Decodability is Not Causality: Dissociating Probe Readouts from Behavioral Drivers via SAE Decomposition
A feature-level diagnostic is introduced that decomposes a deployed True/False probe into sparse-autoencoder (SAE) features, ranks those features by both probe alignment and by gradient sensitivity of the model's behavior, and ablates the resulting shared, probe-only, and random feature sets under a coherence gate.
Devesh Tiwari, Camille Davis, Shivank Sinha et al.
· 1 citation