OBU constructs a calibration set of paired benign boundary queries, computes the activation covariance, and extracts a low-rank subspace via singular value decomposition.
Orthogonal Boundary Unlearning preserves benign skills while removing dangerous ones
By restricting weight updates to the null space of benign representations, the method cuts attack success rates by over 90% without degrading scientific accuracy.
Independent
Niccolò Ronchetti
Research Digest··2 min read
Ronchetti et al.
§