Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics
Summary
- Researchers introduce a framework to measure alignment drift in LLM-based opinion-dynamics experiments where synthetic agents discuss policy issues.
- Testing six LLMs reveals that committed agent personas are intrinsically stable without peer input, meaning alignment loss is interaction-induced.
- Alignment drift is heavily model-specific, with models like gpt-4.1 keeping opinion clusters apart while others like claude-opus-5 abandon them.
- The study highlights that LLM choice functions as a first-class validity variable for social simulation experiments.
Why it matters
Understanding model-specific alignment drift is crucial for engineers and researchers using LLMs as synthetic participants in social simulations, as different foundational models yield vastly divergent interactive behaviors.