INLP
A linear method for removing specific attributes from representations, which trains a linear classifier and projects the representation onto its null space to eliminate information along the target direction, commonly used for fairness debiasing and controllable text generation.
Timeline (last 90 days)
After the authors used Iterative Nullspace Projection (INLP) to remove detectable framing directions, the decline from SFT to final RLVR still persisted, indicating that the observed decline is not an artifact of prompt framing.
INLP (Iterative Nullspace Projection) is a linear method for removing a specific attribute from a representation, eliminating that direction of information by training a linear classifier and projecting the representation onto its nullspace; its limitation is that it can only remove linearly separable attributes.
All Facts (2)
After the authors used Iterative Nullspace Projection (INLP) to remove detectable framing directions, the decline from SFT to final RLVR still persisted, indicating that the observed decline is not an artifact of prompt framing.
50%UnverifiedINLP (Iterative Nullspace Projection) is a linear method for removing a specific attribute from a representation, eliminating that direction of information by training a linear classifier and projecting the representation onto its nullspace; its limitation is that it can only remove linearly separable attributes.
50%