Nationality encoding in language model hidden states: Probing culturally differentiated representations in persona-conditioned academic text

Personas, identity, and agents2026arXivApproved editorial review

Authors: Paul Jackson, Ruizhe Li, Elspeth Edelstein

Keywords: Persona conditioning, Activation steering

Source: Open primary source (opens in a new tab)

3
Authors
10
Findings
23
Limitations
5
Evidence

Editorial summary

English

Jackson, Li and Edelstein ask whether Gemma-3-4b-it hidden states distinguish academic text generated under British and Chinese personas. They cross 45 templates with six conditions combining nationality, instructional medium, and role, producing 270 introductions of 149-261 words. Logistic probes use generated-token states from all 35 layers; a fixed 80/20 split places 216 texts in cross-validated layer selection and 54 in one test evaluation. Controls include shuffled labels, TF-IDF, cross-prompt-family transfer, and probes for medium and role.

The main result is very high decodability: nationality reaches .968 ± .031 cross-validated accuracy at layer 18 and 54/54 on the holdout. Shuffled labels yield .499, while TF-IDF reaches .371 in cross-validation and .463 on test. Cross-family transfer ranges from .589 to .933, with five of six directions above .844. This shows that a signal not reducible to surface unigrams is linearly recoverable in this run. British or Chinese, however, is explicitly written in every instruction. The probe may recover contextual persistence of that label and repeated persona-clause wording rather than culture independently learned by the model.

The linguistic analysis selects positions by their own discriminatory scores: the top 5% of five-token windows and top 2.5% of individual tokens, with thresholds calibrated at layer 24. Filtering leaves 6,961 windows and 2,353 tokens. Selected windows show more premodification under Chinese personas and more postmodification, hedging, boosting, and passive voice under British personas; highlighted lexical domains also differ. Structural effects are small, and the tests treat thousands of overlapping rows repeated across layers and nested in only 270 texts as independent. Selecting rows with a nationality classifier and then testing those rows for nationality differences also creates selection circularity.

The decisive check is the full-surface baseline: across 1,734 sentences there are no significant nationality differences in phrase, UPOS, predicate, hedges, boosters, modals, or passive voice. This does not prove cultural neutrality, but it limits practical inference: effects occur at probe-selected positions rather than as general differences in measured outputs. There is no activation intervention, human-authored comparator, cultural-authenticity rating, pedagogical outcome, or direct measure of cultural hollowness. Linear decodability does not establish causality or educational harm.

Cultural validity needs further caution. Two nationality labels compress heterogeneous populations, institutions, disciplines, and language histories. The six profiles mix nationality with medium and role, omit their exact clauses, and probes also recover medium (.870 test) and role (.963), showing shared prompt structure remains available. Nationality is null within the English-medium cohort and appears within the Chinese-medium cohort with small cells. Several odds ratios also lack a clear reference category and appear directionally inconsistent with displayed proportions.

The work is a 43-page v2 preprint using one checkpoint, greedy generation, and one output per prompt. It omits the split seed, exact model revision, and a reproducible environment. The paper mentions a run stored locally or on Google Drive but releases no corpus, prompts, persona clauses, activations, splits, code, Stanza corrections, lexicons, or outputs. Its supported contribution is an exploratory demonstration that an explicit nationality condition remains decodable in one model's hidden states, not evidence that validated British and Chinese cultural mechanisms exist inside the model.

Español

Jackson, Li y Edelstein preguntan si los estados ocultos de Gemma-3-4b-it permiten distinguir textos académicos generados bajo personas británicas y chinas. Cruzan 45 plantillas con seis condiciones que combinan nacionalidad, medio de instrucción y rol, y producen 270 introducciones de 149 a 261 palabras. Sobre los estados de los tokens generados en las 35 capas entrenan probes logísticos; una partición fija 80/20 usa 216 textos para selección por validación cruzada y 54 para un único test. También comparan etiquetas permutadas, TF-IDF, transferencia entre familias de prompts y probes de medio y rol.

El resultado principal es una decodificación muy alta: nacionalidad alcanza accuracy de validación cruzada .968 ± .031 en la capa 18 y 54/54 en el holdout. El control con etiquetas permutadas queda en .499 y TF-IDF obtiene .371 en validación y .463 en test. La transferencia entre familias va de .589 a .933, con cinco de seis direcciones por encima de .844. El resultado muestra que una señal no reducible a unigramas superficiales es linealmente recuperable en este run. Sin embargo, British o Chinese está escrito explícitamente en cada instrucción: el probe puede estar leyendo persistencia contextual de esa etiqueta y de las mismas cláusulas de persona, no una cultura aprendida de forma independiente.

