Control questions matter: introspection, refusal ablation, and gates in open Gemma models

Researchers analyzed open Gemma AI models to study how control questions affect their introspection and refusal behaviors, focusing on mechanisms like refusal ablation and gating.