Change a model's behavior at inference time by adding a direction to its residual stream — contrastive activation addition, steering vectors, persistent edits, and function vectors extracted from in-context-learning prompts. Use to induce or suppress a behavior (sentiment, refusal, style, a task), to test whether a direction found by a probe is causal, or to build always-on model modifications. Covers the two things that decide whether steering works: scaling the vector relative to the activation norm, and sweeping the coefficient to find the band where behavior changes before fluency collapses.
