Probing Classifiers: Finding What a Layer Knows

Train a classifier on hidden states to test what a layer encodes, with the control task that separates the representation's content from the probe's own capacity.