EMNLP 2026 Tutorial · Causal Foundations for Interpretability

What you can vary or observe narrows the equivalence class

CRL, causal representation learning, pins z down with contrastive pairs, interventions on concepts and natural experiments.

Your affordances: what you can vary or observe

Equivalence class

the explanations the data cannot tell apart

18 of 18 candidate axis pairs fit

Your axes, angle from h₁40°

✓Activations alone: your axes fit
?Interventions on concepts: patching sets h; learning z from patched runs is open.

z: the concepts, here gender and tense; g: the mixing that writes z into the activation h.
Dots: inputs, in two coordinates of h; lines: candidate concept axes, 18 pairs at right angles, 5° apart; each pair is one explanation.
Arrows: contrastive pairs, two inputs that differ in one concept.
Hollow dots: the same inputs in a second checkpoint, a natural experiment whose g writes gender in twice as strongly.