Traces of Meaning and Binding in LLMs

Is there a difference between a concept and a word for an LLM? Large language models are a remarkable form of intelligence that learn about the world only through text. Do they represent concepts, or just patterns of words? Using causal probes as a tool, I will walk through what we know about concept formation in LLMs, tracing the progression from tokens to language-independent concept representations, to knowledge association, function representation, and finally to theory of mind. We consider the classic variable-binding puzzle and the surprising solution LLMs have learned. We will discuss the new research opportunities opened by our ability to conduct experiments directly on artificial neural networks.