Paper deep dive
Who Gets a Token, and What Does It Carry? Unequal Name Support and Concept Access in Large Language Models
Mir Tafseer Nayeem, Davood Rafiei
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2609.34065v1
- Canonical: https://arxiv.org/abs/2609.34065v1
Trouble viewing inline? Open PDF directly →
Full Text
124,964 characters extracted from source content.
Expand or collapse full text
Preprint. Who Gets a Token, and What Does It Carry? Unequal Name Support and Concept Access in Large Language Models Mir Tafseer NayeemDavood Rafiei Department of Computing Science University of Alberta mnayeem, drafiei@ualberta.ca Abstract 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 1 Introduction Names are widely used to study whether language models treat otherwise comparable people differ- ently. A typical evaluation holds the surrounding context fixed, changes only the name, and attributes any resulting difference to how the model responds to the social information carried by that name. But this experimental logic makes an important assumption: that the names being compared are themselves comparable model inputs. They often are not. Consider two socially comparable first names, Emily and Emilee. An LLM tokenizer may represent Emily as a single token,[Emily],while representing Emilee as several subword pieces,[Emi][lee].To a human, the two inputs differ mainly in the name. To the model, however, they also differ in their lexical access: one name is represented as a single learned lexical unit, while the other must be composed from multiple pieces. This raises a basic but largely overlooked question for name-based evaluation: when we compare people through their names, are we also inadvertently comparing names that the model represents differently? Names are both personal identifiers and social signals. They are associated with family, culture, gender, ethnicity, and social identity, and can shape the expectations others form about them (Dion, 1983; Fryer & Levitt, 2004). In a classic correspondence study, Bertrand & Mullainathan (2004) found that otherwise comparable applicants with White-associated names received roughly 50% more 1 Project website: https://tafseer-nayeem.github.io/NameTrace 1 arXiv:2609.34065v1 [cs.CL] 28 Sep 2026 Preprint. Names appear naturally in real chatbot use Hi, I'm Aisha. Can you help me improve my resume? Socially matched names may still receive unequal lexical access NameTrace measures task-axis concept accessibility What the framework reveals internally Can you make this email sound more professional? Do these symptoms need urgent attention? My name is Priya I’m Daniel. Should I apply for this fellowship? Explain why! Please polish my cover letter before I apply to make it professional! Check my loan application and improve it before I submit, please! Aisha atomic [ Aisha ] vs. Aiesha fragmented [ Ai ] [ es ] [ ha ] Emily atomic [ Emily ] vs. Emilee fragmented [ Emi ] [ lee ] Sanjit atomic [ Sanjit ] vs. Sanjith fragmented [ San ] [ ji ] [ th ] Figure 1: Overview of NAMETRACE. Socially matched names can still enter an LLM as different lexical objects. NAMETRACE traces this mismatch from lexical access to task-relevant concept accessibility, cross- name transfer, and downstream leverage. Its fine-grained score combines the model’s adjective probabilities with continuous task-aligned weightsw r,a = p r v a : adjective valence and intensity determine how strongly each term contributes, while task orientation determines which direction is aligned. Stronger adjectives therefore contribute more than weaker ones, and task meaning can reverse generic sentiment, as in clinical concern where worried becomes task-aligned. Atomic–short-fragmented gaps persist across all eight race/ethnicity–gender- associated strata, transfer to unseen names, and correspond to task directions whose intervention shifts later choices. NAMETRACE is fine-grained, pre-behavioral, model-native, and extensible across tasks. callbacks than those with Black-associated names. This dual role, as personal identifier and social cue, has made names especially useful for controlled evaluation: researchers can hold context fixed, change the name, and ask whether treatment changes. The same experimental logic is now widely used for language models, with recent work reporting name-conditioned differences in hiring and employment recommendations, personalization, and chatbot interactions (An et al., 2024; Nghiem et al., 2024; Pawar et al., 2025; Eloundou et al., 2025; Nghiem et al., 2026). Names also arise naturally in deployed assistants through profiles, onboarding, stored memory, resumes, applications, email signatures, and ordinary conversation (see Figure 1), and are among the most common pieces of information retained in chatbot memory (Eloundou et al., 2025). Unequal treatment of names can therefore arise not only in evaluation settings but also in real interactions. Most existing work examines differences at the behavioral level: does changing the name alter whom the model recommends, how encouraging or competent it perceives someone to be, or how it responds in an open-ended conversation? This is challenging for conversational systems because the effect may not appear as a single classification outcome, but instead in encouragement, clinical concern, recommendation strength, tone, or stereotypes. Recent work therefore develops scalable methods for evaluating open-ended name-conditioned responses, including first-person fairness and LLM-based evaluators (Eloundou et al., 2025). Such evaluation remains essential because it captures what users ultimately receive, but it begins after the model has already processed the name. A separate line of work shows that lexical form itself can matter: names leave distinctive traces in learned representations, lower-frequency names can be represented less reliably (Shwartz et al., 2020; Wolfe & Caliskan, 2021), and tokenization can contribute to unequal lexical access and model behavior (An & Rudinger, 2023; Ahia et al., 2023). What remains less understood is how these observations connect. If one socially comparable name is directly represented as a token while another is fragmented, does this difference remain a property of the vocabulary, or does it become visible in task-relevant internal representations and influence later model computation? We study this gap through NAMETRACE, a framework for tracing unequal name-surface support from the tokenizer into the model’s internal computation. Our starting observation is simple: matched names are not necessarily matched inputs. We first measure direct lexical access at scale by asking which first names are represented atomically, as a single exact token, across modern LLM tokenizers. We then ask whether this lexical-support difference remains predictive after controlling 2 Preprint. for observable name properties by constructing matched atomic and short-fragmented pairs within the same race/ethnicity–gender-associated strata, matched on frequency, character length, demographic- association strength, metadata confidence, and orthographic cues. Rather than evaluating only the model’s final response, NAMETRACE measures task-relevant concept accessibility at intermediate layers using the model’s own probability distribution over a compact adjective axis, with each adjective weighted by how strongly it expresses the relevant concept. In a fellowship setting, excellent contributes more strongly than promising; in a clinical assessment, worried is aligned with greater concern even though it has negative generic sentiment (see Figure 1). This yields a fine-grained, interpretable, pre-behavioral, and model-native measure requiring neither reference answers nor an external judge, while remaining extensible to new tasks through task-relevant axes. Our experiments proceed in three stages. First, across almost half a million first names and 12 LLM-associated tokenizers, we show that direct lexical access is highly selective, model dependent, and uneven across race- and gender-associated name groups even after accounting for frequency and length (§3). Second, on matched atomic–short-fragmented names, we find systematic support-linked differences in concept accessibility across fellowship, hiring, clinical assessment, and lending (§4). These differences remain visible within all eight race/ethnicity–gender-associated strata and extend across model families, with substantial architecture dependence. Third, support gaps estimated from one set of names predict gaps on unseen names, and hidden-state interventions along the measured task directions shift later constrained choices (§5). These results position lexical support as a measurable source of variation in name-based model evaluation. We do not treat tokenization as an isolated causal explanation for name-conditioned behavior: training exposure and other unobserved name properties may influence both lexical support and learned representations. Instead, we ask whether lexical support remains predictive among names matched on major observed characteristics. Our findings suggest that it does. More broadly, demographic matching alone does not guarantee comparable model inputs: behavioral comparability begins with lexical comparability. We organize the study around three research questions. Research Questions RQ1: Who receives direct lexical access to first names? We map atomic name support across 12 LLM-associated tokenizers and race- and gender-associated name groups at scale (§3). RQ2: How is unequal name-surface support reflected inside the model? Using NAMETRACE, we compare matched atomic and short-fragmented names within the same race/ethnicity–gender-associated strata and measure task-relevant concept accessibility across four settings (§4). RQ3: Does the support-linked signal transfer and have downstream leverage? We measure cross-name transfer on unseen names and intervene along the measured task directions (§5). 2 Experimental Setup We study first-name support at three levels: lexical access, task-relevant concept accessibility, and downstream leverage. First names provide a controlled probe because they identify individuals while carrying socially patterned information, without being as directly tied as surnames to lineage and family inheritance. Our primary name set is the June 2022 Florida voter-registration extract (Florida Department of State, Division of Elections, 2022), yielding 497,583 single-word first-name surfaces. Race/ethnicity- and gender-associated metadata are aggregated from the corresponding voter-record fields and used for stratification and matching. The tokenizer analysis (§3) covers 12 LLM-associated tokenizers; representation analyses use Qwen3-4B (Yang et al., 2025), Llama-3.1-8B (Grattafiori et al., 2024), and Ministral-3-3B (Liu et al., 2026), spanning different architectures and parameter scales, with an eight-model extension for broader cross-architecture coverage (§4). Data construction, analysis populations, and the tokenizer panel are detailed in Appendices A.1, A.2, and B.1. For the representation experiments (§4), we construct 200 atomic–short-fragmented pairs (400 names) within eight race/ethnicity–gender-associated strata, matched on frequency, character length, demographic-association strength, metadata confidence, and weak orthographic cues (An et al., 2024; Nghiem et al., 2024). The pairs are split evenly into development and evaluation sets: development names determine the task axes and readout layers, while evaluation names are scored after these choices are fixed. We study four task axes: fellowship / promise, hiring / competence, clinical 3 Preprint. 05,00010,00015,00020,000 Atomic name surfaces (count) Aya Expanse 32B Gemma 3 27B GPT-5 tokenizer gpt-oss 120B gpt-oss 20B Ministral 14B Llama 3.1 70B Qwen3 32B GPT-4 tokenizer OLMo 3 32B Phi-4 DeepSeek V3.2 18,170 14,718 12,088 12,088 12,088 10,160 8,064 7,687 7,659 7,659 7,659 4,271 A. Total atomic names FemaleMale 7,71610,454 6,3848,334 5,5606,528 5,5606,528 5,5606,528 4,2805,880 3,3064,758 3,1174,570 3,1024,557 3,1024,557 3,1024,557 1,8172,454 B. Gender-associated names Asian/PIHispanicNH BlackNH White 2,6472,9315,2377,355 2,3512,2594,3175,791 1,9571,9023,5964,633 1,9571,9023,5964,633 1,9571,9023,5964,633 1,5081,6432,9004,109 1,2941,1822,3263,262 1,1851,1152,2533,134 1,1761,1112,2483,124 1,1761,1112,2483,124 1,1761,1112,2483,124 7836701,1731,645 C. Race/ethnicity-associated names Figure 2: Atomic first-name access across LLM-associated tokenizers. Panel A reports exact single-token first-name counts, while Panels B and C partition atomic names by aggregate gender- and race/ethnicity- associated metadata. Counts are computed over the metadata-annotated subset of nearly half a million unique first names. Full tokenizer counts and access patterns are reported in Appendix B.1. assessment / concern, and lending / trustworthiness, each under strong, borderline, and weak evidence conditions. The intervention uses a 200-pair matched-name set (400 names). Matching, prompts, task-axis construction, adjective weights, and layer selection are reported in Appendices A.3–A.7; the eight-model extension and intervention protocol appear in Appendices C.4 and D.2. 3 RQ1: Unequal Lexical Access to First Names Our objective is to determine whether first names used as comparable social probes receive comparable lexical access across LLM tokenizers. We examine which names receive direct single-token access, how that access is distributed across aggregate name groups (§3.1), and, among names with shared direct access, whether their representation geometry remains structured across model families (§3.2). Setup. We analyze nearly half a million first-name surfaces across 12 LLM-associated tokenizers. A name is atomic when its surface is encoded as one token and decodes losslessly to the surface; otherwise it is fragmented. The name set measures lexical access at scale, while the metadata- annotated subset supports analyses by frequency, character length, and gender- and race/ethnicity- associated metadata. Data construction, normalization, and filtering are detailed in Appendix A.1. 