Encodes movement direction — up, down, left, right, toward, away.
What It Does
Space.Direction neurons activate on directional movement references: compass directions (north, southeast), relative directions (up, down, left, right, forward, backward), movement toward or away ('approaching', 'retreating', 'converging'), and implied directionality ('rising', 'falling', 'expanding'). They encode motion trajectory in space.
How It Behaves
Direction neurons are the smallest Space sub-type by count but show a distinctive early-layer concentration, suggesting that directional parsing is among the most primitive spatial processing operations — the brain (and the model) establishes movement direction before integrating it with complex contextual information. They co-activate with Time.Sequence neurons for processes that evolve directionally over time ('the market moved upward throughout Q3').
Research Example
In Mistral 7B, Space.Direction neurons activate on 'the stock price rose' and 'temperatures are falling' — both metaphorical uses of directional language where no literal movement in physical space is occurring. The model has generalized directional space neurons to encode directional change across abstract domains, which is why models can reason consistently about 'upward trends' and 'downward pressure' in economic contexts.