Publication of the Paper “Angular Steering: Behavior Control via Rotation in Activation Space”
A method for steering large language model behavior by rotating vectors in activation space.
Torilab published the paper “Angular Steering: Behavior Control via Rotation in Activation Space” on arXiv (arXiv:2510.26243). The authors are Minh Hieu Vu (Torilab) and Tan M. Nguyen (NUS).
Whenever AI processes language, internal representations emerge within a high-dimensional space, resembling arrows pointing in different directions. Certain behaviors, such as refusing a request or expressing emotion, can be associated with particular directions in this space. Rather than adding to or removing a direction, Angular Steering rotates the arrow by an angle. This provides a way to adjust behavior continuously, much like turning a dial.
The paper also introduces Adaptive Angular Steering, a selective method that acts only on activation vectors aligned with the target feature, helping make the steering process more stable and coherent.