Subjectivity Coefficient for Evaluating LLM Social Simulations

This research introduces a subjectivity coefficient based on entropy to distinguish objective tasks from subjective social simulations. It argues that accuracy-based evaluation and hard-label training are unreliable for human behavior modeling, which is inherently probabilistic.