Discriminates positive vs. negative sentiment and evaluative direction.
What It Does
Boolean.Polarity neurons encode the directional valence of a statement — whether the overall evaluation is moving in a positive or negative direction. Unlike Affirmation and Negation which operate at the token level, Polarity captures the semantic trajectory of phrases, sentences, and arguments. 'The product is impressive but overpriced' has conflicting Polarity signals that these neurons must balance.
How It Behaves
Polarity neurons activate strongly in early layers on explicit sentiment words, then again in later layers when the overall trajectory of an argument or sentence is being resolved. They interact heavily with Relation.Contrast neurons — the 'but', 'however', and 'although' constructions that flip polarity direction.
Research Example
In Mistral 7B, Boolean.Polarity neurons distinguish between 'unfortunately correct' and 'happily wrong' — the modifier has inverted the expected polarity of the adjacent word. This makes them critical for detecting sarcasm and negated sentiment patterns where surface-level tokens mislead simpler models.