Refusal Beyond a Single Direction: A Preliminary Comparison of Diff-in-Means and INLP

In three linesComparison of two intervention methods on safety refusals in chat models: Difference-in-Means (DiM) vs Iterative Nullspace Projection (INLP). On 5 open-weight models, INLP counterfactual flipping matches DiM for refusal suppression, while nullspace projection is consistently weaker. The two approaches operate differently in activation space.

Summary generated by Claude — human-verified