Training-Free Cultural Alignment of Large Language Models via Persona Disagreement

Society, culture, and collective behavior2026arXivApproved editorial review

Authors: Huynh Trung Kiet, Dao Sy Duy Minh, Tuan Nguyen, Chi-Nguyen Tran, Phu-Hoa Pham, Nguyen Lam Phu Quy, The Anh Han, Long Tran-Thanh

Keywords: DISCA, Cultural alignment, Persona disagreement, World Values Survey, MultiTP, Moral Machine, AMCE, Inference-time steering, Prospect-Theory importance sampling, LLM-as-judge

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

8
Authors
9
Findings
18
Limitations
3
Evidence

Editorial summary

English

The preprint presents DISCA, an inference-time controller intended to move LLM moral decisions toward country averages from Moral Machine/MultiTP without changing weights. It runs the same model under a base prompt and four personas, young, middle-aged, older, and country aggregate, symmetrizes A/B order, searches for a scalar logit correction with importance sampling under a Prospect-Theory-inspired asymmetric utility, and shrinks it when two independent passes disagree. Cultural alignment here means lower L2 distance between six model AMCEs and the country's mean human AMCE vector. It does not mean moral legitimacy, individual fidelity, complete cultural representation, or safety.

The personas do consume country-level data. They are built from World Values Survey Wave 7 using unweighted cohort means over ten dimensions, normalization, and deterministic quartile descriptors. No per-country labels is therefore faithful only if it means no country-specific MultiTP or human-AMCE labels during inference; it does not mean country-data-free alignment. Survey weights, cohort Ns, and uncertainty are not reported. The manual fallback is not validated in the headline panel: Saudi Arabia, the named example, is excluded from the 20 countries.

Proposition 1 establishes a narrower result than the implemented product. If each persona were an unbiased, i.i.d. observation with common variance around an unobserved human logit, between-agent variance would estimate noise and an MSE-optimal scalar shrinkage factor would exist. The paper acknowledges that personas share the model and scenario. The oracle factor also depends on the unobserved human target and cannot be computed. DISCA replaces it with PT-IS and an exponential two-pass gate. The proposition motivates qualitative shape, but does not prove that the full controller is MSE-optimal, that prompt variation estimates human uncertainty, or that two passes measure demographic disagreement rather than Monte Carlo noise.

In the main 20-country table, the six selected models at or above 3.8B reduce MIS by 10.3%-23.6%; the selected 2B Gemma-4-E2B improves by 3.4%. Llama-3.3-70B and Qwen3-VL-8B improve in 20/20 countries; Phi-4 in 18/20 and has the lowest MIS, 0.346. Phi-4 plus DISCA is below vanilla Llama-3.3-70B, 0.346 versus 0.849, and Llama plus DISCA, 0.668. On Phi-4, DISCA beats PRISM prompting 0.384, MC-Dropout 0.403, activation steering 0.430, WVS Profile 0.453, and vanilla 0.454. This is strong descriptive benchmark evidence, not a general law that calibration beats scale.

Selection constrains generalization. The seven models are explicitly chosen for the most robust gains after 28 combinations; the broader table omits nine models with non-positive gain. The BRA/CHN/DEU/JPN/USA prototyping panel is included in the final 20 countries. The headline describes a favorable selection rather than an architecture-wide success rate. Three seeds establish controller-sampling stability, macro standard deviation up to 0.006, but not WVS, human-target, or scenario-sampling uncertainty.

For open-ended responses, Claude 3 Opus returns A/B and confidence, which the study converts into a pseudo-logit. All four models improve on average: 6.85% Llama, 6.67% Phi-4, 4.55% Qwen2.5, and 2.13% Phi-3.5. The table nevertheless contains seven negative country-model cells rather than the six stated in the text; the worst is -3.10%. There is no human validation of judge choice, confidence calibration, or cultural equivalence. Gaussian noise on cached pseudo-logits tests numeric noise, not semantic bias. On factual BLEnD, seven of 16 countries improve, but aggregate SEM-B falls by 4.4 points.

The WVS-to-AMCE link is overstated. Ten predictors are fit to only 20 countries, leaving nine residual degrees of freedom with an intercept. The text emphasizes raw R-squared of 0.55-0.69, but the paper's own adjusted values are only 0.05-0.35, 0.05 for Age and 0.09 for Species, without cross-validation, intervals, or multiplicity correction. The so-called causal impact matrix is a leave-one-descriptor-out controller ablation, not causal identification of how WVS values produce human preferences. One language per country also confounds culture with translation, tokenization, and language proficiency.

