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

I am posting some highlights from my master's thesis I thought could be interesting to this audience.SummaryIntrospection can be separated from model confusion but requires careful controlsRefusal direction ablation can increase model introspection performance, but looking at the logit differences on 'Yes'/'No' control questions shows the result of Mechanisms of Introspective Awareness (+53% detection) occurs with significant model degradationInjecting a concept vector and inserting the concept ...