Encodes part-of and membership relationships.
What It Does
Relation.PartWhole neurons activate on part-whole and membership relationships: meronymy ('the engine is part of the car'), set membership ('France is a member of the EU'), hierarchical inclusion ('mammals are a subclass of animals'), and structural composition ('the novel consists of three parts'). They encode the relationship of something being a component of or belonging to a larger whole.
How It Behaves
PartWhole neurons show a middle-layer concentration, where compositional structure is being integrated with semantic content. They interact with Space.Nesting neurons (parts are spatially nested in wholes), Number.Fraction neurons (fractions are parts of wholes), and Relation.Conjunction neurons (wholes are composed by conjoining parts). PartWhole neurons are particularly important for models handling technical documentation, biological classification, and organizational structure.
Research Example
In Gemma 2B, Relation.PartWhole neurons activate on 'the left ventricle is a chamber of the heart' and 'chapter 3 is part of the first section' with similar patterns — the same PartWhole relational structure applies regardless of whether the domain is anatomy or document organization. This domain-general representation of part-whole relationships is why models can transfer reasoning about composition across very different domains.