--- title: Dynamically Scaled Activation Steering | daily.dev description: Apple researchers introduce Dynamically Scaled Activation Steering (DSAS), a method-agnostic framework that adaptively modulates the strength of activation... canonical: https://daily.dev/posts/dynamically-scaled-activation-steering-tynobwqem twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Dynamically Scaled Activation Steering | daily.dev og:description: Apple researchers introduce Dynamically Scaled Activation Steering (DSAS), a method-agnostic framework that adaptively modulates the strength of activation... og:url: https://daily.dev/posts/dynamically-scaled-activation-steering-tynobwqem og:image: https://api.daily.dev/og/posts/TYNobWQeM.png og:image:alt: Dynamically Scaled Activation Steering og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Dynamically Scaled Activation Steering **[Apple Machine Learning Research](https://daily.dev/sources/mlapple)** · 1 min read · 0 upvotes · 0 comments ## Summary Apple researchers introduce Dynamically Scaled Activation Steering (DSAS), a method-agnostic framework that adaptively modulates the strength of activation steering interventions per input and layer rather than applying uniform steering across all inputs. By intervening strongly only when undesired behavior is detected, DSAS improves the trade-off between toxicity mitigation and utility preservation compared to static steering methods, and can be jointly optimized end-to-end with the steering function. The approach generalizes beyond language models to text-to-image diffusion models, enabling modulation of specific concepts, and adds minimal computational overhead while improving interpretability by pinpointing which tokens need steering and by how much. Code is planned for release on GitHub. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is Dynamically Scaled Activation Steering (DSAS)? DSAS is a method-agnostic framework that adaptively modulates the strength of activation steering interventions, intervening strongly only when undesired model behavior is detected instead of applying uniform steering to every input. It computes context-dependent scaling factors at generation time to adjust any existing steering method's strength across layers and inputs, improving the trade-off between toxicity mitigation and utility preservation. _daily.dev surfaces research like this for engineers building safer, more controllable generative models._ ### Can activation steering methods be applied to text-to-image diffusion models, not just language models? Yes, activation steering combined with dynamic scaling has been demonstrated on a text-to-image diffusion model, showing that adaptive steering can modulate specific concepts within image generation. This suggests the technique generalizes beyond language models to other generative architectures, since the scaling framework is designed to be method-agnostic. _Explore generative model control techniques across modalities via daily.dev for teams evaluating diffusion pipelines._ ## Similar posts on daily.dev - [DeepSeek-V4-Flash makes LLM steering interesting again](https://daily.dev/posts/deepseek-v4-flash-makes-llm-steering-interesting-again-0rujyanjx) · sean goedecke · 7 upvotes · 0 comments - [NDSS 2025 – Safety Misalignment Against Large Language Models](https://daily.dev/posts/ndss-2025-safety-misalignment-against-large-language-models-hpoay68cq) · Security Boulevard · 0 upvotes · 0 comments --- Tags: [#diffusion-models](https://daily.dev/tags/diffusion-models) [View this post on daily.dev](https://daily.dev/posts/dynamically-scaled-activation-steering-tynobwqem) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Dynamically Scaled Activation Steering","url":"https://daily.dev/posts/dynamically-scaled-activation-steering-tynobwqem","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/dynamically-scaled-activation-steering-tynobwqem"},"datePublished":"2026-09-18T14:50:30.949Z","dateModified":"2026-09-18T14:50:49.808Z","description":"Apple researchers introduce Dynamically Scaled Activation Steering (DSAS), a method-agnostic framework that adaptively modulates the strength of activation...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8dda20dea5f9a8a0e25a1b7bb36a1a20?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8dda20dea5f9a8a0e25a1b7bb36a1a20?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Apple Machine Learning Research","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Apple Machine Learning Research","logo":"https://media.daily.dev/image/upload/s--mD-1uSe0--/f_auto,q_auto/v1785654191/logos/mlapple","url":"https://daily.dev/sources/mlapple"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/dynamically-scaled-activation-steering-tynobwqem","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"diffusion-models","timeRequired":"PT1M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Apple Machine Learning Research","item":"https://daily.dev/sources/mlapple"},{"@type":"ListItem","position":3,"name":"Dynamically Scaled Activation Steering"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/dynamically-scaled-activation-steering-tynobwqem#faq","mainEntity":[{"@type":"Question","name":"What is Dynamically Scaled Activation Steering (DSAS)?","acceptedAnswer":{"@type":"Answer","text":"DSAS is a method-agnostic framework that adaptively modulates the strength of activation steering interventions, intervening strongly only when undesired model behavior is detected instead of applying uniform steering to every input. It computes context-dependent scaling factors at generation time to adjust any existing steering method's strength across layers and inputs, improving the trade-off between toxicity mitigation and utility preservation. daily.dev surfaces research like this for engineers building safer, more controllable generative models."}},{"@type":"Question","name":"Can activation steering methods be applied to text-to-image diffusion models, not just language models?","acceptedAnswer":{"@type":"Answer","text":"Yes, activation steering combined with dynamic scaling has been demonstrated on a text-to-image diffusion model, showing that adaptive steering can modulate specific concepts within image generation. This suggests the technique generalizes beyond language models to other generative architectures, since the scaling framework is designed to be method-agnostic. Explore generative model control techniques across modalities via daily.dev for teams evaluating diffusion pipelines."}}]} ```