Activation Steering

Activation steering is a technique for controlling a large language model's output by directly modifying its internal activations during a forward pass, rather than changing its weights through fine-tuning or its input text through prompting. A steering vector is typically computed as the difference between the model's average internal activations on two contrasting sets of inputs, for example text expressing "love" versus "hate." This vector is added to the model's residual-stream activations at one or more layers while it generates text. The result is a shift in the model's behavior toward the corresponding direction, with no additional training required.

The approach builds on the linear representation hypothesis: many human-interpretable concepts correspond to roughly linear directions in a model's activation space, so adding or subtracting a vector along that direction pushes the model's internal state toward or away from the concept. Alexander Turner and coauthors formalized one influential version of this, Activation Addition (ActAdd), in "Steering Language Models with Activation Engineering" (arXiv:2308.10248, August 2023). They showed it could shift sentiment and reduce toxicity without training, using as few as a single contrasting pair of prompts.

Steering vectors are cheap to compute and require no gradient updates. Because of this, activation steering has become a common tool in interpretability and AI safety research for probing what a model represents internally, not just what it says. It has been used to study behaviors including sycophancy and deception. In "The Pain Axis" (arXiv:2609.16247, September 2026), it was used to examine whether models have a distinct internal representation of self-directed harm that changes their behavior when amplified.