PLURAL: A Global Dataset for Value Alignment

Aditya Vashistha; Anya Shukla; Dhruv Agarwal; Tanya Goyal

REVIEW 3 major objections 8 minor 18 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

Survey-grounded synthetic data steers LLMs toward national values

2026-07-10 01:15 UTC pith:SRD6EAIT

load-bearing objection PLURAL is a solid dataset paper with a real resource released, but the human evaluation is not as independent from GLOBE as the paper implies. the 3 major comments →

arxiv 2607.08034 v1 pith:SRD6EAIT submitted 2026-07-09 cs.CL cs.AIcs.CY

PLURAL: A Global Dataset for Value Alignment

Dhruv Agarwal , Anya Shukla , Tanya Goyal , Aditya Vashistha This is my paper

classification cs.CL cs.AIcs.CY

keywords pluralvaluesurveyalignmentcountriesdatasetpreferencevalues

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Large language models disproportionately reflect Western values, yet are deployed worldwide. This paper argues that the missing ingredient for pluralistic alignment is not better training algorithms but better data: specifically, preference data grounded in rigorous, nationally representative social-science surveys rather than crowdsourced annotations or persona-based generation. The authors build PLURAL, a dataset of roughly 500,000 synthetic preference triplets covering 20 countries, by feeding individual survey responses from the Integrated Values Survey (a merged World and European Values Study spanning 92 countries) through a two-stage LLM pipeline that converts terse fixed-choice answers into realistic advice-seeking scenarios with preferred and dispreferred responses. The key claim is that this pipeline preserves both the differences between countries and the diversity within each country present in the original survey, and that models fine-tuned on country-specific subsets of PLURAL shift measurably toward the cultural profiles of those countries, as verified by both an automated benchmark using an independent cultural framework and blind human evaluation by 176 evaluators in India, Brazil, and Japan.

Core claim

The central discovery is that synthetic preference data generated from real, nationally representative survey responses carries enough country-specific value signal to steer LLM behavior, and that this signal survives the transformation from fixed-choice survey answers to naturalistic conversational scenarios. Fine-tuning on PLURAL reduced mean absolute error against an external cultural benchmark by 15.7–27.7 percent across five countries, and human evaluators in three countries preferred PLURAL-aligned responses over baseline responses 57–72 percent of the time. Crucially, a control experiment showed that grounding in actual survey responses outperformed persona-conditioned generation (18.

What carries the argument

The pipeline works in two stages. Stage 1 takes a participant's demographics and their answers to a group of related survey questions, and generates five preference triplets (prompt, preferred response, dispreferred response) that embed the participant's value orientation into a realistic scenario. Stage 2 expands terse responses into natural assistant-like answers while preserving the value stance. The pipeline uses multiple frontier LLMs with stochastic decoding for diversity, operates over question groups rather than individual items so the model can condition on multiple related answers, and filters survey questions to retain only prescriptive beliefs (what people ought to do) ratherthan

Load-bearing premise

The automated evaluation treats scores from the GLOBE cultural framework (derived from middle managers across 62 societies) as ground truth for each country's cultural profile. If those scores do not faithfully represent national values, the measured improvements could reflect convergence to a flawed target rather than genuine cultural alignment.

What would settle it

If models fine-tuned on PLURAL showed no improvement over demographic-prompting baselines on the GLOBE benchmark, or if human evaluators in India, Brazil, and Japan did not prefer PLURAL-aligned responses over vanilla model responses, the core claim that PLURAL contains learnable value signal would be undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

The same pipeline could be extended to all 92 countries in the Integrated Values Survey, providing value-alignment data at global scale without additional human annotation.
The finding that post-training compresses cross-country diversity (DPO retains only ~18 percent of the variation between country profiles) suggests that the bottleneck for pluralistic alignment may shift from data scarcity to training-method design.
Country-specific adapters trained on PLURAL could enable a single base model to serve culturally adapted behavior on demand, or could be composed into a single model that conditions on user nationality.
The tension between faithful cultural representation and safety norms (e.g., Gender Egalitarianism barely moved, possibly because matching some countries would require generating less egalitarian responses) defines a boundary condition for how far pluralistic alignment can or should go.
Survey-grounded synthetic generation could be applied to other large-scale social-science instruments beyond the IVS to create value-alignment data for sub-national communities or demographic groups.

