Dynamically Scaled Activation Steering
Alex Ferrando, Xavier Suau, Jordi Gonz\`alez, Pau Rodriguez
TL;DR
Dynamically Scaled Activation Steering (DSAS) is a versatile framework that adaptively modulates steering interventions in generative models, improving control over undesired behaviors like toxicity while preserving utility and maintaining low computational costs.
Contribution
DSAS introduces a context-dependent, adaptive scaling mechanism for steering in generative models, enhancing effectiveness and interpretability without significant overhead.
Findings
DSAS improves the trade-off between toxicity mitigation and utility preservation.
It enhances steering effectiveness across different models and tasks.
DSAS maintains low computational overhead while increasing interpretability.
Abstract
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 introduce Dynamically Scaled Activation Steering (DSAS), a method-agnostic steering framework that decouples when to steer from how to steer. DSAS adaptively modulates the strength of existing steering transformations across layers and inputs, intervening strongly only when undesired behavior is detected. At generation time, DSAS computes context-dependent scaling factors that selectively adjust the strength of any steering method. We also show how DSAS can be jointly optimized end-to-end together with the steering function. When combined with existing steering methods, DSAS consistently improves…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The controller is method-agnostic and composes with CAA, ITI, and LINEAS. The paper shows drop-in gains across three families on two LLMs. 2. Pareto-front reporting instead of single operating points is good practice for safety-utility trade-offs. End-to-end variant with a principled control-set regularizer improves or matches vanilla DSAS when trained with LINEAS shows joint optimization is viable. 3. Low overhead with simple linear classifiers and small FLOP cost per token provides prac
1. Relatively small training sets for toxicity risk overfitting and unstable layer classifiers, and results may not hold with broader distributions. Using only 32 examples per split should be justified, with variance analyses beyond 4 seeds. 2. The paper seems to train classifier on average embeddings and then applying per-token, and this mismatch may blunt localization. Per-token supervision would better support token-wise scaling. 3. Given scaling phenomena in steering, showing results on
The paper is well-written and plots are clear and well-designed. The results in Figure 2 are strong at showing that they improve the Pareto frontier and Table 1 convincingly shows improvement over existing methods. The authors also show that their method extends beyond LLMs to diffusion models as well.
Some results give concerns about the sample quality for samples generated after steering and it is unclear how diffusion model performance is affected by steering. Figure 2 should be clarified to ensure that the comparison is 1-1.
The paper clearly identified pain points in existing conditional steering methods and addressd this gap with a universal conditioning mechanism that adapts steering strength per input (e.g., per token or spatial feature) and shows consistent improvements over prior conditional steering baselines. The framework extended DSAS to end-to-end training with joint optimization, and demonstrated cross-modal generality by successfully applying it to text-to-image diffusion models. The diffusion results
While DSAS shows promising results on toxicity mitigation, validating its generalisability beyond this single behavior would strengthen the work, e.g. extending experiments to other safety-related behaviors such as bias mitigation and factuality changes. Similarly, experiments are limited to 1B–2B parameter models; evaluating DSAS on larger-scale LLMs would help confirm scalability across sizes. The banana case study in the text-to-image diffusion experiments is interesting, but appears too nar
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsGenerative Adversarial Networks and Image Synthesis · Cell Image Analysis Techniques · Explainable Artificial Intelligence (XAI)