Mechanistic interpretability

A branch of interpretability that tries to reverse-engineer the specific internal circuits a model uses to do a task.