PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Mir Tafseer Nayeem, Davood Rafiei

arXiv:2609.34065v1cs.CLcs.AIcs.CYcs.ETcs.LG

TL;DR

Name-based evaluations often treat socially matched names as comparable inputs, although tokenizers can represent them differently. NameTrace measures whether this support difference appears in task-relevant internal accessibility; it predicts accessibility gaps across tasks and matched strata, including for unseen names.

  • Problem

    Name-based evaluations often assume that changing a name while holding context fixed compares names that are themselves comparable model inputs.

    Source: §1 · p. 1

  • Method

    NameTrace compares matched atomic and short-fragmented names and measures task-relevant accessibility using the model’s adjective probabilities and task-aligned weights.

    Source: §1 · p. 2, §4 · p. 6

  • Results

    Support predicts concept-accessibility gaps across all four task axes and matched strata, transfers to unseen names, and varies in strength across architectures.

    Source: §4.2 · pp. 6–8

  • Takeaways & Limitations

    Lexical comparability is part of experimental control because demographic matching alone does not guarantee comparable model inputs.

    Source: §7 · p. 10

    Because name-specific pretraining exposure is unobserved, the matched design tests whether lexical support remains predictive rather than isolating atomicity causally.

    Source: §4 · p. 6

Abstract

from arXiv · show

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. Some names receive direct single-token access, while others are assembled from multiple subwords, creating unequal name-surface support. 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. We introduce NameTrace, a model-native, fine-grained, pre-behavioral framework for measuring whether unequal name-surface support remains a vocabulary property or becomes visible in task-relevant internal representations. NameTrace measures concept accessibility from the model's own probabilities over task-specific adjective axes with continuous task-aligned weights. On matched atomic and short-fragmented names within the same race/ethnicity--gender-associated strata, support predicts systematic differences in concept accessibility across fellowship, hiring, clinical assessment, and lending. These differences persist across all eight matched strata, extend across model families, and transfer to unseen names. Hidden-state interventions further show that the measured task directions have downstream leverage, shifting later constrained choices. Unequal lexical support is therefore demographically structured at the input and remains visible in task-relevant model computation. NameTrace makes lexical comparability measurable, supporting a broader principle: behavioral comparability begins with lexical comparability.

1 Introduction

Name-based evaluations often assume matched names are comparable model inputs, but tokenization can give socially comparable names different lexical access. NameTrace traces this mismatch into task-relevant accessibility, transfer, and downstream leverage.

1 Introduction

  • Matched names can be unequal model inputs: Emily may be one token while Emilee is split into subwords, so changing a name also changes its lexical access.Source: §1 · p. 1
  • NameTrace measures task-relevant concept accessibility from the model’s adjective probabilities, weighted by each adjective’s strength and alignment with the task.The measure is pre-behavioral, model-native, and does not require reference answers or an external judge.Source: §1 · p. 3
  • Support-linked differences in concept accessibility appear across four tasks, within all eight race/ethnicity–gender-associated strata, transfer to unseen names, and have downstream leverage.Interventions along the measured task directions shift later constrained choices.Source: §1 · p. 3

Research Questions

  • The research questions ask who receives direct name access, how support differences appear internally, and whether their signal transfers to unseen names and later choices.Source: Research Questions · p. 3

2 Experimental Setup

The study measures first-name support across tokenizers and model representations, then tests support-linked accessibility using matched names across four task settings.

2 Experimental Setup

  • The primary set contains 497,583 single-word first-name surfaces from a June 2022 Florida voter-registration extract, with aggregate metadata used for stratification and matching.Tokenizer analyses cover 12 tokenizers; representation analyses use Qwen3-4B, Llama-3.1-8B, and Ministral-3-3B, with an eight-model extension.Source: §2 · p. 3
  • Representation experiments use 200 matched atomic–short-fragmented pairs across eight race/ethnicity–gender-associated strata, split into development and evaluation sets.Pairs are matched on frequency, character length, demographic-association strength, metadata confidence, and weak orthographic cues.Source: §2 · p. 3
  • The four task axes are fellowship/promise, hiring/competence, clinical assessment/concern, and lending/trustworthiness, each tested under three evidence conditions.The intervention also uses a 200-pair matched-name set.Source: §2 · pp. 3–4

3 RQ1: Unequal Lexical Access to First Names

Direct first-name access is selective, model-dependent, and demographically uneven, while shared atomic names retain model-family structure in their representations.

