Glossary Entry

Activation Patching

A causal interpretability technique that transplants internal activations from one forward pass into another to identify which components actually drive a behaviour.

Interpretability LLMs

Also called: causal tracing, interchange intervention

Seed source: Locating and Editing Factual Associations in GPT (Meng et al., 2022)

Activation patching runs a model twice, on a clean prompt and a corrupted or contrasting one, then copies chosen activations from one run into the other and observes which downstream answers flip. If patching a particular layer and position restores or redirects the behaviour, that site is causally implicated in producing it. The ROME paper’s causal tracing used this to localize factual recall to specific mid-layer MLP sites.

Patching sits at the top of the interpretability evidence hierarchy alongside ablation and steering: unlike probes or lenses, which show information is present, an intervention shows the model actually consumes it. The concept-swap experiments in Anthropic’s 2026 global workspace paper (replacing one concept’s pattern in J-space and watching answers track the implant) are activation patching aimed at a specific representational target.