Abliteration Mitigation via Refusal Aliases

Nathan Truong

Abliteration, the removal of refusal capabilities from large language models by projecting weight matrices orthogonal to an extracted refusal direction, has emerged as a prominent…