Abliteration
Editing an open-weight model's internals to strip out its ability to refuse requests, without retraining it.
Definition
Abliteration is a technique for removing a language model's refusal behavior by editing its weights directly, rather than retraining it. It builds on a research finding that refusal is largely carried by a single consistent direction inside the model's internal representations: identify that direction, mathematically subtract it from the model's weights, and the model largely loses the ability to say no. Because it needs only a modest amount of compute and no training data, abliteration became a common way to produce 'uncensored' versions of open-weight models. It is contentious — supporters frame it as user control over models they run themselves and a way to remove over-refusal, while critics note it strips safety guardrails wholesale, and the process can also dent a model's general quality.