The Failure Happens Before the Drift: The Social Dynamics of Values in LLM Agent Societies

Large Language Model (LLM)-based agents are increasingly used as proxies for human participants in social science research, yet it remains unclear whether they can faithfully simulate diverse and conflicting human value systems. We present a World Values Survey (WVS)-grounded simulation framework where culturally diverse agents with different communication styles engage in longitudinal, value-laden discussions. Across approximately 4,000 conversations involving 1,200 personas, 15 topics, and three models (GPT-4o, Gemini-2.5-Flash, and Gemma-4-E4B), we evaluate value faithfulness, value drift, and conversational realism. We find that more than 50% of personas fail to express their assigned WVS profiles from the outset, while 2-7% drift after repeated conversations. Ablations removing demographic details improve faithfulness for some models but do not change the broader trend: simulated value distributions still systematically deviate from the assigned WVS profiles. Compared to human discussions, simulated dialogues show a different trade-off between stylistic consistency and semantic diversity, often producing content-wise varied but stylistically repetitive exchanges.

Introduction. There is a growing interest in using LLMs as stand-ins for human populations. Prior work has used LLMs to simulate individuals with specific demographic attributes or personality traits (Argyle et al. 2023; Tjuatja et al. 2024; Wang et al. 2025), as well as collective phenomena such as group dynamics and organizational behavior (Park et al. 2024; Zhou et al. 2024). In these settings, LLMs are increasingly positioned as scalable proxies for human judgment, opinion, and behavior (Choi, Young, and Ferrara 2026; Tjuatja et al. 2024; Park et al. 2024). However, real-world societies are shaped not only by demographic diversity but also by diverse and often conflicting value systems(Schwartz 1992). Therefore, the validity of such simulations depends on whether LLM agents can reliably instantiate and sustain culturally grounded value profiles during interaction.

Discussion / Conclusion. While LLM-based simulations increasingly model individuals and populations, our results show that conversational plausibility can mask failures of representational fidelity. They may therefore reproduce long-standing external validity problems, including overgeneralization from WEIRD populations (Henrich, Heine, and Norenzayan 2010). These risks might get amplified in synthetic populations, digital twins, and policy simulations, where unfaithful value profiles can distort estimates of public opinion, intervention effects, or group behavior. Such distortions may especially affect marginalized or culturally underrepresented groups and contribute to algorithmic monoculture, where downstream studies inherit the same model-specific biases (Kleinberg and Raghavan 2021; Saha and Choudhury 2025; Pandey, Saha, and Choudhury 2025; Saha, Pandey, and Choudhury 2025; Saha et al. 2025). Also, since richer demographic prompting is not a sufficient safeguard, LLM-based simulations should be treated as virtual respondents, but as modelmediated instruments requiring validation against the populations they claim to represent (Bisbee et al. 2024; Ziems et al. 2024; Madden 2025).