This neuron activates on tokens in definitional or factual-answer contexts—particularly on proper names (e.g. “Tokyo,” “Japan,” “graphene”) and the linking words around them (e.g. “known,” “as,” “form of”) when the model is stating a definition or specific factual information.