Where Pith is reading between the lines

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

If PLURAL's pipeline preserves within-country diversity as claimed, one could train individual-specific or community-specific adapters rather than country-level ones, enabling personalization along value dimensions rather than just nationality.
The compression problem (18 percent retention under DPO, 30 percent under SFT) suggests that diversity-preserving preference optimization methods, such as those that explicitly penalize collapse toward a mode, might recover substantially more cross-country variation from the same data.
If the GLOBE framework used for automated evaluation does not faithfully represent national values (it samples middle managers, not nationally representative populations), the true cultural alignment gains from PLURAL may be larger or smaller than measured, depending on how GLOBE scores diverge from population-level values.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

PLURAL is a solid dataset paper with a real resource released, but the human evaluation is not as independent from GLOBE as the paper implies.

read the letter

The main thing to know: PLURAL is a genuinely useful dataset contribution. The authors take the Integrated Values Survey (nationally representative, 92 countries), filter to prescriptive value questions using Rokeach's hierarchy, and convert individual survey responses into ~500K synthetic preference triplets via a two-stage LLM pipeline. The dataset is public on HuggingFace. That alone makes the paper worth reading for anyone working on pluralistic alignment or cultural adaptation of LLMs. The PersonaHub ablation is the most novel empirical finding: holding the generation pipeline constant, IVS-grounded data reduces MAE by 18.1% versus 7.0% for persona-conditioned generation at equal scale. That's real evidence that survey grounding matters beyond generic persona diversity. The automated evaluation uses GLOBE, a framework distinct from the training data (IVS), which avoids the worst circularity trap. DPO fine-tuning reduces MAE by 15.7–27.7% across five countries, and general capabilities don't degrade (MMLU holds steady). The paper is also honest about limitations, which I appreciate. The stress-test concern about human evaluation independence is partially valid but overstated. Yes, the human eval uses GLOBE-derived prompts and selects the 36 prompts with the largest automated score gaps. This means the human eval tests agreement with GLOBE on GLOBE's strongest-signal items, not fully independent validation. That's a real methodological soft spot. But it's not contamination in the circular sense—the human evaluators are judging cultural typicality, not GLOBE scores, and they don't see the automated scores. The concern that GLOBE itself may be a flawed target (middle managers, not nationally representative) is acknowledged by the authors and is the more substantive issue. If GLOBE's dimensions are biased, then both the MAE metric and the prompt selection inherit that bias. The dataset-level validation (embedding-based country prediction at 78% accuracy) provides some independent signal but may capture lexical variation rather than deep value preservation. The DPO compression finding—only 18% of cross-country variation retained—is underplayed but important. It suggests current post-training methods can't fully exploit the signal PLURAL contains, which limits practical utility even if the dataset itself is strong. Minor issues: the LLM judge isn't validated against human annotators (though the authors use a pre-validated prompt), and human evaluators skew young and male. Overall, this is a well-executed dataset paper with a released resource, a reasonable evaluation, and honest limitations. The GLOBE dependency is the main weakness but doesn't undermine the core contribution. Recommend serious peer review, likely accept with revisions addressing the human eval independence concern and the GLOBE validity question more directly.

Referee Report

3 major / 8 minor

Summary. The manuscript introduces PLURAL, a large-scale synthetic preference dataset for value alignment grounded in the Integrated Values Survey (IVS), which provides nationally representative samples across 92 countries. The authors sample 100 respondents from each of 20 culturally diverse countries and use a two-stage LLM pipeline to convert terse survey responses into ~500,000 preference triplets (prompt, preferred response, dispreferred response). The pipeline uses Rokeach's hierarchy of beliefs to isolate prescriptive (normative) values from descriptive and primitive beliefs. The authors evaluate PLURAL via (i) dataset-level validation showing preservation of cross-country differences and within-country diversity, (ii) automated evaluation using the LLM-GLOBE benchmark showing MAE reductions of 15.7%–27.7% over baselines across five countries, and (iii) blind human evaluation with 176 evaluators in India, Brazil, and Japan. The dataset is publicly released.

Significance. The paper addresses an important gap in pluralistic alignment: existing preference datasets are either Western-skewed (PRISM) or limited in geographic coverage (Community Alignment). Grounding synthetic data generation in a rigorous, nationally representative social-science survey (IVS) is a sensible and scalable approach. The public release of ~500K triplets spanning 20 countries is a tangible contribution to the community. The PersonaHub ablation (§F.5) is a well-designed control that isolates the contribution of IVS grounding by holding the generation pipeline constant. The human evaluation, while limited in scope, provides complementary evidence beyond the automated metrics. The finding that DPO compresses cross-country diversity (§6, §F.6) is an honest and useful observation for future work.

major comments (3)

