Tag
This paper presents the first systematic study of how additive activation steering, calibrated in single-turn chat, transfers to tool-using ReAct agents. It finds that while the injected direction reaches late layers at near-full strength across settings, the behavioral coupling varies unpredictably between models and contexts, with amplification up to 2x or attenuation, posing immediate safety concerns.
This paper introduces a general formulation of non-linear intervention for large language models, extending beyond the Linear Representation Hypothesis to manipulate features encoded along non-linear manifolds, and validates the approach on refusal bypass steering.