Feature steering changes a neural network’s behaviour by deliberately increasing, decreasing or replacing an internal representation feature during inference. It is a causal experiment and, in some applications, a control mechanism: identify a direction associated with a variable, intervene on that direction, and observe how downstream computation and output change.
Feature steering sits at the meeting point of representation learning, mechanistic interpretability and model control. A feature can come from a supervised probe, a contrast vector, principal direction, sparse autoencoder, dictionary latent or another representation-discovery method. The scientific question is not whether the direction has an attractive label. It is whether interventions on it change the predicted behaviour selectively, at realistic magnitudes, across relevant contexts.
This owner covers steering vectors, activation addition, coefficient editing, sparse-autoencoder latents, dose-response curves, layer and position choice, normalization, feature entanglement, off-target effects, nonlinear saturation, context dependence, ablation, natural-use evidence, adversarial robustness, monitoring and the boundary between causal leverage and mechanistic explanation.
Quick Read
- Feature direction: a vector or latent associated with a represented variable.
- Steering: intervene on that feature during model execution.
- Coefficient α: controls intervention strength.
- Dose-response: measures behaviour over several α values.
- Selectivity: target behaviour changes more than unrelated behaviours.
- Ablation: reduce or remove a naturally active feature.
- Natural scale: compare intervention magnitude with activations the model normally experiences.
- Entanglement: one direction can affect several correlated variables.
- Off-manifold: strong edits can create unnatural hidden states.
- Steerability: causal leverage, not proof of unique feature identity.
1. The basic steering equation
Let h be a hidden activation and d a feature direction. A simple additive intervention is:
h' = h + αd.
Positive α strengthens the direction. Negative α suppresses or reverses it. The downstream network receives h′ instead of h.
If behaviour changes systematically with α, the direction has causal leverage at that intervention site. The meaning of that leverage depends on how d was obtained and what else d changes.
2. A contrast vector is the simplest steering direction
Collect activations for examples with property A and examples without it. Subtract the mean vectors. The difference can define a direction associated with the property.
This can work surprisingly well, but the direction inherits every systematic difference between the two groups. If “formal writing” examples are also longer and more technical, the steering vector can mix formality, length and topic.
3. Probe directions are predictive, not automatically native
A linear probe can learn a weight vector that separates a labelled variable. Steering along that vector tests whether the decodable variable can causally influence behaviour.
But a probe can combine many signals the model never naturally reads with that exact vector. Strong steering therefore establishes an intervention route, not necessarily a native computational feature.
4. Sparse-autoencoder features provide an unsupervised steering vocabulary
The Sparse Autoencoders owner explains learned overcomplete dictionaries. A latent coefficient can be increased, clamped, suppressed or replaced before decoding back into the model activation.
This can produce semantically crisp interventions, but SAE features are non-orthogonal, approximate and conditional on the dictionary. Steering one latent can alter neighbouring represented variables.
5. Layer choice changes what steering means
An early-layer feature can encode lexical or local sensory structure. A middle-layer feature can encode contextual state. A late-layer feature can be close to output decisions.
The same semantic label at different layers can therefore produce different interventions. Steering “French” early may change token-level language cues; steering it late may change output language choice or answer content.
6. Token position matters
Applying a direction at every token is a different intervention from applying it only at the final token, subject token or generated positions. Repeated steering can accumulate and create stronger distribution shift.
State the position policy as part of the method. “We steered layer 12” is incomplete for a sequence model.
7. Dose-response is the minimum causal curve
Do not test only α=0 and one dramatic α. Evaluate a range around the natural activation scale.
A useful feature often shows a graded region: small edits produce small target changes, moderate edits strengthen the effect, and very large edits eventually saturate or destabilize the model.
A response that appears only at enormous α is weaker evidence of natural mechanistic relevance than one that changes behaviour inside the distribution of naturally observed feature magnitudes.
8. Selectivity is more important than spectacle
A steering direction that makes every output incoherent can also change the target metric dramatically. That is not a clean control feature.
Measure target change against off-target changes: fluency, unrelated task accuracy, perplexity, neighbouring feature activations, style, safety behaviour and general loss. The best steering feature changes what it claims to change while preserving what it claims not to change.
