Steer large language models while preserving unrelated abilities
LocUS: Head Selection and Subspace Projection for Targeted Activation Steering
Computation and LanguageMachine Learning
Summary
Large language models can be guided to change how they respond without needing extra training, but existing methods affect many parts of the model and can unintentionally reduce skills the model already has. The authors present LocUS, a way to focus interventions on specific parts of the model related to certain words or ideas, and only on a few attention heads. This targeted approach keeps the model’s general abilities better while changing only what’s needed, like reducing toxic or biased responses.
What this means in practice
- •For ai safety teams: Control harmful or biased language outputs from AI models by steering activations with minimal effect on overall model performance.
- •For ai chatbot developers: Reduce sycophantic or undesirable responses in conversational AI by applying focused activation steering to select model components.
Authors
Irene Tallini, Lorenzo Basile, Valentino Maiorca, Francesco Locatello, Alberto Cazzaniga
Abstract
Activation steering is a powerful training-free paradigm for controlling large language models at inference time. However, standard approaches estimate a per-layer steering direction from contrastive data and apply it on the layer's entire representation space, which may couple the intervention to off-target properties present in the contrastive data and degrade unrelated capabilities. To mitigate this issue, we introduce LocUS (Localized Unembedding Steering), a method which grounds activation steering to the model's own output vocabulary subspace. By identifying a property-specific linear subspace within the unembedding matrix, LocUS enforces a geometric constraint that restricts the steering transformation to a specific subspace and at the same time localizes its application to a sparse subset of attention heads. Extensive evaluations across three model families on toxicity mitigation, sentiment redirection and sycophancy suppression show that LocUS matches or outperforms state-of-the-art baselines while intervening on under 6% of parameters and better preserving general capability.