AppleAnalysisAI ModelsSeptember 18, 2026

Apple researchers introduce Dynamically Scaled Activation Steering

DSAS decouples when to steer from how to steer, adaptively scaling existing steering transformations across layers and inputs so intervention is strong only when undesired behavior is detected. Apple reports it improves the Pareto front for toxicity mitigation versus utility preservation and also works on text-to-image diffusion models.

1 source

More stories today

Open the live feed