§3.4: The human evaluation is described as independent of GLOBE, but it is not. The prompts shown to evaluators are drawn from the LLM-GLOBE benchmark, and the 36 prompts per comparison are selected specifically because they exhibit the largest automated GLOBE score gap between models. This means the human evaluation tests whether humans agree with GLOBE on the subset of GLOBE prompts where GLOBE itself detects the strongest signal. If GLOBE's cultural dimensions are biased (a limitation acknowledged in §6), the selected prompts may be exactly those where GLOBE's biases are most pronounced, and human agreement on those prompts would not establish genuine cultural alignment independent of GLOBE. The authors should either (a) run a supplementary human evaluation on non-GLOBE prompts (e.g., from PLURAL's own held-out data or from a separate cultural benchmark) to provide a truly independent
§3.3 and §6: The GLOBE ground-truth scores are derived from middle managers across 62 societies, not nationally representative samples. The authors acknowledge this limitation in §6 but do not discuss its implications for the central MAE results. If GLOBE scores systematically differ from national population values (e.g., because middle managers are more urban, educated, or globally oriented), then reducing MAE to GLOBE scores may measure convergence to a managerial subpopulation rather than to the target country's cultural profile. The authors should add a sensitivity analysis or at least a quantitative argument for why GLOBE scores are a reasonable proxy for national values despite the sampling difference, particularly given that the IVS training data IS nationally representative.
§4.1: The dataset-level validation uses embedding-based representations (google/embeddinggemma-300m) to test whether cross-country value differences and within-country diversity are preserved. However, embeddings may capture surface-level lexical and stylistic variation rather than genuine value preservation. The country-prediction accuracy of 78.0% (Table 2) could partly reflect country-specific vocabulary or phrasing artifacts introduced during synthetic generation rather than preserved value structure. The authors should include a control that tests whether embeddings of responses with shuffled or neutralized value content still yield above-chance country prediction, which would indicate that lexical artifacts rather than value signals drive the result.

minor comments (8)

§2.3, Table 1: The table states 'Comparisons ~500K' but the exact count is not given. Providing the precise number would be helpful.
§2.4: The pool of frontier LLMs used for generation includes 'GPT-5 Mini, Mistral Large 3, DeepSeek-v3.2.' These model names should be verified for accuracy; some appear to be future versions that may not yet exist or may be named differently.
Figure 2: The examples are labeled 'abridged' but the degree of abridgment is unclear. Including at least one full, unabridged example in the main text (in addition to Table 12 in the appendix) would help readers assess data quality without navigating to the appendix.
§3.4: The human evaluation compensates evaluators at $8/hour. It would be useful to report the average time taken per evaluation session to confirm that this rate is fair and that evaluators were not rushed.
Table 3: The 'vs. CA' row for Japan shows '–' but the text does not clearly explain why CA was not available for Japan. The appendix (§E.5) explains that CA covers only five countries, but this should be noted in the main text for clarity.
§F.4: The additional model experiments (Qwen-2.5-7b and 14b) are conducted only for Brazil. While the justification is reasonable, testing at least one additional country would strengthen the generalization claim across model families.
Appendix A: The inter-coder agreement (Cohen's κ = 0.65) is described as 'substantial agreement,' which is correct per Landis & Koch (1977). However, the reference for this benchmark is missing.
§6: The discussion of tension between cultural alignment and safety norms (re: Gender Egalitarianism) is important but brief. Expanding this discussion, perhaps with a concrete example of a conflict, would strengthen the paper's contribution to the pluralistic alignment agenda.

Circularity Check

0 steps flagged

No significant circularity: the paper's central claims are validated against external benchmarks (GLOBE) and independent human evaluation, with no self-definitional or fitted-input-as-prediction patterns.

full rationale

The paper's derivation chain is straightforward and self-contained. PLURAL is generated from IVS survey responses (§2.4), then evaluated via three methods: (1) dataset-level validation using embedding-based country prediction (§4.1), (2) automated evaluation using the GLOBE framework—a distinct cultural framework from the training data IVS (§3.3, §4.2), and (3) blind human evaluation (§3.4, §4.3). The central claim—that PLURAL contains learnable value signal—is tested against GLOBE, which is explicitly described as 'entirely distinct from the one used for training' (§3.3). The human evaluation uses independent human judges. The PersonaHub ablation (§4.2, App. F.5) holds the generation pipeline constant while varying only the grounding source, providing a fair methodological comparison. No step in the derivation chain reduces to its own inputs by construction. The dataset-level validation (§4.1) checks whether synthetic generation preserves IVS structure, but this is a validation step, not a prediction claim. The skeptic's concern that the human evaluation uses GLOBE-derived prompts is a methodological independence concern (correctness risk), not circularity—the human judgments are not forced by construction to agree with GLOBE scores, and the paper does not claim the human evaluation independently validates GLOBE itself. The paper cites prior work by some of the same authors (Agarwal et al., 2025a,b) for context on Western-centric bias and 'fluent but foreign' regional models, but these citations provide motivation rather than load-bearing mathematical or empirical premises. No self-citation chain forces the central result. Score: 2, reflecting minor self-citations that are not load-bearing and a central claim with independent empirical content validated against external benchmarks and human judgment.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper introduces no new theoretical entities or postulated objects. The free parameters are standard ML hyperparameters and design choices, none fitted to the evaluation target. The axioms are domain assumptions from social science (Rokeach, GLOBE) plus the core assumption that LLM-based synthetic generation preserves value signal, which is empirically tested rather than assumed.