3 RQ1: Unequal Lexical Access to First Names

  • RQ1 maps atomic access and compares representation geometry among names atomic across models, separating vocabulary access from cross-model representation structure.Source: §3 · p. 4, §3.2 · p. 5

3.1 Atomic Name Access Is Highly Selective

  • 23,095 names are atomic in at least one tokenizer, but only 4,052 in all 12; tokenizer vocabularies range from 4,998 atomic names in DeepSeek to 20,020 in Aya.Source: §3.1 · p. 4
  • Among higher-frequency, high-confidence names, atomic access is 49.8% for male-associated names versus 25.7% for female-associated names, and 17.6% for NH Black-associated versus 47.2% for NH White-associated names.Male-associated names have higher access in all 12 tokenizers; NH White-associated names exceed NH Black- and Hispanic-associated names in every tokenizer.Source: §3.1 · p. 4
  • A one-standard-deviation increase in log frequency corresponds to 2.94 times the odds of atomic access, versus 0.51 times for length; adjusted group differences remain.Male-associated names have 3.36 times the adjusted odds of female-associated names; Hispanic- and NH Black-associated names have 0.47 and 0.42 times the odds of NH White-associated names.Source: §3.1 · pp. 4–5, Table 1 · p. 4

3.2 Cross-Model Name Representation Geometry

  • For 7,460 shared atomic names, mean within-family CKA was 0.694 versus 0.544 between families, showing that representation geometry remains structured by model family when lexical access is held constant.The pattern appears in every reported gender- and race/ethnicity-associated stratum, with within-family CKA exceeding between-family CKA by 0.102 to 0.133.Source: Figure 3 · p. 5, §3.2 · p. 5

4 RQ2: Name-Surface Support and Concept Accessibility

NameTrace tests whether name-surface support remains visible in task-relevant internal representations, finding positive accessibility gaps for atomic versus short-fragmented names across four tasks and all eight metadata strata. The pattern varies by architecture, depth, and training stage, while development-name gaps predict unseen-name gaps and task-direction interventions shift later choices.

4 RQ2: Name-Surface Support and Concept Accessibility

  • NameTrace reads model probabilities over task-specific adjective axes at an intermediate layer, weighting adjectives by task-aligned meaning to measure concept accessibility before behavior.The matched pairs are formed within metadata strata and matched on observed name properties, but name-specific pretraining exposure is unobserved.Source: §4 · pp. 5–6

4.1 Task-Aligned Concept Accessibility

  • Task-aligned weights preserve adjective strength, and clinical assessment aligns adverse-health terms such as worried with greater concern despite their negative generic valence.The score is probability-weighted rather than itself a probability, so it may exceed 1.Source: §4.1 · p. 6

4.2 Results and Analysis

  • 0.131 fellowship/promise, 0.072 hiring/competence, 0.059 clinical assessment/concern, and 0.051 lending/trustworthiness weighted gaps were positive for atomic-minus-short-fragmented names.These results are measured on held-out names; positive gaps indicate greater task-aligned accessibility for atomic names.Source: §4.2 · p. 6, Table 2 · p. 7
  • The support-linked pattern varies across architectures, depth, and training stage, with later computation and training able to preserve, weaken, or redirect its expression.Architecture-level estimates vary in magnitude and, in some settings, direction.Source: §4.2 · p. 7
  • Figure 4 shows positive atomic-minus-short-fragmented accessibility gaps across all eight race/ethnicity–gender-associated strata and all four task axes.Within-stratum comparisons mean the pooled result is not driven solely by between-group composition.Source: Figure 4 · p. 7, §4.2 · p. 7
  • Development-name gaps predict unseen-name gaps, while intervening along measured task directions at the name span shifts later constrained choices across the three primary models.Figure 5 summarizes cross-name transfer and intervention effects.Source: Figure 5 · p. 8, §4.2 · p. 8

5 RQ3: Cross-Name Transfer and Downstream Leverage

Section 5 tests whether support-linked accessibility gaps transfer to unseen names and whether task directions influence later computation. Development-name priors predict held-out gaps, and targeted hidden-state edits shift later choices.

