AI roleplay personas reportedly retain an assistant-associated core

An analysis shared by a user examines internal features in Gemma and Llama using sparse autoencoders. The author reports that roleplay personas retain an assistant-associated core and progressively differentiate from it across model layers, while story characters lack that core. They also describe features associated with what they call “Immersive Simulation Mode,” distinguishing immersive roleplay and story generation from the default assistant. In certain cases, the author says, those features activate in the default assistant context too, making its behavior “bizarre.”