Mechanistic Interpretability

We introduce a more principled approach for evaluating the quality of mechanistic interpretations via behavior-preserving resampling ablations—converting hypotheses into classes of activations inside a neural network can be resampled without affecting behavior.