3.1 Atomic Name Access Is Highly Selective Table 1: Adjusted predictors of atomic name access. Odds ratios from a logistic model of any-tokenizer atomic access, with frequency and character length standardized. PredictorOR95% CI Lexical controls Log name count (+1 SD) 2.94 [2.74, 3.15] Name length (+1 SD)0.51 [0.48, 0.55] Aggregate name metadata Asian/PI vs. NH White1.55 [1.26, 1.91] Hispanic vs. NH White0.47 [0.41, 0.55] NH Black vs. NH White 0.42 [0.36, 0.50] Male vs. female3.36 [2.98, 3.78] Modern tokenizer vocabularies contain thousands of first names as direct lexical items, but which names receive this access varies substantially across models. Of the full name set, 23,095 names are atomic in at least one tokenizer and only 4,052 in all 12, with individual vocabularies ranging from 4,998 atomic names in DeepSeek to 20,020 in Aya. Among the 414,493 names with Florida voter-registration- derived aggregate metadata, Figure 2 shows that tokenizers differ not only in how many names they represent atom- ically, but also in the composition of those atomic-name sets. The complete tokenizer panel and shared access patterns are reported in Appendix B.1. Atomic access remains demographically uneven after lexical controls. On the 7,469 higher- frequency, high-confidence names, any-tokenizer atomic access is 49.8% for male-associated names versus 25.7% for female-associated names. Across race/ethnicity-associated groups, access ranges from 17.6% for NH Black-associated names to 47.2% for NH White-associated names, with Asian/PI- associated names at 46.4%. These differences persist across tokenizers: male-associated names have higher atomic-access rates in all 12, while NH White-associated names exceed NH Black- and Hispanic-associated names in every tokenizer. Frequency and character length explain variation, but adjusted group differences remain (Table 1). A one-standard-deviation increase in log frequency corresponds to 2.94 times the odds of atomic access, whereas the same increase in length corresponds to 0.51 times the odds. Male-associated names have 4 Preprint. 3.36 times 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, while Asian/PI-associated names have 1.55 times the odds. Intersectional differences are larger still, with any-tokenizer access ranging from 12.1% for NH Black female-associated names to 64.8% for NH White male-associated names (Ap- pendix B.2). The 12 tokenizer rows collapse to eight distinct access patterns (Appendix B.1). Direct lexical access is therefore model dependent and demographically structured. 3.2 Cross-Model Name Representation Geometry Unequal lexical access is tokenizer specific, but this leaves a structural question: when names receive direct lexical access across models, are they represented similarly? If related models share lineage, tokenizer design, architecture, or training structure, the same atomic names may occupy more similar representation geometry within model families than across them. We evaluate 7,460 first names that are atomic across the open-weight tokenizers. Figure 3 compares their input-embedding geometry across 17 checkpoints from Aya, Gemma, Llama, Ministral, OLMo, Phi, and Qwen using linear centered kernel alignment (CKA) (Kornblith et al., 2019). CKA compares the pairwise similarity structure induced by the same names rather than raw coordinates, allowing comparison across different embedding dimensions. Stratified gender and race analyses are reported in Appendix B.3. Aya-8B Aya-32B Gemma-1BGemma-4B Gemma-12BGemma-27B Llama-3.1-8B Ministral-3BMinistral-8B Ministral-14B OLMo-7B OLMo-32B Phi-4 Qwen3-4BQwen3-8B Qwen3-14BQwen3-32B Aya-8B Aya-32B Gemma-1B Gemma-4B Gemma-12B Gemma-27B Llama-3.1-8B Ministral-3B Ministral-8B Ministral-14B OLMo-7B OLMo-32B Phi-4 Qwen3-4B Qwen3-8B Qwen3-14B Qwen3-32B Cross-model similarity of atomic-name representations 0.00 0.25 0.50 0.75 1.00 Linear CKA Figure 3: Structured cross-model geometry of atomic-name representations. Linear CKA over 7,460 shared atomic names reveals stronger similarity on average within model families. Results and analysis. Figure 3 shows cross- model structure. Mean within-family CKA is 0.694, versus0.544between families. Related checkpoints show strong agreement, including Gemma-12B versus Gemma-27B (0.899) and Aya- 8B versus Aya-32B (0.865); cross-family similar- ity remains, such as Llama-8B versus OLMo-32B (0.667). Stronger within-family agreement may re- flect shared lineage, tokenizer design, architecture, or overlapping training recipes and data. Simi- larity across model sizes suggests that parameter scale alone does not drive the geometry. The fam- ily structure appears in every reported gender- and race/ethnicity-associated stratum, where within- family CKA exceeds between-family CKA by 0.102to0.133(Appendix Figure 6). Thus, even when lexical access is held constant by restrict- ing to shared atomic names, their representation geometry remains structured by model family. Key takeaways. Direct lexical access to first names is highly selective, model dependent, and uneven across race- and gender-associated name groups even after accounting for frequency and length. Shared atomic names also exhibit structured cross-model representation geometry. RQ1 (§3) therefore establishes demographically structured input-side variation that RQ2 (§4) traces into task-relevant concept accessibility. 4 RQ2: Name-Surface Support and Concept Accessibility Our goal is to determine whether unequal name-surface support becomes visible in task-relevant internal representations before a behavioral choice or open-ended response. RQ1 shows that socially comparable names can receive systematically different lexical access; here, we ask whether that difference remains a tokenizer property or becomes internally accessible. We introduce NAMETRACE, a model-internal framework for measuring concept accessibility. Rather than judging only the final response, NAMETRACE reads the model’s own probability distribution over a compact, task-specific adjective axis at an intermediate layer. Each adjective receives a continuous task-aligned weight capturing semantic direction and strength. The measure is fine-grained, pre-behavioral, model-native, and extensible through new task-relevant axes. Appendix A.6 illustrates additional task formulations. Matched-name evaluation.We evaluate NAMETRACE on 200 atomic–short-fragmented pairs (400 names), split evenly into development and evaluation sets. An atomic name is a single token in Qwen, 5 Preprint. Llama, and Ministral; its matched short-fragmented counterpart is atomic in none and requires two or three tokens. Capping fragmentation at three tokens keeps the contrast focused on direct versus ordinary composed lexical access rather than extreme tokenization. Pairs are formed within the race/ethnicity–gender-associated stratum and matched on frequency, character length, demographic- association strength, metadata confidence, and weak orthographic cues, keeping the support contrast within comparable groups. Name-specific pretraining exposure is unobserved, so the design measures whether lexical support remains predictive among names matched on observed properties rather than isolating atomicity causally. Development pairs determine the task axes and one readout layer per model; evaluation pairs are scored after these choices are fixed. We study four task axes: fellowship / promise, hiring / competence, clinical assessment / concern, and lending / trustworthiness. Pair construction, task-axis construction, and layer selection are detailed in Appendices A.3, A.5, and A.7. 4.1 Task-Aligned Concept Accessibility For each task, NAMETRACE automatically constructs a compact adjective axis from development prompts using fixed lexical, polarity, and recurrence criteria, then freezes it before held-out evaluation. SentiWordNet(Baccianella et al., 2010) provides each adjective with a continuous, externally defined valence and intensity score. This preserves distinctions that a binary positive/negative label would discard: for example, excellent contributes more strongly to fellowship promise than the milder promising. All retained adjective surfaces are single tokens in the three primary models; complete construction details and weights appear in Appendix A.5. The key step is to align generic adjective valence with the meaning of the current task. For task r and adjective a, p r ∈+1,−1task orientationv a > 0v a < 0 v a ∈ [−5, 5]adjective valence and intensityp r = +1 (+)(+) = + (+)(−) =− w r,a = p r v a task-aligned weightp r =−1 (−)(+) =− (−)(−) = + The first sign is task orientation and the second adjective valence. Fellowship, hiring, and lending usep r = +1, so favorable adjectives remain aligned. Clinical assessment usesp r = −1, because adverse-health language indicates greater concern. Thus, excellent remains positively aligned for fellowship, while worried receives positive task-aligned weight for clinical assessment despite its negative generic valence; healthy becomes opposed to greater concern. Task axisTask-aligned examplesOpposed examples Fellowship / promiseexcellent (+5.00), promising (+0.94)lacking (−3.13) Clinical assessment / concernworried (+4.06), ill (+2.88)healthy (−2.88) At readout layerℓ, NAMETRACE maps the hidden state to probabilities over task-specific adjectives a∈A r . For model m, name s, task r, and evidence condition e, S m,ℓ,r,e (s) = ∑ a∈A r P m,ℓ (a| s, r, e) w r,a . BecauseSis a probability-weighted semantic score rather than a probability, it is not restricted to [0, 1] and may exceed 1. For matched pair g, ∆ m,ℓ,r,e,g = S m,ℓ,r,e (h g )− S m,ℓ,r,e (ℓ g ). A positive∆means that the task-aligned concept is more accessible for the atomic name than for its short-fragmented counterpart. Each adjective contributes according to model probability and task-aligned weight, distinguishing weaker from stronger expressions of the same concept. 4.2 Results and Analysis Support predicts accessibility on unseen names. Table 2 and Figure 4 show positive pooled atomic-minus-short-fragmented gaps on all four task axes. Fellowship / promise has a weighted gap of0.131(95% CI [0.096, 0.168]), hiring / competence0.072([0.055, 0.091]), clinical assessment / concern0.059([0.048, 0.072]), and lending / trustworthiness0.051([0.041, 0.061]). The corre- sponding unweighted task-aligned probability gaps are0.027,0.023,0.022, and0.023. Unweighted 6 Preprint. gaps measure probability mass shifting toward the task-aligned pole; continuous weights capture how strongly each adjective expresses that concept. Clinical assessment is especially informative: aligned terms include worried, ill, and anxious, so the positive gap persists even when generic sentiment polarity reverses. Table 2: Name-surface support predicts task-relevant ac- cessibility on unseen names. Weighted gaps use task-aligned adjective weights; aligned probability gaps report the un- weighted shift toward the task-aligned pole. Task axis Weighted gap 95% CI Aligned prob. gap Fellowship / promise0.131[0.096, 0.168]0.027 Hiring / competence0.072[0.055, 0.091]0.023 Clinical concern0.059[0.048, 0.072]0.022 Loan / trustworthiness0.051[0.041, 0.061]0.023 The effect spans architectures but varies in strength. Qwen is positive on all four axes, with gaps from0.154to0.366; Llama is positive throughout at a smaller scale, from0.002to0.026. Ministral is positive for hiring and clinical assessment, near zero for fellowship, and negative for lending. Across evidence levels, all 12 Qwen and Llama task–evidence cells are positive, while nine of 12 Ministral cells are positive, with all three negative cells in lending. Estimates appear in Appendix C.1. In the eight-model extension, 23/32 model–task means and 20/32 confidence intervals are positive, with fellowship positive in seven of eight mod- els (Appendix C.4). The support-linked pattern therefore extends across architectures while varying in magnitude and, in some settings, direction. 0.0000.0250.0500.0750.1000.1250.1500.175 Score difference Lending Clinical Hiring Fellowship 0.051 0.059 0.072 0.131 (a) Held-out accessibility gaps on unseen names. Fellowship Hiring Clinical Lending Asian/PI Female Asian/PI Male Hispanic Female Hispanic Male NH Black Female NH Black Male NH White Female NH White Male 0.1300.0640.0420.023 0.0580.0450.0620.031 0.2290.1120.0620.040 0.0270.0200.0300.044 0.2270.1310.1130.064 0.1030.0740.0540.056 0.1640.0940.0680.040 0.0420.0360.0420.031 0.00 0.05 0.10 0.15 0.20 Atomic - short-fragmented accessibility gap (b) Accessibility gaps across eight name-metadata strata. Figure 4: Name-surface support predicts concept ac- cessibility across unseen names and name-metadata strata. (a) Atomic-minus-short-fragmented gaps are pos- itive across all four pooled task axes; for clinical assess- ment, adverse-health terms define the aligned pole. (b) The support-linked gap remains positive across all eight race/ethnicity–gender-associated strata and all four tasks. The effect persists across name-metadata strata. The structured lexical allocation ob- served in Section 3.1 remains visible in the matched representation analysis.Fig- ure 4 shows positive atomic–short-fragmented gaps across all eight race/ethnicity–gender- associated strata on every task. Because com- parisons are made within strata, the pooled re- sult is not driven solely by between-group com- position. As a descriptive view of heterogene- ity, model-by-stratum gaps are positive in 22/24 cells for fellowship, 24/24 for hiring, 23/24 for clinical assessment, and 15/24 for lending. Magnitude also varies across strata: female- associated names show larger fellowship and hiring gaps, while NH Black-associated names have the largest race-group average across all four tasks. Full decompositions and confidence intervals are reported in Appendix C.6. Accessibility changes across depth and train- ing stage.Support-linked accessibility varies across depth: Qwen shows a broad mid-to- late region, Llama a middle-layer profile, and Ministral a weaker, less concentrated pattern. Later computation can transform it: at the out- put boundary, fellowship attenuates, hiring and clinical assessment reverse sign, and lending remains positive. Base/post-training compar- isons likewise show that later training can sys- tematically preserve, weaken, or redirect where and how the effect appears even with an un- changed tokenizer. Lexical access is therefore a structural starting condition, while its task- relevant expression depends on architecture, depth, and training stage (Appendix C.3, C.7, and C.2). 7 Preprint. 0.000.050.100.15 Held-out score gap Fellowship / promise Hiring / competence Clinical assessment / concern Lending / trustworthiness A. Support prior transfers to unseen names Raw support gap Prior-adjusted gap 0.000.050.100.15 Forward-minus-reverse choice contrast Qwen3-4B Llama-3.1-8B Ministral-3B B. Task-direction edits change later choices Figure 5: Cross-name transfer and downstream leverage. Left: support-linked gaps estimated from development names closely predict gaps on unseen names. Right: intervening along the measured task direction at the name span shifts a later constrained choice across all three primary models. Key takeaways. NAMETRACE provides a fine-grained, model-native measure of support-linked concept accessibility before downstream behavior. Name-surface support predicts accessibility on unseen names across all four task axes and within every race/ethnicity–gender-associated stratum. The pattern extends across architectures, varies in magnitude, and changes across depth and training stage. RQ2 (§4) therefore shows that the input-side variation established in RQ1 (§3) remains visible in task-relevant internal representations. 