Identity neurons not fitting a specific sub-type.
What It Does
Identity.General neurons activate on named entities that resist clean categorization into the six specific sub-types: abstract named concepts ('Newtonianism', 'the Renaissance'), unnamed but specific entities ('the man in the blue hat', 'the woman who invented the vaccine'), and entities that span multiple sub-types. They form the general entity-recognition substrate that underlies all more specific Identity sub-types.
How It Behaves
Identity.General is the largest Identity sub-type and shows even distribution across middle and late layers. It represents the model's general capacity to recognize and track specific entities through discourse, independent of their type. Models with better named entity recognition benchmarks consistently show more precisely specialized Identity neurons (higher proportions of Person, Place, and Organization vs. General) — specialization is a mark of entity-recognition quality.
Research Example
In Llama 3.1 8B, Identity.General neurons activate on references to abstract concepts treated as entities: 'Neoliberalism argued that markets...', 'Quantum mechanics predicts that...' These personified abstractions activate General identity neurons, not the more specific sub-types — the model recognizes that something entity-like is being tracked without classifying it into a standard category.