Power Steering: Behavior Steering via Layer-to-Layer Jacobian Singular Vectors

Researchers have developed 'Power Steering,' a new technique for controlling large language model (LLM) behavior by analyzing the mathematical relationships between neural network layers. Created by Omar Ayyub, the method computes layer-to-layer Jacobians—matrices that show how small changes in one layer's activations affect later layers—and extracts their most influential components via singular value decomposition. Crucially, Power Steering uses power iteration to approximate these top components in just ~15 forward passes through the model, making it vastly more efficient than previous approaches.

This efficiency enables researchers to systematically map every possible source/target layer pair in a model, creating a comprehensive 'sensitivity map' of behavioral controls. Unlike contrastive activation addition (CAA), which requires carefully crafted prompt pairs, Power Steering can discover latent behaviors and steering vectors from single prompts. The technique produces results comparable to more computationally expensive methods like MELBO (which uses non-linear optimization) while being orders of magnitude faster, opening new possibilities for interpretability research and controlled AI behavior modification.