Why it matters

Probes reveal encoded features. Understanding shapes basic interpretability.

Advertisement

The architecture

Extract activations at layer.

Train linear classifier for property.

Accuracy indicates presence.

Linear probingActivationsat target layerLinear classifierfor propertyAccuracyfeature encoded?Cheap to run; caveat: probe learns to combine features, may find without encoding
Probes.
Advertisement

How it works end to end

Layer-wise probes reveal representation evolution.

Sparse probes (Gurnee 2023).

Diagnostic classifier tradition.