Base Models Outperform Post-Trained Models in Human Opinion Emulation

Research distinguishes between emulation (generating individual responses) and estimation (predicting population distributions) in opinion simulation. Findings show base models are superior emulators that preserve demographic structures, whereas post-trained models are more effective at direct distributional estimation.