Linear Representation Hypothesis
A central hypothesis in mechanistic interpretability research, positing that AI models encode high-level semantic concepts as linear directions in activation space. It forms the theoretical foundation for activation steering techniques.
Core Facts
Timeline (last 90 days)
The Linear Representation Hypothesis (LRH) assumes that models encode high-level concepts as linear directions in the activation space.
Linear steering relies on the linear representation hypothesis; when the concept distribution itself is curved, shifting activations along a fixed direction deviates from the true concept manifold, leading to control distortion or unnatural intermediate states.
A linear probe is a simple linear classifier trained on model activations, used to detect whether specific behaviors or concepts exist inside the model.
All Facts (3)
A linear probe is a simple linear classifier trained on model activations, used to detect whether specific behaviors or concepts exist inside the model.
75%UnverifiedThe Linear Representation Hypothesis (LRH) assumes that models encode high-level concepts as linear directions in the activation space.
50%UnverifiedLinear steering relies on the linear representation hypothesis; when the concept distribution itself is curved, shifting activations along a fixed direction deviates from the true concept manifold, leading to control distortion or unnatural intermediate states.
50%