Internal conflicts remain. Magistral has MIS 0.350 in the main table and 0.334 in an all-20-country ablation. The fourth-persona comparison reports 0.4252 for the aggregate and 0.4049 for a utilitarian anchor on Phi-4/20 countries, incompatible with the main 0.346; 0.4252 also matches N=2 in another sweep. Tail safety reports mean Delta MIS 0.096 across 120 cells, while the six headline means imply 0.1015. A suite called nine checks displays eight rows. The arXiv page says 57 pages while the v2 PDF renders as 48. Oversampling with replacement duplicates scenarios and changes weights; it adds no independent information.

Public reproducibility cannot resolve these issues. The package contains TeX, bibliography, and five figures, but no code, derived WVS profiles, complete prompts, scenario IDs, logits, IS candidates, gates, responses, judge records, run-level results, full sweep, environment, or exact model revisions. Active text says reference implementations are in a code archive, while a source comment says code will be released after acceptance; no archive or repository is linked. A faithful reading is that DISCA lowers distance to AMCE averages for a favorably selected model set and often does so in a Claude-mediated open-ended track. It does not establish general cultural alignment, optimality of the full algorithm, normative legitimacy, minority protection, or independent reproducibility.

Español

El preprint presenta DISCA, un controlador de inferencia para acercar decisiones morales de un LLM a promedios nacionales de Moral Machine/MultiTP sin cambiar pesos. Ejecuta el mismo modelo con un prompt base y cuatro personas, jóvenes, mediana edad, mayores y agregado nacional, simetriza A/B, busca una corrección escalar de logits mediante importance sampling con utilidad asimétrica inspirada en Prospect Theory y la reduce cuando dos pasadas independientes discrepan. Cultural alignment significa aquí menor distancia L2 entre seis AMCE del modelo y el vector AMCE humano medio del país; no significa legitimidad moral, fidelidad individual, representación completa de una cultura ni seguridad.

Las personas sí consumen datos por país. Se construyen desde World Values Survey Wave 7: medias no ponderadas por cohorte en diez dimensiones, normalización y descriptores por cortes deterministas de cuartil. Por tanto, «sin etiquetas por país» sólo es fiel si significa sin etiquetas MultiTP/AMCE por país durante la inferencia; no significa alineación sin datos nacionales. No se publican pesos de encuesta, N por cohorte ni incertidumbre. El fallback manual tampoco queda validado en el panel principal: Arabia Saudí, el ejemplo citado, está excluida de los 20 países.

La Proposición 1 demuestra un resultado más estrecho que el producto. Si cada persona fuera una observación insesgada, i.i.d. y con varianza común alrededor de un logit humano no observado, la varianza entre agentes estimaría el ruido y existiría un shrinkage escalar óptimo en MSE. El paper reconoce que las personas comparten modelo y escenario. El factor óptimo depende además del target humano no observado y no puede calcularse. DISCA lo sustituye por PT-IS y una compuerta exponencial entre dos runs. La prueba motiva la forma cualitativa, pero no demuestra que el controlador completo sea MSE-óptimo, que la variación de prompts estime incertidumbre humana ni que dos pasadas midan desacuerdo demográfico en vez de ruido Monte Carlo.

En la tabla principal de 20 países, los seis modelos seleccionados de al menos 3,8B reducen MIS 10,3%-23,6%; Gemma-4-E2B 2B mejora 3,4%. Llama-3.3-70B y Qwen3-VL-8B mejoran en 20/20 países; Phi-4 en 18/20 y obtiene la menor MIS, 0,346. Phi-4+DISCA queda por debajo de Llama-3.3-70B vanilla, 0,346 frente a 0,849, y de Llama+DISCA, 0,668. En Phi-4, DISCA supera PRISM prompt 0,384, MC-Dropout 0,403, activation steering 0,430, WVS Profile 0,453 y vanilla 0,454. Es un resultado descriptivo fuerte en ese benchmark, no una ley de que calibración venza a escala.

La selección limita la generalización. Los siete modelos se eligen explícitamente por los gains más robustos tras 28 combinaciones; la tabla ampliada omite nueve modelos con ganancia no positiva. El panel de prototipado BRA/CHN/DEU/JPN/USA está incluido en los 20 países finales. El headline describe una selección favorable, no la tasa de éxito sobre arquitecturas no seleccionadas. Las tres semillas prueban estabilidad del muestreo del controlador, std macro hasta 0,006, pero no incertidumbre de WVS, del target humano o de escenarios.

