Modifying the internal representations at specific layers to test causal effects on model behavior and outputs.