Activation Steering
Injecting a computed direction vector into a model's residual stream at inference time to bias its output toward a target behaviour — without any weight updates or retraining.
Imagine being able to change how an artificial intelligence behaves without retraining it, rewording your prompt, or touching its underlying weights. This is activation steering. It works by modifying the model’s internal activations on the fly.
To do this, researchers find a specific direction vector in the model’s internal representation space. One common way to find this vector is by comparing two contrasting prompts—one that shows the desired behavior, and one that does not. By subtracting their activation patterns at a specific layer, you get a steering vector. Adding this vector back into the model during a live run immediately biases the output toward that target behavior.
An even more precise method uses sparse autoencoders. These tools break down a model's complex activations into individual, interpretable features, like a specific brand or concept. When Anthropic amplified a feature for the Golden Gate Bridge in one of their models, it began mentioning the bridge in response to almost any question.
This technique relies heavily on mechanistic interpretability. Essentially, interpretability finds the lever inside the model, and steering pulls it.
For search engine optimization, activation steering is a valuable research tool. It shows that brand associations are not permanently locked in. Instead, they are localized, adjustable patterns. While you cannot directly steer external, proprietary models today, experimenting with open-weight models shows how targeted content can shift where a brand sits in an AI’s mental map.
What activation steering is
Activation steering is a technique for controlling what a language model outputs by modifying its internal activations during a forward pass, rather than by changing its weights or crafting a prompt. A direction vector representing a desired behaviour is computed, then added to the model's residual stream at a chosen layer during inference. The effect is immediate and does not require retraining.
How the steering vector is found
The standard approach involves running the model on two contrasting prompts — one that exhibits the target behaviour and one that does not — and subtracting the resulting activation patterns at a specific layer. The difference captures the direction in the model's internal representation space that corresponds to the behaviour. Scaling that vector up and injecting it during a subsequent generation biases the model toward the target without touching the underlying weights.
Sparse autoencoders offer a more precise route: they decompose the model's activations into interpretable features, each corresponding to a concept or property. Amplifying a specific feature — a brand, a topic, a quality signal — produces targeted shifts in output. Anthropic demonstrated this by amplifying a "Golden Gate Bridge" feature in Claude, causing the model to reference the bridge in answers to completely unrelated questions.
Relationship to mechanistic interpretability
Activation steering depends on the insights from mechanistic interpretability: you need to know which layer encodes the feature of interest before you can steer it. The two practices are complementary — interpretability locates the lever; steering pulls it.
AI SEO relevance
For AI SEO, activation steering is primarily a research tool for understanding how models encode brand information and testing whether that encoding can be modified. It demonstrates that brand associations are not hard-coded but are localised, adjustable patterns — which supports the broader argument that content and citation strategies can shift where a brand sits in a model's representational space. Direct production use of steering on external models is not currently possible, since it requires access to the model's internal activations, but open-weight models like Gemma make experimentation straightforward.