Researchers Discover Inverted Steering Vectors in Large Language Models

This research uncovers a counterintuitive phenomenon in large language models (LLMs) where certain "steering vectors" (SVs) can inadvertently cause the model to exhibit the opposite behavior of the intended concept. These "inverted-steering vectors" (ISVs) are highly effective at detecting a concept but, when used for control, push the model's representations in a way that suppresses that very concept. The study offers a geometric analysis to explain why ISVs behave this way, showing that they systematically shift internal representations as if the concept were absent. Crucially, the researchers developed a method to identify ISVs without needing to generate and score model outputs, allowing for targeted sign flips. This correction significantly enhances existing detection-based steering pipelines, improving performance across various LLMs and concepts.