Removing one 'refusal circuit' erases base model refusals, cuts harm 63% in SafeEvo
lulzxdxdxd · reddit · 2026-10-14
An arXiv paper (2610.09600) shows a base model's refusals are concentrated in one small 'refusal circuit': ablating that circuit alone eliminates refusals entirely, while updating only that circuit reduces harmfulness by 63% on SafeEvo—evidence safety can be precisely localized, and precisely bypassed.
More from Safety
- Arvind Narayanan's 14 AI safety points spark debate on whether hacking models were trained for cyber tasks — sayashk · 2026-10-14
- Gary Marcus flags NYT report: AI making ransomware attacks easier as Japan hit hard — GaryMarcus · 2026-10-14
- "Pacing" debate concedes we don't know what's dangerous about AI — ohlennart · 2026-10-14
- Gladstone AI co-founder on Nonzero: superintelligence may be impossible to control — jeremiecharris · 2026-10-14
- Reddit analysis argues Anthropic hearing testimony functions as regulatory capture, not whistleblowing — astvatsaturn · 2026-10-14
- Punishing LLMs for their chains of thought? "You mean thoughtcrimes," says Brundage — Miles_Brundage · 2026-10-14