Where to Steer: Input-Dependent Layer Selection for Steering Improves LLM Alignment
Where to Steer (W2S) is a framework that automatically chooses which layer within a large language model should be adjusted to guide its behavior, based on the specific input it receives. Traditional methods use fixed layers, assuming one best layer works for everything. W2S learns this mapping from input embeddings to optimal layers, consistently outperforming fixed-layer steering across various models and behaviors.