Causal Foundations for Interpretability — Interactive Demos

A control task shows what a probe can score by memorisingIn Hewitt and Liang's control task, each word type is given a fixed random tag, so the only way to score well is to memorise; the same linear probe still reaches 71.2%, against 97.2% on part of speech.Open demo →

Pearl's ladder sorts questions into rungsCounterfactual: what this one input would have produced had a been different.Open demo → The same gender question reads differently at each rungRung 1: the probe scores the same in both hidden worlds. Rung 2: INLP moves the predictions in both. Rung 3: one bio, and which rerun equals the target.Open demo → Re-running a forward pass answers counterfactuals about heads but not about conceptsBoth reruns use one bio; each keeps different nodes fixed.Open demo →

A rotated basis fits the same subspaceA symmetry: a change that leaves the loss unchanged.Open demo → Steering along a mixed axis moves two conceptsA mixed axis: steering along it moves both concepts.Open demo → SAEs trained on the same activations disagreeTwo seeds, two dictionaries: consistent with several sparse solutions.Open demo → Same Loss, Different StoryReconstruction cannot see the rotation, and sparsity can pick the wrong one.Open demo → What you can vary or observe narrows the equivalence classCRL, causal representation learning, pins z down with contrastive pairs, interventions on concepts and natural experiments.Open demo → Will the Edit Survive Deployment?Transport: a finding still holds on new inputs. Our answer to “when will it generalise?”Open demo →

Diagnose the claimA tool for writing: put any claim through it before it goes in a paper.Open demo →