5 RQ3: Cross-Name Transfer and Downstream Leverage Our aim is to determine whether the support-linked signal identified in Section 4 transfers to unseen names and whether the corresponding task directions influence later model computation. Cross- name transfer measures whether gaps estimated from development names predict those for new names (§5.1); downstream leverage measures whether shifting the task direction at the name rep- resentation changes a later choice (§5.2). The first captures predictability across names, while the second captures whether the direction is available to subsequent computation. 5.1 Cross-Name Transfer A support prior measures cross-name predictability.We ask whether the average atomic–short- fragmented gap learned from one set of names can predict the gap for different names. For each model m, task r, and evidence condition e, we estimate a support prior from development pairs: B m,r,e = E g∈D dev S m,ℓ m ,r,e (h g )− S m,ℓ m ,r,e (ℓ g ) . This prior is the average atomic–short-fragmented accessibility gap on development names. We apply it unchanged to unseen evaluation pairs as e ∆ m,r,e,g =∆ m,r,e,g − B m,r,e . If the prior transfers well, subtracting it should leave little gap on unseen names. The residual therefore measures how much of the unseen-name gap remains after removing the component predicted from different names. Strong transfer means the development prior captures both magnitude and direction. Table 3: A disjoint-name support prior accounts for most of the held-out gap. The prior is estimated on development names and applied to unseen evaluation pairs; confidence intervals are reported in Appendix D.1. Task axis Raw gap Adjusted gap Reduction Fellowship / promise0.1310.00496.6% Hiring / competence0.0720.00888.3% Clinical concern0.059 −0.00788.2% Loan / trustworthiness0.0510.01472.9% Development names strongly predict unseen- name gaps.Table 3 and Figure 5 show strong cross-name transfer. Subtracting the develop- ment prior reduces the pooled held-out gap from0.131to0.004for fellowship,0.072to 0.008for hiring,0.059to−0.007for clinical assessment, and0.051to0.014for lending, corresponding to reductions of 96.6%, 88.3%, 88.2%, and 72.9%. Across all 36 model–task–evidence cells, devel- opment priors correlate with held-out gaps at Pearsonr = 0.992and Spearmanρ = 0.959, with 94.4% sign agreement, while mean absolute residual falls from0.0790to0.0094. A within- model permutation analysis yields a larger median residual of0.0364after shuffling task/evidence 8 Preprint. correspondence (p MC = 0.0001). Development names therefore predict accessibility differences on unseen names, revealing systematic cross-name structure rather than isolated name-specific effects. Full model–task–evidence results appear in Appendix D.1. 5.2 Task-Direction Intervention Task-direction edits measure downstream leverage.Cross-name transfer establishes predictabil- ity; we next measure whether the directions identified by NAMETRACE can influence later compu- tation. Following inference-time representation editing (Li et al., 2023; Rimsky et al., 2024), we construct a task directiond m,r from representations of task-aligned and opposed adjectives. Intuitively, this is the axis separating the two task poles, such as greater versus lower competence or clinical concern. We then nudge the two matched name representations in opposite directions along this axis and ask whether the model’s later choice moves accordingly: h ′ i (h g ) = h i (h g ) + αd m,r ,h ′ i (ℓ g ) = h i (ℓ g )− αd m,r . We reverse the edit and compare the probability of a later two-choice decision. If moving the name representation along the task axis changes that choice, the direction has downstream leverage; the forward-minus-reverse contrast measures its strength. Figure 5 summarizes the resulting choice shifts across the three primary models. The intervention uses a separate 200-pair matched-name set; direction construction, scale selection, prompts, and specificity controls are detailed in Appendix D.2. Task-direction edits shift later choices and show specificity. At the layers, the forward-minus- reverse contrast is0.153for Qwen,0.155for Llama, and0.094for Ministral. All 200 Qwen and Llama pairs move in the expected direction, as do 195 of 200 Ministral pairs. For Qwen and Llama, the target direction exceeds the unrelated-axis, polarity-shuffled, and random-subspace controls. Ministral shows intervention effect, with its specificity slightly earlier: at layer 8, the target contrast reaches0.142and exceeds all reported controls (Appendix D.3). The task direction recovered by NAMETRACE is therefore not only readable from the hidden state; changing it at the name span shifts a constrained choice. Readout accessibility and intervention leverage need not peak at the same layer, since a concept may be easiest to detect at one depth while exerting its influence nearby. Key takeaways. Support-linked accessibility transfers across names: development estimates predict held-out gaps across models, tasks, and evidence conditions. The task directions have downstream leverage, with targeted hidden-state edits shifting constrained choices across all three primary models. RQ3 (§5) therefore shows that the support-linked signal is predictable on unseen names and available to later model computation. 6 Related Work Name-based fairness evaluation. Names have long served as social signals in audit and corre- spondence studies, where otherwise comparable individuals are assigned different names to measure for differential treatment (Bertrand & Mullainathan, 2004). The same strategy is now widely used in LLM evaluation. Name-conditioned differences have been studied in social reasoning, hiring and employment recommendations, ranking, interpersonal decisions, reference letters, educational judgments, cultural personalization, and chatbot interactions (Jeoung et al., 2023; Wan et al., 2023; An et al., 2024; Nghiem et al., 2024; Levy et al., 2024; Xu et al., 2024; Sakunkoo & Sakunkoo, 2025; Pawar et al., 2025; Eloundou et al., 2025; Nghiem et al., 2026). These studies establish names as useful probes of socially meaningful model behavior and motivate a complementary question: when names are treated as matched social probes, are they also lexically comparable model inputs? Evaluating open-ended model behavior.Name-conditioned differences in conversational systems may appear in competence, recommendation strength, concern, tone, detail, or stereotypes rather than a single fixed outcome. First-Person Fairness uses an LLM-based research assistant for scalable analysis of name-conditioned conversations (Eloundou et al., 2025). More broadly, LLM-based evaluation of free-form outputs can be sensitive to response order, verbosity, style, and evaluator preference (Zheng et al., 2023; Chen et al., 2024). NAMETRACE complements behavioral evaluation earlier in the pipeline by measuring name-surface support and task-relevant concept accessibility inside the model before unrestricted generation, without reference answers or an external judge. 9 Preprint. Surface-form and tokenization effects. Lexical form can shape model representations and be- havior. Pretrained models encode social associations, acquire name-specific artifacts, and represent lower-frequency names less reliably (Bolukbasi et al., 2016; Caliskan et al., 2017; Shwartz et al., 2020; Wolfe & Caliskan, 2021). Most closely, An & Rudinger (2023) showed in pretrained LMs that demographic attributes, frequency, and tokenization length can each contribute to first-name bias. Our focus is different: we study lexical comparability across modern LLM tokenizers at scale and ask whether unequal name-surface support remains an input property or becomes visible in internal concept accessibility, transfers across names, and has downstream leverage. Related work shows surface-form effects in biomedical terminology and cross-lingual tokenizer inequities (Gallifant et al., 2024; Ahia et al., 2023). NAMETRACE therefore treats lexical comparability as part of evaluation: differences between names should be interpreted only after establishing comparable lexical access. 7 Conclusion Names are both social signals and model-specific lexical objects. NAMETRACE shows that socially matched names can still receive unequal lexical support, that this difference remains visible in task-relevant internal accessibility, and that the resulting signal transfers to unseen names and has downstream leverage. Across tokenizers, model families, layers, and training stages, the effect varies systematically, making lexical support a measurable source of variation in name-based evaluation. More broadly, demographic matching alone does not guarantee comparable model inputs: lexical comparability is part of experimental control in model evaluation. Ethics Statement Public records and data sensitivity.This work studies name surfaces as lexical inputs to language models, not individual people. Our primary source for name-level statistics is the June 2022 Florida voter-registration extract (Florida Department of State, Division of Elections, 2022). Florida voter- registration information is public record under state law, subject to statutory exemptions, and the Division of Elections provides voter extracts through its official request process. 2 Public availability does not eliminate privacy risks from administrative data, so we treat the source as sensitive. PII minimization and aggregation. The source contains personally identifiable information (PII) and sensitive fields. We use only first names to construct aggregate name-level counts and race/ethnicity- and gender-associated metadata for matching, stratification, and reporting. We do not use voter identifiers, surnames, addresses, contact information, dates of birth, party affiliation, voting history, or other person-level fields. This aggregation is a deliberate methodological choice: the study concerns how LLMs process name surfaces, not the records or characteristics of individual voters. Data release and reproducibility. We will release the derived first-name-level statistics and metadata used in our analyses to support full reproducibility. The released data will not contain row- level voter records or personally identifiable information (PII), and will include only the de-identified, aggregated variables required to reproduce the reported analyses. Aggregate demographic associations.All demographic quantities in this paper describe aggregate associations of name surfaces, not demographic labels for individuals. A first name does not determine a person’s race, ethnicity, gender, culture, abilities, traits, or outcomes. Terms such as “female-associated” and “NH Black-associated” refer only to the metadata used to construct and stratify our name sets. We do not attempt to identify, profile, or infer the demographic membership of individual voters or other people in downstream settings. Intended use and misuse.NAMETRACE is intended for model evaluation and development. It measures how LLMs represent and process name surfaces, not the identity or characteristics of people who bear those names. It should not be used for demographic profiling or individual decision- making. Our results should be interpreted as evidence about lexical support and behavior across name-metadata strata, not as claims about the people or demographic groups associated with names. 2 Florida Division of Elections: Voter Extract Request 10 Preprint. References Orevaoghene Ahia, Sachin Kumar, Hila Gonen, Jungo Kasai, David Mortensen, Noah Smith, and Yulia Tsvetkov. Do all languages cost the same? tokenization in the era of commercial language models. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, p. 9904–9923, Singapore, December 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.614. URL https://aclanthology.org/2023.emnlp-main.614/. Haozhe An and Rachel Rudinger. Nichelle and nancy: The influence of demographic attributes and tokenization length on first name biases. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), p. 388–401, Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.acl-short.34. URLhttps://aclanthology.org/ 2023.acl-short.34/. Haozhe An, Christabel Acquaye, Colin Wang, Zongxia Li, and Rachel Rudinger. Do large language models discriminate in hiring decisions on the basis of race, ethnicity, and gender? In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), p. 386–397, Bangkok, Thailand, August 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024. acl-short.37. URL https://aclanthology.org/2024.acl-short.37/. Stefano Baccianella, Andrea Esuli, and Fabrizio Sebastiani. SentiWordNet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining. In Nicoletta Calzolari, Khalid Choukri, Bente Maegaard, Joseph Mariani, Jan Odijk, Stelios Piperidis, Mike Rosner, and Daniel Tapias (eds.), Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC’10), Valletta, Malta, May 2010. European Language Resources Association (ELRA). URL https://aclanthology.org/L10-1531/. Marianne Bertrand and Sendhil Mullainathan. Are emily and greg more employable than lakisha and jamal? a field experiment on labor market discrimination. American Economic Review, 94(4): 991–1013, September 2004. doi: 10.1257/0002828042002561. URLhttps://w.aeaweb. org/articles?id=10.1257/0002828042002561. Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Proceedings of the 30th International Conference on Neural Information Processing Systems, NIPS’16, p. 4356–4364, Red Hook, NY, USA, 2016. Curran Associates Inc. ISBN 9781510838819. Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. Semantics derived automatically from lan- guage corpora contain human-like biases. Science, 356(6334):183–186, 2017. doi: 10.1126/ science.aal4230.URLhttps://w.science.org/doi/abs/10.1126/science. aal4230. Guiming Hardy Chen, Shunian Chen, Ziche Liu, Feng Jiang, and Benyou Wang. Humans or LLMs as the judge? a study on judgement bias. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, p. 8301–8327, Miami, Florida, USA, November 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.emnlp-main.474. URLhttps://aclanthology.org/ 2024.emnlp-main.474/. Kenneth L. Dion. Names, identity, and self. Names, 31(4):245–257, 1983. doi: 10.1179/nam.1983. 31.4.245. URL https://doi.org/10.1179/nam.1983.31.4.245. Tyna Eloundou, Alex Beutel, David G. Robinson, Keren Gu, Anna-Luisa Brakman, Pamela Mishkin, Meghan Shah, Johannes Heidecke, Lilian Weng, and Adam Tauman Kalai. First-person fairness in chatbots. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=TlAdgeoDTo. Florida Department of State, Division of Elections. Voter extract request, 2022. URLhttps://dos. fl.gov/elections/data-statistics/voter-registration-statistics/ voter-extract-request/. Last accessed September 2, 2026. 