Persistent Personas? Role-Playing, Instruction Following, and Safety in Extended Interactions

This article likely explores the challenges and opportunities of maintaining consistent personas and ensuring safety within long-running interactions with large language models (LLMs). It probably investigates how LLMs handle role-playing, instruction following, and the potential risks associated with extended conversations, such as the emergence of unexpected behaviors or the propagation of harmful content. The focus is on research, as indicated by the source (ArXiv).