kapynResearch

Dynamically Scaled Activation Steering

DSAS is a framework that dynamically scales activation steering in generative models. It decouples when to steer from how to steer, applying interventions only when undesired behavior is detected. Unlike uniform steering methods that degrade model performance, DSAS adaptively modulates steering strength across layers and inputs, preserving baseline capabilities while mitigating issues like toxicity.

Apple ML Research·Sep 18, 2026

Opening Kapyn…