AI
Who Gets a Token, and What Does It Carry? Unequal Name Support and Concept Access in Large Language Models
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 04:01:47 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 17:18:24 UTC.
Briefings
Research indicates that name support in large language models is not uniform across all names. The study reveals that matched names often receive unequal lexical access, where some names are accessed as single tokens while others are assembled from multiple subwords. This inequality is demographically structured, varying by race and gender-associated name metadata. The findings suggest that unequal lexical support at the input remains visible in task-relevant model computation, affecting concept accessibility across various real-world tasks like hiring and clinical assessment. The study introduces NameTrace as a framework to measure this phenomenon, showing that lexical comparability is a prerequisite for behavioral comparability.
Key points
- NameTrace is a model-native, fine-grained framework that measures unequal name-surface support by analyzing concept accessibility from model probabilities over task-specific adjective axes.
- Across nearly half a million first names and 12 LLM-associated tokenizers, direct lexical access is highly selective, model dependent, and uneven across race- and gender-associated name metadata.
Context and limitations
The study covers only first names and 12 specific LLM-associated tokenizers, limiting generalizability to other name types or models. Results are based on pre-behavioral lexical support measurements and do not account for post-processing interventions or long-term behavioral outcomes.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 04:01:47 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifjtqojeeebz4zpo3dmwglpspzq6vjqm2xm4t6sfri2zhev2rih2e