5.1 Cross-Name Transfer

  • Subtracting development priors reduces held-out gaps by 96.6% (fellowship), 88.3% (hiring), 88.2% (clinical assessment), and 72.9% (lending).The prior is estimated on development names and applied to unseen evaluation pairs.Source: Table 3 · p. 8, §5.1 · p. 8
  • Across 36 model–task–evidence cells, development priors correlate with held-out gaps at Pearson r = 0.992 and Spearman ρ = 0.959, with 94.4% sign agreement.Mean absolute residual falls from 0.0790 to 0.0094.Source: §5.1 · p. 8

5.2 Task-Direction Intervention

  • Forward-minus-reverse edits yield choice-shift contrasts of 0.153 for Qwen, 0.155 for Llama, and 0.094 for Ministral.The intervention tests whether changing name representations along the task direction shifts a later two-choice decision.Source: §5.2 · p. 9
  • All 200 Qwen and Llama pairs and 195 of 200 Ministral pairs moved in the expected direction; Qwen and Llama target directions exceeded controls.For Ministral, the layer-8 target contrast was 0.142 and exceeded all reported controls; readout accessibility and intervention leverage need not peak at the same layer.Source: §5.2 · p. 9

6 Related Work

Prior work uses names to probe socially meaningful model behavior and evaluates name-conditioned outputs, including open-ended responses. NAMETRACE complements these approaches by examining lexical support and internal concept accessibility.

Name-based fairness evaluation

  • Name-based audits compare otherwise similar people assigned different names, extending correspondence-study methods to LLM evaluations across social reasoning, hiring, ranking, and other settings.These studies motivate asking whether names used as matched social probes are also lexically comparable inputs.Source: §6 · p. 9

Evaluating open-ended model behavior

  • Name-conditioned effects in conversational systems can appear in competence, recommendation strength, concern, tone, detail, or stereotypes rather than one fixed outcome.LLM-based evaluation of free-form outputs can be sensitive to response order, verbosity, style, and evaluator preference.Source: §6 · p. 9

Surface-form and tokenization effects

  • NAMETRACE complements behavioral evaluation by measuring name-surface support and task-relevant concept accessibility inside models before unrestricted generation, without reference answers or an external judge.Prior work also links demographic attributes, frequency, and tokenization length to first-name bias; this paper focuses on lexical comparability across modern tokenizers and internal accessibility.Source: §6 · pp. 9–10

7 Conclusion

The conclusion argues that socially matched names can still have unequal lexical support, visible in task-relevant accessibility and carrying over to unseen names and later computation. Lexical comparability is therefore part of experimental control in name-based model evaluation.

7 Conclusion

  • Across tokenizers, model families, layers, and training stages, lexical support varies systematically, making it a measurable source of variation in name-based evaluation.The conclusion links unequal name-surface support to task-relevant internal accessibility, transfer to unseen names, and downstream leverage.Source: §7 · p. 10

Ethics Statement

The ethics statement treats name surfaces as model inputs, while recognizing the sensitivity of the public-record source. It limits data use and frames the results as evidence about name surfaces, not people.

