Encodes individual characters and alphabetical sequences.
What It Does
Symbol.Letter neurons activate at the character level — individual letters, digits, and special characters as discrete symbols. They fire most strongly on contexts where individual characters carry meaning independent of the word they form: acronyms (DNA, NATO), spelled-out words ('N-A-T-O'), single-letter identifiers in mathematics ('x', 'y', 'n'), and cases where a model must reason about character sequences explicitly.
How It Behaves
Letter neurons appear with higher concentration in early layers, alongside Lexical and Morphology neurons, as part of the fundamental token processing pipeline. They are more active in models used for tasks involving character-level manipulation (spelling correction, acronym expansion, anagram solving) and show consistent response signatures across all alphabets and writing systems represented in training data.
Research Example
In Qwen3 1.7B, Symbol.Letter neurons are strongly active when the model processes 'DNA' in 'DNA sequencing' differently from 'DNA' as an acronym expanded as 'Deoxyribonucleic Acid' — the same three characters produce different Letter neuron patterns depending on whether they are being processed as a word-token (whole acronym) or as individual letter-symbols being spelled out.