When LLMs Imagine People: A Human-Centered Persona Brainstorm Audit for Bias and Fairness in Creative Applications

Applications, bias, and safety2026arXivApproved editorial review

Authors: Hongliu Cao, Eoin Thomas, Rodrigo Acuna Agost

Keywords: Large Language Models, Personality, Bias, Fairness, Persona

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

3
Authors
20
Findings
70
Limitations
11
Evidence

Editorial summary

English

The paper proposes the Persona Brainstorm Audit (PBA), a procedure for detecting potentially stereotypical associations in LLM-generated persona lists. This review uses arXiv v2, revised on 24 February 2026: a 27-page ACM manuscript that still contains the template text “Make sure to enter the correct conference title” and a placeholder DOI. No definitive publication or official repository was found; the paper itself postpones release of profiles, mappings, and code until publication.

The base prompt asks for “20 diverse user profiles” in JSON per call, with eight fields: name, gender, ethnicity, sexual orientation, social class, education, occupation, and top personal interest. The study nominally generates 10,000 profiles from each of 12 models: nine OpenAI aliases, GPT-3.5, GPT-4, GPT-4o, three GPT-4.1 variants, and three GPT-5 variants, and three Mistral releases. Temperature is 1. Values are lowercased, stemmed, deduplicated, and consolidated into categories by GPT-5, reportedly followed by human validation. The reduction is extreme: for example, GPT-5's 1,128 ethnicity strings become six categories; 689 education strings become five; 1,384 occupations become 18; and 1,501 interests become 15. Only the 50 most frequent names across models are retained for name-based analysis.

PBA crosses four identity axes, name, gender, ethnicity, and sexual orientation, with four social dimensions, class, education, occupation, and interest, builds 16 contingency tables, and computes Cramér's V. It then transforms V through degree-of-freedom-dependent thresholds into severity scores: 0–0.33 small, 0.33–0.66 medium, 0.66–1 high, and >1 very high. The exact algorithm for this second normalization is not specified mathematically. Cramér's V measures association, not harm, discrimination, or unfairness; an association may be semantically expected or vary with margins and analyst-created categories. The paper also calls the normalized metric “bounded” although its results exceed 1.

In the main analysis, GPT-4o has the lowest mean score (0.678), followed by Ministral-3B (0.699), GPT-4 and GPT-5 nano (both 0.707). GPT-4.1 mini (0.969), GPT-5 (0.950), GPT-5 mini (0.920), and Mistral-medium (0.893) rank highest. Name associations are largest on average (1.113), while occupation and interest concentrate high scores. Heatmaps show non-heterosexual identities clustering in Creative & Design and appearing less in Engineering, gender-linked occupations, and differences when gender × sexual orientation × social class are combined. The nurse drill-down assigns more than 90% of nurses to women in almost every model; GPT-4.1 nano reaches 100%. These distributions document generated patterns worth inspecting, but the aggregate score alone does not determine which patterns are harmful.

The paper presents GPT-3.5 → GPT-4 → GPT-4o → GPT-4.1 → GPT-5 as a longitudinal trajectory and concludes that bias declines and resurfaces. Without exact snapshots, inference dates, or comparable checkpoints, this is a cross-sectional comparison of distinct products, not controlled longitudinal tracking of one system. Paired t-tests treat the 16 correlated dimensions as replicates when comparing models, with unclear independence. The analyzed sample is also not exactly 120,000 after deduplication: 3,778 profiles are removed, leaving 116,222; Mistral-medium loses 2,060 (20.6%) compared with only 3 for GPT-5. Unequal deduplication changes the distributions and removes a direct signal of model collapse or concentration.

The sample-size analysis compares 5k, 10k, 15k, and 20k profiles on six models and reports high ICCs and correlations for most dimensions, but it does not state whether samples are independent or nested, and provides no repeated samples or intervals. Under the UX-researcher role, most severity labels persist. Under the debiasing prompt, stability drops sharply: recomputing the 16 rows in Table 5 gives means ICC(C,1)=0.381, ICC(A,1)=0.361, Spearman=0.422, Kendall=0.306, and severity difference=0.312. The printed “Mean” row, 0.320, 0.319, −0.200, −0.200, 0.167, is not the mean of the rows. Nevertheless, the abstract and conclusion claim stability under debiasing prompts. The base prompt already requires “diverse” profiles, so the intervention is not compared with a neutral condition; and a lack of p<0.05 with few models does not demonstrate no effect or equivalence.

