← 研踪 PaperTrailarXiv:2611.03980 · v1
Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics
首次提交 2026-11-03 05:20
版本更新 2026-11-03 05:20
雷达观测 2026-11-04 23:40
摘要
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.
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-03 05:20 · 最新观测版本
Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics
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.
查看此版本解读 →查看此版本原文 ↗