Logistic Regression Probe

A classical technique in interpretability research that extracts hidden state vectors from intermediate Transformer layers and trains lightweight linear classifiers on them, in order to test whether specific information is encoded in the representations of a given layer.