Human validation is exploratory: nine evaluators compare only two Gender × Occupation charts selected because GPT-5 mini scores 1.134 and GPT-5 nano 0.779. Everyone judges the former more biased. This provides face validity for one obvious contrast, but does not validate all 16 dimensions, the thresholds, degree-of-freedom comparability, or regulatory use. The paper does not report ethics review, consent, exact DEI/non-DEI composition, compensation, order randomization, or inter-rater agreement.

The defensible contribution is a simple protocol for generating open categorical data and locating associations for human auditors to investigate, with useful macro-to-micro drill-down examples. It does not establish that the normalized scale validly and comparably measures fairness, that it is robust to debiasing prompts, or that it supports model rankings, deployment thresholds, or certification. PBA should be presented as an exploratory signal-detection instrument rather than a validated fairness metric.

Español

El artículo propone Persona Brainstorm Audit (PBA), un procedimiento para detectar asociaciones potencialmente estereotípicas en listas de personas generadas por LLM. Esta revisión usa arXiv v2, revisado el 24 de febrero de 2026: un manuscrito ACM de 27 páginas que todavía conserva el texto de plantilla «Make sure to enter the correct conference title» y un DOI ficticio. No se encontró publicación definitiva ni repositorio oficial; el propio artículo pospone la liberación de perfiles, mappings y código hasta la publicación.

El prompt base pide en cada llamada «20 diverse user profiles» en JSON, con ocho campos: nombre, género, etnia, orientación sexual, clase social, educación, ocupación e interés principal. Se generan nominalmente 10.000 perfiles por cada uno de 12 modelos: nueve aliases de OpenAI, GPT-3.5, GPT-4, GPT-4o, tres GPT-4.1 y tres GPT-5, y tres versiones Mistral. La temperatura es 1. Los valores se pasan a minúsculas, se someten a stemming, se deduplican y luego GPT-5 los consolida en categorías, supuestamente con validación humana. La reducción es extrema: por ejemplo, 1.128 etiquetas de etnia de GPT-5 terminan en seis categorías; 689 niveles educativos, en cinco; 1.384 ocupaciones, en 18; y 1.501 intereses, en 15. Para nombre solo se conservan los 50 más frecuentes del conjunto de modelos.

PBA cruza cuatro ejes tratados como identidad, nombre, género, etnia y orientación sexual, con cuatro dimensiones sociales, clase, educación, ocupación e interés, construye 16 tablas de contingencia y calcula Cramér V. Después transforma V mediante umbrales dependientes de los grados de libertad en una escala de severidad: 0–0,33 pequeña, 0,33–0,66 media, 0,66–1 alta y >1 muy alta. El algoritmo exacto de esta segunda normalización no está especificado matemáticamente. Cramér V mide asociación, no perjuicio, discriminación ni falta de equidad; una asociación puede ser semánticamente esperable o variar por los márgenes y por las categorías creadas. Además, el artículo llama «bounded» a la métrica normalizada aunque sus resultados superan 1.

En el análisis principal, GPT-4o obtiene la media menor (0,678), seguido de Ministral-3B (0,699), GPT-4 y GPT-5 nano (ambos 0,707). GPT-4.1 mini (0,969), GPT-5 (0,950), GPT-5 mini (0,920) y Mistral-medium (0,893) quedan arriba. Las asociaciones con nombre son las mayores en promedio (1,113), y ocupación e interés concentran los valores altos. Los heatmaps muestran concentración de identidades no heterosexuales en Creative & Design y baja presencia en Engineering, asociaciones por género en ocupaciones y diferencias al combinar género × orientación sexual × clase. El ejemplo fino de enfermería asigna más del 90 % a mujeres en casi todos los modelos; GPT-4.1 nano llega al 100 %. Estas distribuciones sí documentan patrones generados que merecen inspección, pero la puntuación agregada no determina por sí sola cuáles son dañinos.

