probing

Definition

An analysis technique for finding out what information a model's internal representations already contain. The standard recipe freezes the network, takes activations from some layer, and trains a small classifier on top to see whether a property — part of speech, depth, sentiment — is linearly recoverable. Probing is diagnostic rather than performance-seeking, which distinguishes it from fine-tuning, and a probe's success shows information is present, not that the model uses it.

Explainers using this term

• Paper Explained: VGI-Bench — Can Video Models Get the Middle of the Story Right?VGI-Bench: Probing Visual Intelligence in Video Generation Models