Encodes specific dates, years, and calendar references.
What It Does
Time.Date neurons activate on specific calendar references: explicit dates ('June 14, 1946'), years ('in 1989', '2026'), months and days, historical date references ('the 18th century', 'the 1960s'), and relative calendar expressions ('last Tuesday', 'next March'). They encode a point or period on the calendar axis.
How It Behaves
Date neurons show an early-to-middle layer distribution, with the model parsing temporal references early and then integrating them with event context through the middle layers. They co-activate strongly with Identity neurons (linking dates to the entities whose histories they describe) and Time.Event neurons (linking dates to the events that occurred on them). Date neuron suppression is one of the primary causes of temporal hallucination — models generating plausible-sounding but incorrect dates.
Research Example
In Mistral 7B, Time.Date neurons fire strongly on '1989' in the context of the Berlin Wall but show attenuated firing when '1989' is embedded in a mathematical expression ('data from 1989 samples'). The model's Date neurons are sensitive to calendar-reference context vs. numerical context, though this disambiguation is imperfect and a source of temporal errors when numbers are misread as dates.