El artículo presenta GPT-3.5 → GPT-4 → GPT-4o → GPT-4.1 → GPT-5 como una trayectoria longitudinal y concluye que el sesgo baja y reaparece. Sin snapshots exactos, fechas de inferencia ni checkpoints comparables, esto es una comparación transversal de productos distintos, no seguimiento controlado del mismo sistema. Las pruebas t pareadas tratan las 16 dimensiones correlacionadas como réplicas para comparar modelos, lo que no satisface claramente la independencia. El tamaño publicado tampoco es exactamente 120.000 tras deduplicar: se eliminan 3.778 perfiles y quedan 116.222; Mistral-medium pierde 2.060 (20,6 %) frente a solo 3 de GPT-5. La deduplicación desigual cambia las distribuciones y elimina precisamente una señal de colapso o concentración del modelo.

La sensibilidad a tamaño compara 5k, 10k, 15k y 20k en seis modelos y obtiene ICC y correlaciones altas en la mayoría de dimensiones, pero no se indica si son muestras independientes o subconjuntos anidados, no hay réplicas ni intervalos. Con el rol de UX researcher, casi todas las severidades se mantienen. Con el prompt de debiasing, la estabilidad cae mucho: recalculando las 16 filas de la Tabla 5, las medias son ICC(C,1)=0,381, ICC(A,1)=0,361, Spearman=0,422, Kendall=0,306 y diferencia de severidad=0,312. La fila «Mean» impresa, 0,320, 0,319, −0,200, −0,200, 0,167, no es la media de las filas. Pese a ello, abstract y conclusión afirman estabilidad bajo prompts de debiasing. Además, el prompt base ya exige perfiles «diversos», por lo que la intervención no contrasta una condición neutral; y ausencia de p<0,05 con pocos modelos no demuestra ausencia de efecto ni equivalencia.

La validación humana es exploratoria: nueve evaluadores comparan solo dos gráficos de Gender × Occupation seleccionados porque GPT-5 mini puntúa 1,134 y GPT-5 nano 0,779. Todos ven más sesgo en el primero. Esto aporta validez aparente para ese contraste evidente, pero no valida las 16 dimensiones, los umbrales, la comparabilidad entre grados de libertad ni el uso regulatorio. No se reportan protocolo ético, consentimiento, composición exacta DEI/no-DEI, compensación, aleatorización del orden o acuerdo entre evaluadores.

La contribución defendible es un protocolo sencillo para producir datos categóricos abiertos y localizar asociaciones que un auditor humano puede investigar, con ejemplos útiles de zoom macro/micro. No queda establecido que la escala normalizada mida fairness de forma válida y comparable, que sea robusta a prompts de debiasing ni que sustente rankings de modelos, umbrales de despliegue o certificación. La ficha debe presentar PBA como instrumento exploratorio de señalización, no como métrica de equidad ya validada.

Research question

Can an open generation of structured profiles reveal associations between identities and social roles, compare them across 12 LLMs, show intersectional patterns, and track changes between versions using a Cramér V normalization?

Method

Each call requests 20 "diverse" profiles in JSON with eight attributes. A nominal 10,000 are generated per model across 12 LLMs at temperature 1. After lowercasing, stemming, and deduplication, GPT-5 consolidates values to 3 genders, 6 ethnicities, 5 sexual orientations, 3 classes, 5 education levels, 18 occupations, and 15 interests; name uses the top 50. 16 Cramér V values and an unformalized severity normalization are computed. Models are compared using paired t tests and BH-FDR, heatmaps and intersections are inspected, and sizes 5k to 20k, a UX role, a debiasing prompt, and a human validation n=9 are tested.

Sample: 120,000 nominal profiles before deduplication; 116,222 after. The loss ranges from 3 in GPT-5 to 2,060 in Mistral-medium. Each call co-generates 20 profiles and the unit of independence is not modeled. The sensitivity analysis uses six models and four sizes with no described replications. The human validation includes nine professionals of undisclosed background, mixing DEI experts and non-experts without a breakdown.

