Controlling Large Language Models Through Concept Activation Vectors