LLM Personas as a Substitute for Field Experiments in Method Benchmarking

Personas, identity, and agents2025arXivApproved editorial review

Authors: Enoch Hyunwook Kang

Keywords: LLM personas, Field-experiment benchmarking, Aggregate-only observation, Method-blind evaluation, KL discriminability, Sample complexity, TextBO, Digital twins

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

1
Authors
12
Findings
19
Limitations
5
Evidence

Editorial summary

English

The paper asks when an LLM-persona benchmark can occupy the same position as a field experiment inside a method-development loop. Its answer is not that personas reproduce humans. It defines a much narrower equivalence: for a method that submits an artifact and receives feedback, replacing human evaluation with persona evaluation is merely a panel change if two conditions hold. Aggregate-only observation (AO) requires the method to see only the final score, without individual votes, stable identities, ordering, or side channels. Method-blind evaluation (MB) requires the score distribution to depend only on the submitted artifact, not on the algorithm that produced it, its identity, provenance, or interaction history. Under this definition, every adaptive transcript law factorizes through an artifact-conditioned kernel. The paper proves that this property, called just panel change (JPC), is equivalent to AO+MB and supplies a counterexample in which leaking individual votes distinguishes two benchmarks with the same aggregate score. The text itself acknowledges that the theorem can look almost tautological: JPC is defined through the factorization supplied by AO and MB. Its useful contribution is to turn an ambiguous substitution claim into two auditable protocol requirements. The converse auxiliary-panel construction proves observational equivalence, not that personas and humans share a mechanism or score distribution.

The paper then separates interface validity from statistical usefulness. For artifact pairs separated by at least a user-chosen distance r, it fixes a pair distribution and defines discriminability through a lower quantile of KL divergence. Under Gaussian scores with common variance, this becomes a q-robust signal-to-noise quantity, kappa_Q(q). If each artifact is evaluated independently L times and sample means are compared, the misordering probability for a random pair is bounded by q + exp(-L*kappa/2); choosing L at least 2/kappa times log(1/delta) makes the bound q+delta. This is a distributional guarantee over the chosen pairs after tolerating a q fraction of poorly separated pairs, not a uniform guarantee for every artifact. L is per prompt, so a comparison uses 2L persona evaluations. The main rule assumes fresh independent panels, Gaussianity, constant variance, and a non-adaptive pairwise comparison. An appendix derives heteroscedastic Gaussian KL but does not redo the reported calibration with it. With q=0.05 and delta=0.05, the total bound is 0.10; it should not be described as a 95% guarantee.

The only applied study is a TextBO proof of concept, not a human experiment. TextBO improves image-generation prompts for eight synthetic advertising campaigns. Each ad is scored by sampling 200 Twin-2K-500 profiles, conditioning Gemini 2.5 Flash on each profile's survey responses and the image, converting log-probabilities over 1-5 to an expected score, and averaging. The appendix uses ten improvement steps and ten variants per step, yielding 100 pairs per scenario. At q=delta=0.05, the table gives per-prompt sizes of 1,180, 259, 46, 637, 203, 46, 1,303, and 80. Five of eight exceed the deployed panel of 200 and three exceed 500. The paper also contains an internal error: the prose says a 46-1,180 range and 460.25 mean, while the table and formula give 46-1,303 and 469.25. It does not report how many independent repetitions estimated each mean and variance, how the lower quantile was estimated from only 100 pairs, or intervals for kappa and L. It also does not account for adaptive selection, multiple comparisons, or lower-tail quantile uncertainty, which is driven by roughly the five least-discriminable cases in each scenario.

Artifact auditing narrows the interpretation further. The current complete Twin-2K-500 full-persona snapshot has 2,058 unique PIDs and 2,058 unique nonempty summaries; summaries range from 11,602 to 18,482 characters and encode demographics, attitudes, and many prior survey answers. Hugging Face exposes only a split named data. TextBO creates its own 80/20 split by lexicographically sorting pid_<number> filenames: at the audited revision this yields 1,646 train and 412 test profiles, places 1,111 PIDs beginning with 1 in train, and leaves only prefixes 6-9 in test. It is not random or stratified. Each evaluation samples 200 profiles without replacement, about 12.2% of the pool, whereas the derivation assumes i.i.d. evaluators. The download script always selects the latest dataset revision, and the paper does not pin the revision used for its table.

