arxiv: 2604.12851 · v1 · submitted 2026-04-14 · 💻 cs.CY

Recognition: unknown

Can Persona-Prompted LLMs Emulate Subgroup Values? An Empirical Analysis of Generalisability and Fairness in Cultural Alignment

Bryan Chen Zhengyu Tan , Zhengyuan Liu , Xiaoyuan Yi , Jing Yao , Xing Xie , Nancy F. Chen , Roy Ka-Wei Lee

Authors on Pith no claims yet

Pith reviewed 2026-05-10 14:11 UTC · model grok-4.3

classification 💻 cs.CY

keywords cultural alignmentLLM fairnesspersona promptingsubgroup valuesWorld Values SurveySingaporefine-tuning

0 comments

The pith

Even advanced LLMs reach only 57.4 percent accuracy when predicting the cultural preferences of Singapore demographic subgroups.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tests whether LLMs can be made to match the separate value patterns of different demographic groups inside one country rather than a single average set of values. Using Singapore and World Values Survey data, the authors show that even GPT-4.1 predicts subgroup modal answers correctly only 57.4 percent of the time. A simple fine-tuning step on structured preference numbers raises accuracy on entirely new subgroups by 17.4 percent on average and carries some benefit into open-ended answers. At the same time the training step increases the size of performance differences across subgroups when measured by distance-aware scores, and models already handle young, male, Chinese, and Christian personas more easily than others.

Core claim

State-of-the-art LLMs achieve only 57.4 percent accuracy in predicting the modal preferences of demographic subgroups drawn from the World Values Survey in Singapore. Fine-tuning on a dataset of over 20,000 structured numerical preference examples lifts accuracy on unseen, out-of-distribution subgroups by an average of 17.4 percent, with partial transfer to open-ended generation. Pre-existing biases favor emulation of young, male, Chinese, and Christian personas, and the same fine-tuning step that raises average accuracy also widens disparities between subgroups under distance-aware metrics.

What carries the argument

Persona prompts paired with numerical World Values Survey responses used to score and then fine-tune model predictions of subgroup modal preferences.

If this is right

Fine-tuning on structured numerical preferences raises accuracy on unseen subgroups by an average of 17.4 percent.
Models already emulate young, male, Chinese, and Christian personas more accurately than other subgroups.
Fine-tuning improves average performance yet widens subgroup disparities when distance-aware metrics are used.
The accuracy gains from fine-tuning partially carry over to open-ended generation tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Developers should track distance-aware disparity metrics during alignment work instead of reporting only average accuracy.
The same fine-tuning approach could be tried on value surveys from other multicultural countries to check whether the accuracy-disparity trade-off repeats.
Additional regularization or fairness constraints may be needed during fine-tuning to keep gains from increasing group-level differences.
Real-world deployment tests with live users from different subgroups would show whether the measured improvements survive outside the survey format.

Load-bearing premise

World Values Survey answers collected in Singapore correctly capture the true modal preferences of its demographic subgroups and persona prompts can make models produce outputs that match those preferences.

What would settle it

Collect fresh survey responses from the same Singapore subgroups on a new set of questions and test whether the fine-tuned models still show the reported accuracy gains and transfer effects over baseline models.

Figures

Figures reproduced from arXiv: 2604.12851 by Bryan Chen Zhengyu Tan, Jing Yao, Nancy F. Chen, Roy Ka-Wei Lee, Xiaoyuan Yi, Xing Xie, Zhengyuan Liu.

**Figure 1.** Figure 1: Overview of the experimental framework. (A) We first map the societal value landscape using our Modal Diversity Score. (B–C) We fine-tune models on structured preferences to test i) compositional generalisation to unseen intersectional personas (e.g., age × religion) and ii) transfer to open-ended generation. (D) While SFT yields substantial accuracy gains on out-of-distribution (OOD) subgroups, (E) our an… view at source ↗

