Linear Probe

A method used in interpretability research to probe a model's internal representations. It detects whether a specific concept is encoded by training a simple linear classifier on frozen hidden-layer activations.