---
title: "Dynamically Scaled Activation Steering"
description: "Activation steering has emerged as a way to guide generative models toward desired behaviors, but uniform interventions can reduce performance when steering is unnecessary. DSAS, or Dynamically Scaled Activation Steering, addresses this by separating when to steer from how to steer and adapting intervention strength across layers and inputs. At generation time, it computes context-dependent scaling factors that modify existing steering transformations, and it can also be jointly optimized end-to-end with the steering function. Combined with existing methods, DSAS improves the reported Pareto front between toxicity mitigation and utility preservation, while adding minimal computational overhead; experiments also apply it to a text-to-image diffusion model for modulating specific concepts and show token-level interpretability through the strength and location of steering."
---

# Dynamically Scaled Activation Steering

[Apple](https://yomu.fyi/company/apple) · Sep 18, 2026

**Type:** Explainer

## Summary

Activation steering has emerged as a way to guide generative models toward desired behaviors, but uniform interventions can reduce performance when steering is unnecessary. DSAS, or Dynamically Scaled Activation Steering, addresses this by separating when to steer from how to steer and adapting intervention strength across layers and inputs. At generation time, it computes context-dependent scaling factors that modify existing steering transformations, and it can also be jointly optimized end-to-end with the steering function. Combined with existing methods, DSAS improves the reported Pareto front between toxicity mitigation and utility preservation, while adding minimal computational overhead; experiments also apply it to a text-to-image diffusion model for modulating specific concepts and show token-level interpretability through the strength and location of steering.

## Context

Existing activation-steering methods generally apply interventions uniformly across inputs, which can degrade model performance when steering is unnecessary. The work considers toxicity mitigation, utility preservation, and modulation of specific concepts in a text-to-image diffusion model.

## Approach / What changed

DSAS decouples when to steer from how to steer. At generation time, it computes context-dependent scaling factors that adapt the strength of an existing steering transformation across layers and inputs, intervening strongly when undesired behavior is detected. The scaling mechanism can also be jointly optimized end-to-end with the steering function.

## Takeaways

- DSAS adaptively scales existing steering transformations across layers and inputs instead of applying one intervention strength uniformly.
- When combined with existing steering methods, DSAS improves the reported trade-off between toxicity mitigation and utility preservation.
- The framework adds minimal computational overhead and identifies which tokens require steering and by how much.

## FAQ

### How does DSAS decide when and how strongly to steer?

At generation time, DSAS computes context-dependent scaling factors for an existing steering transformation. These factors adapt intervention strength across layers and inputs, with strong intervention applied when undesired behavior is detected.

### Can DSAS work with steering methods beyond one specific implementation?

Yes. DSAS is described as method-agnostic and adjusts the strength of any steering method; its scaling mechanism can also be jointly optimized end-to-end with the steering function.

### Does DSAS apply only to language models?

No. The work also applies DSAS to a text-to-image diffusion model, where adaptive steering is used to modulate specific concepts.

**Tags:** [LLMs](https://yomu.fyi/topic/llm), [Machine Learning](https://yomu.fyi/topic/machine-learning)

- Source: [Apple](https://machinelearning.apple.com/research/dynamically-scaled-activation-steering)
- Source URL: https://machinelearning.apple.com/research/dynamically-scaled-activation-steering
- Ingested by Yomu: 2026-09-18T15:00:06.924Z

[Read original post](https://machinelearning.apple.com/research/dynamically-scaled-activation-steering)