New work on AI Theory of Mind: We ask: *How capable are models at inducing belief states without using conversation?* This is important to track as it unlocks many AI agent use-cases, but also potential harms. In our task, recent generations of models exceeded human performance 🧵