feature steering
In plain wordsSteering an AI model by turning up or down specific concepts that interpretability tools have found inside it — for example, making a model obsessed with a bridge by maxing out its 'bridge' concept.
Activation steeringactivation steering, n.: Changing an AI model's behavior by adjusting the internal numbers it computes while it runs — like turning a hidden dial — instead of rewording the prompt or retraining the model. performed on interpretable featuresfeature (interpretability), n.: A concept or property that a neural network represents internally — such as 'is French', 'mentions a bridge' or 'this code has a bug' — usually corresponding to a direction in its internal activity. learned by a sparse autoencodersparse autoencoder, n.: A small add-on network trained to break an AI model's tangled internal signals into a much larger set of simpler pieces, most of which can be labeled with an understandable concept. (SAE), transcodertranscoder, n.: A stand-in for one part of an AI model that does the same job using a large set of simple, understandable pieces, so researchers can see what that part computes. or similar dictionary, rather than on directions found by contrasting examples. The usual operation clamps a feature's activation to a fixed value — high to amplify the concept, zero or negative to suppress it — and substitutes the edited SAE reconstruction plus the original reconstruction error, so that only the targeted feature changes; a simpler variant adds the feature's decoder vector as a steering vectorsteering vector, n.: A direction inside an AI model's internal activity that, when added while the model runs, nudges its behavior — say, more upbeat, more honest or fixated on a topic — without retraining it.. Because features carry human-readable labels and are found without labeled data, feature steering offers fine-grained, nameable control and a causal check on feature interpretations. Evaluations found off-target effects and a narrow range of useful strengths before output quality collapses.
Clamping the 'inner conflict' feature to five times its maximum made every answer agonize over trade-offs.
Origin
Clamping dictionary features to steer a production-scale model was demonstrated in Anthropic's 'Scaling Monosemanticity' study (Adly Templeton and colleagues, May 2024), whose Golden Gate ClaudeGolden Gate Claude, prop. n.: A one-day public demo in 2024 in which Anthropic turned up a single internal concept inside its Claude chatbot — the Golden Gate Bridge — so it steered almost every conversation back to the bridge. demo popularized the idea. Anthropic's report 'Evaluating feature steering' (October 25, 2024), on Claude 3 Sonnet, and Goodfire's public Ember feature-steering API (December 2024) helped fix the term.
Usage note
Feature steering inherits the limits of the dictionary: feature splittingfeature splitting, n.: What happens when a concept-finding tool is given more room: a broad concept it found before breaks into several narrower ones, much as 'animal' might split into 'dog', 'cat' and 'bird'., feature absorptionfeature absorption, n.: A flaw in concept-finding tools where a general concept goes missing in some cases because a more specific concept swallows it — like a 'starts with S' detector staying silent on the word 'snake'. and reconstruction error constrain how cleanly one concept can be controlled. The AxBench benchmark (2025) found SAE features uncompetitive for both steering and concept detection, trailing prompting and simple difference-of-means directions; later work found that selecting features by their effect on outputs narrows the gap.
Related
- sparse autoencodersparse autoencoder, n.: A small add-on network trained to break an AI model's tangled internal signals into a much larger set of simpler pieces, most of which can be labeled with an understandable concept.
- Golden Gate ClaudeGolden Gate Claude, prop. n.: A one-day public demo in 2024 in which Anthropic turned up a single internal concept inside its Claude chatbot — the Golden Gate Bridge — so it steered almost every conversation back to the bridge.
- feature (interpretability)feature (interpretability), n.: A concept or property that a neural network represents internally — such as 'is French', 'mentions a bridge' or 'this code has a bug' — usually corresponding to a direction in its internal activity.
- activation steeringactivation steering, n.: Changing an AI model's behavior by adjusting the internal numbers it computes while it runs — like turning a hidden dial — instead of rewording the prompt or retraining the model.
- steering vectorsteering vector, n.: A direction inside an AI model's internal activity that, when added while the model runs, nudges its behavior — say, more upbeat, more honest or fixated on a topic — without retraining it.
- transcodertranscoder, n.: A stand-in for one part of an AI model that does the same job using a large set of simple, understandable pieces, so researchers can see what that part computes.
- contrastive activation additioncontrastive activation addition, n.: A way to make a steering signal more reliable by averaging the difference in a model's internal state across hundreds of example answers that do and don't show a behavior, such as flattery.
Mentioned in
activation steeringactivation steering, n.: Changing an AI model's behavior by adjusting the internal numbers it computes while it runs — like turning a hidden dial — instead of rewording the prompt or retraining the model., feature (interpretability)feature (interpretability), n.: A concept or property that a neural network represents internally — such as 'is French', 'mentions a bridge' or 'this code has a bug' — usually corresponding to a direction in its internal activity., Golden Gate ClaudeGolden Gate Claude, prop. n.: A one-day public demo in 2024 in which Anthropic turned up a single internal concept inside its Claude chatbot — the Golden Gate Bridge — so it steered almost every conversation back to the bridge., steeringsteering, n.: Guiding what an AI model does — its topic, tone or behavior — while it runs, by changing its instructions, its internal signals or how its next word is chosen, rather than by retraining it., GoodfireGoodfire, prop. n.: A San Francisco company, founded in 2024, that builds tools to look inside AI models, understand what they have learned and adjust how they behave., NeuronpediaNeuronpedia, prop. n.: A free website where anyone can browse, search and test the internal concepts that researchers have found inside AI models, with examples of the text that switches each one on.