Selective Activation Steering for Robust Model Alignment
July 3, 2026
Three activation steering methods—SwFC, StTP, and StMP—provide lightweight runtime defenses against misalignment. By using logistic regression boundaries to intervene only on specific tokens, these methods mitigate dishonesty and dismissiveness without sacrificing linguistic coherence.
HOW THIS AFFECTS YOU
●
researcherYou can implement projection-aware steering to selectively intervene on activations rather than applying uniform additive shifts.
●
policyThis provides a potential technical mechanism for real-time safety enforcement during model inference.