Comparing Methods for Steering LLM Refusal Behavior

This research investigates methods for controlling "refusal" behavior in safety-tuned large language models (LLMs), which is when a model declines to answer certain prompts. Building on prior work suggesting refusal is mediated by a single linear direction in the residual stream, this study compares two intervention techniques: Difference-in-Means (DiM) and Iterative Nullspace Projection (INLP). The comparison involved applying DiM-based interventions (activation addition and directional ablation) and INLP-derived interventions (nullspace projection and counterfactual flipping) to five open-weight chat models. The findings indicate that INLP's counterfactual flipping method performs comparably to DiM's directional ablation in suppressing refusal. However, INLP's nullspace projection proved less effective. The study also highlighted geometric differences in how these interventions affect the model's activation space. Nullspace projection tends to collapse activations between harmful and harmless clusters, whereas counterfactual flipping moves them into the opposite cluster. This suggests that LLMs might encode the absence of a concept differently from its opposite, offering intriguing avenues for future research into model interpretability and control.