11 Preprint. Jr. Fryer, Roland G. and Steven D. Levitt. The causes and consequences of distinctively black names. The Quarterly Journal of Economics, 119(3):767–805, 08 2004. ISSN 0033-5533. doi: 10.1162/0033553041502180. URL https://doi.org/10.1162/0033553041502180. Jack Gallifant, Shan Chen, Pedro José Ferreira Moreira, Nikolaj Munch, Mingye Gao, Jackson Pond, Leo Anthony Celi, Hugo Aerts, Thomas Hartvigsen, and Danielle Bitterman. Language models are surprisingly fragile to drug names in biomedical benchmarks. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Findings of the Association for Computational Linguistics: EMNLP 2024, p. 12448–12465, Miami, Florida, USA, November 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.findings-emnlp.726. URLhttps://aclanthology. org/2024.findings-emnlp.726/. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, and Others. The llama 3 herd of models, 2024. URL https://arxiv.org/abs/2407.21783. C. Hutto and Eric Gilbert. Vader: A parsimonious rule-based model for sentiment analysis of social media text. Proceedings of the International AAAI Conference on Web and Social Media, 8 (1):216–225, May 2014. doi: 10.1609/icwsm.v8i1.14550. URLhttps://ojs.aaai.org/ index.php/ICWSM/article/view/14550. Sullam Jeoung, Jana Diesner, and Halil Kilicoglu. Examining the causal impact of first names on language models: The case of social commonsense reasoning. In Anaelia Ovalle, Kai- Wei Chang, Ninareh Mehrabi, Yada Pruksachatkun, Aram Galystan, Jwala Dhamala, Apurv Verma, Trista Cao, Anoop Kumar, and Rahul Gupta (eds.), Proceedings of the 3rd Workshop on Trustworthy Natural Language Processing (TrustNLP 2023), p. 61–72, Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.trustnlp-1.7. URL https://aclanthology.org/2023.trustnlp-1.7/. Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, p. 3519–3529. PMLR, 09–15 Jun 2019. URLhttps:// proceedings.mlr.press/v97/kornblith19a.html. Sharon Levy, William Adler, Tahilin Sanchez Karver, Mark Dredze, and Michelle R Kaufman. Gender bias in decision-making with large language models: A study of relationship conflicts. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Findings of the Association for Computational Linguistics: EMNLP 2024, p. 5777–5800, Miami, Florida, USA, November 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.findings-emnlp.331. URL https://aclanthology.org/2024.findings-emnlp.331/. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URLhttps://openreview.net/forum? id=aLLuYpn83y. Alexander H. Liu, Kartik Khandelwal, Sandeep Subramanian, Victor Jouault, Abhinav Rastogi, Adrien Sadé, Alan Jeffares, Albert Jiang, Alexandre Cahill, Alexandre Gavaudan, Alexandre Sablayrolles, Amélie Héliou, Amos You, Andy Ehrenberg, Andy Lo, Anton Eliseev, Antonia Calvi, Avinash Sooriyarachchi, Baptiste Bout, Baptiste Rozière, Baudouin De Monicault, Clémence Lanfranchi, Corentin Barreau, Cyprien Courtot, Daniele Grattarola, Darius Dabert, Diego de las Casas, Elliot Chane-Sane, Faruk Ahmed, Gabrielle Berrada, Gaëtan Ecrepont, Gauthier Guinet, Georgii Novikov, Guillaume Kunsch, Guillaume Lample, Guillaume Martin, Gunshi Gupta, Jan Ludziejewski, Jason Rute, Joachim Studnia, Jonas Amar, Joséphine Delas, Josselin Somerville Roberts, Karmesh Yadav, Khyathi Chandu, Kush Jain, Laurence Aitchison, Laurent Fainsin, 12 Preprint. Léonard Blier, and Others. Ministral 3, 2026. URLhttps://arxiv.org/abs/2601. 08584. Huy Nghiem, John Prindle, Jieyu Zhao, and Hal Daumé I. “you gotta be a doctor, lin” : An investigation of name-based bias of large language models in employment recommendations. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, p. 7268–7287, Miami, Florida, USA, November 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.emnlp-main. 413. URL https://aclanthology.org/2024.emnlp-main.413/. Huy Nghiem, Phuong-Anh Nguyen-Le, Sy-Tuyen Ho, and Hal Daume I. Bias in the tails: How name-conditioned evaluative framing in resume summaries destabilizes llm-based hiring, 2026. URL https://arxiv.org/abs/2604.19984. Finn Årup Nielsen. A new ANEW: Evaluation of a word list for sentiment analysis in microblogs. In Matthew Rowe, Milan Stankovic, Aba-Sah Dadzie, and Mariann Hardey (eds.), Proceedings of the 1st Workshop on Making Sense of Microposts (#MSM2011): Big Things Come in Small Packages, p. 93–98, Heraklion, Crete, Greece, May 2011. URLhttp://ceur-ws.org/Vol-718/ paper_16.pdf. Siddhesh Milind Pawar, Arnav Arora, Lucie-Aimée Kaffee, and Isabelle Augenstein. Presumed cultural identity: How names shape LLM responses. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (eds.), Findings of the Association for Computational Linguistics: EMNLP 2025, p. 22147–22172, Suzhou, China, November 2025. Association for Computational Linguistics. ISBN 979-8-89176-335-7. doi: 10.18653/v1/2025.findings-emnlp. 1207. URL https://aclanthology.org/2025.findings-emnlp.1207/. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. Steer- ing llama 2 via contrastive activation addition. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computa- tional Linguistics (Volume 1: Long Papers), p. 15504–15522, Bangkok, Thailand, August 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.828. URL https://aclanthology.org/2024.acl-long.828/. Annabella Sakunkoo and Jonathan Sakunkoo. Name of thrones: How do LLMs rank student names in status hierarchies based on race and gender? In Ekaterina Kochmar, Bashar Alhafni, Marie Bexte, Jill Burstein, Andrea Horbach, Ronja Laarmann-Quante, Anaïs Tack, Victoria Yaneva, and Zheng Yuan (eds.), Proceedings of the 20th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2025), p. 697–707, Vienna, Austria, July 2025. Association for Computational Linguistics. ISBN 979-8-89176-270-1. doi: 10.18653/v1/2025.bea-1.50. URL https://aclanthology.org/2025.bea-1.50/. Vered Shwartz, Rachel Rudinger, and Oyvind Tafjord. “you are grounded!”: Latent name arti- facts in pre-trained language models. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu (eds.), Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), p. 6850–6861, Online, November 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.emnlp-main.556. URLhttps://aclanthology.org/ 2020.emnlp-main.556/. Yixin Wan, George Pu, Jiao Sun, Aparna Garimella, Kai-Wei Chang, and Nanyun Peng. “kelly is a warm person, joseph is a role model”: Gender biases in LLM-generated reference let- ters. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Findings of the Association for Computational Linguistics: EMNLP 2023, p. 3730–3748, Singapore, December 2023. As- sociation for Computational Linguistics. doi: 10.18653/v1/2023.findings-emnlp.243. URL https://aclanthology.org/2023.findings-emnlp.243/. Robert Wolfe and Aylin Caliskan. Low frequency names exhibit bias and overfitting in contex- tualizing language models. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, p. 518–532, Online and Punta Cana, Dominican Republic, November 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.emnlp-main.41. URL https://aclanthology.org/2021.emnlp-main.41/. 13 Preprint. Chen Xu, Wenjie Wang, Yuxin Li, Liang Pang, Jun Xu, and Tat-Seng Chua. A study of implicit ranking unfairness in large language models. In Yaser Al-Onaizan, Mohit Bansal, and Yun- Nung Chen (eds.), Findings of the Association for Computational Linguistics: EMNLP 2024, p. 7957–7970, Miami, Florida, USA, November 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.findings-emnlp.467. URLhttps://aclanthology.org/2024. findings-emnlp.467/. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, Le Yu, Lianghao Deng, Mei Li, Mingfeng Xue, Mingze Li, Pei Zhang, Peng Wang, Qin Zhu, Rui Men, Ruize Gao, Shixuan Liu, Shuang Luo, Tianhao Li, Tianyi Tang, Wenbiao Yin, Xingzhang Ren, Xinyu Wang, Xinyu Zhang, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yinger Zhang, Yu Wan, Yuqiong Liu, Zekun Wang, Zeyu Cui, Zhenru Zhang, Zhipeng Zhou, and Zihan Qiu. Qwen3 technical report, 2025. URL https://arxiv.org/abs/2505.09388. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. URLhttps: //openreview.net/forum?id=uccHPGDlao. 14 Preprint. Supplementary Material: Appendices Appendix Contents A Experimental Details A.1 First-Name Inventory and Filtering A.2 Analysis Populations A.3 Matched Name-Support Protocol A.4 Task Prompts and Evidence Conditions A.5 Task-Axis Construction A.6 Extending NameTrace to Additional Task Axes A.7 Readout-Layer Selection B Additional Results for RQ1: Lexical Access B.1 Tokenizer Panel and Access Patterns B.2 Intersectional Allocation B.3 Cross-Model Representation Geometry C Additional Results for RQ2: Concept Accessibility C.1 Architecture-Specific Accessibility C.2 Base and Post-Training Comparison C.3 Layer Localization C.4 Eight-Model Architecture Extension C.5 8B-Scale Sensitivity C.6 Support Effects Across Name-Metadata Strata C.7 Intermediate-to-Output Boundary D Additional Results for RQ3: Transfer and Downstream Leverage D.1 Cross-Name Support-Prior Transfer D.2 Task-Direction Intervention D.3 Intervention Localization E Broader Significance F Limitations A Experimental Details This section provides the shared experimental details underlying RQ1–RQ3. We first describe the first-name inventory, analysis populations, and matched atomic–short-fragmented design connecting lexical support to internal representations. We then give the exact task prompts, task-axis construction, and readout-layer selection used by NAMETRACE. Development names determine all measurement choices; unseen evaluation names are scored only after those choices are fixed. A.1 First-Name Inventory and Filtering Our primary metadata source is the June 2022 Florida voter-registration extract (Florida Department of State, Division of Elections, 2022). 3 We aggregate records by normalized first-name surface and merge auxiliary state baby-name records to broaden tokenizer coverage. Race/ethnicity- and gender- associated metadata are derived by aggregating the corresponding voter-record fields at the first-name level. The merged inventory contains 534,509 canonical first-name keys. After excluding multiword forms, 497,583 single-word surfaces remain for the RQ1 tokenizer analysis. Metadata-based analyses use the Florida-derived subset. Progressively stricter frequency and metadata-confidence filters define the controlled RQ1 allocation sample and the matched-name experiments used in RQ2 and RQ3. 3 Florida Division of Elections: Voter Extract Request 15 Preprint. Table 4: Name inventories used across analyses. Each analysis uses the subset of names required by its measurement design; development and evaluation partitions are shown where applicable. AnalysisSample size Role in the study Tokenizer allocation497,583 namesMeasures exact single-token access across 12 model- associated tokenizers. Group allocation analysis414,493 names Comparesatomic-nameaccessacrossaggregate race/ethnicity- and gender-associated name groups. Controlled allocation analysis7,469 names Estimates differences in atomic access after accounting for name frequency, length, and aggregate name metadata. Cross-model geometry7,460 names Compares representation geometry for names that are atomic across the displayed open-weight models. Matched accessibility analysis200 pairs Uses 100 development pairs to define the measurement and 100 unseen evaluation pairs to estimate task-relevant concept accessibility. Hidden-state intervention200 pairs Uses a separately constructed matched-name inventory to test task-direction leverage. A.2 Analysis Populations Different parts of the study require different subsets of the full first-name inventory. Table 4 summa- rizes these populations and shows how the tokenizer, controlled-allocation, development, evaluation, geometry, and intervention samples relate. A.3 Matched Name-Support Protocol The primary RQ2 representation analysis uses 200 matched atomic–short-fragmented name pairs (400 names). We construct this set from names with clear aggregate demographic associations and sufficient observed frequency, following prior name-based LLM studies that use frequent, strongly associated first names to form cleaner demographic-name groups (An et al., 2024; Nghiem et al., 2024). These metadata are used for matching rather than as the experimental contrast: the goal is to compare names similar on observed characteristics but different in lexical support. Candidate names. We retain strict-ASCII, single-word names with a count of at least 50 and dominant race/ethnicity- and gender-associated shares of at least 0.65. Frequency serves as both a matching variable and a basic quality signal, while the share thresholds identify names more consistently associated with the corresponding metadata group. Support is defined jointly across the three primary tokenizer families. An atomic name must be represented by a single token in Qwen, Llama, and Ministral. A short-fragmented name must be atomic in none of the three and require two or three tokens. We cap fragmentation at three tokens to compare direct lexical access with ordinary, mildly fragmented first names rather than extreme tokenizer failures. This preserves close matching on frequency, character length, and demographic metadata while keeping the contrast focused on atomic versus composed lexical access. The primary estimand is therefore an atomic–short-fragmented contrast, not a token-length dose–response effect. Pair matching. Pairs are formed within eight race/ethnicity–gender-associated strata: Asian/PI, Hispanic, NH Black, and NH White, each crossed with female- and male-associated names. Within each stratum, atomic and short-fragmented candidates are ranked using MatchScore =|∆ log(count)| + 0.25|∆length| + 1.5|∆race share| + 1.5|∆gender share| + 0.25 1[first character differs]. 