Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics

DOI10.48550/ARXIV.2611.03980MaRDI QIDQ7420478arXivFDO

Authors Natasa Djurdjevac Conrad, Sebastian Pokutta, Christof Schütte, Nataša Djurdjevac Conrad

Publication date 2 November 2026

Published in arXiv

Tldr We investigate how large language model (LLM) agents adapt their opinions when discussing a contested policy issue with each other, quantifying opinion dynamics using a fixed psychometric instrument independent of any dynamics model. The observed loss of alignment is interaction-induced and model-specific; the choice of LLM significantly affects simulation results.

Full work available at URL https://doi.org/10.48550/ARXIV.2611.03980

arXiv classification

cs.AIphysics.soc-ph

This page was built for publication: Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7420478)