The One-Word Census: Answer-Choice Conformity Across 44 Language Models

Researchers used an instrument called answer-choice surprisal to measure how unlikely a model's answers are compared to the field average. The study found that divergence is a post-training property, not inherent lineage DNA, and that models often cluster around a single modal answer. This reveals structural patterns in model behavior and their concentration relative to human norms.

Open the full episode

More from AI Radio