steerability
In plain wordsHow easily an AI model's behavior — its tone, style, persona or rules — can be adjusted by the people using or deploying it, for example through instructions, without retraining.
- 1
Language models
The degree to which a model's behavior can be directed at inference time by developers and users — through system promptssystem prompt, n.: Background instructions given to a chatbot before the conversation starts, setting its role, rules, tone and tools. Users usually never see it, but it shapes every reply., instructions, personas, examples, style or effort settings — and the reliability with which it maintains the requested behavior over a long interaction. High steerability lets one model serve many products and user preferences (verbosity, tone, format, values within policy) and is evaluated by measuring adherence to such specifications. It is deliberately bounded: a model should not be steerable into violating its provider's policies, so steerability must be balanced against robustness to jailbreaksjailbreak, n.: A trick for getting an AI chatbot to break its own rules — for example, to give dangerous instructions it would normally refuse — usually through cleverly worded requests. through an instruction hierarchyinstruction hierarchy, n.: A rule built into AI assistants about whose instructions win when they conflict: the developer's beat the user's, and the user's beat any instructions hidden in a web page or document..
The new model's steerability is better: it keeps the terse persona for the whole session.
- 2
Interpretability and generative models
The extent to which outputs can be controlled by directly manipulating internal representations — moving through a generative model's latent space or adding steering vectorssteering 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. to activations — without changing weights or prompts. See 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..
Origin
Ali Jahanian, Lucy Chai and Phillip Isola studied the 'steerability' of generative adversarial networks' latent spaces in 2019. OpenAI's GPT-4 announcement (March 2023) made steerability a headline property of language models, describing how developers could prescribe a model's style and task through the system message.
Usage note
Overlaps with 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., which usually names specific techniques, especially activation-based ones; steerability is the property those techniques exploit.
Related
- system promptsystem prompt, n.: Background instructions given to a chatbot before the conversation starts, setting its role, rules, tone and tools. Users usually never see it, but it shapes every reply.
- instruction followinginstruction following, n.: An AI's ability to do exactly what it is asked, including details such as length, format, language and things to avoid, rather than something merely related.
- 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.
- 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.
- role promptingrole prompting, n.: Telling an AI to take on a part — 'You are an experienced tax accountant' — so that its answers adopt that role's knowledge, vocabulary and tone.
- instruction hierarchyinstruction hierarchy, n.: A rule built into AI assistants about whose instructions win when they conflict: the developer's beat the user's, and the user's beat any instructions hidden in a web page or document.
- model specificationmodel specification, n.: A public document in which an AI company spells out how its models are supposed to behave — what they should do, refuse and prioritize, and whom to obey when instructions conflict.
Sources
- OpenAI (2023). GPT-4. OpenAI research announcement, 14 March 2023. openai.com/index/gpt-4-research
- Jahanian, A., Chai, L. & Isola, P. (2020). On the 'steerability' of generative adversarial networks. ICLR 2020. arxiv.org/abs/1907.07171
Mentioned in
instruction followinginstruction following, n.: An AI's ability to do exactly what it is asked, including details such as length, format, language and things to avoid, rather than something merely related., 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., system promptsystem prompt, n.: Background instructions given to a chatbot before the conversation starts, setting its role, rules, tone and tools. Users usually never see it, but it shapes every reply.