Exploring Unsupervised Methods for Recovering the Refusal Direction in LLMs

Researchers explore unsupervised techniques to identify the refusal direction in large language models (LLMs), aiming to better understand and potentially control when models refuse certain prompts.