Causal Foundations for Interpretability

About

Interpretability findings regularly fail to generalise: a circuit found under one ablation strategy disappears under another, a probe that detects a concept scores just as high when the labels are randomly shuffled, two sparse autoencoders trained on the same activations yield different, both-plausible feature dictionaries. This tutorial maps interpretability methods onto Pearl's causal hierarchy — association, intervention, counterfactual — so a result can be matched to the rung its evidence actually reaches, and asks the prior question of identifiability: what a method can recover at all from the experiments one can run.

Tutorial T05: Wednesday, 28 October 2026, 14:00–17:30, room P3, EMNLP 2026 in Budapest.