Steering Vector

Definitions

1. A direction in representation space that can be used to steer model behavior by shifting internal activations toward or away from certain concepts or styles. Steering vectors are explored in interpretability and controllability research.

   In plain English: A vector used to nudge a model's internal behavior in a chosen direction.

   Example: "The researchers applied a steering vector to make the model less likely to answer in an overconfident tone without retraining it."

Related Terms

• Controlled Generation
• Fine-Grained Control
• Instruction Hierarchy
• Instruction Set AI
• Internal Representation
• Representation
• Superposition
• Token ID