Findings

  • GPT-4o has the lowest mean PBA (0.678) and GPT-4.1 mini the highest (0.969) in Table 2.
  • GPT-5 obtains 0.950, with nine of 16 dimensions in the very high category and none low.
  • Associations with name are the highest on average (1.113); Name x Occupation reaches 1.195.
  • Occupation and Interest concentrate high associations for the four identities.
  • Ethnicity x Occupation averages 1.091 and Gender x Interest, 1.013.
  • Gender x Education and Gender x Social Class are lower than Gender x Occupation and Interest.
  • The heatmaps show concentration of queer identities in Creative & Design and low presence in Engineering across many models.
  • The GPT-5 family increases assignment to IT for several non-heterosexual orientations compared with GPT-3.5/GPT-4.x.
  • Mistral patterns concentrate lesbian profiles in Healthcare in some versions.
  • The intersectional analysis shows class differences across gender x sexual orientation combinations that do not appear in isolated axes.
  • In nursing, female representation exceeds 90% in almost all models; GPT-4.1 nano reports 100%.
  • The L1 male-female gap by occupation ranges from 54 in GPT-4/GPT-5 nano to 90 in Mistral-medium.
  • The plot of five OpenAI aliases shows a decline from GPT-3.5 to GPT-4o and a subsequent rebound across several dimensions.
  • The 5k to 20k sensitivity analysis reports ICC and high correlations in the majority of the 16 dimensions.
  • Severity does not change in ten of 16 dimensions in the sample size analysis.
  • The UX role maintains the majority of patterns, except for very low concordance in Sexual Orientation x Education.
  • The debiasing prompt notably reduces concordance: recalculated ICC means of 0.381/0.361, Spearman 0.422, and Kendall 0.306.
  • None of the 16 paired t tests of the debiasing prompt reaches p<0.05; this is not a test of equivalence.
  • The nine evaluators rank the GPT-5 mini plot as more biased than GPT-5 nano in Gender x Occupation.
  • The results locate generated associations that can guide human inspection, but do not by themselves validate harm or unfairness.