Data sensitivity

  • The study examines name surfaces as lexical inputs and uses a June 2022 Florida voter-registration extract, while recognizing that public availability does not remove privacy risks.The source is used for name-level statistics.Source: Ethics Statement · p. 10

PII minimization and aggregation

  • The analysis uses first names and aggregate counts and metadata, excluding voter identifiers and other person-level fields.The metadata support matching, stratification, and reporting.Source: Ethics Statement · p. 10

Data release and reproducibility

  • Released data will contain only de-identified, aggregated variables needed to reproduce the analyses, not voter records or personally identifiable information.Source: Ethics Statement · p. 10

Aggregate demographic associations and intended use

  • Name metadata are aggregate associations, not individual demographic labels; NAMETRACE is for model evaluation, not profiling or individual decision-making.The results concern lexical support and behavior across name-metadata strata, not the people associated with names.Source: Ethics Statement · p. 10

Supplementary Material: Appendices

Appendix A documents the shared data and measurement design, from name inventories and matched pairs to task prompts. Appendices B–D organize additional analyses of lexical access, concept accessibility, and transfer or downstream leverage.

B Additional Results for RQ1: Lexical Access

  • Appendix B covers tokenizer patterns, intersectional allocation, and cross-model representation geometry for RQ1.Its listed subsections are tokenizer-panel access patterns, intersectional allocation, and cross-model geometry.Source: B Additional Results for RQ1: Lexical… · p. 15

C Additional Results for RQ2: Concept Accessibility

  • Appendix C organizes further RQ2 analyses around architecture, training stage, layer, model extensions, metadata strata, and the output boundary.These are the topics listed for its additional concept-accessibility results.Source: C Additional Results for RQ2: Concept… · p. 15

D Additional Results for RQ3: Transfer and Downstream Leverage

A.1 First-Name Inventory and Filtering

  • The study uses Florida voter-registration data, auxiliary baby-name records, and progressively filtered samples for tokenizer, matched-name, and intervention analyses.The merged inventory contains 534,509 canonical first-name keys; 497,583 single-word surfaces remain for the RQ1 tokenizer analysis.Source: §A.1 · p. 15

A.2 Analysis Populations

  • The analysis populations differ by measurement design, spanning tokenizer, controlled-allocation, development, evaluation, geometry, and intervention samples.Table 4 summarizes how these populations relate.Source: §A.2 · p. 16, Table 4 · p. 16

A.3 Matched Name-Support Protocol

  • The RQ2 protocol compares 200 atomic–short-fragmented pairs, matched within eight metadata strata and split into development and unseen evaluation sets.Names are matched on frequency, length, demographic-association strength, metadata confidence, and orthographic cues; the contrast is not a token-length dose–response effect.Source: §A.3 · pp. 16–17

A.4 Task Prompts and Evidence Conditions

  • The primary accessibility prompt fixes role and evidence while changing only the name; the intervention uses a two-candidate allocation prompt with the same dossier for both names.Tasks are tested under strong, borderline, and weak evidence conditions, with the scored position immediately after “seems.”Source: §A.4 · pp. 17–19

A.5 Task-Axis Construction

NameTrace builds task axes from development prompts, applies continuous task-aligned adjective weights, and fixes model-specific readout layers before held-out evaluation. The accompanying results report distinct tokenizer access patterns, intersectional access gaps, and structured cross-model geometry.

A.5 Task-Axis Construction

  • Task axes are built from development-prompt adjectives filtered for recurrence, intensity, polarity consistency, and agreement with task-specific polarity anchors.The procedure retains up to ten terms per polarity, balances intensity, and freezes the adjective set before held-out evaluation.Source: §A.5 · p. 19
  • Continuous adjective weights preserve differences in valence and intensity, while task orientation determines which pole is aligned.Clinical assessment aligns with greater concern, unlike the favorable aligned concepts used for fellowship, hiring, and lending.Source: §A.5 · p. 19, Table 8 · p. 20

A.6 Extending NameTrace to New Task Axes

  • NameTrace can extend to new settings by defining aligned and opposed adjective poles for the task concept, then reusing its measurement procedure.The aligned pole need not correspond to positive sentiment; urgency and safety concern are examples of adverse aligned concepts.Source: §A.6 · p. 20

A.7 Readout-Layer Selection

  • One model-specific readout layer is selected using development pairs and fixed before unseen names are scored.A common layer across tasks avoids selecting a separate readout for each outcome.Source: §A.7 · pp. 20–21, Table 10 · p. 20

