Tag
This paper distinguishes between two tasks in LLM opinion simulation—emulation (generating individual responses) and estimation (directly predicting population distributions)—and finds that base models are better emulators while post-trained models are better estimators, using the Pew American Trends Panel.