raghu@dark-factory :~/kb/path-patching $ cat

Path patching

A causal-interpretability technique that intervenes on specific computational paths (edges) between model components rather than whole activation sites, isolating which connections carry a behavior.

grounded in: trend theme 'Interpretability rigor: applying formal causal methods to actually explain LLM internals'; refines the existing activation-patching / causal-scrubbing methods

Connected concepts

Activation patching, Causal scrubbing, Circuit tracing, Causal abstraction

Explore it live in the knowledge graph →