ML//alignment//mechanistic interpretability//activation steering

2026-04-08

Activation steering is an inference-time intervention on a language model that adds or subtracts concept vectors to or from the residual stream without modifying any weight, and it is used to amplify or suppress a concept in the model's behavior.

The concept vector comes from contrastive prompt pairs, one mentioning the concept and one without: both run through the model and their activations are subtracted, and the difference points in the direction of that concept in latent space. Adding the vector amplifies the concept (positive steering); adding the negated vector suppresses it (negative steering, or inhibition).

Injection happens at the residual stream level, at all token positions, often across multiple layers. No weights, attention heads or MLPs are touched; only the shared communication channel between layers is altered.

Vectors are calibrated so that a force of 1.0 corresponds to the mean activation norm at the model's middle layer. The typical steering force is 0.1, and higher forces cause degenerate or incoherent output.

The Mythos system card used three distinct vectors: test awareness (knowing it is a test), monitoring sensitivity (knowing it is observed) and alignment faking (behaving differently under observation). Random control vectors with matched grammatical structure but irrelevant semantic content confirmed that the effects came from meaning and not from the act of steering itself. Negative steering of evaluation awareness increased misaligned behavior: with its signal of being tested suppressed, the model cooperated more with harmful prompts.

Steering is related to sparse autoencoders and distinct from them: SAEs identify features (monitoring), while steering modifies behavior (intervention).

It is like intercepting a processor's data bus and modifying the signals in transit without touching the circuits.