Conceptrefusal direction
Refusal Direction
A concept in mechanistic interpretability research referring to the low-dimensional direction onto which the activation difference between refusing and complying with requests is projected in aligned language models. Moving activations along this direction can alter the model's tendency to refuse.
Timeline (last 90 days)
Nov 5
Refusal direction is a concept in mechanistic interpretability research: in aligned language models, the activation difference between refusing harmful requests and complying with requests can be projected onto a low-dimensional direction, and moving activations along this direction can significantly change the model's refusal tendency.
Unverified50%