The related TextBO repository compiles, but it contains no kappa calculation, Table 1 scores, logs, images, persona inputs, exact seeds, or locked environment. It also has no tests, CI, tags, or repository license file; Ruff reports 27 findings. The README acknowledges missing fixed-path folders and assets. API or log-probability failures are silently replaced by a uniform distribution with expected score 3, flattening the benchmark without exposing the error count. The base seed is randomly generated, Gemini uses an undated alias, and dependencies have only minimum versions. The table arithmetic can therefore be checked, but the experiment cannot be reproduced end to end. The supported contribution is conceptual: AO and MB are useful benchmark-hygiene questions, and the required panel size can vary widely with the synthetic channel. There are no human outcomes, clicks, conversions, or causal effects. Larger persona panels reduce sampling noise around the synthetic judge; they do not correct systematic bias, unrepresentative personas, or construct invalidity.

Español

El artículo pregunta cuándo un benchmark de personas LLM puede ocupar el mismo lugar que un experimento de campo dentro del bucle de desarrollo de un método. Su respuesta no es que las personas reproduzcan a los humanos. Define una equivalencia mucho más estrecha: para el método que envía un artefacto y recibe feedback, cambiar evaluación humana por evaluación con personas es un simple cambio de panel si se cumplen dos condiciones. La observación solo agregada (AO) exige que el método vea únicamente el score final, sin votos individuales, identidades estables, orden ni canales laterales. La evaluación ciega al método (MB) exige que la distribución del score dependa solo del artefacto enviado, no del algoritmo que lo produjo, su identidad, procedencia o historial. Bajo esta definición, las leyes de todas las transcripciones adaptativas factorizan mediante un kernel condicionado por el artefacto. El paper prueba que esta propiedad, llamada just panel change (JPC), equivale a AO+MB y ofrece un contraejemplo donde filtrar votos individuales permite distinguir dos benchmarks con el mismo score agregado. El propio texto reconoce que el teorema puede parecer casi tautológico: JPC se define mediante la factorización que AO y MB proporcionan. Su aportación útil es convertir una afirmación ambigua de sustitución en dos requisitos de protocolo auditables. La construcción recíproca con un panel auxiliar demuestra equivalencia observacional, no que personas y humanos compartan mecanismo ni distribución de scores.

Después separa validez de interfaz y utilidad estadística. Para pares de artefactos separados al menos por una distancia r elegida por el usuario, fija una distribución de pares y define discriminabilidad mediante un cuantil inferior de divergencia KL. Bajo scores gaussianos con la misma varianza, esa cantidad se reduce a una relación señal-ruido robusta, kappa_Q(q). Si cada artefacto se evalúa L veces de forma independiente y se comparan medias, la probabilidad de invertir un par aleatorio queda acotada por q + exp(-L*kappa/2); elegir L mayor o igual que 2/kappa por log(1/delta) deja el límite en q+delta. Es una garantía distribucional sobre los pares elegidos, después de tolerar una fracción q de pares poco distinguibles, no una garantía uniforme para cualquier artefacto. L es el número por prompt, de modo que comparar dos requiere 2L evaluaciones. La fórmula principal supone paneles frescos e independientes, gaussianidad, varianza constante y comparación pareada no adaptativa. El anexo deriva la KL heterocedástica, pero no rehace con ella la calibración. Con q=0,05 y delta=0,05, el error total acotado es 0,10; no debe presentarse como una garantía del 95%.

El único estudio aplicado es una prueba de concepto sobre TextBO, no un experimento con personas. TextBO mejora prompts para generar anuncios en ocho campañas sintéticas. Cada anuncio se puntúa muestreando 200 perfiles de Twin-2K-500, condicionando Gemini 2.5 Flash con las respuestas de encuesta de cada perfil y la imagen, convirtiendo las log-probabilidades de 1 a 5 en un score esperado y promediando. El anexo usa diez pasos de mejora y diez variantes por paso: 100 pares por escenario. Para q=delta=0,05, la tabla da tamaños por prompt de 1.180, 259, 46, 637, 203, 46, 1.303 y 80. Cinco de ocho superan las 200 personas empleadas y tres superan 500. Hay además un error interno: el texto afirma rango 46-1.180 y media 460,25, pero la tabla y la fórmula dan 46-1.303 y media 469,25. El paper no informa cuántas repeticiones independientes se usaron para estimar medias y varianzas, cómo se obtuvo el cuantil con solo 100 pares, ni intervalos para kappa o L. Tampoco corrige selección adaptativa, comparaciones múltiples o incertidumbre del cuantil inferior, que depende aproximadamente de los cinco casos menos discriminables de cada escenario.

