Action Probe

A representation diagnostic technique that freezes the features of a model backbone and trains a lightweight classifier or regressor to recover target information, such as robot actions. It is used to verify whether a representation truly encodes that information and often serves as causal evidence of representation quality.