CLAS: Context-Dependent Activation Steering Dynamically Adapts LLM Behavior Per Token

arXiv·medium signal

Hsu, Beaglehole, and Radhakrishnan introduce Contextual Linear Activation Steering (CLAS), which dynamically adjusts steering strength based on input context rather than applying a fixed steering vector uniformly to all tokens. Existing activation steering methods suffer from inconsistent quality because a single steering strength fails to generalize across diverse prompts. CLAS makes activation steering practical for production use cases where input distributions vary significantly.