[KongchangAI]
ConceptEAP

Edge Attribution Patching

A causal attribution technique developed from activation patching ideas in mechanistic interpretability. It quantifies the causal contribution of individual components to task outputs by replacing activations along information-passing paths between model components and observing the resulting changes in output.

Timeline (last 90 days)

Sep 21

The study uses EAP (Edge Attribution Patching) to identify task-relevant components such as attention heads and logit-level activations.

Unverified50%

All Facts (1)

Source Articles