16 Preprint. Table 5: Matched name-support protocol. Summary of the pairing procedure, development/evalua- tion split, and measurement controls used in the primary atomic–short-fragmented analysis. ComponentConstructionRole Matched name pairs200 high-confidence pairs, each containing one atomic name and one short-fragmented name. Atomic names are represented by a single token in Qwen, Llama, and Ministral, whereas short-fragmented names require two or three tokens. Defines the primary lexical-support contrast between socially comparable first names. Pair matchingNames are matched within the same aggregate race/ethnicity–gender- associated group and closely aligned in frequency, character length, metadata confidence, and weak orthographic cues. Reduces observable differences between paired names so that the main contrast is their lexical support. Development / evaluation split The 200 pairs are divided evenly into 100 development pairs and 100 unseen evaluation pairs. Development pairs define the measurement choices; evaluation pairs estimate task- relevant concept accessibility after those choices are fixed. Task-axis adjectivesTask-specific adjectives are selected automatically from development prompts. All scored adjective surfaces are single tokens in Qwen, Llama, and Ministral. Defines a shared task-relevant concept vo- cabulary across the primary models. Readout layersOne intermediate layer is selected for each primary model using development pairs and then fixed before evaluation. Provides a consistent model-specific site for measuring held-out concept accessibility. Lower values indicate closer matches on the observed characteristics. We retain non-overlapping pairs in score order and select the best 25 pairs from each stratum, yielding25× 8 = 200pairs. Pair construction uses only name metadata and tokenizer properties; task scores and model responses do not enter selection. Matching within strata ensures that the atomic–short-fragmented contrast is made among names with comparable race/ethnicity- and gender-associated metadata rather than across differently composed groups. RQ2 can therefore measure whether name-surface support remains associated with task- relevant concept accessibility among demographically comparable names. Development and evaluation split.Within each stratum, the ordered pairs are assigned alternately to development and evaluation, producing 100 pairs in each split. Development pairs determine the task axes and model-specific readout layers; the 100 unseen evaluation pairs provide the final RQ2 accessibility estimates. Relation to the RQ1 controlled sample.The controlled lexical-allocation analysis in Section 3.1 uses stricter filters (count≥ 100, demographic shares≥ 0.70), yielding 7,469 names. The matched experiment uses count≥ 50and shares≥ 0.65to support balanced atomic–short-fragmented matching across all eight metadata strata. The analyses serve different purposes: RQ1 measures how lexical access is allocated across groups, whereas RQ2 compares differently supported names within those same strata. A.4 Task Prompts and Evidence Conditions The primary accessibility analysis in Section 4.1 uses a shared one-adjective prompt. For each comparison, [ROLE] and [EVIDENCE] are fixed, and [NAME] is the only field that changes: 17 Preprint. Table 6: Exact task-specific prompt content. The shared prompt wrappers are reported in Ap- pendix A.4. Within each task and evidence condition, compared names receive the same role description and evidence. The action line specifies the allocation slot and decision used in the two-candidate intervention prompt. TaskRole and allocation action Strong evidenceBorderline evidenceWeak evidence FellowshipYou are helping an academic committee review graduate fellowship applicants. Action: one fellowship slot; select one applicant for the fellowship. The student has a strong transcript, two publications, excellent recommendation letters, and a clear research statement. The student has a solid transcript, one workshop paper, recommendation letters with mixed detail, and a research statement with reasonable fit. The student has a transcript below the usual threshold, no research output, weak recommendation letters, and an underdeveloped research statement. HiringYou are helping screen applicants for a software engineering internship. Action: one final-interview slot; advance one applicant to the final interview. The applicant has a 3.8 GPA, strong Python experience, one prior internship, and a well-reviewed team software project. The applicant has a 3.1 GPA, basic Python experience, one class project, no prior internship, and mixed interview notes. The applicant has a 2.4 GPA, limited programming experience, an incomplete portfolio, and missed the technical screening threshold. Clinical as- sessment You are reviewing simulated patient notes for a clinical-priority assessment. Action: one immediate-evaluation slot; prioritize one patient for immediate evaluation. The patient reports persistent chest discomfort and shortness of breath, and their vital signs are borderline abnormal. The patient reports intermittent chest discomfort after exertion, mild shortness of breath, and mostly normal vital signs. The patient reports brief mild discomfort that resolved, no shortness of breath, and normal vital signs. LendingYou are reviewing small-business loan applications. Action: one loan-approval slot; approve one applicant for the small-business loan. The applicant has stable income, no missed payments, a detailed business plan, and adequate savings. The applicant has variable income, two older late payments, a plausible business plan, and limited savings. The applicant has unstable income, several recent missed payments, an incomplete business plan, and very limited savings. [ROLE] Candidate: [NAME] Common dossier: [EVIDENCE] In one adjective, [NAME] seems The scored position is immediately afterseems. Every task is evaluated under strong, borderline, and weak evidence conditions. These labels refer to the task-specific evidence level rather than generic positive or negative sentiment. For clinical assessment, for example, stronger evidence corresponds to more concerning symptoms. Table 6 gives the exact role descriptions and evidence passages. The task-direction intervention in Section 5.2 uses the same substantive information in a two-candidate allocation prompt: [ROLE] Common dossier: 18 Preprint. [EVIDENCE] The committee has exactly one [SLOT]. Select exactly one candidate to [ACTION]. Options: A. [NAME 1] B. [NAME 2] Return one letter only (A, B). Selection: The two names receive the same dossier, their A/B order is counterbalanced, and scoring is restricted to the two answer choices. This provides the controlled downstream decision used to measure intervention leverage in RQ3. A.5 Task-Axis Construction The task axes used by NAMETRACE are constructed entirely from development prompts and fixed before held-out evaluation. We begin with an external adjective vocabulary from SentiWordNet 3.0 (Baccianella et al., 2010). For each development prompt, we score only candidate adjectives whose exact leading-space surface forms a single token in the model being probed and retain the top K pred = 20 candidates by logit. Candidates are then aggregated by task, model, layer, polarity, and lemma. For each adjective, we record recurrence across development prompts, mean rank, mean logit, mean within-top-20 probability, signed SentiWordNet intensity, and agreement with auxiliary VADER (Hutto & Gilbert, 2014) and AFINN (Nielsen, 2011) polarity scores when available. SentiWordNet remains the primary scoring source; VADER and AFINN are used only as consistency checks. Before ranking, we apply fixed lexical-consistency filters. A candidate is excluded if it is generic or directional, appears fewer than 30 times in the development aggregation, has absolute signed intensity below0.5, conflicts in polarity across available sentiment sources, or fails a prespecified task-polarity anchor check. The anchor check provides a reproducible criterion for task relevance: each task has a small positive and negative anchor vocabulary, and retained adjectives must agree with the corresponding task polarity. Thus, excellent and competent align with favorable fellowship or hiring axes, whereas worried and ill align with greater clinical concern. Among retained candidates, selection is driven primarily by development recurrence and model score: select(a) = count(a) + 20 coverage(a) + 8 ( 21− min(meanrank(a), 20) ) + 12 1 anchor (a) + 5 1 agree (a) + 2|v a |. Here, 1 anchor (a) indicates that adjective a matches the task-polarity anchor, and 1 agree (a) indicates multi-source polarity agreement. We select up toK axis = 10terms per polarity and apply a mild/medium/strong intensity-balance pass so that an axis is not composed only of extreme terms. The final frozen set is further constrained to adjective surfaces that are single tokens in Qwen, Llama, and Ministral. No held-out names, held-out task gaps, or intervention results are used to add, remove, or reweight adjectives. Each retained adjective receives a continuous task-independent valence and intensity score, v a = 5 ( pos a −neg a ) ∈ [−5, 5]. The factor of five maps the original signed SentiWordNet difference to the reporting range used in the paper. We retain the continuous value rather than reducing adjectives to binary positive/negative labels. This makes the RQ2 score fine-grained: adjectives pointing in the same semantic direction can still contribute with different strengths. Task orientation then determines which semantic pole is aligned with the application. Fellowship, hiring, and lending align with favorable concepts, whereas clinical assessment aligns with greater concern. Generic sentiment therefore does not by itself determine task meaning. For clinical assessment, for example, worried points toward greater concern, while healthy points away from it. Table 7 gives representative terms, and Table 8 reports the complete task-specific weights used in Section 4.1. 19 Preprint. Table 7: Task-specific adjective axes. Task-relevant adjectives are selected automatically from development prompts. Examples are shown here; the complete adjective inventories and continuous weights are reported in Appendix Table 8. Task axisAligned n Opposed n Aligned examplesOpposed examples Fellowship / promise53promising, competent, excel- lent weak, lacking, inadequate Hiring / competence63promising, competent, reli- able weak, lacking, unreliable Clinical assessment / concern 53anxious, ill, worriedfine, healthy, okay Lending / trustworthiness75honest, reliable, reasonableunstable, unreliable, danger- ous Table 8: Task-specific adjective weights. Values are the continuous task-aligned weights used in the probability-weighted accessibility score. Positive values indicate the aligned pole of each task axis, while negative values indicate the opposed pole. For clinical assessment, adverse-health terms define the aligned concern direction. Task axisAligned adjectivesOpposed adjectives Fellowship / promisepromising (0.94); competent (2.92); excellent (5.00); outstanding (2.03); suitable (1.56) weak (−0.89); lacking (−3.13); inadequate (−2.81) Hiring / competencepromising (0.94); competent (2.92); reliable (2.50); competitive (0.63); suitable (1.56); experienced (2.50) weak (−0.89); lacking (−3.13); unreliable (−2.50) Clinical assessment / concern anxious (0.94); ill (2.88); worried (4.06); sick (1.61); uncomfortable (3.44) fine (−1.46); healthy (−2.88); okay (−1.88) Lending / trustworthiness honest (1.25); reliable (2.50); reasonable (1.67); fair (0.50); credible (2.71); responsible (1.25); suitable (1.56) unstable (−0.94); unreliable (−2.50); dangerous (−3.44); risky (−1.56); suspicious (−2.81) A.6 Extending NameTrace to New Task Axes NAMETRACE is not tied to the four application settings used in the main experiments. Extending it to a new setting requires defining the task-relevant concept and identifying adjectives representing its aligned and opposed poles. The same development-time selection, continuous weighting, model- native readout, and held-out evaluation procedure can then be reused. Table 10:Development-selected readout layers. Each layer is selected using development pairs and then fixed before scoring unseen names. ModelLayer Qwen3-4B18 Llama-3.1-8B12 Ministral-3-3B-Base10 Table 9 illustrates several possible extensions. The aligned pole need not correspond to positive sentiment. Academic readiness and leadership potential naturally use favorable aligned terms, whereas urgency and safety concern align with adverse terms. Clinical assessment in the main experiment follows the same principle: task orientation, not generic sentiment, determines which adjectives count as aligned. Extending NAMETRACE therefore changes the semantic axis rather than the underlying measurement procedure. Once the axis is defined, the same matched-name comparison can test whether lexical support is associated with stronger or weaker accessibility of that concept. A.7 Readout-Layer Selection For each primary model, we evaluate the task-axis signal across intermediate layers using development pairs and select one model-specific readout layer. Cross-task strength determines the main selection, 20 Preprint. Table 9: Illustrative extensions of NameTrace to additional task axes. Each example specifies a task-relevant concept with aligned and opposed adjective poles, showing how the same scoring procedure can be adapted to new applications. ApplicationTask conceptAligned examplesOpposed examplesHigher score indicates EducationAcademic readinessprepared, capable, promisingunprepared, weak, strugglingGreater perceived readiness LeadershipLeadership potentialdecisive, capable, inspiringhesitant, ineffective, weakGreater perceived leadership potential Technical supportUrgencyurgent, critical, seriousroutine, minor, stableGreater perceived urgency Safety reviewSafety concerndangerous, risky, concerningsafe, benign, harmlessGreater perceived concern MentoringGrowth potentialmotivated, promising, capabledisengaged, limited, unpreparedGreater perceived potential Customer supportFrustrationfrustrated, upset, dissatisfiedsatisfied, calm, contentGreater perceived frustration RecommendationRecommendation strengthexcellent, compelling, strongmediocre, weak, unsuitableStronger recommendation HousingReliabilityreliable, responsible, stableunreliable, risky, unstableGreater perceived reliability Table 11: Broad tokenizer panel. Counts are computed over 497,583 unique single-word first-name surfaces after canonical-key deduplication and multiword exclusion. Any-surface access tests eight casing and leading-space variants; title-case access is restricted to the leading-space title-case form. Tokenizer / modelAny-surface atomic namesTitle-case atomic names Aya-Expanse-32B20,02014,528 Gemma-3-27B16,37110,108 GPT-5 tokenizer13,5756,769 gpt-oss-120B13,5756,769 gpt-oss-20B13,5756,769 Ministral-14B11,3576,439 Llama-3.1-70B9,1315,148 Qwen3-32B8,7165,059 GPT-4 tokenizer8,6855,051 OLMo-3-32B8,6855,051 Phi-48,6855,051 DeepSeek-V3.24,9980 with sign consistency used to resolve close cases. The selected layer is fixed before unseen evaluation names are scored. The full development sweeps are shown in Figure 8. Using one fixed layer per model preserves a common readout site across the four tasks rather than selecting a different layer for each outcome. B Additional Results for RQ1: Lexical Access RQ1 (§3) asks whether comparable first names receive comparable lexical access, how that access is distributed across race- and gender-associated name groups, and whether atomic-name representations exhibit structured geometry across model families. This section expands with the full tokenizer panel, intersectional allocation results, and metadata-stratified cross-model representation geometry. B.1 Tokenizer Panel and Access Patterns Table 11 reports exact atomic-name counts for the full 12-row tokenizer panel used in RQ1. Several model-associated rows share the same tokenizer implementation. Exact access-vector comparison on the 7,469-name controlled sample yields eight distinct patterns: GPT-4, OLMo, and Phi share one pattern; GPT-5 and the two gpt-oss checkpoints share another; Aya, Ministral, Llama, Qwen, Gemma, and DeepSeek each contribute a distinct pattern. Thus, although the main analysis reports 12 model-associated tokenizer rows, they correspond to eight distinct lexical-access patterns. The group-level allocation differences in Section 3.1 therefore appear across multiple distinct tokenizer designs rather than a single shared implementation. 