Post-training makes large language models less human-like

TL;DR

This empirical study finds that post-training, the process of making LLMs useful assistants, degrades their performance in modeling human behavior.

Problem

LLMs are increasingly used as human surrogates, but it is unclear which models best capture human behavior and why.

Approach

The researchers introduce a novel dataset, Psych-201, to measure behavioral alignment with humans at scale across various model families, sizes, and objectives.

Results & Contribution

Post-training consistently reduces alignment with human behavior, and this misalignment widens in newer model generations. Furthermore, persona-induction techniques do not improve individual-level predictions. The results suggest that the very processes used to make LLMs useful assistants may also make them less accurate models of human behavior.