La auditoría de artefactos estrecha aún más la lectura. La versión completa actual de Twin-2K-500 contiene 2.058 PIDs y 2.058 resúmenes únicos, sin nulos ni duplicados; los resúmenes miden entre 11.602 y 18.482 caracteres y codifican demografía, actitudes y numerosas respuestas previas. Hugging Face solo ofrece un split llamado data. El código de TextBO crea su propio 80/20 ordenando lexicográficamente nombres pid_<número>: con la revisión auditada produce 1.646 perfiles train y 412 test, concentra 1.111 PID que empiezan por 1 en train y deja test únicamente con prefijos 6-9. No es un split aleatorio ni estratificado. Dentro de cada evaluación toma 200 perfiles sin reemplazo, aproximadamente 12,2% del pool, mientras la derivación supone evaluadores i.i.d. El script descarga siempre la revisión más reciente y el paper no fija la usada para la tabla.

El repositorio relacionado de TextBO compila, pero no incluye el cálculo de kappa, los scores de la Tabla 1, logs, imágenes, inputs de personas, seeds exactas ni un entorno bloqueado. Tampoco tiene tests, CI, tags o archivo de licencia; Ruff detecta 27 incidencias. El README reconoce que faltan carpetas y assets de rutas fijas. Los fallos de API o log-probabilidades se sustituyen silenciosamente por una distribución uniforme cuyo score es 3, aplanando el benchmark sin exponer el número de errores. El seed base se genera al azar, Gemini usa un alias no fechado y las dependencias solo tienen mínimos. Por tanto, la aritmética de la tabla es comprobable, pero el experimento no es reproducible de extremo a extremo. La evidencia sólida es conceptual: AO y MB son buenas preguntas de higiene y el tamaño necesario puede variar mucho según el canal sintético. No hay datos humanos, clicks, conversiones ni efectos causales. Más personas reducen ruido de muestreo alrededor del juez sintético; no corrigen sesgo sistemático, población no representativa ni invalidez del constructo.

Research question

Under what interface conditions can a human panel be replaced by LLM persons without changing the feedback structure observed by an adaptive method, and how many persona evaluations does that synthetic channel need to distinguish pairs of artifacts at a chosen resolution?

Method

Theoretical work with an applied calibration. Models the benchmark as panel, microinstrument, aggregation, and observable kernel; defines AO, MB, and JPC; tests JPC if and only if AO+MB and presents a counterexample by vote leakage. Defines a lower KL quantile and, under homoscedastic Gaussian scores, derives a paired inversion bound and a sample size rule. The annex applies the rule to 800 TextBO pairs, 100 for each of eight synthetic advertising scenarios, evaluated with Twin-2K-500 profiles and Gemini 2.5 Flash.

Sample: There are no human participants or field results. The calibration declares eight scenarios, ten steps, and ten variants per step: 100 pairs per scenario and 800 in total. Each TextBO score averages 200 profiles. The current complete review of Twin-2K-500 has 2,058 profiles; the related code creates 1,646 train and 412 test by lexicographic order of PID, although the exact review used in the table is not fixed.

Findings

  • JPC is equivalent to AO+MB in the defined interface: the method only sees an aggregate and the score depends on the artifact, not on the method or provenance.
  • The equivalence allows persons and humans to produce completely different score distributions; it does not prove agreement.
  • Vote or identity leakage can break JPC even if the aggregate kernel matches.
  • Under homoscedastic Gaussianity, L grows as 2/kappa times log(1/delta), with distributional error bounded by q+delta.
  • With q=delta=0.05, the total bound is 0.10, not 0.05.
  • The table sizes are 1,180, 259, 46, 637, 203, 46, 1,303, and 80 per prompt.
  • Five scenarios require more than 200 and three more than 500 according to the table itself.
  • The correct range and mean are 46-1,303 and 469.25; the prose 46-1,180 and 460.25 is erroneous.
  • The proof of concept only calibrates a synthetic ad judge; it contains no human outcomes.
  • The complete audited dataset has 2,058 unique PIDs and non-empty summaries, but no official train/test split.
  • The TextBO 80/20 split reproduces 1,646/412 profiles and is strongly structured by the PID text.
  • The public artifacts do not allow reproducing the kappa calculation or Table 1.