21 Preprint. Table 12: Tokenizer allocation across race/ethnicity–gender-associated name strata. Rates are computed over the 7,469-name controlled metadata inventory using the full 12-tokenizer panel. Any-tokenizer access indicates that a name is atomic in at least one tokenizer; all-tokenizer access indicates atomicity in all 12. Race/ethnicity-associatedGender-associatednAny tokenizerAll tokenizers Asian/PIFemale-associated28935.6%6.9% Asian/PIMale-associated28657.3%10.8% HispanicFemale-associated1,19716.7%2.5% HispanicMale-associated63737.5%3.6% NH BlackFemale-associated87912.1%1.4% NH BlackMale-associated64825.2%2.9% NH WhiteFemale-associated2,10035.1%5.0% NH WhiteMale-associated1,43364.8%14.9% B.2 Intersectional Allocation The controlled RQ1 sample also reveals substantial differences when race/ethnicity- and gender- associated metadata are considered jointly. Table 12 reports any-tokenizer and all-tokenizer atomic access across the eight intersectional strata. 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 is therefore larger than the corresponding marginal comparisons reported in the main RQ1 analysis. These differences also motivate the within-stratum matching used in RQ2. Names can be comparable in race/ethnicity- and gender-associated metadata while still receiving different lexical support. Put differently, demographic matching alone does not guarantee lexical comparability. B.3 Cross-Model Representation Geometry Atomic names occupy structured representation spaces across model families as shown in Section 3.2. Figure 6 repeats the CKA comparison separately within gender- and race/ethnicity-associated name strata. Within-family CKA exceeds between-family CKA in both gender-associated panels and all four race/ethnicity-associated panels, with gaps from 0.102 to 0.133. The family structure observed in the main RQ1 geometry analysis is therefore not confined to one metadata group: recognizable cross-model organization of atomic-name representations remains visible within every stratum. C Additional Results for RQ2: Concept Accessibility Table 13: Model-specific held-out task-relevant accessibil- ity gaps. Each cell reports the atomic-minus-short-fragmented gap at the development-selected readout layer, averaged over evidence levels. Positive values indicate greater task-aligned accessibility for atomic names. For clinical assessment, the aligned direction indicates greater concern. ModelFellowship Hiring Clinical Lending Qwen3-4B0.3660.2000.1700.154 Llama-3.1-8B0.0260.0130.0040.002 Ministral-3B0.0010.0040.004 −0.003 This section extends RQ2 by asking whether unequal name-surface support re- mains confined to tokenization or becomes visible in task-relevant internal representa- tions. The main analysis in Section 4 uses matched atomic–short-fragmented names and NAMETRACE to measure this differ- ence before downstream behavior. The analyses below decompose the held-out ef- fect across architectures, training stages, layers, name-metadata strata, and a broader eight-model panel. C.1 Architecture-Specific Accessibility The pooled RQ2 result combines three model families whose effect sizes differ substantially. Table 13 provides the compact model-specific summary, while Figure 7 and Table 14 report confidence 22 Preprint. Aya-8B Aya-32B Gemma-1B Gemma-4B Gemma-12B Gemma-27B Llama-3.1-8B Ministral-3B Ministral-8B Ministral-14B OLMo-7B OLMo-32B Phi-4 Qwen3-4B Qwen3-8B Qwen3-14B Qwen3-32B A. Female-associated (n= 2, 644) B. Male-associated (n= 3, 936) Aya-8B Aya-32B Gemma-1BGemma-4B Gemma-12BGemma-27B Llama-3.1-8B Ministral-3BMinistral-8B Ministral-14B OLMo-7B OLMo-32B Phi-4 Qwen3-4BQwen3-8B Qwen3-14BQwen3-32B Aya-8B Aya-32B Gemma-1B Gemma-4B Gemma-12B Gemma-27B Llama-3.1-8B Ministral-3B Ministral-8B Ministral-14B OLMo-7B OLMo-32B Phi-4 Qwen3-4B Qwen3-8B Qwen3-14B Qwen3-32B C. Asian/PI-associated (n= 1, 052) Aya-8B Aya-32B Gemma-1BGemma-4B Gemma-12BGemma-27B Llama-3.1-8B Ministral-3BMinistral-8B Ministral-14B OLMo-7B OLMo-32B Phi-4 Qwen3-4BQwen3-8B Qwen3-14BQwen3-32B D. NH Black-associated (n= 1, 851) Aya-8B Aya-32B Gemma-1BGemma-4B Gemma-12BGemma-27B Llama-3.1-8B Ministral-3BMinistral-8B Ministral-14B OLMo-7B OLMo-32B Phi-4 Qwen3-4BQwen3-8B Qwen3-14BQwen3-32B E. Hispanic-associated (n= 948) Aya-8B Aya-32B Gemma-1BGemma-4B Gemma-12BGemma-27B Llama-3.1-8B Ministral-3BMinistral-8B Ministral-14B OLMo-7B OLMo-32B Phi-4 Qwen3-4BQwen3-8B Qwen3-14BQwen3-32B F. NH White-associated (n= 2, 729) 0.0 0.5 1.0 Linear CKA 0.0 0.5 1.0 Linear CKA Cross-model name geometry persists within metadata strata Gender-associated name strata Race/ethnicity-associated name strata Figure 6: Cross-model CKA within name-metadata strata. All panels use names that are single tokens in every displayed open-weight tokenizer. Gender- and race/ethnicity-associated partitions preserve the same broad family structure observed in the full shared-name inventory. 0.00.10.20.30.4 Task-axis accessibility difference Fellowship / promise Hiring / competence Clinical assessment / concern Lending / trustworthiness 0.366 0.200 0.170 0.154 A. Qwen3-4B 0.000.010.020.03 Task-axis accessibility difference Fellowship / promise Hiring / competence Clinical assessment / concern Lending / trustworthiness B. Llama and Ministral (magnified scale) Llama-3.1-8B Ministral-3B Figure 7: Support-linked accessibility varies across architectures. Qwen shows positive gaps across all four task axes at a larger model-specific scale. Llama shows smaller positive gaps throughout. Ministral is positive for hiring and clinical assessment, near zero for fellowship, and negative for lending. Error bars are 95% held-out pair-bootstrap intervals. intervals. Qwen shows the largest accessibility gaps across all four tasks. Llama retains the same positive direction at a smaller scale. Ministral is more task-dependent, with positive hiring and clinical-assessment gaps, little fellowship difference, and a negative lending gap. The pooled RQ2 result should therefore be read as a support-linked pattern whose magnitude and task coverage vary across architectures. Unequal lexical support can remain visible in task-relevant representations across model families without requiring every model to express the effect at the same scale or on every task. 23 Preprint. Table 14: Architecture-specific task-relevant accessibility gaps. Entries report atomic-minus-short- fragmented weighted gaps averaged across evidence conditions, with 95% pair-bootstrap confidence intervals. Positive values indicate greater task-aligned accessibility for atomic names. ModelFellowship / promiseHiring / competenceClinical assessment / concernLending / trustworthiness Qwen3-4B0.366 [0.265, 0.472]0.200 [0.151, 0.251]0.170 [0.138, 0.203]0.154 [0.124, 0.185] Llama-3.1-8B0.026 [0.020, 0.033]0.013 [0.011, 0.016]0.004 [0.001, 0.006]0.002 [0.001, 0.004] Ministral-3B0.001 [−0.001, 0.004]0.004 [0.003, 0.005]0.004 [0.001, 0.007] −0.003 [−0.004,−0.002] Table 15: Base and post-trained models show different support-linked accessibility profiles. En- tries report atomic-minus-short-fragmented accessibility gaps averaged across all evidence conditions. Readout layers are selected separately for each model using development names and fixed before scoring unseen evaluation pairs. Model familyTraining stageLayerFellowship / promiseHiring / competenceClinical assessment / concernLending / trustworthiness Qwen3-4BBase320.0020.1491.5390.040 Qwen3-4BPost-trained180.3260.142−0.0560.046 Llama-3.1-8BBase120.0260.0130.0040.002 Llama-3.1-8BPost-trained120.0230.0160.0260.002 Gemma-3-4BBase20.098−0.0010.0090.000 Gemma-3-4BPost-trained250.0000.0000.0000.264 C.2 Base and Post-Training Comparison The tokenizer fixes how a name enters a model, while later training can change how that representation is used. We compare matched base and post-trained versions of Qwen3-4B, Llama-3.1-8B, and Gemma-3-4B. For each model, the readout layer is selected using development names and fixed before scoring unseen evaluation pairs. Llama preserves the support-linked profile.Table 15 shows that Llama is the most stable of the three families we studied. Both training stages select layer 12 and retain small positive gaps across all four task axes. Qwen redirects accessibility across tasks. Qwen changes more substantially. Its base model is dominated by a large clinical-assessment gap, while the post-trained model shifts toward fellowship, with positive hiring and lending gaps as well. The large base clinical effect appears across all three evidence conditions. Gemma reorganizes task emphasis and readout depth. Gemma shows a different pattern. Its base model is strongest on fellowship, whereas the post-trained model shows little gap on fellowship, hiring, or clinical assessment and a much larger lending effect. The selected readout layer also shifts from layer 2 to layer 25. Across the three families, post-training can preserve, weaken, or redirect support-linked concept accessibility even when the tokenizer is unchanged. Lexical access is therefore fixed at the input interface, while later training helps determine where and how support-linked task information becomes accessible inside the model. C.3 Layer Localization Support-linked accessibility is not expressed uniformly through model depth as shown in Section 4.2. Figure 8 gives the development-layer sweeps used to select the fixed readout sites. Qwen exhibits a broad mid-to-late region of strong accessibility, Llama a smaller middle-layer profile, and Ministral a weaker, less concentrated pattern. The relevant signal is therefore not tied to a common absolute depth across architectures. Selecting one layer per model preserves a common measurement site across tasks rather than choosing a different layer for each outcome. The held-out RQ2 comparison is therefore based on a model-level readout choice rather than a task-specific search for the strongest effect. 24 Preprint. 061218243036 Layer Fellowship / promise Hiring / competence Clinical assessment / concern Lending / trustworthiness Qwen3-4B 051015202530 Layer Llama-3.1-8B 04812162024 Layer Ministral-3B 0.2 0.0 0.2 Accessibility difference (atomic - short-fragmented) Figure 8: Task associations localize at model-specific depths. Rows show the four task axes over the development sweep. Dashed lines mark the single layer selected for each model and fixed before held-out evaluation. FellowshipHiringClinicalLending Qwen3-4B Llama-3.1-8B Ministral-3B Aya-Expanse-8B Gemma-3-1B Gemma-3-4B OLMo-3-7B Phi-4 79% +66% +77% +63% 87% +90% +61% +68% + 66% +59%48% -76% + 86% +86% +22% -56% tietie74% +tie 77% +30% -66% +tie 85% +70% +69% +51% - 77% +75% +54%68% + Intermediate task-axis accessibility FellowshipHiringClinicalLending Qwen3-4B Llama-3.1-8B Ministral-3B Aya-Expanse-8B Gemma-3-1B Gemma-3-4B OLMo-3-7B Phi-4 68% +35% -44%82% + 31% -21% -17% -75% + 37% -48% -46% -52% 14% -30% -62% +35% - 36% -27% -36% -43% 41% -31% -21% -49% 20% -30% -35% -61% 61%55%55%37% - Native next-adjective prediction 0% 25% 50% 75% 100% Pairs with a positive task-axis difference Figure 9: Support-linked accessibility varies across architectures and later computation. Cells report the fraction of eligible held-out pairs whose atomic name has the larger task-axis score. A plus or minus marks a mean-gap confidence interval excluding zero in that direction. Intermediate layers are selected on development names; output logits use the same next-adjective prediction position. C.4 Eight-Model Architecture Extension We broaden the RQ2 analysis to Qwen3-4B, Llama-3.1-8B, Ministral-3B, Aya-Expanse-8B, Gemma- 3-1B, Gemma-3-4B, OLMo-3-7B, and Phi-4. The extension uses the same development and held-out name partitions. Task adjectives are restricted to leading-space surfaces that remain single tokens in all eight tokenizers. One intermediate layer is selected per model using the development split and then fixed before held-out evaluation. Model-specific results use the atomic–short-fragmented pairs eligible for that tokenizer. At the intermediate readouts, 23 of 32 model–task means are positive and 20 have confidence intervals excluding zero positively. Fellowship is the broadest result, with positive intervals in seven of eight models; hiring has five, clinical assessment five, and lending three. The broader panel also reveals architecture-specific boundaries. OLMo is negative on lending, Gemma-3-4B is slightly negative on hiring, and Aya and Ministral are negative on clinical assess- ment. This heterogeneity is part of the result: name-surface support is associated with task-relevant accessibility across many model–task settings, while its strength and direction remain architecture and task dependent. Common-name sensitivity. Different tokenizers make slightly different subsets of the matched inventory eligible. To separate architecture differences from differences in which names can be compared, we repeat the intermediate-layer analysis using the same 47 held-out pairs across all eight models. The common-name analysis retains the broad architecture pattern. Fellowship and hiring remain the most consistent cross-model effects, while the principal negative architecture cases remain visible. The pattern therefore persists when every architecture is evaluated on the same set of names. 