What is Causal Scrubbing?

Detailed Deep Dive

Causal Scrubbing is a quantitative framework introduced by Redwood Research for rigorously testing mechanistic interpretability hypotheses. Rather than relying on qualitative feature visualizations, Causal Scrubbing formalizes a hypothesis about a neural network circuit as a computational graph mapping. It then performs behavior-preserving resampling ablations, replacing internal activations with resampled equivalents that maintain the hypothesized invariants. If the model's output performance remains intact, the hypothesis is confirmed; if performance drops, the hypothesis is proven false.