Evaluating Alignment of Behavioral Dispositions in LLMs
The research evaluated how well Large Language Models (LLMs) reflect human behavioral dispositions like empathy and assertiveness. They transformed self-reported questionnaire answers into situational judgment tests (SJTs) to measure actual behavior in realistic scenarios. Findings show models struggle with low human consensus by being overconfident, and with high consensus, smaller models drift significantly while frontier models still fail to match human preferences.