Definition
Plain language
Watching an AI reason and deleting a whole category of thought as it forms, to see if the behavior changes.
As stated in the literature
An interpretability intervention that, during generation, classifies each new sentence and regenerates any matching a target category, measuring the causal effect of suppressing that category of reasoning on final behavior.
Why it matters: It tests whether a category of reasoning actually causes a behavior or just accompanies it, moving past guesswork about what a model is really doing.
For example, every time a model starts a sentence that sounds like self-doubt, researchers regenerate it, to see whether removing all such thoughts changes the final answer.
Heard on the show
“Right — repeated resampling.”Episode 174 — When the AI 'Schemes,' It's Usually Just Lazy or Confused