Probing Classifiers

A probe asks: is property P decodable from the activations at layer k? Take a trained model, run inputs through it, harvest activations at a chosen layer, and train a small classifier — usually linear logistic regression — to predict P. If the probe learns the property, some feature at that layer encodes it linearly.

The setup is deliberately cheap: frozen model, off-the-shelf labels, a classifier trivially small next to the network it's reading. That cheapness is the point — probes are a survey instrument, meant to scan 'what's represented where' across layers, tasks, and concepts.