El análisis lingüístico selecciona posiciones por su propio score discriminativo: top 5% de ventanas de cinco tokens y top 2,5% de tokens individuales, con umbrales calibrados en la capa 24. Tras filtros quedan 6.961 ventanas y 2.353 tokens. En las ventanas seleccionadas aparecen más premodificación bajo persona china y más postmodificación, hedging, boosting y pasiva bajo persona británica; los dominios léxicos destacados también difieren. Pero la magnitud estructural es pequeña y las pruebas tratan miles de filas solapadas, repetidas entre capas y anidadas en solo 270 textos como si fueran independientes. Además, elegir filas por un clasificador de nacionalidad y después probar diferencias por nacionalidad introduce circularidad de selección.

La comprobación decisiva es el baseline de superficie completa: sobre 1.734 oraciones no hay diferencias significativas en frase, UPOS, predicado, hedges, boosters, modales o pasiva. Esto no prueba neutralidad cultural, pero sí limita la inferencia práctica: los efectos aparecen en posiciones elegidas por el probe, no como diferencias generales en los textos medidos. Tampoco hay intervención sobre activaciones, comparación con autores humanos, ratings de autenticidad cultural, outcomes pedagógicos ni evaluación de «cultural hollowness». La decodificación lineal no demuestra causalidad ni daño educativo.

La validez cultural exige todavía más cautela. Dos etiquetas nacionales sintetizan poblaciones, instituciones, disciplinas e historias lingüísticas heterogéneas. Los seis perfiles mezclan nacionalidad con medio y rol, no publican sus cláusulas exactas y los probes también recuperan medio (.870 en test) y rol (.963), señal de que la estructura compartida del prompt sigue presente. Dentro del cohort inglés la diferencia por nacionalidad es nula; dentro del chino aparece con celdas pequeñas. Los odds ratios de algunas tablas tampoco dejan clara su categoría de referencia y parecen invertir la dirección de las proporciones.

El trabajo es un preprint v2 de 43 páginas y usa un solo checkpoint, generación greedy y una única salida por prompt. No informa seed de la partición, revisión exacta del modelo ni entorno reproducible. El paper menciona un run guardado localmente o en Google Drive, pero no publica corpus, prompts, cláusulas, activaciones, splits, código, correcciones Stanza, lexicones ni outputs. La contribución real es una demostración exploratoria de que la condición nacional explícita permanece decodificable en estados ocultos de un modelo; no evidencia de que «lo británico» y «lo chino» sean mecanismos culturales internos validados.

Research question

Can a linear probe recover the British or Chinese national condition written in the prompt from hidden states of Gemma-3-4b-it during the generation of academic introductions, and what linguistic features appear at the positions with the highest discriminative score?

Method

Probing study on a synthetic corpus of 270 texts: 45 templates across six person conditions. Activations of generated tokens are averaged in each of 35 layers and L2 logistic regression with standardization is fitted. A fixed 216/54 partition selects the layer by cross-validation and evaluates once on test. Controls: permutation, selectivity, TF-IDF, transfer between families, medium and role probes, and sentence-level baseline. The highest-score positions are annotated with Stanza and manual lexicons and compared with frequency and rank tests.

Sample: A single Gemma-3-4b-it generates 270 texts: 45 templates crossed with six conditions, 135 per national label. Length 149-261 words, mean 210. The split contains 216 training/validation texts and 54 test texts, but likely shares template identities. From 8,015 windows and 2,921 candidate tokens, 6,961 and 2,353 remain after filters; these are nested and overlapping observations, not independent samples.

Findings

  • The nationality probe reaches .968 ± .031 accuracy in cross-validation at layer 18 and 54/54 in the single holdout of 54 texts.
  • The two-sided 95% exact binomial interval for 54/54 is approximately .934-1.000; the observed perfect result retains uncertainty.
  • The control with permuted labels obtains .499 and selectivity .469; TF-IDF obtains .371 in CV and .463 in test.
  • Transfer between prompt families ranges from .589 to .933; five of six directions exceed .844 and base-to-theory is the weak exception.
  • Medium of instruction is also decodable, .884 CV and .870 test, and role reaches .940 CV and .963 test.
  • In selected windows, the Chinese condition shows more premodification; the British condition shows more postmodification, hedging, boosting, and passive voice, with small effects and several adjusted p close to .05.
  • In individual tokens, pre- and postmodification do not repeat; more nominal predicates appear under the Chinese condition and more adverbial slots under the British condition.
  • The baseline of 1,734 sentences finds no significant differences in any structural or stance family measured.
  • Peaks per layer do not coincide: centroids at 18, token and domain separation earlier, and some structural features in early or late layers.
  • The published corpus, window, token, sentence, layer, and cohort counts close arithmetically.

