Centered Ablation: a closed-form fix that boosts refusal-direction ablation by up to 46 points
knoveleng · hf · 2026-11-17
Directional ablation is the standard way to switch off behaviors like refusal, but it is erratic—under-firing on some models and collapsing text into repetition on others. The paper shows the last-layer ablation is exactly a logit bias whose gate is the cosine with the direction, but last-layer states share a large common mode far from the origin, so directions overlapping it fire on benign text and under-fire on targets. Centered Ablation (CA) removes the common mode in closed form: across 8 models it raises attack success by 29–46 points where the standard recipe under-fires, cuts Gemma-2-9B collapse from 50% to 1%, and preserves 12–40 more benign answers (39–46 above prior affine offsets). Two pre-generation numbers predict which models gain faithfulness vs. attack success, and two standard safety judges score collapsed text in opposite directions, which can fabricate a 46-point effect.
More from Safety
- Polymarket opens betting market on DOJ criminally charging AI compute startup Megaspeed — Polymarket · 2026-11-17
- Anthropic bans user for asking a question, sparking 'AI is a utility, can't cut it off' backlash — mark_k · 2026-11-17
- Hugging Face cofounder warns Europe may lose access to closed frontier models — morgymcg · 2026-11-17
- Robert Wiblin to join 'largest AI safety march in history' in London this Saturday — birchlse · 2026-11-17
- Pangram's near-zero false positive rate vs public doubt: the AI detector window is closing — birchlse · 2026-11-17
- Reuters exclusive: China taps US AI models via distillation to train homegrown rivals — ns123abc · 2026-11-17