization and algorithmic-monoculture points are stated as risks ("may therefore", "might get amplified"), not measured results, and the effect on marginalized groups is likewise a stated concern. What follows at this strength is narrow: a longitudinal drift check cannot stand in for a check at the start of the run, and a persona's values need validating against the target population before its conversations are used as evidence about that population.