Steering at the Source: Style Modulation Heads for Robust Persona Control

Researchers have identified a method to control Large Language Model behavior by targeting only three specific attention heads called 'Style Modulation Heads' rather than the entire residual stream. This approach maintains model coherency while enabling precise persona and style control, offering a more efficient alternative to fine-tuning.