Glossary Entry

Activation Steering

Modifying a model's behaviour at inference time by adding, amplifying, or suppressing concept directions in its internal activations rather than changing its weights or prompt.

Interpretability LLMs

Also called: steering vector, steering vectors, feature steering

Seed source: Scaling Monosemanticity (Templeton et al., 2024)

If a concept is represented as a direction in activation space, you can push on it: add a scaled concept vector into the residual stream during the forward pass and the model’s behaviour shifts toward (or, with a negative scale, away from) that concept. No retraining, no prompt changes, just an edit to the intermediate computation. The most famous demonstration is Golden Gate Claude, where clamping a sparse-autoencoder feature made Claude obsess over the bridge in every conversation.

Beyond the party trick, steering matters as evidence: it upgrades a direction from “correlated with a concept” to “causally moves behaviour,” and it previews practical applications like dialling persona traits or suppressing unwanted behaviours. Its mirror image is targeted ablation, projecting a direction out to see which capabilities depend on it.