Definition

Representation engineering (RepE) (Zou et al., 2023) works top-down on the model’s internal representations: reading (linear probes recover directions for honesty, hallucination, power, morality) and control (adding or ablating these directions steers behavior).

Many concepts admit approximately linear directions in the residual stream: honesty, deception, sycophancy, hallucination, power-seeking, fact recall, refusal. Used defensively it can disrupt harmful outputs; used adversarially it is abliteration.

Appears in