Steering Vector
Definition
A control vector alters internal activations of a neural network to influence its behavior at output.
A control vector is a direction in the activation space of a neural network that is added or amplified during inference to influence behavior.
Simply explained
Like a controller inside the model: it amplifies a specific pattern and can thereby change, for example, the tone or content of the response.
Why it matters in practice
The term helps to select, develop, evaluate AI systems in a well-founded manner and realistically assess their technical risks.
Example from business
A research team alters an activation direction and systematically measures effects on tone, content, and side effects.
Limits and pitfalls
- An observed effect is not automatically causal or stable; interventions can unpredictably change capabilities.
Origin & history
Research on Activation Addition and Representation Engineering made such interventions known from 2023 as interpretability and control methods.
Use in mid-sized companies
- Possible application: A research team alters an activation direction and systematically measures effects on tone, content, and side effects.
Frequently asked questions
What does Control Vector mean?
A control vector is a direction in the activation space of a neural network that is added or amplified during inference to influence behavior. What should you watch out for? An observed effect is not automatically causal or stable; interventions can unpredictably change capabilities.
Sources
- Activation Addition: Steering Language Models Without Optimization · Turner et al., Activation Addition: Steering Language Models Without Optimization
Source checked on 10.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on: