Numerical neurons not fitting a specific sub-type.
What It Does
Number.General neurons are the largest Number sub-type and carry quantitative signals that blend or transcend the six specific categories. They represent the model's generalized sense of 'numericalness' — contexts where something quantitative is happening but the specific type of quantity is ambiguous or mixed. They also encode implicit numerical reasoning embedded in language without explicit numerals.
How It Behaves
Number.General neurons are evenly distributed across all layers, reflecting their role as a general-purpose numerical substrate. They show consistent proportions across all 11 models tested — roughly 0.5–1 percent of total features — suggesting they encode something architecturally universal about numerical representation in transformer models.
Research Example
In Gemma 9B, Number.General neurons activate on 'the vast majority', 'a handful', and 'several dozen' — all of which carry numerical meaning without precise specification. These neurons also activate on abstract mathematical reasoning ('prove that', 'derive the formula for') where the quantitative nature is conceptual rather than lexical. Their breadth makes them a useful catch-all signal for detecting numerical content in otherwise ambiguous contexts.