Activation steering – Definition & Meaning

Definition of activation steering

: a method for changing an AI model's behavior during inference by modifying selected internal activations, usually without changing the model's trained weights
- The researchers used activation steering to make truthful responses more likely while leaving the model weights unchanged.
- A steering vector derived from contrasting prompts was added to the transformer's hidden states during generation.

Two meanings, one word

In everyday English

Steering is the act of guiding something toward a chosen direction.

In AI

A method for changing an AI model's behavior during inference by modifying selected internal activations, usually without changing the model's trained weights.

Origin & history

first known use2023

The modern language-model technique took shape in 2023 through work on inference-time interventions and activation engineering. Turner and colleagues introduced Activation Addition, or ActAdd, as a lightweight way to compute and add steering vectors, while other researchers used related interventions to elicit more truthful answers.

Test yourself

Which of these is the meaning of activation steering?

Related terms

inference the act of running a trained model to produce an output, such as answering a prompt or… interpretability research into understanding what is happening inside a model, such as which internal features… latent space the hidden, many-dimensional space of numbers in which a model represents concepts, where… fine-tuning further training of an existing model on a smaller, focused dataset so it gets better at a… transformer a neural network design that processes a whole sequence at once and uses attention to work out… alignment the work of making an AI system's goals and behavior match what its designers and users…

Cite this entry

"Activation steering." AI Dictionary, Dadgogo, https://dadgogo.com/dictionary/activation-steering/. Accessed 6 Oct. 2026.