Controlling Chat Style in Language Models via Single-Direction Editing

Researchers developed a training-free method to control stylistic attributes in large language models by identifying that different styles are encoded as linear directions in the model's activation space. The approach enables precise style control while preserving core capabilities and supports linear style composition across over a dozen tested models.

• →Stylistic attributes in LLMs are encoded as linear directions in the model's activation space, providing empirical evidence for representation engineering approaches.
• →The new method offers training-free style control that maintains model performance while enabling precise stylistic adjustments.
• →Linear style composition is possible, allowing multiple stylistic attributes to be combined effectively.
• →The approach enhances safety by enabling ablation of undesirable behaviors through representation manipulation.
• →Testing across over a dozen models demonstrates broad applicability and high style adherence at minimal computational cost.