Relational neurons not fitting a specific sub-type.
What It Does
Relation.General neurons carry relational signals that blend or transcend the six specific categories. They activate on generic relational language ('related to', 'associated with', 'connected to'), complex multi-type relationships (causal-contrastive, analogical-partial), and implicit structural relationships embedded in discourse without explicit relational markers.
How It Behaves
Relation.General neurons are the largest Relation sub-type and show even distribution across all layers. They grow proportionally with model scale more reliably than any other element — the larger the model, the higher the proportion of Relation.General neurons. This scaling property is consistent with the hypothesis that larger models build richer associative networks, with Relation.General neurons as the substrate for cross-domain conceptual linking.
Research Example
In cross-model comparison, the proportion of Relation.General neurons increases monotonically from GPT-2 Small (12.9 percent) to Llama 3.1 8B (27 percent) to even larger models. This scaling relationship is one of the strongest quantitative patterns in our dataset, and supports the view that relational reasoning capacity scales with model size in a measurable way at the neuron level.