Apple Machine Learning Research1 min readpaperintermediate
Dynamically Scaled Activation Steering
Summary
Dynamically Scaled Activation Steering (DSAS) is a method‑agnostic framework that learns per‑token, per‑layer scaling factors to turn existing activation‑steering interventions on only when a model is likely to produce undesired output (e.g., toxic text). The scaling can be optimized jointly with any steering function, improves the toxicity‑utility trade‑off on language models, transfers to text‑…
- DSAS separates *when* to steer from *how* to steer by learning context‑dependent scaling factors for any existing activation‑steering transformation.
- The scaling factors are computed at generation time per token and per layer, allowing strong intervention only on inputs flagged as risky.
- Joint end‑to‑end optimization of the scaling network and the steering function yields a better Pareto front between toxicity mitigation and utility preservation.
- Experiments (details in the full paper) show consistent improvements over baseline steering on language models and a text‑to‑image diffusion model, with <5 % extra compute overhead.
Uniform steering degrades model quality on benign inputs. DSAS offers a plug‑in layer that can be added to any existing steering pipeline, preserving performance on safe content while still suppressing harmful generations, a practical need for deployed LLM services.
6/10
