Encodes geographic locations and spatial identities.
What It Does
Identity.Place neurons activate on named geographic entities: countries ('France', 'Brazil'), cities ('Tokyo', 'Lagos'), geographic features ('the Sahara', 'the Pacific'), administrative regions, landmarks, and addresses. They encode place as a specific named location that can be placed on a map, distinct from Space neurons which encode relational spatial concepts like direction and position.
How It Behaves
Place neurons show roughly equal distribution across middle and late layers, with a slight late-layer peak for precise geographic localization. They co-activate strongly with Time neurons (historical events in specific places), Relation neurons (spatial relationships between places), and occasionally with Number neurons (population statistics, geographic measurements). Models trained on more multilingual or geographically diverse corpora show broader Place neuron firing across a wider range of location names.
Research Example
In Gemma 2B, Identity.Place neurons fire on 'Paris' differently depending on whether it appears in 'Paris, France' (geographic reference, full Place.Place activation) vs. 'Paris Hilton' (person name containing a place name, suppressed Place.Place activation). This disambiguation is imperfect in all models, contributing to occasional entity confusion in place-name contexts.