Steering LLM Personality via Latent Feature Interventions

This study introduces a novel method for manipulating the personality traits of large language models (LLMs) at a mechanistic level. Instead of relying on prompt engineering or fine-tuning, the approach directly intervenes within the model's latent feature space. By utilizing sparse autoencoders and contrastive activation analysis, researchers identify specific latent directions within the residual stream that correlate with target OCEAN personality traits. The core of the method involves applying a small, additive steering vector to the LLM's hidden states. This intervention is shown to enhance desired personality traits without compromising the model's overall language modeling performance. The researchers also explore optimization techniques, such as a linear weighting heuristic with grid search, to balance personality expression with task performance. This work demonstrates a promising path towards more controllable and interpretable AI systems.