Controllable and explainable personality sliders for LLMs at inference time

Researchers propose Sequential Adaptive Steering (SAS), a new framework for controlling Large Language Model personalities at inference time without retraining. The method uses orthogonalized steering vectors to enable precise, multi-dimensional personality control by adjusting coefficients, validated on Big Five personality traits.

• →Sequential Adaptive Steering (SAS) enables multi-dimensional personality control in LLMs without parameter updates or retraining.
• →The method orthogonalizes steering vectors to prevent destructive interference when controlling multiple personality traits simultaneously.
• →This approach offers a parameter-efficient alternative to expensive Supervised Fine-Tuning or RLHF methods.
• →Users can instantly synthesize complex personality profiles by adjusting coefficients rather than training distinct models.
• →Validation on Big Five personality traits shows superior performance compared to naive baseline approaches.