Limitations

  • JPC is an interface factorization property and not an empirical test on persons or humans.
  • The necessity of AO is built into the observational definition of JPC itself.
  • The auxiliary panel construction does not preserve the real structure of the human and persona protocols.
  • The main rule assumes Gaussianity and constant variance across artifacts.
  • The guarantee is over a random pair from a chosen distribution and tolerates a fraction q of bad pairs.
  • The distance r and the pair distribution are task- and method-dependent decisions.
  • A prompt edit does not validate a stable distance between stochastically generated images.
  • The 100 pairs per scenario come from the TextBO adaptive trajectory and not from all pairs of an edit.
  • The 5% quantile depends on very few cases and has no interval or sensitivity.
  • The estimator of means, variances, and kappa is not explained, nor is the number of independent repetitions.
  • Uncertainty of kappa is not propagated to the size L.
  • Multiple comparisons, adaptive selection, or benchmark reuse are not corrected.
  • The real TextBO sampling is without replacement within a finite pool, not exactly i.i.d.
  • The split by PID name is neither random nor stratified.
  • The dataset and models are not versioned in the experiment.
  • API errors become a neutral score and may artificially reduce discriminability.
  • There are no logs, inputs, images, seeds, scores, or Table 1 script published.
  • There are no tests, CI, release, repository license, or locked environment in TextBO.
  • The results are limited to eight synthetic ad briefs and one multimodal judge.

What the study does not establish

  • That LLM persons validly reproduce or represent human participants.
  • That person and human scores match, correlate, or have the same ranking.
  • That a person benchmark estimates causal effects or results of an A/B test.
  • That AO and MB hold empirically in TextBO, Twin-2K-500, or human evaluation.
  • That humans and LLMs share internal evaluation mechanisms.
  • That increasing the panel corrects systematic judge bias or population representativeness.
  • That Table 1 is a 95% guarantee.
  • That 200 persons suffice in the eight scenarios.
  • That 500 persons are a uniformly conservative choice.
  • That the prose range 46-1,180 or mean 460.25 are correct.
  • That the estimated quantile and L have small uncertainty.
  • That the paired bound transfers without adjustment to adaptive optimization.
  • That a textual edit measures a minimum significant difference between ads.
  • That the TextBO split is official, random, or stratified.
  • That the published artifacts reproduce the calibration.
  • That the preprint has passed peer review or has confirmed archival publication.

Traceability

Scope: Full text

Version: arXiv:2512.21080v3, 20 pages; related TextBO code and complete Twin-2K-500 full-persona snapshot also audited

Consulted source: https://arxiv.org/abs/2512.21080v3

Review: Codex 20-page full-text visual, theorem/proof, sample-complexity, Table 1 arithmetic, TextBO code, complete Twin-2K-500 Parquet and claim-boundary audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Gemini 2.5 Flash multimodal persona-conditioned judge
  • TextBO prompt-optimization system
  • Imagen 4.0 Ultra preview in the related public TextBO implementation

Instruments and metrics

  • Panel-instrument-aggregation benchmark model
  • Aggregate-only observation (AO)
  • Method-blind evaluation (MB)
  • Just panel change (JPC) transcript factorization
  • Pairwise KL discriminability
  • q-robust signal-to-noise kappa_Q(q)
  • Homoscedastic Gaussian pairwise sample-complexity bound
  • Heteroscedastic Gaussian KL extension
  • Persona-conditioned 1-5 ad-effectiveness log-probability score
  • TextBO clause-level prompt-edit distance

Data used

  • LLM-Digital-Twin/Twin-2K-500 full_persona
  • TextBO-generated synthetic ad campaigns and prompt-improvement pairs (not released with the calibration)

Evidence and location

  • Version, authorship, dates, categories, arXiv DOI, and license: Official arXiv record 2512.21080v3 checked 2026-07-16
  • Definitions, theorems, proofs, counterexample, assumptions, formula, calibration, table, and declared limits: arXiv:2512.21080v3 PDF, 20 pages; every page rendered and visually inspected
  • Code, split, sampling, fallbacks, dependencies, reproducibility, and absent artifacts: Enoch-H-Kang/TextBO commit d95178651c99f3006605336f486433f6bc1037fe; compileall, AST and Ruff audit on 2026-07-16
  • Rows, PIDs, nulls, duplicates, profile sizes, and reproduced split: Complete seven-Parquet audit of LLM-Digital-Twin/Twin-2K-500@f883165a3026fde855dfd448e0cd16443ab257b6
  • Table 1 arithmetic, independence, selection, uncertainty, and claim limits: reports/verification/article-272-arxiv-interface-equivalence-ao-mb-kl-sample-complexity-textbo-persona-independence-and-claim-audit.json