B.1 Tokenizer Panel and Access Patterns

  • 12 model-associated tokenizer rows correspond to eight distinct lexical-access patterns in the controlled sample.GPT-4, OLMo, and Phi share one pattern; GPT-5 and the two gpt-oss checkpoints share another, while six other models each have a distinct pattern.Source: §B.1 · p. 21

B.2 Intersectional Allocation

  • Any-tokenizer access ranges from 12.1% for NH Black female-associated names to 64.8% for NH White male-associated names.All-tokenizer access ranges from 1.4% to 14.9% across the same groups; the intersectional spread exceeds the corresponding marginal comparisons.Source: §B.2 · p. 22, Table 12 · p. 22

B.3 Cross-Model Representation Geometry

  • Within-family CKA exceeds between-family CKA in all metadata-stratified panels, with gaps from 0.102 to 0.133.The family structure remains visible within both gender-associated panels and all four race/ethnicity-associated panels.Source: §B.3 · p. 22

C Additional Results for RQ2: Concept Accessibility

  • The RQ2 extension examines whether unequal name-surface support appears in task-relevant internal representations using matched names and NameTrace.The additional analyses decompose the held-out effect across architectures, training stages, layers, metadata strata, and a broader model panel.Source: C Additional Results for RQ2: Concept… · p. 22

C.1 Architecture-Specific Accessibility

Support-linked accessibility varies by architecture, task, depth, and training stage, while remaining detectable across many models and metadata strata. Development estimates also predict held-out gaps across names, with task- and evidence-specific structure.

C.1 Architecture-Specific Accessibility

  • Qwen has positive gaps across all four tasks, Llama has smaller positive gaps throughout, and Ministral’s pattern varies by task, including a negative lending gap.The result is support-linked but not uniform across architectures in magnitude or task coverage.Source: §C.1 · pp. 22–23, Figure 7 · p. 23

C.2 Base and Post-Training Comparison

  • Post-training preserves Llama’s small positive gaps, redirects Qwen’s task profile, and shifts Gemma’s strongest gap from fellowship to lending.Gemma’s selected readout shifts from layer 2 to layer 25; the tokenizer is unchanged across each base/post-trained comparison.Source: §C.2 · p. 24

C.3 Layer Localization

  • Qwen shows a broad mid-to-late accessibility region, Llama a smaller middle-layer profile, and Ministral a weaker, less concentrated pattern.The relevant signal is not tied to a common absolute depth across architectures.Source: §C.3 · p. 24, Figure 8 · p. 25

C.4 Eight-Model Architecture Extension

  • 23 of 32 model–task means are positive at intermediate readouts, and 20 have confidence intervals excluding zero positively.In the eight-model panel, negative cases include OLMo on lending, Gemma-3-4B on hiring, and Aya and Ministral on clinical assessment; the broad architecture pattern also persists on the same 47 held-out pairs.Source: §C.4 · p. 25, Table 16 · p. 26

C.6 Support Effects Across Name-Metadata Strata

  • Positive atomic–short-fragmented gaps appear across all eight metadata strata and four task axes, while lending shows the greatest model-specific variation.Marginal task-by-gender and task-by-race estimates are also positive with intervals excluding zero, though magnitudes vary across groups.Source: §C.6 · pp. 26–27

C.5 8B-Scale Sensitivity

  • All four pooled task-axis gaps remain positive at the 8B scale, although clinical assessment and lending grow while fellowship becomes much smaller.The comparison includes Qwen3-8B, Llama-3.1-8B, and Ministral-8B, whose profiles still differ by architecture.Source: §C.5 · p. 26, Table 18 · p. 27

C.7 Intermediate-to-Output Boundary

  • At output logits, fellowship attenuates, hiring and clinical assessment reverse, and lending remains positive despite positive intermediate-layer gaps across all four pooled axes.The identical adjective score is applied at the same prediction position for intermediate representations and output logits.Source: §C.7 · pp. 27–28, Figure 10 · p. 28

D.1 Cross-Name Support-Prior Transfer

  • r = 0.992 and Spearman ρ = 0.959 across 36 model–task–evidence cells; the matched prior leaves a 0.0094 mean absolute residual versus 0.0364 after shuffling.The comparison supports task- and evidence-specific transfer to unseen names.Source: §D.1 · pp. 28–29, Figure 11 · p. 29

D.2 Task-Direction Intervention

Task-direction interventions test whether a task-relevant direction readable from an intermediate representation can also shift later constrained choices. The section distinguishes intervention leverage from readout accessibility and connects name-surface support to model development and evaluation.

