Activation Steering
A model control and interpretability technique that injects an activation offset, equal to a direction vector multiplied by a coefficient, into the residual stream at an intermediate layer of the forward pass in order to guide the model's internal representations and generation behavior.
Timeline (last 90 days)
When the input itself is normal and does not involve behavior that needs to be suppressed, forced activation steering intervention drags down the model's normal performance, causing performance loss.
Activation steering methods were first systematized by institutions such as Anthropic and have since been widely applied to alignment tasks such as reducing bias and suppressing harmful outputs.
All Facts (2)
When the input itself is normal and does not involve behavior that needs to be suppressed, forced activation steering intervention drags down the model's normal performance, causing performance loss.
50%UnverifiedActivation steering methods were first systematized by institutions such as Anthropic and have since been widely applied to alignment tasks such as reducing bias and suppressing harmful outputs.
50%