Encodes logical, mathematical, and relational operators.
What It Does
Symbol.Operator neurons activate on operator symbols that specify relationships or transformations: mathematical operators (+, -, ×, ÷, =, ≠, >, <), logical operators (AND, OR, NOT, →, ↔), comparison operators (==, !=, >=), and assignment operators. They encode the operational relationship between operands, not the operands themselves.
How It Behaves
Operator neurons show broad layer distribution and are particularly active in models with strong mathematical and code training. They interact heavily with Number neurons (mathematical operators) and Symbol.Code neurons (programming operators). A notable property: Operator neurons encode operator precedence context — '2 + 3 × 4' activates different Operator neuron patterns than '(2 + 3) × 4', reflecting the model's representation of operator hierarchy.
Research Example
In Llama 3.1 8B, Symbol.Operator neurons activate on natural language equivalents of operators: 'greater than', 'not equal to', 'divided by' produce similar firing signatures to their symbolic counterparts '>', '≠', '÷'. This cross-modal encoding is one mechanism by which language models can translate between natural-language and symbolic mathematical notation.