Omer Hofman

New method detects LLM policy violations without training

Researchers have developed a new training-free method for detecting policy violations in large language models (LLMs). This approach, called Activation-Space Whitening, operates directly on the LLM's internal representa…