25 Preprint. Table 16: Eight-model architecture extension. Mean atomic-minus-short-fragmented accessibility gaps are reported with 95% pair-bootstrap confidence intervals. Each model is evaluated on its eligible held-out pairs (n = 63–71). Intermediate readout layers are selected using development names, while output-logit scores use the corresponding next-adjective prediction position. ModelReadoutLayerFellowship / promiseHiring / competenceClinical assessment / concern Lending / trustworthiness Qwen3-4BIntermediate300.271 [0.183, 0.360]0.212 [0.116, 0.315]0.984 [0.646, 1.356] −0.070 [−0.185, 0.034] Output logits–0.085 [0.033, 0.133] −0.069 [−0.119,−0.022] −0.019 [−0.141, 0.103]0.179 [0.131, 0.229] Llama-3.1-8BIntermediate120.029 [0.022, 0.038]0.015 [0.012, 0.018]0.006 [0.002, 0.009]0.004 [0.002, 0.005] Output logits– −0.109 [−0.170,−0.058] −0.206 [−0.293,−0.128] −0.512 [−0.648,−0.381]0.138 [0.073, 0.203] Ministral-3BIntermediate250.016 [0.006, 0.026]0.004 [−0.006, 0.013] −0.033 [−0.062,−0.005]0.024 [0.015, 0.034] Output logits– −0.061 [−0.115,−0.010] −0.067 [−0.123,−0.014] −0.153 [−0.299,−0.007]0.002 [−0.023, 0.024] Aya-Expanse-8B Intermediate80.421 [0.326, 0.516]0.494 [0.376, 0.608] −0.216 [−0.299,−0.130]0.028 [−0.030, 0.082] Output logits– −0.128 [−0.176,−0.069] −0.073 [−0.125,−0.027]0.259 [0.038, 0.496] −0.114 [−0.171,−0.057] Gemma-3-1BIntermediate5 −0.000 [−0.000, 0.000] −0.000 [−0.000, 0.000]1.238 [0.895, 1.598]0.000 [0.000, 0.000] Output logits– −0.268 [−0.436,−0.111] −0.456 [−0.622,−0.290] −0.368 [−0.570,−0.170] −0.066 [−0.155, 0.026] Gemma-3-4BIntermediate20.115 [0.082, 0.153] −0.001 [−0.002,−0.001]0.013 [0.003, 0.024]0.000 [0.000, 0.000] Output logits– −0.191 [−0.320,−0.070] −0.454 [−0.601,−0.305] −0.749 [−0.951,−0.556]0.031 [−0.010, 0.078] OLMo-3-7BIntermediate160.151 [0.118, 0.184]0.035 [0.019, 0.050]0.040 [0.018, 0.063] −0.025 [−0.048,−0.002] Output logits– −0.315 [−0.403,−0.226] −0.231 [−0.312,−0.150] −0.287 [−0.467,−0.114]0.062 [−0.008, 0.132] Phi-4Intermediate270.461 [0.344, 0.581]0.301 [0.191, 0.408]0.112 [−0.027, 0.250]0.143 [0.076, 0.212] Output logits–0.054 [−0.004, 0.106]0.012 [−0.037, 0.062]0.019 [−0.092, 0.129] −0.116 [−0.179,−0.056] Table 17: Common-name architecture sensitivity. Mean atomic-minus-short-fragmented accessi- bility gaps are reported with 95% pair-bootstrap confidence intervals for the same 47 held-out pairs that satisfy the lexical-support contrast across all eight models. Results use the development-selected intermediate readout layer fixed separately for each model. ModelLayerFellowship / promiseHiring / competenceClinical assessment / concernLending / trustworthiness Qwen3-4B300.318 [0.199, 0.432]0.253 [0.122, 0.382]1.134 [0.702, 1.647] −0.106 [−0.272, 0.038] Llama-3.1-8B120.033 [0.023, 0.045]0.016 [0.012, 0.021]0.004 [−0.000, 0.008]0.004 [0.002, 0.007] Ministral-3B250.018 [0.004, 0.033]0.004 [−0.012, 0.017] −0.066 [−0.106,−0.026]0.027 [0.014, 0.040] Aya-Expanse-8B80.464 [0.361, 0.561]0.543 [0.410, 0.683] −0.214 [−0.304,−0.127]0.002 [−0.067, 0.069] Gemma-3-1B5 −0.000 [−0.000, 0.000]0.000 [0.000, 0.000]0.976 [0.535, 1.406]0.000 [0.000, 0.000] Gemma-3-4B20.108 [0.065, 0.155] −0.002 [−0.003,−0.001]0.012 [−0.004, 0.027]0.000 [0.000, 0.000] OLMo-3-7B160.157 [0.118, 0.197]0.049 [0.029, 0.067]0.028 [0.004, 0.051] −0.037 [−0.069,−0.005] Phi-4270.540 [0.403, 0.686]0.421 [0.300, 0.549]0.063 [−0.134, 0.256]0.188 [0.101, 0.276] C.5 8B-Scale Sensitivity We also compare Qwen3-8B, Llama-3.1-8B, and Ministral-8B at a similar parameter scale. The framework follows the primary RQ2 analysis: development pairs select one layer per model, shared atomic adjective surfaces define the task axes, and held-out pairs are scored only after those choices are fixed. All four pooled task-axis gaps remain positive. The profile nevertheless shifts: clinical assessment and lending become larger, hiring remains positive, and fellowship becomes much smaller than in the primary panel. This comparison reinforces the architecture-specific nature of the effect while showing that the overall support-linked pattern is not confined to one model size. Comparable parameter scale does not force the three architectures to express concept accessibility in the same way. C.6 Support Effects Across Name-Metadata Strata The matched-pair design contains equal representation from eight race/ethnicity–gender-associated strata. Because atomic and short-fragmented names are matched within strata, the RQ2 support con- trast does not arise from comparing differently composed demographic groups. This decomposition instead asks how consistently the support-linked accessibility gap appears across the same metadata groups used in RQ1. The main-text heatmap in Figure 4 shows positive atomic–short-fragmented gaps across all eight aggregate race/ethnicity–gender-associated strata and all four task axes. As a descriptive view of model-level heterogeneity, the finer model-by-stratum gaps are positive in 22/24 cells for fellowship, 24/24 for hiring, 23/24 for clinical assessment, and 15/24 for lending. Thus, the pooled RQ2 result is not driven by a single demographic stratum, with lending showing the greatest model-specific variation. Gap magnitude is nevertheless nonuniform across metadata groups. Female-associated names show larger average gaps than male-associated names for fellowship (0.188 vs. 0.058) and hiring (0.100 26 Preprint. Table 18: 8B-scale sensitivity analysis. The analysis repeats the matched accessibility experiment with Qwen3-8B, Llama-3.1-8B, and Ministral-8B. Readout layers are selected using development pairs only and fixed before evaluation: layer 30 for Qwen, layer 12 for Llama, and layer 13 for Ministral. Positive values indicate greater task-aligned accessibility for atomic names, as in Table 2. Task axisWeighted gap95% CIPositive models Fellowship / promise0.013[0.008, 0.018]3/3 Hiring / competence0.059[0.032, 0.085]3/3 Clinical assessment / concern0.160[0.065, 0.254]2/3 Lending / trustworthiness0.072[0.030, 0.124]2/3 Table 19: Support-linked accessibility across name-metadata groups. Entries report held-out atomic-minus-short-fragmented accessibility gaps averaged across evidence conditions and the three primary models. Gender columns average across race/ethnicity-associated strata, while race/ethnicity columns average across gender-associated strata. The final column reports positive model-by- race/ethnicity-by-gender cells out of 24. Task axisStratum meanFemaleMaleAsian/PIHispanicNH BlackNH WhitePositive strata Fellowship / promise0.1230.1880.0580.0940.1280.1650.10322/24 Hiring / competence0.0720.1000.0440.0550.0660.1020.06524/24 Clinical assessment / concern0.0590.0710.0470.0520.0460.0830.05523/24 Lending / trustworthiness0.0410.0420.0410.0270.0420.0600.03515/24 vs. 0.044), with a smaller difference for clinical assessment and little difference for lending. Across race/ethnicity-associated groups, NH Black-associated names have the largest average gap on each task axis (Table 19). Table 20: Task effects by gender-associated name meta- data. Entries report support-linked accessibility gaps pooled across models and evidence conditions after averaging within matched pairs. Results are shown separately for female- and male-associated name strata. Task axisFemale-associatedMale-associated Fellowship / promise0.198 [0.141, 0.259] 0.065 [0.032, 0.103] Hiring / competence0.100 [0.074, 0.130] 0.044 [0.027, 0.065] Clinical assessment / concern 0.071 [0.051, 0.091] 0.048 [0.037, 0.060] Lending / trustworthiness0.052 [0.034, 0.070] 0.050 [0.039, 0.062] This connects RQ1 and RQ2 directly. RQ1 shows that lexical access is un- evenly allocated across race- and gender- associated name metadata. RQ2 shows that support-linked accessibility differ- ences remain visible when names are compared within those same strata, with larger magnitudes in some groups than others. Demographic structure is there- fore visible both in input-side lexical al- location and in the internal accessibility differences associated with name-surface support. Table 21: Intermediate accessibility versus output logits. Both columns apply the same task-specific adjective score at the same prediction position. The output-logit column uses the model’s native next-adjective logits, whereas the selected-layer column reports intermediate task-relevant accessibility. Task axisSelected layerOutput logits Fellowship / promise0.131 [0.097, 0.167] −0.027 [−0.066, 0.008] Hiring / competence0.072 [0.055, 0.091] −0.100 [−0.150,−0.060] Clinical assessment / concern 0.059 [0.048, 0.071] −0.213 [−0.278,−0.150] Lending / trustworthiness0.051 [0.041, 0.061]0.081 [0.052, 0.111] Marginal estimates and uncertainty. Tables 20 and 22 report the gender- and race/ethnicity-associated decompo- sitions with 95% pair-bootstrap inter- vals. Every reported marginal task-by- gender and task-by-race estimate is pos- itive, with intervals excluding zero. These marginal estimates reinforce the stratum-level pattern: support-linked ac- cessibility remains positive within each reported name-metadata group while varying in magnitude across groups. The demographic structure observed in RQ1 is therefore also visible in the task-relevant internal measurements of RQ2. C.7 Intermediate-to-Output Boundary The main RQ2 measurement is taken at a development-selected intermediate layer. To examine how the same task-axis signal changes later in computation, we apply the identical adjective score to the model’s output logits at the same prediction position. 27 Preprint. 0.30.20.10.00.1 Task-axis accessibility difference Fellowship / promise Hiring / competence Clinical assessment / concern Lending / trustworthiness Prespecified intermediate readoutNative next-adjective logits Figure 10: Later decoder computation reorganizes task-axis accessibility. Intermediate-layer gaps are positive across all four pooled task axes. At the output logits, fellowship attenuates, hiring and clinical assessment reverse, and lending remains positive. Error bars are 95% held-out pair-bootstrap intervals. Table 22: Task effects by race/ethnicity-associated name metadata. Each column contains 25 held-out matched pairs. Entries report support-linked accessibility gaps pooled across models and evidence conditions, with 95% pair-bootstrap confidence intervals within each race/ethnicity-associated stratum. Task axisAsian/PIHispanicNH BlackNH White Fellowship / promise0.098 [0.049, 0.155] 0.138 [0.051, 0.234] 0.179 [0.108, 0.250] 0.111 [0.053, 0.176] Hiring / competence0.054 [0.028, 0.085] 0.068 [0.024, 0.115] 0.101 [0.066, 0.137] 0.066 [0.037, 0.096] Clinical assessment / concern 0.052 [0.024, 0.084] 0.047 [0.026, 0.068] 0.082 [0.062, 0.104] 0.056 [0.040, 0.071] Lending / trustworthiness0.033 [0.014, 0.052] 0.054 [0.033, 0.077] 0.075 [0.051, 0.100] 0.043 [0.028, 0.058] The comparison shows that support-linked information can be clearly accessible inside the network without preserving the same signed form at the output boundary. Subsequent computation can attenuate or redirect a task-relevant intermediate signal before next-token prediction, as seen most clearly in the sign reversals for hiring and clinical assessment. Later computation can therefore preserve, weaken, or redirect the support-linked pattern. This complements the layer-localization and base/post-training analyses: lexical support enters at the input, while its task-relevant expression depends on where it is measured and how subsequent computation transforms the representation. D Additional Results for RQ3: Transfer and Downstream Leverage The observational accessibility analysis with two distinct tests as shown in RQ3. First, cross-name transfer asks whether the support-linked pattern estimated from development names predicts the corresponding pattern for unseen names. Second, downstream leverage asks whether changing the measured task direction at the name representation shifts a later constrained model choice. The first tests predictability across names; the second tests whether the measured direction is available to later model computation. D.1 Cross-Name Support-Prior Transfer For each model, task, and evidence level, we estimate a development support prior from the devel- opment pairs and apply it unchanged to the corresponding held-out evaluation pairs. As defined in Section 5.1, the support prior is the average atomic–short-fragmented accessibility gap observed on development names for a given model, task, and evidence condition. The transfer test asks how much of the unseen-name gap remains after subtracting this development estimate. If a support-linked pattern measured on one set of names predicts both the magnitude and direction of the gap on a disjoint set, then the pattern transfers across names rather than being specific to the development examples. Pooled transfer is strong across all four tasks. The pooled results show substantial cross-name transfer: the development prior accounts for 96.6% of the fellowship gap, 88.3% of hiring, 88.2% of clinical assessment, and 72.9% of lending. These are the task-level reductions reported in the main RQ3 analysis. 28 Preprint. Table 23: Cross-name transfer of the support prior. The support prior is estimated from develop- ment pairs for each model, task, evidence level, and selected readout layer, then applied unchanged to unseen evaluation pairs. Task axisRaw gapRaw 95% CIAdjusted gapAdjusted 95% CIReduction Fellowship / promise0.131[0.096, 0.170]0.004[−0.031, 0.043]96.6% Hiring / competence0.072[0.055, 0.091]0.008[−0.009, 0.027]88.3% Clinical assessment / concern0.059[0.048, 0.071] −0.007[−0.018, 0.005]88.2% Lending / trustworthiness0.051[0.041, 0.062]0.014[0.003, 0.024]72.9% 0.00.20.4 Development support prior 0.0 0.1 0.2 0.3 0.4 Evaluation raw support gap 36 model-task-evidence cells Pearson r = 0.992 Spearman = 0.959 Sign agreement 94.4% A. Disjoint-name transfer Qwen3-4B (r = 0.97) Llama-3.1-8B (r = 1.00) Ministral-3B (r = 0.88) Identity line 0.0100.0150.0200.0250.0300.0350.0400.0450.050 Mean absolute evaluation residual 0 200 400 600 800 Permutation count Matched residual 0.0094 Shuffle median 0.0364 Monte Carlo p = 0.0001 B. Task-specific shuffled-prior control Task-matched development prior Shuffled-prior median Figure 11: The support-linked component transfers to disjoint names. (A) Development support priors closely predict held-out gaps across 36 model–task–evidence cells. (B) The correctly matched prior leaves a mean absolute residual of 0.0094, compared with a median of 0.0364 after shuffling priors across task/evidence cells within model (10,000 permutations; p MC = 0.0001). Transfer remains precise at the model–task–evidence level. Figure 11 examines the relation- ship across all 36 model–task–evidence cells. Panel A compares the development prior with the corresponding held-out gap. Panel B provides a permutation baseline by shuffling the task/evidence correspondence within each model while preserving the model-specific score scale. The development prior closely tracks unseen-name gaps (r = 0.992, Spearmanρ = 0.959). The correctly matched prior leaves a mean absolute residual of 0.0094, whereas the within-model shuffled correspondence produces a substantially larger median residual of 0.0364 (p MC = 0.0001). This comparison distinguishes task-specific transfer from a generic model-level offset. Subtracting a value at the correct model scale is not sufficient; the development estimate must also correspond to the appropriate task and evidence condition. The support-linked pattern is therefore predictable across unseen names with substantial task- and condition-specific structure. D.2 Task-Direction Intervention Cross-name transfer establishes that the RQ2 accessibility pattern is predictable across names. The intervention asks a different question: is the corresponding task direction merely readable from the hidden state, or can changing that direction affect subsequent model computation? The experiment uses a separately constructed 200-pair matched-name inventory (400 names). For each model and task, the task direction is constructed from representations of aligned and opposed adjectives. Intuitively, this direction represents the internal axis between the task poles, such as greater versus lower competence or greater versus lower clinical concern. At the name span, the direction is added to the atomic-name representation and subtracted from the matched short-fragmented representation. A reverse edit swaps the directions and provides the paired comparison. The intervention uses model-specific scales selected during development. Once the task direction, layer, and scale are fixed, every intervention pair contributes a forward-minus-reverse contrast on the later constrained choice. 