Limitations

  • The manuscript retains a conference title and template DOI, so it is not a definitive ACM publication.
  • No peer-reviewed version or verifiable acceptance was found.
  • Code, profiles, mappings, and outputs are promised after publication and are not available.
  • There is no official repository located by title, arXiv ID, or method name.
  • Normalization, deduplication, tables, figures, and end-to-end statistical tests cannot be reproduced.
  • GPT-3.5, GPT-4, and GPT-4o are aliases without snapshot identifiers.
  • Exact snapshots of GPT-4.1/GPT-5, inference dates, or API region/provider are not fixed either.
  • No system prompt, max tokens, top_p, seed, timeout, retries, JSON errors, or discarded responses are reported.
  • Only temperature 1 is reported; stochasticity across models is not controlled.
  • The base prompt already demands "diverse" profiles, so the baseline is not neutral with respect to fairness.
  • The debiasing prompt partially repeats the diversity instruction from the baseline.
  • Age, capabilities, and region are requested in the debiasing instruction but are not output fields, and their effect is not measured.
  • Each call generates 20 profiles jointly; these rows are not iid observations and may be balanced within the batch.
  • The number of calls, order, context between calls, and isolation of conversations are not detailed.
  • The text claims 10,000 profiles without duplicates per model, but Table 7 finds 3,778 duplicates.
  • After deduplication, 116,222 remain, not 120,000 analyzed observations.
  • Deduplication is highly uneven: 20.6% in Mistral-medium, 7.76% in Ministral-3B, and 0.03% in GPT-5.
  • Removing duplicates can hide collapse or concentration, which are relevant behaviors for diversity.
  • It is not clarified whether samples are backfilled to 10,000 unique profiles or different final sizes are compared.
  • GPT-5 defines the semantic consolidation used to audit GPT models as well, introducing evaluator dependence.
  • The human validation of the mappings does not report evaluators, rubric, disagreements, changes, or reliability.
  • The reductions 1,128 to 6 ethnicities, 689 to 5 levels, and 1,501 to 15 interests imply massive normative decisions.
  • The mappings may merge non-equivalent identities and fabricate or erase associations.
  • Missingness or unknown and ambiguous categories are not reported.
  • For names, only the global top 50 is used without reporting what fraction of each model is excluded.
  • Names encode culture, language, gender, and ethnicity; their association with other attributes does not automatically equate to bias.
  • Cramér V quantifies symmetric association, not direction, valence, harm, or discrimination.
  • Statistical independence is not a sufficient normative definition of fairness; some associations may be legitimate and others harmful even when V is low.
  • Cramér V depends on margins, cardinality, sparsity, and grouping decisions.
  • Expected frequencies are not reported; the cited statistical source requires checking cells <5 and <1 for the chi-square approximation.
  • Tables such as 50 names x 18 occupations can be sparse even with thousands of profiles after exclusions.
  • No Cramér V bias correction for large or sparse tables is used or discussed.
  • The exact formula for threshold normalization does not appear in the article.
  • The Kim reference only shows thresholds up to five degrees of freedom; extrapolation for larger tables is not documented.
  • The article alternates between the degrees of freedom of the test (r-1)(c-1) and the minimum used in Cramér V without specifying which governs the threshold.
  • The normalized metric exceeds 1, contradicting the description of bounded labels or scores in the conclusion.
  • Averaging heterogeneous normalized scores across 16 constructs has no calibrated interpretation.
  • The cutoffs 0.33/0.66/1 are not empirically validated for fairness or harm.
  • The proposed deployment threshold <1.0 lacks risk, domain, or population validation.
  • Paired t tests between models use 16 correlated dimensions as replicates and do not justify independence or normality of differences.
  • The 16 dimensions share the same profiles and categories, generating dependence.
  • Effect sizes or intervals for differences between models are not published.
  • OpenAI aliases are not controlled checkpoints of a single evolutionary line.
  • The supposed longitudinal trajectory is a cross-sectional comparison of products with unknown training, architecture, and dates.
  • The rebound cannot be attributed to model scale, alignment, or generation without a causal design.
  • The size sensitivity does not say whether 5k, 10k, 15k, and 20k are independent samples or nested subsets.
  • There are no replications per size, seeds, bootstrap, or intervals; a single series does not measure sample variation.
  • Size robustness is evaluated in only six of 12 models.
  • ICC and correlations with six models as units offer unstable estimates and are not accompanied by intervals.
  • The role and debiasing tests appear to use few models; the effective n and sample construction are not fully specified.
  • p>0.05 does not demonstrate absence of effect, equivalence, or general insufficiency of prompt engineering.
  • No equivalence test is applied and no non-inferiority margin is defined.
  • Table 5 prints an arithmetically false Mean row: it also does not match the recalculated means of its 16 rows.
  • The text says that the majority of debiasing metrics exceed 0.40, but many ICC and Kendall values are clearly below.
  • Abstract and conclusion claim stability under debiasing despite low and negative concordances in several dimensions.
  • The severity difference uses coarse bins and can hide large changes within a category.
  • The human validation selects two systems already separated by the metric and a single dimension.
  • Nine evaluators cannot calibrate 16 dimensions or multiple cultures or forms of harm.
  • It is not reported how many were DEI experts, their profiles, recruitment, compensation, or conflicts.
  • Consent, ethical review, or exemption for the human study are not documented.
  • Anonymizing brands does not by itself eliminate order effects; randomization of plot order is not reported.
  • Inter-evaluator agreement or uncertainty of ratings is not calculated.
  • Providing the four-category scale may induce the same discretization that is intended to be validated.
  • The validation shows apparent validity of an obvious contrast, not convergent, predictive, or regulatory validity.
  • The claim of lower data leakage risk is plausible, but is not tested against memorization of patterns or prompts.
  • The supposed cultural generalization is not evaluated: everything is in English and there are only US/European providers.
  • The format is structured and categorical; it does not audit narrative, tone, agency, dignity, or semantic stereotypes of complete persons.
  • Real outcomes or downstream harm in design, hiring, education, or recommendation are not evaluated.
  • The certification and compliance proposals exceed the available exploratory validation.
  • Reference [38] erroneously assigns arXiv:2401.12345 to Prompting Fairness; that ID corresponds to Distributionally Robust Receive Combining.

What the study does not establish

  • It does not establish that Cramér V is a direct measure of fairness or harm.
  • It does not demonstrate that independence between every identity and every role is the desirable just state.
  • It does not validate that the normalized scores are comparable across cardinalities and constructs.
  • It does not justify the small/medium/high/very high thresholds as risk levels.
  • It does not demonstrate that GPT-4o is globally fairer than other models outside of this prompt and mapping.
  • It does not demonstrate that newer models cause a resurgence of bias.
  • It does not test robustness to debiasing prompts; its own concordances fall substantially.
  • It does not test that a debiasing prompt cannot reduce bias in other designs or with adequate power.
  • It does not validate the method in languages or cultures other than Western English.
  • It does not audit open narrative persons beyond eight structured categorical fields.
  • It does not demonstrate real impact on opportunities, hiring, or user decisions.
  • It does not allow certification, compliance, or deployment gating with the proposed <1 threshold.
  • It does not establish causality between name, gender, ethnicity, or orientation and the generated roles.
  • It does not allow reproducing rankings or errors while code, mappings, and profiles remain unpublished.

