Adversarial Robustness of Activation Steering in Large Language Models

Kien Le, Thai Le

Activation steering has become a popular training-free method to control LLMs' behavior by injecting precomputed direction vectors into the model's residual stream at inference tim…