Encodes human names and personal identities.
What It Does
Identity.Person neurons activate when the model encounters or generates references to specific human beings: proper names (both historical and contemporary), pronouns that have been linked to named individuals through coreference, and descriptive references that uniquely identify a person ('the inventor of the telephone', 'the 44th US president'). They distinguish person-entities from other entity types.
How It Behaves
Person neurons are distributed across middle and late layers, reflecting that person-identity requires contextual processing — 'Michael Jordan' the basketball player activates differently from 'Michael Jordan' the philosopher. The disambiguation happens in the middle layers, and the stable person-entity representation emerges in the later layers. Person neurons interact strongly with Time neurons (biographical dates), Space neurons (birthplaces, residences), and Role neurons (professions, titles).
Research Example
In Mistral 7B, Identity.Person neurons show measurably different firing signature for living vs. deceased public figures, reflecting the different distributional patterns in training data — contemporary figures appear in news and social media contexts, historical figures in biographical and educational contexts. This contextual difference means models can subtly confuse what a historical figure said vs. what they have been quoted as saying in secondary sources.