Traceability

Scope: Full text

Version: arXiv:2602.00044v2, revised 24 February 2026; 27-page ACM-submission manuscript with placeholder venue and DOI

Consulted source: https://arxiv.org/pdf/2602.00044v2

Review: Codex full-text, bilingual-fidelity, arXiv-v2, 27-page visual, Cramer-normalization, contingency-sparsity, batch-independence, category-consolidation, deduplication, table-arithmetic, prompt-robustness, human-validation, citation and artifact-availability audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI GPT-3.5 alias; exact API model snapshot not reported
  • OpenAI GPT-4 alias; exact API model snapshot not reported
  • OpenAI GPT-4o alias; exact API model snapshot not reported
  • OpenAI GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano aliases; exact dated snapshots not reported
  • OpenAI GPT-5, GPT-5 mini, and GPT-5 nano aliases; exact dated snapshots not reported
  • Ministral-3B model identifier ministral-3b-2410
  • Mistral Small model identifier mistral-small-2501
  • Mistral Medium model identifier mistral-medium-2505
  • GPT-5 as the semantic category-consolidation model, creating evaluator dependence on one audited family

Instruments and metrics

  • Base JSON prompt requesting 20 diverse profiles with eight categorical attributes
  • Role prompt instructing the model to act as an expert UX researcher
  • Debiasing instruction requesting diversity in gender, age, ethnicity, socioeconomic background, abilities, and regions
  • Lowercasing, stemming, exact-profile deduplication, and GPT-5 semantic consolidation
  • Sixteen identity × social-dimension contingency tables
  • Raw Cramér's V and an incompletely specified degree-of-freedom-aware normalized severity score
  • Small/medium/high/very-high bins at 0.33, 0.66, and 1.0
  • Paired t-tests across dimensions or models and BH-FDR for pairwise model comparisons
  • ICC(C,1), ICC(A,1), Spearman, Kendall, and mean severity difference
  • Heatmap drill-downs for sexual orientation × occupation, gender × occupation, nurse × gender, and intersectional social class
  • Exploratory 1–4 human bias-severity rating with written justifications

Data used

  • Nominally 120,000 generated profiles: 10,000 per model before deduplication
  • 116,222 unique profiles after removing 3,778 exact duplicates according to Table 7
  • Sixteen normalized contingency tables per model over four identity and four social dimensions
  • Sample-size sensitivity sets of 5,000, 10,000, 15,000, and 20,000 profiles for six recent models
  • Role-playing and debiasing generations for six recent GPT-5/Mistral models; exact sample construction not fully specified
  • Nine-person exploratory comparison of two selected Gender × Occupation distributions

Evidence and location

  • Version, status, and metadata: arXiv:2602.00044v2 abstract and submission history; revised 24 February 2026
  • Prompt, models, temperature, and pipeline: arXiv v2, Sections 3.1 to 4.1 and Appendix A.1 to A.2, Tables 6 to 7
  • Metric, thresholds, and lack of normalization formula: arXiv v2, Section 3.2, Equation 1 and severity scale
  • Scores and comparisons of 12 models: arXiv v2, Sections 4.2 to 4.4, Table 2 and Figures 1 to 5
  • Sensitivity to size, role, and debiasing: arXiv v2, Section 5, Tables 3 to 5
  • Error in the debiasing Mean row: arXiv v2 Table 5; independent arithmetic recomputation across the 16 printed rows
  • Zoom on gender, occupation, and nursing: arXiv v2, Appendices A.3 to A.4, Figures 6 and Table 8
  • Exploratory human validation: arXiv v2, Appendix A.5, Table 9
  • Assumptions and thresholds for chi-square/Cramér V: Kim 2017, Statistical notes for clinical researchers: Chi-squared test and Fisher's exact test, DOI 10.5395/rde.2017.42.2.152
  • Erroneous bibliographic reference [38]: arXiv v2 reference [38] versus arXiv:2401.12345 official record, Distributionally Robust Receive Combining
  • Absence of code and definitive publication: arXiv Code, Data and Media section plus targeted title/method/arXiv GitHub searches, checked 15 July 2026