En abierto, Claude 3 Opus lee cada respuesta, devuelve A/B y confianza y el estudio la transforma en pseudo-logit. Los cuatro modelos mejoran en media: 6,85% Llama, 6,67% Phi-4, 4,55% Qwen2.5 y 2,13% Phi-3.5. Sin embargo, la tabla contiene siete celdas país-modelo negativas, no seis como dice el texto; la peor es -3,10%. No hay validación humana de la decisión del juez, calibración de confianza ni equivalencia cultural. Añadir ruido gaussiano a pseudo-logits comprueba ruido numérico, no sesgo semántico. En BLEnD factual, siete de 16 países mejoran pero SEM-B agregado cae 4,4 puntos.

El vínculo WVS→AMCE está sobreinterpretado. Se ajustan diez predictores con sólo 20 países; con intercepto quedan nueve grados de libertad residuales. El texto destaca R² crudo 0,55-0,69, pero la tabla da R² ajustado 0,05-0,35, 0,05 en Age y 0,09 en Species, sin cross-validation, intervalos ni corrección por multiplicidad. La llamada causal impact matrix es una ablación leave-one-descriptor-out del controlador, no identificación causal de cómo WVS produce preferencias humanas. Una lengua por país también confunde cultura con traducción, tokenización y competencia lingüística.

Hay conflictos internos. Magistral tiene MIS 0,350 en la tabla principal y 0,334 en la ablación de los mismos 20 países. La comparación de la cuarta persona informa 0,4252 para el agregado y 0,4049 para el ancla utilitaria en Phi-4/20 países, incompatibles con el 0,346 principal; 0,4252 coincide con N=2 en otro sweep. Tail safety da ΔMIS 0,096 en 120 celdas, mientras las seis medias headline implican 0,1015. La batería llamada nueve checks muestra ocho filas. El arXiv dice 57 páginas y el PDF v2 renderiza 48. El oversampling con reemplazo duplica escenarios y cambia pesos; no añade información independiente.

La reproducibilidad pública no permite resolverlo. El paquete contiene TeX, bibliografía y cinco figuras, pero no código, perfiles WVS derivados, prompts completos, IDs de escenarios, logits, candidatos IS, gates, respuestas, juicios, resultados por run, sweep completo, entorno o revisiones exactas de modelos. El texto activo afirma que hay implementaciones en un archivo de código, mientras un comentario del fuente dice que se publicarán tras aceptación; no hay archivo ni repositorio enlazado. La lectura fiel es que DISCA reduce la distancia a promedios AMCE en una selección favorable de modelos y suele hacerlo en un track abierto mediado por Claude. No demuestra alineación cultural general, optimalidad del algoritmo completo, legitimidad normativa, protección de minorías ni reproducibilidad independiente.

Research question

Can disagreement among four national personas derived from WVS guide, without fine-tuning, a logit correction that brings LLM moral decisions closer to the mean human AMCEs of each country?

Method

DISCA compares base prompt and four personas per country, symmetrizes A/B, generates scalar candidates around consensus, weights them with Prospect-Theory utility and importance sampling, and contracts the correction according to the discrepancy between two passes. It is evaluated with MIS L2 against six AMCEs of MultiTP across 20 countries, seven selected models, and an open track of four models judged by Claude 3 Opus; baselines, ablations, BLEnD, and sweeps are added.

Sample: Headline binary: 20 countries, 7 selected models, and 6 MultiTP dimensions, with equal weight per country; three seeds except Gemma. Open track: 20 × 310 scenarios × 4 models, mediated by Claude. WVS→AMCE: 20 countries and 10 predictors. Selection: 28 combinations across 5 countries. Many sweeps: USA/VNM/DEU, Phi-4, N=250 per country.

Findings

  • The six selected models >=3.8B improve MIS 10.3%-23.6%; Gemma 2B improves 3.4%.
  • Llama-3.3-70B and Qwen3-VL-8B improve in 20/20 countries; Phi-4 in 18/20.
  • Phi-4+DISCA achieves MIS 0.346 versus 0.849 Llama-70B vanilla and 0.668 Llama+DISCA.
  • In Phi-4, DISCA 0.346 outperforms PRISM 0.384, MC-Dropout 0.403, activation steering 0.430, and vanilla 0.454.
  • The four open models improve 2.13%-6.85% on average, but there are 7 negative cells, not 6.
  • Aggregated BLEnD worsens by 4.4 pp despite seven countries with a gain.
  • Raw R² WVS→AMCE 0.55-0.69 reduces to adjusted R² 0.05-0.35 with N=20 and p=10.
  • Three seeds yield macro std <=0.006, limited to controller randomness.
  • Magistral, the fourth persona, tail-safety, and several counts present incompatible figures.