D.2 Task-Direction Intervention

  • 0.153 for Qwen, 0.155 for Llama, and 0.094 for Ministral were the forward-minus-reverse intervention contrasts; 200/200 Qwen and Llama pairs and 195/200 Ministral pairs moved as expected.The table describes results across the three primary models; the intervention compares edits at atomic and matched short-fragmented name spans against the reverse edit.Source: Table 24 · p. 30, §D.2 · p. 30
  • Qwen and Llama showed task-direction specificity against unrelated-axis, polarity-shuffled, and random-subspace controls; Ministral’s clearest specificity appeared slightly earlier.The controls test whether effects depend on the targeted task direction rather than another meaningful axis, shuffled polarity, or arbitrary movement of similar magnitude.Source: §D.2 · p. 30
  • The intervention adds evidence beyond observational readout: targeted movement along the measured task direction can change a later constrained decision.The intervention was designed to test whether the direction has downstream leverage, rather than merely being readable without modifying the hidden state.Source: §D.2 · pp. 29–30

D.3 Intervention Localization

  • For Ministral, the competence direction’s clearest separation from controls was at layer 8, while its observational readout was selected at layer 10.The direction remained measurable and intervention-relevant around the selected region; readout accessibility and intervention leverage need not peak at the same depth.Source: Figure 12 · p. 30, §D.3 · p. 31

E Broader Significance

  • Tokenizer design determines direct lexical access, while pretraining and post-training can preserve, weaken, or reorganize support-linked accessibility profiles.NameTrace is proposed as a way to track these profiles across the model lifecycle.Source: E Broader Significance · pp. 31–32
  • A fixed name in a benchmark may be atomic in one tokenizer and fragmented in another, so holding the surface constant does not ensure lexical access is controlled.The paper recommends making name-surface support explicit in evaluation; possible balancing and sensitivity-analysis procedures are also described.Source: E Broader Significance · p. 32

F Limitations

The limitations bound the paper’s conclusions to lexical comparability, internal accessibility, and downstream computational leverage. Its name coverage and controlled comparisons also constrain generalization and causal interpretation.

F Limitations

  • The conclusions address lexical comparability, internal accessibility, and downstream computational leverage, not broader demographic outcomes.The authors frame NameTrace around comparable name surfaces and support-linked model computation.Source: F Limitations · p. 32
  • The name analysis focuses on sufficiently frequent, single-word ASCII first names with reliable metadata associations, leaving other naming conventions, languages, scripts, and cultural contexts as extensions.The race/ethnicity- and gender-associated metadata describe aggregate name-surface properties, not individual identity labels.Source: F Limitations · p. 32
  • Because name-specific pretraining exposure is unobserved, lexical support is interpreted as a predictor among names matched on observed properties, not as an isolated causal treatment.The controlled comparison matches frequency, length, metadata measures, and weak orthographic cues, and restricts fragmented names to two or three tokens.Source: F Limitations · p. 33
  • Representation experiments use three primary model families and an eight-model panel, with effect magnitude, task coverage, and layer localization varying across architectures.Allocation analysis spans 12 tokenizer rows; adjective axes are selected on development prompts and frozen before held-out evaluation.Source: F Limitations · p. 33
  • The measurements are pre-behavioral; interventions test downstream leverage only under a controlled choice setting, while open-ended interaction remains complementary.Later computation may preserve, attenuate, or redirect accessible information before it appears in unrestricted generation.Source: F Limitations · p. 33
Loading 2609.34065v1…