Limitations

  • Nationality is an explicit label of the prompt; recovering it may reflect instruction retention and not internalized culture.
  • Controls with neutral behavior, arbitrary or swapped labels, omitted nationality, prompt-state baseline, and human texts are missing.
  • A linear probe demonstrates decodability, not causal use of the direction; there is no activation intervention.
  • The split is by texts and not grouped by template, so the same template may appear in train and test under different conditions.
  • Transfer between families retains the same person clauses and does not prove generalization to novel national formulations.
  • Seed, exact number of folds, split identifiers, and partition repetitions are not published.
  • The holdout contains only 54 texts and is evaluated once; 54/54 does not imply universal accuracy of 1.
  • Positions are selected by the score of a nationality classifier and then tested by nationality, introducing circular selection.
  • The 6,961 windows and 2,353 tokens are nested within 270 texts, overlap, and repeat across six layers, but the tests do not model this dependence.
  • There is no bootstrap by text/template or mixed model; significance may be overestimated by pseudoreplication.
  • It is not clear that threshold selection and linguistic inference are performed exclusively on untouched test data.
  • 35 layers, six token layers, two samples, multiple families, and thresholds are explored; Bonferroni covers only subsets of the analysis.
  • The top 2.5% and top 5% thresholds are calibrated at layer 24 without sensitivity analysis and despite the best centroid being at layer 18.
  • Some odds ratios appear to use different or undeclared orientations and contradict the direction of the proportions shown.
  • Global structural effects are small and several stance results barely exceed the correction.
  • The ordinary sentence-level analysis is completely null for the variables measured, so the visible implications are uncertain.
  • British and Chinese reduce heterogeneous cultures and populations to two synthetic cues without participatory or expert validation.
  • The cohorts combine medium and role and do not publish the six exact clauses; the design does not cleanly isolate each factor.
  • The manual Stanza audit does not report sample, annotators, agreement, accuracy before/after, or a complete correction log.
  • The semantic lexicons incorporate researcher knowledge and are not public.
  • Only one model, one checkpoint, and one greedy output per prompt are tested; there is no replication by seed, sampling, model, or version.
  • Exact review of Gemma and versions of Python, Transformers, PyTorch, CUDA, Stanza, and their models is missing.
  • Corpus, prompts, clauses, activations, splits, code, annotations, lexicons, and intermediate results are not published.

What the study does not establish

  • It does not demonstrate that Gemma has learned authentic or independent representations of British and Chinese cultures.
  • It does not separate the retention of an explicit national label from an emergent cultural representation.
  • It does not demonstrate that the linear direction of the probe causes linguistic decisions or is used by the model.
  • It does not establish general surface differences between the texts; the sentence-level baseline is null.
  • It does not validate that the selected features are culturally authentic, desirable, or representative of real people.
  • It does not measure cultural hollowness, academic quality, learning, pedagogical harm, or user outcomes.
  • It does not allow inferring causality of nationality because person, medium, role, and prompt wording are intertwined.
  • It does not justify treating thousands of tokens and windows as independent replicates of 270 texts.
  • It does not prove generalization to other models, checkpoints, domains, languages, prompts, seeds, or human authors.
  • It does not allow independent reproduction of generations, activations, split, annotations, or statistics.

Traceability

Scope: Full text

Version: arXiv:2604.10151v2, 43-page preprint; no peer-reviewed venue, DOI or public artifact release identified

Consulted source: https://arxiv.org/abs/2604.10151

Review: Codex 43-page visual full-text, prompt-retention, probing, template-split, selection-circularity, clustered-dependence, annotation, cultural-construct, artifact and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Gemma-3-4b-it

Instruments and metrics

  • L2 logistic-regression linear probes over 35 hidden layers
  • Shuffled-label permutation baseline with 100 permutations
  • Selectivity score
  • TF-IDF unigram-bigram logistic classifier
  • Cross-prompt-family transfer tests
  • Five-token window and single-token decision scores
  • Stanza structural and stance annotation
  • Manually curated semantic-domain lexicons
  • Chi-square and Cramer's V
  • Fisher exact test
  • Mann-Whitney U and rank-biserial effect size
  • Within-family Bonferroni correction

Data used

  • 270 synthetic research-article introductions
  • 45 prompt templates across base, alternate and theory families
  • Six British/Chinese persona conditions combining medium and role
  • 6,961 retained high-signal five-token windows
  • 2,353 retained high-signal single tokens
  • 1,734 sentences in the full-surface baseline

Evidence and location

  • Metadata, design, results, limitations, ethics, and appendices: arXiv:2604.10151v2; 43 pages rendered and inspected; PDF SHA-256 eebf3508fbba4f76a6050a3c67cdfaee1d62bd1a7ad513132b73229e526f07dc
  • Extracted text and count checks: Local full-text SHA-256 86b56d479ab4ea57ffbb3ed4f45ee5574ed628b2d6a5fde72535c0577d04b375; corpus, layer, cohort, token and sentence counts independently summed
  • Absence of public repository and artifacts: Exact-title GitHub repository search and exact run-20260325-110501 code search returned zero matches on 2026-07-17; no artifact URL appears in the paper
  • Uncertainty of 54/54: Exact two-sided 95% Clopper-Pearson interval computed as .933968-1.000000 for 54 successes in 54 trials
  • Integral audit of explicit prompting, selection, dependence, cultural construct, and reproducibility: reports/verification/article-369-nationality-prompt-retention-probe-selection-pseudoreplication-cultural-construct-artifact-and-reproducibility-audit.json