protologue

Activation Steering

Also called activation addition, activation engineering, steering vectors.

Activation steering is a technique for controlling a language model's behavior at inference time by adding a steering vector, typically the difference between the model's internal activations on contrasting inputs, to its hidden activations during the forward pass, without changing its weights.

Description

A steering vector can be computed from as little as one pair of contrasting prompts, such as "Love" and "Hate", or by averaging activation differences across many positive and negative examples, a variant called contrastive activation addition. Scaling the vector up or down, or subtracting it, strengthens or weakens the targeted behavior or topic. It needs no gradient-based optimization, but it requires access to the model's internal activations, which hosted models do not expose.

Sources

  1. Turner et al. (2023). Steering Language Models With Activation Engineering.
  2. Panickssery et al. (2023). Steering Llama 2 via Contrastive Activation Addition.

Cite this entry

Protologue. (2026). Activation Steering. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0219). https://protologue.com/t/activation-steering/

BibTeX
@misc{protologue_activation_steering,
  title = {Activation Steering},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0219},
  url = {https://protologue.com/t/activation-steering/}
}

Markdown JSON