Who Gets a Token, and What Does It Carry? Unequal Name Support and Concept Access in Large Language Models
A study finds unequal name tokenization in LLMs predicts demographic gaps in task-relevant concept access.
Researchers show that matched personal names are often unequal LLM inputs because some are single tokens while others are split into subwords. Across nearly half a million first names and 12 LLM tokenizers, direct lexical access varies by model and by race- and gender-associated name metadata. Their NameTrace framework finds that this support predicts concept accessibility in fellowship, hiring, clinical assessment, and lending tasks across eight matched strata and multiple model families. Hidden-state interventions show the measured directions can change later constrained choices.
- Matched names often differ in tokenizer support.
- Study covers nearly 500,000 names and 12 tokenizers.
- NameTrace links lexical support to concept access.
- Gaps persist across eight race and gender strata.
- Hidden-state interventions shift later constrained choices.
From the original article
1 min read at huggingface.co
Names are personal identifiers, but they also carry social meaning and are widely used to evaluate how language models treat different people. Such evaluations typically assume that matched names are comparable model inputs. We show that this assumption often fails at the lexical interface: matched names are not necessarily matched inputs.
By Mir Tafseer Nayeem, Davood Rafiei · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.