**Figure 2.** Figure 2: Full prompt templates used in our experiments. These are dynamically populated with variables (e.g., [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗

read the original abstract

Despite their global prevalence, many Large Language Models (LLMs) are aligned to a monolithic, often Western-centric set of values. This paper investigates the more challenging task of fine-grained value alignment: examining whether LLMs can emulate the distinct cultural values of demographic subgroups. Using Singapore as a case study and the World Values Survey (WVS), we examine the value landscape and show that even state-of-the-art models like GPT-4.1 achieve only 57.4% accuracy in predicting subgroup modal preferences. We construct a dataset of over 20,000 samples to train and evaluate a range of models. We demonstrate that simple fine-tuning on structured numerical preferences yields substantial gains, improving accuracy on unseen, out-of-distribution subgroups by an average of 17.4%. These gains partially transfer to open-ended generation. However, we find significant pre-existing performance biases, where models better emulate young, male, Chinese, and Christian personas. Furthermore, while fine-tuning improves average performance, it widens the disparity between subgroups when measured by distance-aware metrics. Our work offers insights into the limits and fairness implications of subgroup-level cultural alignment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Fine-tuning on WVS subgroup data lifts average accuracy but widens distance-aware gaps, with the main limit being the unvalidated assumption that survey responses equal true modal values.

read the letter

The main thing to know is that even GPT-4.1 only hits 57% accuracy predicting modal preferences for Singapore demographic subgroups on WVS items, simple fine-tuning raises out-of-distribution accuracy by 17 points on average, yet the same step increases disparity when measured by distance-aware metrics. The paper also notes pre-existing advantages for young, male, Chinese, and Christian personas and some carry-over to open-ended generation. They built a 20k-sample dataset from the Singapore WVS wave and tested held-out subgroups, which is the concrete new piece. The numbers are reported clearly and the OOD split plus the fairness delta give a practical signal that average gains can come at the cost of uneven performance. That part is useful for anyone tracking how alignment techniques behave across groups. The soft spot is exactly the one the stress-test flags: the work treats WVS item responses as ground-truth modal values and persona prompts as a reliable readout of internalized preferences. Neither is checked against other sources or prompt variants in the reported experiments. Survey data can shift with wording, order, or social-desirability effects, and we already know persona prompts often surface instruction following rather than stable representations. If those premises are off, the accuracy figures and the widening-disparity claim lose force. The paper is Singapore-specific, so broader claims about cultural alignment need that caveat. Readers working on multicultural LLM deployment or fairness audits will find the empirical measurements worth looking at; the dataset size and the before-after fairness comparison are the parts that stand up on their own. It is coherent enough and grounded enough in actual runs to deserve a serious referee, though the review should press on validation of the preference labels and on whether the distance-aware metric is the right one for the fairness question. I would send it out for review with those two points as required revisions.

Referee Report

2 major / 2 minor

Summary. The paper investigates whether LLMs can emulate distinct cultural values of demographic subgroups, using Singapore and the World Values Survey (WVS) as a case study. It reports that even GPT-4.1 achieves only 57.4% accuracy in predicting subgroup modal preferences, constructs a dataset of over 20,000 samples, and shows that simple fine-tuning on structured numerical preferences improves accuracy on unseen out-of-distribution subgroups by an average of 17.4%. These gains partially transfer to open-ended generation, but pre-existing biases favor young, male, Chinese, and Christian personas, and fine-tuning widens subgroup disparities under distance-aware metrics.

Significance. If the empirical results hold under the stated assumptions, the work provides concrete evidence on the limits of persona-based cultural alignment and the fairness trade-offs of fine-tuning for subgroup values. The large-scale dataset (>20,000 samples) and reporting of specific accuracy numbers plus OOD improvement deltas are strengths that enable measurable assessment of generalisability and disparity effects.

major comments (2)

[§3 (Data Construction)] §3 (Data Construction): The headline results (57.4% baseline accuracy, 17.4% OOD gain, and widened distance-aware disparities) rest on the unvalidated premise that WVS Singapore responses accurately encode the 'true' modal preferences of the defined demographic subgroups. No sensitivity checks for survey artifacts (question order, social-desirability bias, translation effects) are reported, rendering the accuracy and fairness claims dependent on this premise.
[§4 (Persona Prompting Experiments)] §4 (Persona Prompting Experiments): The evaluation treats persona-prompted next-token distributions as faithful readouts of internalized subgroup values, but provides no ablations on prompt phrasing variations or controls to separate instruction-following from stable value emulation. This distinction is load-bearing for the claim that fine-tuning improves emulation and for the partial transfer to open-ended generation.

minor comments (2)

