CSAIL FORUM with Jacob Andreas: Training Language Models to Talk About Themselves
bis
Massachusetts, Massachusetts
Title: Training Language Models to Talk About Themselves Abstract: Current language models (LMs) can converse knowledgeably and in depth about a wide range of topics. But these same models are surprisingly bad at talking about themselves: They struggle to describe their internal decision-making processes, judge the correctness of their own outputs, or answer user questions about their behavioral tendencies. I'll present some procedures we've recently developed for training LMs to introspect, and observations we've made about how introspection training shapes model behavior, and show that better introspection improves both predictability and user control.