We find that deleting and swapping interventions retain 72-95% of the original model’s prediction accuracy without fine-tuning, and hypothesize the existence of four universal stages of inference across eight different models.

We present MIPS, a novel method for program synthesis based on automated mechanistic interpretability of neural networks trained to perform the desired task, auto-distilling the learned algorithm into Python code.

The soft-minimum of the model-estimated likelihoods of each pixel’s annotated class – that is particularly effective to identify images that are mislabeled, across multiple types of annotation error