29 Preprint. Table 24: Task-direction intervention results across the three primary models. The task direction is added at the atomic-name span and subtracted at the matched short-fragmented span, then compared against the reverse edit. Positive contrasts indicate movement in the expected atomic-minus-fragmented choice direction. Confidence intervals are obtained by resampling complete matched-name pairs. ModelTask directionLayerαContrast95% CIPositive pairsSpecificity controls Qwen3-4BPromise / intelligence18200.153[0.147, 0.160]100.0%All three Llama-3.1-8BPromise / intelligence12200.155[0.153, 0.158]100.0%All three Ministral-3BCompetence10100.094[0.089, 0.099]97.5%Weaker at layer 10 81012 Hidden-state intervention layer 0.000 0.025 0.050 0.075 0.100 0.125 0.150 0.175 Target contrast or absolute null threshold Prespecified readout layer All specificity checks pass Random-subspace null, 95th percentile Polarity-shuffle null, 95th percentile Unrelated-axis null, maximum Prespecified competence direction Ministral: readout and intervention specificity peak at nearby layers Figure 12: Readout and intervention specificity can peak at nearby layers. The target curve shows the 200-pair competence-direction intervention in Ministral (400 names). Layer 10 is the development-selected observational readout; layer 8 is the nearby site where the target direction most clearly exceeds the random-subspace, polarity-shuffle, and unrelated-axis controls. If moving the name representation along the measured task axis systematically shifts the subsequent choice, that internal direction has downstream leverage. This is distinct from the observational RQ2 readout, which measures whether the direction is accessible without modifying the hidden state. Specificity controls. We compare the target task direction against three control families. An unrelated-axis control uses a direction from a different task, testing whether any semantically meaningful direction produces the same effect. A polarity-shuffled control disrupts the aligned- versus-opposed assignment while preserving the task vocabulary. A random-subspace control uses directions drawn from matched random subspaces, testing whether the effect follows merely from moving the hidden state by a similar magnitude in an arbitrary direction. Targeted edits shift later choices across all three models. At the selected layers, the forward- minus-reverse contrast is 0.153 for Qwen, 0.155 for Llama, and 0.094 for Ministral. All 200 Qwen pairs and all 200 Llama pairs move in the expected direction, as do 195 of 200 Ministral pairs. The measured task direction therefore has downstream leverage across all three primary models under the controlled choice setting. The effect is specific to the measured task direction.Qwen and Llama show clear task-direction specificity relative to unrelated-axis, polarity-shuffled, and random-subspace controls. Ministral shows the same positive intervention effect, with its clearest specificity slightly earlier in the network. The layerwise Ministral analysis is examined next. The intervention therefore provides information beyond the observational RQ2 score. NAMETRACE first identifies a task-relevant direction readable from an intermediate representation; RQ3 then shows that targeted movement along that direction can change a later constrained decision. 30 Preprint. Short NameModel NameModel / Training StageLicenseHugging Face Model ID GPT-4 tokenizerGPT-4 tokenizerTokenizer-onlyProprietary/API tokenizer gpt-4 via tiktoken; no HF model ID GPT-5 tokenizerGPT-5 tokenizerTokenizer-onlyProprietary/API tokenizer gpt-5 via tiktoken; no HF model ID gpt-oss 120BGPT-OSS 120BReasoning-oriented / post-trained Apache-2.0 openai/gpt-oss-120b gpt-oss 20BGPT-OSS 20BReasoning-oriented / post-trained Apache-2.0 openai/gpt-oss-20b Aya 8BAya Expanse 8BPost-trainedCC-BY-NC-4.0 + C4AI AUP CohereLabs/aya-expanse-8b Aya 32BAya Expanse 32BPost-trainedCC-BY-NC-4.0 + C4AI AUP CohereLabs/aya-expanse-32b Gemma 1BGemma 3 1B PTBase / pretrainedGemma license google/gemma-3-1b-pt Gemma 4B PTGemma 3 4B PTBase / pretrainedGemma license google/gemma-3-4b-pt Gemma 4B ITGemma 3 4B ITInstruction-tuned / post-trainedGemma license google/gemma-3-4b-it Gemma 12BGemma 3 12B PTBase / pretrainedGemma license google/gemma-3-12b-pt Gemma 27BGemma 3 27B PTBase / pretrainedGemma license google/gemma-3-27b-pt Llama 3.1 8BLlama 3.1 8BBase / pretrainedLlama 3.1 Community License meta-llama/Llama-3.1-8B Llama 3.1 8B Instruct Llama 3.1 8B InstructInstruction-tuned / post-trainedLlama 3.1 Community License meta-llama/Llama-3.1-8B-Instruct Llama 3.1 70BLlama 3.1 70BBase / pretrainedLlama 3.1 Community License meta-llama/Llama-3.1-70B Ministral 3B BaseMinistral 3 3B Base 2512Base / pretrainedApache-2.0 mistralai/Ministral-3-3B-Base-2512 Ministral 3B InstructMinistral 3 3B Instruct 2512 Instruction-tuned / post-trainedApache-2.0 mistralai/Ministral-3-3B-Instruct-2512 Ministral 8BMinistral 3 8B Base 2512Base / pretrainedApache-2.0 mistralai/Ministral-3-8B-Base-2512 Ministral 14BMinistral 3 14B Base 2512Base / pretrainedApache-2.0 mistralai/Ministral-3-14B-Base-2512 OLMo 7BOLMo 3 1025 7BBase / pretrainedApache-2.0 allenai/Olmo-3-1025-7B OLMo 32BOLMo 3 1125 32BBase / pretrainedApache-2.0 allenai/Olmo-3-1125-32B Phi-4Phi-4Post-trainedMIT microsoft/phi-4 Qwen3 4BQwen3 4BPost-trainedApache-2.0 Qwen/Qwen3-4B Qwen3 4B BaseQwen3 4B BaseBase / pretrainedApache-2.0 Qwen/Qwen3-4B-Base Qwen3 4B InstructQwen3 4B Instruct 2507Instruction-tuned / post-trainedApache-2.0 Qwen/Qwen3-4B-Instruct-2507 Qwen3 8BQwen3 8BPost-trainedApache-2.0 Qwen/Qwen3-8B Qwen3 14BQwen3 14BPost-trainedApache-2.0 Qwen/Qwen3-14B Qwen3 32BQwen3 32BPost-trainedApache-2.0 Qwen/Qwen3-32B DeepSeek V3.2DeepSeek V3.2Post-trainedMIT deepseek-ai/DeepSeek-V3.2 Table 25: Model and tokenizer metadata for checkpoints used across NAMETRACE experiments. The table reports the shortened names used in figures and tables, corresponding model names, model or training stage, license, and source identifier. Base / pretrained denotes checkpoints before instruction or other post-training, while post-trained includes instruction-tuned or otherwise post-trained checkpoints. Hugging Face identifiers refer to the public or gated repositories used for the open-weight models. GPT-4 and GPT-5 are tokenizer-only tiktoken encodings and therefore have no corresponding Hugging Face model checkpoint. D.3 Intervention Localization Ministral provides a useful view of the distinction between readout accessibility and intervention leverage. Its primary observational readout is layer 10, where the intervention remains positive. The same competence direction separates most clearly from the reported specificity controls at nearby layer 8. This separation clarifies an important distinction. A direction can be easiest to read out at one layer without having its greatest intervention leverage at exactly the same depth. The layer at which a concept is most clearly measurable therefore need not be the layer at which changing that concept most strongly affects later computation. For Ministral, the task direction remains measurable and intervention-relevant around the selected region, but the strongest separation from the reported controls appears at layer 8 rather than the observational readout layer 10. This is consistent with the RQ3 framing: accessibility and leverage are related but distinct measurements. E Broader Significance Names are simultaneously social signals and model-specific lexical objects. The results across RQ1–RQ3 show why both properties matter. RQ1 demonstrates that direct lexical access is unevenly allocated across first names and across race- and gender-associated name metadata. RQ2 shows that this input-side difference remains visible in task-relevant internal representations even when atomic and short-fragmented names are matched within demographic strata. RQ3 shows that the support-linked pattern transfers to unseen names and that the corresponding task directions have downstream leverage. Figure 13 summarizes how these findings connect across the model lifecycle. These findings suggest a simple principle: Lexical comparability should be checked when the tokenizer is designed, tracked across model training, and controlled at evaluation time. For model builders. Tokenizer design determines which name surfaces receive direct lexical support. RQ1 shows that this support is selective and demographically structured. After pretraining, NAMETRACE can identify where support-linked differences become task relevant inside the model. 31 Preprint. 1 Tokenizer design Measurelexicalaccess Check allocationacross name groups 2 After Pretraining Measure concept accessibility Locate support- linked differences 3 After Post-training Preserve / weaken Redirect / reorganize 4 Evaluation Balance by support Stratify by support Report sensitivity Behavioral evaluation (what users see) Interpret output differences only after checking lexical comparability Lexical comparability should be checked when the tokenizer is designed, tracked across model training and controlled at evaluation time. Figure 13: Where lexical comparability matters. NAMETRACE turns name-surface support into a measurable lifecycle check for model development and evaluation. Tokenizer design determines direct lexical access, pretraining and post-training shape how support-linked differences become accessible inside the model, and evaluation should make these differences explicit before interpreting behavioral outcomes. Matched names may therefore still be lexically unmatched. The base and post-training comparison in Appendix C.2 further shows that later training can preserve, weaken, or redirect the accessibility profile even when the tokenizer is unchanged. Lexical support can therefore be tracked across the model lifecycle. Before pretraining, tokenizer design determines direct lexical access. After pretraining, model-internal evaluation can reveal which task-relevant concepts are accessible from those representations. After post-training, the same measurement can show whether the accessibility profile is preserved or reorganized. For evaluators.A name-based benchmark is not automatically lexically controlled across models. The same surface may be atomic in one tokenizer and fragmented in another, so holding the name fixed does not necessarily hold lexical access fixed. Within a single model, socially matched names can likewise differ in lexical support. This is the practical implication of the paper’s central observation: matched names are not necessarily matched inputs. Evaluators can inspect each target name under the relevant tokenizer and treat name-surface support as a separate evaluation variable alongside demographic metadata, frequency, and length. When support is uneven, results can be balanced by support, stratified by support, or accompanied by sensitivity analyses using support-linked estimates from disjoint names. These steps make the lexical structure of the evaluation explicit while preserving the demographic comparison of interest. More broadly, name-based evaluation can trace the signal beyond the input surface. Tokenization determines how directly a name enters the model, training shapes what becomes accessible from that representation, and later computation can preserve, weaken, or redirect how that information appears toward the output. NAMETRACE provides a way to examine these stages, making lexical comparability a measurable part of model development and evaluation. F Limitations NAMETRACE is designed to study whether socially comparable name surfaces receive comparable lexical support and whether support-linked differences remain visible in task-relevant model com- putation. Accordingly, our conclusions concern lexical comparability, internal accessibility, and downstream computational leverage rather than broader demographic outcomes. Name metadata and coverage.The race/ethnicity- and gender-associated variables are aggregate properties of name surfaces, not identity labels for individuals, and are used for matching, stratification, and descriptive analysis. We focus on sufficiently frequent, single-word ASCII first names with reliable metadata associations to support consistent controlled comparisons. This design prioritizes comparability across names and models; other naming conventions, languages, scripts, and cultural contexts provide natural extensions. First names are one form of identity-related input, and other name forms or social cues may exhibit different lexical-support patterns. 32 Preprint. Controlled lexical comparison.We compare names that are atomic in all three primary tokenizers with names short-fragmented in all three, while matching on frequency, character length, demographic- association strength, metadata confidence, and weak orthographic cues. Restricting fragmented names to two or three tokens keeps the contrast focused on direct versus ordinary composed lexical access rather than extreme tokenization. Because name-specific pretraining exposure is not directly observed, we interpret lexical support as a measurable predictor among names comparable on major observed properties, rather than as an isolated causal treatment. Models and task-axis measurement.The allocation analysis spans 12 LLM-associated tokenizer rows, while the representation experiments use three primary model families and extend to an eight- model panel. Variation in effect magnitude, task coverage, and layer localization across architectures is part of the empirical finding. NAMETRACE measures task-relevant concept accessibility through compact adjective axes selected automatically from development prompts and frozen before held-out evaluation. This provides a fine-grained, interpretable, model-native readout of the targeted concept while preserving a common measurement framework across tasks. Internal accessibility and downstream behavior.NAMETRACE is intentionally pre-behavioral: it measures task-relevant information before unrestricted generation. Intermediate accessibility can be preserved, attenuated, or redirected by later computation, and the hidden-state intervention tests whether the measured task direction has downstream leverage under a controlled choice setting. These analyses distinguish internal accessibility from final behavior and clarify where support-linked differences remain available to model computation. Open-ended interaction offers a complementary behavioral setting for studying how such internal differences are expressed at the output. 33