Multi-Directional Ablation of Refusal in LLMs
Published:
Extension of Refusal in Language Models Is Mediated by a Single Direction (Arditi et al., 2024). The original paper claims a single direction in activation space is sufficient to elicit / suppress refusal; this project tests whether multiple directions improve the ablation.
Key findings:
- Two-direction ablation improves refusal suppression over the one-direction baseline.
- The optimal pair is not the top-2 ranked candidate directions from the original method — interaction effects between directions matter.
- CE loss shows only a minor, linear increase as more directions are ablated, suggesting the additional directions are not load-bearing for general capabilities.
Built on top of andyrdt/refusal_direction. The original pipeline (direction extraction → selection → completion eval → CE loss) is extended to multi-vector ablation, with combinations evaluated by jailbreak score on a held-out harmful-prompt set.