Word
activation steering
Definition
n. nudging an AI model's internal signals during thought to make it behave better without retraining its weights.
n. a runtime alignment technique that modifies intermediate residual stream activations using calculated steering vectors to control model tone or truthfulness.
Examples
“Using activation steering made the AI assistant consistently polite without having to train it all over again.”
“Researchers demonstrated activation steering to suppress deceptive tendencies during complex multi-step reasoning tasks.”
“Representation engineering employs activation steering by computing contrastive direction vectors between honest and sycophantic model outputs.”
Examples
simple
“Using activation steering made the AI assistant consistently polite without having to train it all over again.”
contextual
“Researchers demonstrated activation steering to suppress deceptive tendencies during complex multi-step reasoning tasks.”
complex
“Representation engineering employs activation steering by computing contrastive direction vectors between honest and sycophantic model outputs.”
Real-World Examples
“Representation Engineering (aka Activation Steering/Engineering) is a new paradigm for understanding and controlling the behaviour of LLMs.” “Recent evidence suggests that some misalignment behaviors are encoded as linear structure in activation space, making it tractable via activation steering, which could be used as a lightweight runtime defense.” Etymology
Compound technical term coined in the 2020s by representation engineering researchers (Zou et al., Rimsky et al.), combining activation with vehicular steering.
Etymology adapted from Wiktionary, available under CC BY-SA 4.0.
Domains
Scan code
englishreference.com/q/activation-steering