Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics
Abstract
Large language model (LLM) agents are increasingly used as synthetic participants in opinion-dynamics experiments, often assuming that an agent's behaviour is fully determined by its prompt. We test this assumption with a framework in which LLM agents discuss a contested policy issue over repeated rounds while their stance is quantified by a fixed psychometric instrument independent of any dynamics model. Across six LLMs, committed personas are intrinsically stable: in each of the five models tested without peer input, every agent keeps its assigned side for 20 rounds. The observed loss of alignment under discussion is therefore interaction-induced, we call it alignment drift, and model-specific: gpt-4.1 keeps opinion clusters apart, claude-sonnet-5 holds them partially, and gpt-5.6-sol and claude-opus-5 abandon them. Under selective attention, one LLM maintains separated opinion clusters under stance-homophilous attention. Qualitative profiles persist at doubled population and horizon and on a second issue. Such heterogeneity makes the LLM choice a first-class validity variable for social simulation.
Related
- Dynamics-structure interchangeability in binary opinion models on multiplex networks
- Nonlinear Opinion Dynamics of Stubborn Individuals in Time-Varying Cooperation–Competition Networks
- Inferring the microscopic mechanisms of opinion dynamics using a kinetic Ising model
- Opinion dynamics under large-scale exogenous events
- Modelling opinion dynamics during crises as complex contagion with feedback
BibTeX
@misc{conrad2026measuring,
title = {{Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics}},
author = {Nataša Djurdjevac Conrad and Sebastian Pokutta and Christof Schütte},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.03980},
eprint = {2611.03980},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.03980},
}
RIS
TY - UNPB
TI - Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics
AU - Nataša Djurdjevac Conrad
AU - Sebastian Pokutta
AU - Christof Schütte
PY - 2026
DA - 2026/11/02
PB - arXiv
DO - 10.48550/arxiv.2611.03980
UR - https://arxiv.org/abs/2611.03980
AB - Large language model (LLM) agents are increasingly used as synthetic participants in opinion-dynamics experiments, often assuming that an agent's behaviour is fully determined by its prompt. We test this assumption with a framework in which LLM agents discuss a contested policy issue over repeated rounds while their stance is quantified by a fixed psychometric instrument independent of any dynamics model. Across six LLMs, committed personas are intrinsically stable: in each of the five models tested without peer input, every agent keeps its assigned side for 20 rounds. The observed loss of alignment under discussion is therefore interaction-induced, we call it alignment drift, and model-specific: gpt-4.1 keeps opinion clusters apart, claude-sonnet-5 holds them partially, and gpt-5.6-sol and claude-opus-5 abandon them. Under selective attention, one LLM maintains separated opinion clusters under stance-homophilous attention. Qualitative profiles persist at doubled population and horizon and on a second issue. Such heterogeneity makes the LLM choice a first-class validity variable for social simulation.
ER -