Learn · Intermediate

Activation steering: changing a model's behaviour by editing its thoughts

Activation steering is a technique for changing what an AI model does by directly editing its internal numbers while it is running, instead of retraining it or rewriting the prompt. You find a direction in the model's internal space that corresponds to a concept -- honesty, refusal, a particular topic, a tone -- and then add or subtract that direction from the model's activations mid-computation. The effect is immediate, requires no gradient updates, and works at a cost so low it can run on every request.

To see why that is remarkable, it helps to know what the alternatives cost. If you want a model to behave differently, the standard options are fine-tuning, which means collecting data and running training; preference optimization, which means the same thing with more infrastructure; or prompting, which is free but unreliable and consumes context. Activation steering sits outside all three. It treats the model's internal state as something you can reach into and adjust, the way you might turn a knob on a machine that is already running.

Where the direction comes from

The core insight is that language models appear to represent many human-legible concepts as roughly linear directions in their activation space. If you take the model's internal state at some layer while it processes text about honesty, and subtract its state while it processes matched text about deception, the difference vector points, approximately, at "honesty" as the model represents it.

That is the whole recipe in its simplest form, published as the ActAdd method in Steering Language Models With Activation Engineering by Alexander Matt Turner and colleagues in 2023. Take a contrastive pair of prompts. Run both. Subtract one activation from the other. Add the resulting vector, scaled by a coefficient you choose, into the model's residual stream during generation. The model's output shifts toward the concept.

An analogy: imagine a mixing desk for a song that is already playing. You do not re-record the band. You find the fader that happens to control the vocal warmth, and you push it up two decibels. Activation steering is the discovery that a language model has faders like this, and that a surprising number of them correspond to things a person would want to adjust.

The techniques people actually use

Three lines of work built this into something more than a demonstration.

Inference-Time Intervention, from Kenneth Li and colleagues in this 2023 paper, identified attention heads whose activity correlated with truthful answers, then nudged activations along the truthful direction in exactly those heads during generation. It improved truthfulness on a standard benchmark substantially, using a handful of labelled examples rather than a training run.

Representation Engineering, introduced by Andy Zou and colleagues in a 2023 paper, generalized the framing. Rather than treating steering as a trick, it proposed reading and writing model representations as a general method for transparency and control, with recipes for locating directions corresponding to emotions, harmfulness, power-seeking and more.

Refusal as a single direction is the most striking result of the three. Andy Arditi and colleagues showed in this 2024 paper that a model's refusal behaviour -- the whole apparatus of declining harmful requests -- is mediated in large part by one direction in activation space. Add it, and the model refuses harmless requests. Remove it, and safety training stops working.

That last finding is why activation steering matters for safety and not only for control. Whatever refusal training accomplishes, it does not appear to distribute the behaviour robustly across the network. It concentrates it somewhere findable, and anyone with access to the weights can find it.

How it relates to neighbouring ideas

Steering is the write operation for a set of tools whose read operation is mechanistic interpretability. Interpretability asks what a model is representing; steering asks what happens when you change it. The two are tightly coupled in practice, because a steering result is often the strongest available evidence that an interpretability finding is real: if you claim to have found the direction for a concept, adding it should produce the behaviour, and it does or it does not. This is also the relationship the logit lens has to the same problem from the reading side.

Steering is distinct from prompt injection, which manipulates a model through its text input, and from fine-tuning, which changes the weights permanently. It requires access to the model's internals, so it is a tool for people running open weights or working inside a lab, not for API users.

Where it breaks

Three honest limitations. First, the linear-direction assumption is an approximation, and it degrades: concepts that seem clean in one context turn out to be entangled with others, so pushing "formality" up also pushes something you did not want. Second, the scaling coefficient is a hyperparameter with no principled setting -- too small does nothing, too large produces incoherent text, and the usable window varies by model, layer and concept. Third, steering is a blunt instrument compared to training. It does not teach the model anything; it biases what the model already does.

Recent work has made these limits concrete in an adversarial setting. A 2026 paper on covert coordination between agents uses activation-level steering as both an attack -- one agent pushing another's behaviour without leaving anything in the transcript -- and as the basis of a defense, by replacing a suspicious internal handoff with a matched neutral one and measuring whether the behaviour changes. That last move, using steering as a causal probe rather than as a control knob, is probably where the technique is most useful right now: not to make a model do what you want, but to prove what a particular internal state was actually causing.

Key papers
Steering Language Models With Activation Engineering (Turner et al., 2023)
Inference-Time Intervention: Eliciting Truthful Answers from a Language Model (Li et al., 2023)
Representation Engineering: A Top-Down Approach to AI Transparency (Zou et al., 2023)
Refusal in Language Models Is Mediated by a Single Direction (Arditi et al., 2024)

Key questions

What problem does activation steering solve?

It lets you change a model's behaviour without retraining it or rewriting the prompt, by adding a concept direction directly into its internal activations during generation. That makes behavioural adjustment cheap enough to apply per request rather than per training run.

How is activation steering different from fine-tuning?

Fine-tuning permanently changes the model's weights using gradient updates and training data, while activation steering leaves the weights untouched and modifies the internal state at inference time. Steering is reversible, instant and far cheaper, but it only biases behaviour the model can already produce rather than teaching it anything new.

Why does it matter that refusal is a single direction?

Because it means safety training is not robustly distributed across the network but concentrated somewhere findable, so anyone with access to the weights can locate that direction and remove it. It reframes refusal from a learned disposition into a component with an address.

Cite this

APA

Ground Truth. (2026, August 20). Activation steering: changing a model's behaviour by editing its thoughts. Ground Truth. https://groundtruth.day/learn/activation-steering.html

BibTeX

@misc{groundtruth:activation-steering,
  title  = {Activation steering: changing a model's behaviour by editing its thoughts},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/learn/activation-steering.html}
}

Topics: interpretability · safety · control · representations · fundamentals