Post-training reduces human-like behavior in LLMs

A new study introduces Psych-201, a dataset for measuring behavioral alignment between LLMs and humans. The research finds that post-training, which converts base models into assistants, consistently reduces alignment across model families and sizes. This misalignment increases in newer model generations. Persona-induction, a technique for eliciting human-like responses, does not improve individual-level predictions. The results suggest that current methods for making LLMs useful also make them less accurate models of human behavior.