Study Shows NLAs Fail to Interpret Steered Activations Faithfully

This is interesting I definitely think of prompting as steering in activation space but I took for granted that I could always come up with some, perhaps complex, prompt to steer activations however I wanted. Guess I was wrong!