What Can We Actually Steer? A Multi-Behavior Study of Activation Control

Tetiana Bas, Krystian Novak

Large language models (LLMs) require precise behavior control for safe and effective deployment across diverse applications. Activation steering offers a promising approach for LLM…