Articles
-
"Many Are My Names": The Anatomy of the Assistant and Its Personas via Sparse Autoencoders
The paper explores how representations of the Assistant and personas adopted in roleplay evolve through model layers. It shows that personas have the Assistant's core, from which they differentiate with layer depth, while story characters lack it. Also, the paper identifies features which separate the default Assistant from roleplay or story writing: they make personas much more detailed and make the narrative vivid. However, sometimes, these features can activate even in the normal conversations with the Assistant, turning its behavior bizarre!