Limitations

  • arXiv v2 preprint without confirmed peer-reviewed acceptance.
  • Headline models selected after sweeping; nine failures/no-gains omitted.
  • Selection panel of five countries included in the final panel.
  • Crowdsourced target AMCE does not equate to legitimacy, moral truth, or individual fidelity.
  • Prompts use WVS per country; no-per-country-labels only applies to task/AMCE labels.
  • WVS without weights, cohort N, or uncertainty and discretized by quartiles.
  • OLS with 20 countries and 10 predictors, without cross-validation or inference.
  • Leave-one-out ablation labeled causal without causal identification.
  • i.i.d. assumption incompatible with shared model and scenario.
  • Oracle shrinkage incomputable; PT-IS and gate are not proven optimal.
  • One language per country conflates culture and linguistic capability.
  • Open track depends on Claude without human validation.
  • Equal-country macro does not weight by population or protect minorities.
  • Optional utility floor does not establish safety or rights.
  • Oversampling with replacement changes weights without adding observations.
  • Table and count conflicts without artifacts to resolve them.
  • Cost 3.6-3.7× vanilla and A/B logprobs requirement.
  • No code, derived data, outputs, full sweep, or environment.

What the study does not establish

  • General or normative cultural alignment.
  • That national averages represent individuals or minorities.
  • That WVS causes trolley preferences or validates prompts out of sample.
  • That LLM persona disagreement estimates real human disagreement.
  • MSE optimality of the full pipeline.
  • Superiority over fine-tuning or reward modeling not evaluated.
  • Success over non-selected architectures.
  • Transfer to factual knowledge.
  • Cultural validity of the Claude judge.
  • Protection against harmful majorities.
  • Scalability to APIs without logprobs.
  • Correction of the published numerical conflicts.
  • Independent reproduction.
  • Acceptance at NeurIPS 2026.

Traceability

Scope: Full text

Version: arXiv:2605.10843v2

Consulted source: https://arxiv.org/abs/2605.10843v2

Review: Codex 48-page visual full-text, complete TeX, WVS construct, theory-boundary, table arithmetic, selection, judge-validity and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Llama-3.3-70B
  • Magistral-Small-2509 24B
  • Phi-4 14B
  • Qwen3-VL-8B
  • Qwen2.5-7B
  • Phi-3.5-mini 3.8B
  • Gemma-4-E2B 2B
  • Claude-3-Opus-20240229 as open-ended judge
  • Additional checkpoints in the five-country model-selection sweep

Instruments and metrics

  • DISCA loss-averse Prospect-Theory importance-sampling controller
  • Dual-pass exponential reliability gate
  • A/B positional symmetrization and decision-token logits
  • Six-dimensional AMCE and MIS L2 distance
  • Jensen-Shannon distance and Pearson correlation
  • Claude 3 Opus choice-confidence pseudo-logit judge
  • WVS-to-AMCE standardized OLS
  • Leave-one-WVS-descriptor-out ablation
  • BLEnD Soft Exact Match

Data used

  • MultiTP multilingual Moral Machine extension
  • Moral Machine country human AMCE targets
  • World Values Survey Wave 7 individual microdata
  • Author-derived WVS country and age-cohort personas
  • GPT-4 synthetic 200-dilemma tuning pool with three-author proxy labels
  • Open-ended 20-country ethical scenario track
  • BLEnD factual cultural QA
  • No public derived personas, slices, logits, outputs or run-level results

Evidence and location

  • Full text, tables, proofs, prompts, WVS, preprocessing, and limits: arXiv:2605.10843v2; PDF sha256 c79a930ed186f89cd13ee5bd41b1d00f6402b0539906d8875aaab8e7f89afb15; TeX sha256 f60338bbca8df10560758509f4d226e47278ecb3ff4e44720f9611306531490b
  • Metadata and public version: https://arxiv.org/abs/2605.10843v2, submitted 2026-05-11 and revised 2026-05-18
  • Recalculation of adjusted R², counts, selection, construct, and reproducibility: reports/verification/article-341-disca-cultural-alignment-wvs-construct-selection-metric-reporting-and-reproducibility-audit.json