activation steering

Activation steering changes how a model behaves by nudging its internal state at inference time, with no fine-tuning at all. You find a direction in activation space that corresponds to some attribute — honesty, refusal, a happy tone, a topic — and then add that vector into the residual stream as the model runs. The model leans in the chosen direction the way pushing a sail tilts a boat, while everything else continues normally.

The direction is usually built from contrast. Run the model on pairs that differ only in the target attribute — for example matched honest and deceptive completions — and take the difference of their activations to get a steering vector. Adding it amplifies the attribute and subtracting it suppresses it. This connects directly to representation engineering and to SAE feature steering, which derives the direction from a learned dictionary instead of paired prompts.

Steering is appealing because it is cheap, reversible, and interpretable — it treats behavior as directions you can add and remove. But it is blunt: a single vector rarely isolates one clean concept, strong pushes break fluency, and the same handle that promotes a helpful trait could promote a harmful one, so it is a safety tool and a safety risk at once.