9. Feature entanglement creates collateral effects
Representation directions need not be orthogonal. If d_target overlaps d_other, adding d_target changes projection onto d_other too.
This is one geometric reason steering can produce correlated side effects. Feature-level selectivity should be measured, not inferred from a semantic label.
10. Steering can reveal causal leverage without natural use
A direction can change behaviour because downstream weights are sensitive to it even if the model rarely generates that direction naturally. This is analogous to pressing an undocumented control input.
Natural-use claims require observational evidence that the feature activates during the behaviour plus interventions showing that suppressing it changes the behaviour under ordinary execution.
11. Ablation complements positive steering
If a feature is active naturally, reduce its coefficient or remove its decoded contribution. Does the target behaviour weaken?
Positive steering asks whether the feature can push behaviour. Ablation asks whether the model relies on it in the tested context. Agreement strengthens the mechanistic case.
12. Off-manifold steering is a real failure mode
Large vector additions can move h into regions never produced by the model’s upstream computation. Downstream behaviour can become unpredictable or globally unstable.
Compare intervention norms with natural activation statistics, use dose-response curves and inspect model loss. A spectacular result at 100 standard deviations is not equivalent to a precise causal intervention.
13. Context can reverse a steering effect
A feature direction can be read differently depending on other active features. “Politeness” may affect a social dialogue differently from code. “Danger” may trigger refusal only when instruction-following state is also active.
Test across context strata rather than averaging away interactions. A robust steering claim specifies where the direction works and where it does not.
14. Steering and activation patching are cousins
Activation Patching replaces a state with one observed in another run. Feature steering synthesizes a controlled change along a selected direction.
Patching is often more naturalistic because the donor activation occurred naturally. Steering is often more interpretable because one direction can be varied continuously. Combining both gives stronger evidence.
15. Steering and causal tracing
A feature can be steered upstream and its downstream consequences traced through later features or components. This tests not only whether the feature changes output but how its influence propagates.
The Causal Tracing owner owns that multi-step mechanism job.
16. Steering for control is different from steering for science
A production control system cares whether an intervention reliably changes behaviour with acceptable side effects. A scientific study additionally cares whether the direction corresponds to a natural internal variable and explains ordinary computation.
A useful control knob can be scientifically artificial. A natural feature can be too entangled to make a safe control knob.
17. Safety applications need adversarial evaluation
Steering can suppress or amplify safety-relevant internal features, but adaptive prompts can route behaviour through alternative representations. A monitor or controller built on one direction needs tests against paraphrase, multilingual inputs, long context, obfuscation and model updates.
Feature steering is one layer of control, not a guarantee that the underlying capability disappeared.
18. Failure signatures
- One-strength steering: no dose-response curve.
- Wrecking-ball effect: target changes because everything degrades.
- Label leap: a semantic feature name substitutes for causal specificity.
- Natural-use leap: steerability is mistaken for ordinary model reliance.
- Layer ambiguity: same direction label is assumed equivalent across depth.
- Context blindness: effects are averaged across regimes where the feature behaves differently.
- Off-manifold magnitude: intervention far exceeds natural activation range.
- Version blindness: the same steering vector is reused after fine-tuning without revalidation.
19. Evaluation blueprint
- Define the target variable and output metric.
- Specify how the direction is learned.
- Use held-out examples to validate feature association.
- Choose layer and token-position policy.
- Measure natural feature magnitude distribution.
- Run positive and negative α dose-response.
- Measure off-target behaviour and model loss.
- Compare with random and magnitude-matched directions.
- Run ablation where the feature naturally activates.
- Test across contexts, languages and prompt forms.
- Trace downstream consequences for mechanistic claims.
- Revalidate after model updates.
20. World Return
A representation direction is not interesting because it has a name. It becomes interesting when changing it changes the model in a predictable way.
The mature question is not “can we make the model talk more about X?” It is “does this intervention operate at natural scale, selectively, reproducibly, through the downstream mechanism we predicted, without silently changing ten other things?”
Feature steering turns representation geometry into an experiment. Precision comes from measuring what moves with the feature—and what does not.