Encodes containment and hierarchical spatial structure.
What It Does
Space.Nesting neurons activate on containment relationships where one spatial or conceptual entity is inside another: physical containment ('the nucleus inside the cell', 'the file inside the folder'), logical containment ('the subcategory within the category'), and recursive nesting ('the loop within the loop', 'the story within the story'). They encode the spatial logic of things being inside other things.
How It Behaves
Space.Nesting neurons show an extreme middle-layer concentration — among the most layer-localized of any sub-type in our corpus. This concentration reflects the computational demands of parsing nested structure, which requires resolving the containment hierarchy (which thing is inside which?) before committing to a representation. They are particularly relevant for models handling hierarchical data structures, recursive algorithms, or nested legal/contractual language.
Research Example
In Llama 3.1 8B, Space.Nesting neurons activate on JSON nested object structures, HTML tag nesting, and natural language descriptions of hierarchical organization with strikingly similar firing signatures — suggesting the model has learned a domain-general representation of containment hierarchy that applies across code, markup, and natural language.