Definition
Plain language
Rewinding an AI to just before a particular sentence and re-running many times to measure how much that sentence pushed its decision.
As stated in the literature
An interpretability technique that resamples continuations from immediately before and after a given sentence to estimate that sentence's causal contribution to an outcome.
Why it matters: It pinpoints which specific sentences actually drove an outcome, turning vague impressions about a model's reasoning into measurable causes.
For example, researchers rewind a model to just before a key sentence and rerun it many times to see how much that sentence shifted the final decision.
Heard on the show
“The first tool is called sentence resampling.”Episode 174 — When the AI 'Schemes,' It's Usually Just Lazy or Confused