Riemannian-Manifold Steering: Geometry-Aware Generative Autoencoders for Label-Free Steering
Researchers have developed a new method called Riemannian-Manifold Steering to guide language model behavior without requiring labeled data. This approach frames steering as a computation on the geometric structure of activation space, unifying existing linear and nonlinear techniques. The method uses a learned encoder trained on output distances to approximate a specific metric, enabling label-free steering that reliably influences model output across various tasks. AI
IMPACT Introduces a novel geometric framework for controlling LLM behavior, potentially enabling more sophisticated and data-efficient steering techniques.