[Abstract] Abstract: The model reference 'GPT-4.1' is non-standard; specifying the exact version or API identifier would aid reproducibility.
[§5 (Fairness Analysis)] §5 (Fairness Analysis): The distance-aware metrics are introduced without an explicit formula or pseudocode; adding this would strengthen the replicability of the disparity-widening claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive feedback on our manuscript. We address each major comment below, indicating planned revisions where appropriate to improve clarity and robustness.

read point-by-point responses

Referee: §3 (Data Construction): The headline results (57.4% baseline accuracy, 17.4% OOD gain, and widened distance-aware disparities) rest on the unvalidated premise that WVS Singapore responses accurately encode the 'true' modal preferences of the defined demographic subgroups. No sensitivity checks for survey artifacts (question order, social-desirability bias, translation effects) are reported, rendering the accuracy and fairness claims dependent on this premise.

Authors: We acknowledge that the World Values Survey, like all large-scale surveys, is subject to artifacts including social-desirability bias, question-order effects, and translation issues. Our analysis positions WVS responses as the observable reference standard for subgroup modal preferences, following established practice in cross-cultural value research. The paper's claims concern the degree to which LLMs can match these reported preferences rather than asserting they represent ground-truth values. In the revised manuscript we will expand the Limitations section to explicitly discuss these survey characteristics and their implications for interpreting accuracy and disparity metrics. revision: partial
Referee: §4 (Persona Prompting Experiments): The evaluation treats persona-prompted next-token distributions as faithful readouts of internalized subgroup values, but provides no ablations on prompt phrasing variations or controls to separate instruction-following from stable value emulation. This distinction is load-bearing for the claim that fine-tuning improves emulation and for the partial transfer to open-ended generation.

Authors: Our persona prompts follow templates commonly used in the LLM value-elicitation literature and were held fixed across models and subgroups to isolate the effect of demographic specification. While the current version does not include prompt-variation ablations, the consistent patterns observed across multiple base models provide indirect support for robustness. We will add a new subsection with prompt-phrasing ablations and a control condition that removes persona information, allowing clearer separation of instruction-following from value emulation and strengthening the interpretation of fine-tuning gains. revision: yes

Circularity Check

0 steps flagged

Empirical evaluation with no self-referential derivations or fitted predictions

full rationale

The paper conducts an empirical study: it constructs a dataset from World Values Survey responses, measures LLM accuracy on subgroup modal preference prediction via persona prompts (reporting 57.4% for GPT-4.1), evaluates fine-tuning gains on held-out OOD subgroups (17.4% average improvement), and assesses fairness via distance-aware metrics. No equations, derivations, or predictions are presented that reduce by construction to the inputs; all headline numbers are direct experimental measurements on external data splits rather than self-defined or self-cited quantities. The work is therefore self-contained against external benchmarks with no load-bearing self-citation chains or ansatz smuggling.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on the assumption that survey modal preferences serve as valid ground truth and that persona-based prompting plus fine-tuning can be evaluated as emulation of those preferences.

axioms (1)

domain assumption World Values Survey responses accurately capture the modal preferences of demographic subgroups in Singapore
Used directly as the evaluation target for model predictions and fine-tuning.

pith-pipeline@v0.9.0 · 5532 in / 1175 out tokens · 41536 ms · 2026-05-10T14:11:23.251414+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 1 canonical work pages · 1 internal anchor

[1]

Jacy Reese Anthis, Ryan Liu, Sean M

Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Stud- ies. Jacy Reese Anthis, Ryan Liu, Sean M. Richardson, Austin C. Kozlowski, Bernard Koch, Erik Brynjolf- sson, James Evans, and Michael S. Bernstein. 2025. Position: LLM social simulations are a promising research method. InForty-Second International Con- ference on M...

2025
[2]

In First Conference on Language Modeling

Evaluating cultural adaptability of a large lan- guage model via simulation of synthetic personas. In First Conference on Language Modeling. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Effi- cient memory management for large language model serving with PagedAttention....

2023
[3]

William Miner

Phi-4-mini technical report: Compact yet pow- erful multimodal language models via mixture-of- LoRAs. William Miner. 2023. In Singapore, religious diversity and tolerance go hand in hand. MistralAI. 2025. Mistral small 3.1 | mistral AI. https://mistral.ai/news/mistral-small-3-1. Mohammad Alami Musa. 2023. Singapore’s secularism and its pragmatic approach ...

