Refusal Beyond a Single Direction: A Preliminary Comparison of Diff-in-Means and INLP
Researchers have compared two methods, Diff-in-Means (DiM) and Iterative Nullspace Projection (INLP), for controlling refusal behavior in AI chat models. The study found that INLP's counterfactual flipping intervention was as effective as DiM's directional ablation in suppressing model refusal, while its nullspace projection method was less effective. Restricting INLP to key directions maintained most of its suppression capability with minimal impact on model perplexity, offering a tunable approach to controlling AI responses. AI
IMPACT Offers tunable methods for controlling AI refusal, potentially improving safety and reliability in chat models.