A Little Confidence Goes a Long Way

John Scoville, Shang Gao, Devanshu Agrawal +1

We introduce a group of related methods for binary classification tasks using probes of the hidden state activations in large language models (LLMs). Performance is on par with the…