Steering Large Language Model Activations in Sparse Spaces