Who Gets a Token, and What Does It Carry? Unequal Name Support and Concept Access in Large Language Models

Who Gets a Token, and What Does It Carry? Unequal Name Support and Concept Access in Large Language Models. By Mir Tafseer Nayeem and one other, from the University of Alberta. Published on the twenty-eighth of September. Its code repository is public and real: ten source files.

Names are personal identifiers, but they also carry social meaning, and they are widely used to test how language models treat different people. Those tests usually swap one name for a matched one and assume the model receives two comparable inputs. The authors show that this assumption often fails before the model has done anything at all, right at the lexical interface. Matched names are not necessarily matched inputs. Some names get direct access as a single token, while others have to be assembled from several subwords. The paper calls this unequal name-surface support, and its claim is that the inequality is not a harmless quirk of vocabulary. It is demographically structured at the input, and it stays visible inside the computation the model performs on a task.

The first layer of evidence is a census of vocabularies. Across nearly half a million first names and twelve tokenizers associated with large language models, direct lexical access is highly selective, it depends on the model, and it is uneven across the race and gender metadata attached to the names.

To ask whether that unevenness reaches the representations, the authors introduce NameTrace, a framework they describe as model-native, fine-grained and pre-behavioral. It measures how accessible a concept is from the model's own probabilities over adjective axes specific to each task, with continuous weights aligned to that task. The comparison is carefully matched: atomic names against short fragmented names, inside the same stratum of race or ethnicity and gender. Within those strata, support predicts systematic differences in concept accessibility across four settings: fellowship, hiring, clinical assessment and lending. Hidden-state interventions add one more step. Pushing along the measured task directions shifts later constrained choices, so those directions have downstream leverage rather than being a curiosity of measurement.

How far does this reach? The differences persist across all eight matched strata. They extend across model families, and they transfer to names the analysis had not seen. What the abstract establishes is a gap present at the input and in task-relevant internal computation, with measured leverage over later choices. NameTrace is framed as pre-behavioral, so the claim is about the inputs and representations that behavioral tests take for granted, not a new count of biased final outputs.

The argument lands on a principle that reorders the usual audit. Before asking whether a model treats two people alike, check whether it was ever handed two equivalent names. Behavioral comparability, in the authors' words, begins with lexical comparability, and NameTrace is offered as the instrument that makes the second one measurable.