Definition
Plain language
A set of techniques for reading and nudging the internal states of an AI model to understand or change its behavior.
As stated in the literature
A research area focused on identifying and manipulating directions in a model's activation space — including activation steering and hidden-state grafting — to interpret or control behavior without retraining.
Why it matters: It offers a way to understand and steer a model's behavior by adjusting its internal states directly, without the cost of retraining it from scratch.
For example, a researcher might locate the internal pattern that lights up when a model is being deceptive and then turn that pattern up or down to change the behavior.
Heard on the show
“And there has been a quieter line of work asking whether that bottleneck is actually necessary — work on representation engineering, on activation steering, on grafting hidden states between models.”Episode 040 — Two Frozen Models Learn to Whisper: Coupling Through Hidden States