Representation Engineering
An AI safety research paper published in 2023 by Andy Zou and colleagues that systematizes methods for controlling model behavior by manipulating internal representation vectors. It provides a theoretical foundation for activation steering and model unlocking experiments.
Timeline (last 90 days)
Mainstream activation steering methods in academia include CAA (Contrast Activation Addition) based on contrastive activation differences, linear probe-based steering approaches, and various variants under the Representation Engineering framework.
The Representation Engineering (RepE) paper was published in 2023 by Dan Hendrycks' team, demonstrating that high-level concepts in large language models are encoded as linear directions in the internal representation space.
Representation Engineering was systematically proposed by Zou et al. in a 2023 study.
All Facts (3)
Representation Engineering was systematically proposed by Zou et al. in a 2023 study.
60%UnverifiedMainstream activation steering methods in academia include CAA (Contrast Activation Addition) based on contrastive activation differences, linear probe-based steering approaches, and various variants under the Representation Engineering framework.
50%UnverifiedThe Representation Engineering (RepE) paper was published in 2023 by Dan Hendrycks' team, demonstrating that high-level concepts in large language models are encoded as linear directions in the internal representation space.
50%