Steering Vector
A specialized vector added to the internal hidden states of an LLM during inference to actively push its behavior toward a specific concept or tone.
Think of It Like This
Like adding a tiny weight to a bowling ball to ensure it always naturally curves toward the center of the lane no matter how it is thrown.
By isolating the direction in latent space that represents 'politeness' or 'coding ability', researchers can add or subtract this vector during generation. This allows for real-time, zero-shot control over a model's personality or capabilities without any fine-tuning. It is heavily utilized in mechanistic interpretability and AI safety.