A Geometric Account of Activation Steering through Angle-Norm Decomposition

In three linesEmpirical study on activation steering in language models. Authors decompose interventions into angular and radial components, showing concepts reside primarily in angular structure but norm remains important for stability. Recommendation: parameterize steering by interpretable angular and radial components rather than a single additive coefficient.

Summary generated by Claude — human-verified