2023
[4]

Justin Ong

SEA-LION: Southeast asian languages in one network. Justin Ong. 2023. Singaporean youth less likely than older generations to have seen their social status rise: IPS study. https://www.todayonline.com/singapore/youth- social-status-ips-study-2089046. OpenAI. 2024a. Hello GPT-4o. https://openai.com/index/hello-gpt-4o/. OpenAI. 2024b. Introducing GPT-4.1 in...

2023
[5]

Direct Preference Optimization: Your Language Model is Secretly a Reward Model

Direct Preference Optimization: Your Lan- guage Model is Secretly a Reward Model.NeurIPS 2023, abs/2305.18290. Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose opinions do language models reflect? In Proceedings of the 40th International Conference on Machine Learning, volume 202 ofProceedings of ...

work page internal anchor Pith review arXiv 2023
[6]

Taylor Sorensen, Liwei Jiang, Jena D

Role-play with large language models. Taylor Sorensen, Liwei Jiang, Jena D. Hwang, Sydney Levine, Valentina Pyatkin, et al. 2024. Value kalei- doscope: Engaging AI with pluralistic human values, rights, and duties.Proceedings of the AAAI Confer- ence on Artificial Intelligence, 38(18):19937–19947. Nicholas Sukiennik, Chen Gao, Fengli Xu, and Yong Li

2024
[7]

Bryan Chen Zhengyu Tan, Shaun Khoo, Bich Ngoc Doan, Zhengyuan Liu, Nancy F

An evaluation of cultural value alignment in LLM. Bryan Chen Zhengyu Tan, Shaun Khoo, Bich Ngoc Doan, Zhengyuan Liu, Nancy F. Chen, and Roy Ka- Wei Lee. 2026a. Small changes, big impact: De- mographic bias in LLM-based hiring through subtle sociocultural markers in anonymised resumes. Bryan Chen Zhengyu Tan and Roy Ka-Wei Lee
[8]

Unmasking implicit bias: Evaluating persona- prompted LLM responses in power-disparate social scenarios. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Asso- ciation for Computational Linguistics: Human Lan- guage Technologies (Volume 1: Long Papers), pages 1075–1108, Albuquerque, New Mexico. Association for Computation...

2025
[9]

Open source software available from https://github.com/heartexlabs/label-studio

Label Studio: Data labeling soft- ware. Open source software available from https://github.com/heartexlabs/label-studio. Kush R. Varshney. 2024. Decolonial AI alignment: Openness, visesa-dharma, and including excluded knowledges.Proceedings of the AAAI/ACM Confer- ence on AI, Ethics, and Society, 7(1):1467–1481. Yuhang Wang, Yanxu Zhu, Chao Kong, Shuyu We...

2024
[10]

Sub- groups with fewer than 30 respondents (shown in grey) were deemed statistically insignifi- cant and excluded from our study

WVS Respondents (N):The total number of individuals from the original WVS sample of 2,012 who belong to that subgroup. Sub- groups with fewer than 30 respondents (shown in grey) were deemed statistically insignifi- cant and excluded from our study
[11]

WVS Respon- dents

Dataset Samples (S):The number of question-answer pairs generated for that sub- group in our final dataset. The maximum pos- sible is 214, corresponding to one sample for every valid WVS question. This distinction is important. The "WVS Respon- dents" count determines a subgroup’s inclusion, while the "Dataset Samples" count reflects its rep- resentation ...
[12]

Persona Analysis

Analyse Persona Adherence: - Assess how convincingly each response embodies the specified persona. - Consider its tone, linguistic style, likely priorities, and reasoning. Avoid superficial stereotypes. - Write your analysis in the "Persona Analysis" section
[13]

Ground Truth Value

Analyse Value Alignment: - Assess how well the underlying stance of each response aligns with the "Ground Truth Value". - The alignment should be semantic, not necessarily a literal match. - Write your analysis in the "Value Analysis" section
[14]

persona_winner

Provide Final Judgement: - After your analysis, provide a winner for each criterion and an overall winner. The overall winner should be the response you judge to be of higher quality and more useful, considering both criteria. - Format your final decision as a JSON object at the end of your response. — START OF EV ALUATION — Persona Analysis: [Your brief ...

2048