Dynamically Scaled Activation Steering
Apple's new steering method cuts the performance tax of safety interventions by adapting intervention strength per input.
Why it matters
Activation steering has been blunt — applying safety interventions uniformly costs model performance. DSAS decouples *when* to intervene from *how*, letting models run at full capability on safe inputs and only throttle undesired behavior when detected. This is a capability + safety advancement that matters for practitioners deploying safety-critical models.
The key facts
13 to knowMethod-agnostic framework (applicable to existing steering approaches)
Adaptive modulation of steering strength across layers and inputs
Targets toxicity mitigation and other undesired behaviors
Decouples intervention triggers from intervention magnitude
Apple Research publication (Sep 2026)
Addresses performance degradation from uniform steering
Method: Dynamically Scaled Activation Steering (DSAS)
Decouples steering trigger from steering strength
Adapts interventions per-layer and per-input
Targets toxicity mitigation and similar safety goals
Method-agnostic framework (works with existing steering approaches)
Solves performance degradation problem of uniform steering
Source: Apple Machine Learning Research
Go to the source
Apple Machine Learningmachinelearning.apple.com
Publisher excerpt: Activation steering has emerged as a powerful method for guiding the behavior of generative models towards desired outcomes such as toxicity mitigation. However, most existing methods apply interventions uniformly across all inputs, degrading model performance when steering is unnecessary. We…