free parameters (5)

N=100 respondents per country = 100
Selected via Monte Carlo simulation as the smallest N where 99th percentile TVD falls below 0.05. Not fitted to the target result but is a design choice affecting dataset composition.
DPO beta = 0.1
Standard value, not tuned specifically for this dataset.
LoRA rank r = 16
Standard configuration choice.
Temperature for generation = 1.0
Chosen to increase stylistic diversity; standard practice.
5 triplets per question group = 5
Design choice to promote scenario diversity while keeping value orientation fixed.

axioms (4)

domain assumption Rokeach's hierarchy of beliefs correctly separates prescriptive (normative) values from descriptive and primitive beliefs, and only prescriptive beliefs are relevant for AI behavior.
Used in §2.2 to filter 59 question groups down to 41. The 79.7% inter-coder agreement (Cohen's kappa=0.65) provides some validation, but the framework itself is taken as established background.
domain assumption GLOBE cultural dimensions constitute a valid external benchmark for measuring LLM cultural alignment.
Used in §3.3 as the evaluation framework. The authors acknowledge the limitation that GLOBE data comes from middle managers rather than nationally representative samples (§6).
ad hoc to paper LLM-generated synthetic preference triplets faithfully preserve the normative value signals present in the source survey responses.
This is partially validated by the dataset-level analysis (§4.1) showing 78% country-prediction accuracy and low Wasserstein distances, but the assumption that synthetic text preserves value signal is foundational to the entire approach.
domain assumption Country-level typicality judgments from crowd workers are a valid proxy for national cultural alignment.
Used in §3.4 for human evaluation. The authors acknowledge evaluators skew young and male (Table 6) and that asking for country-level typicality mitigates but does not eliminate sampling bias.

pith-pipeline@v1.1.0-glm · 29798 in / 2732 out tokens · 269337 ms · 2026-07-10T01:15:00.745425+00:00 · methodology

0 comments

read the original abstract

Large language models (LLMs) are used worldwide, yet disproportionately reflect Western values, limiting their ability to represent diverse value systems. We introduce PLURAL, a large-scale, value-focused preference dataset grounded in the Integrated Values Survey (IVS), a nationally representative survey spanning 92 countries. Using a two-stage generation pipeline, we transform survey responses into synthetic preference triplets that preserve normative value signals while producing realistic scenarios. We release an initial version of PLURAL containing ~500,000 preference triplets representing people in 20 diverse countries. We evaluate PLURAL in three ways: (i) dataset-level validation showing that it preserves both cross-country value differences and within-country diversity from the original survey; (ii) automated evaluation showing that training on PLURAL improves alignment with target countries' cultural profiles, reducing mean absolute error by up to 27.7% relative to strong baselines; and (iii) blind human evaluation with 176 evaluators in India, Brazil, and Japan, who judge PLURAL-aligned responses as more representative of their national values. Together, these results show that PLURAL contains learnable signal for value steering, offering a scalable resource for pluralistic alignment. Dataset: https://huggingface.co/datasets/agdhruv/plural-alignment

Figures

Figures reproduced from arXiv: 2607.08034 by Aditya Vashistha, Anya Shukla, Dhruv Agarwal, Tanya Goyal.

**Figure 2.** Figure 2: PLURAL dataset examples (abridged). See Table [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Within-country diversity. Procedure. We structure the human evaluation as a comparison between two models (e.g., a baseline model and our fine-tuned model). Using prompts from the LLM-GLOBE benchmark, we first generate responses from both models and select a subset of 36 prompts for which the automated score gap between them is largest. Because GLOBE scores reflect position along a cultural dimension rathe… view at source ↗

**Figure 4.** Figure 4: MAE vs GLOBE ground truth (lower is better) for [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Heatmap (left): Reduction in MAE for each country across the nine GLOBE [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: Example rubric used by the LLM judge to assign 1–7 scores for each GLOBE [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗

**Figure 7.** Figure 7: Example scenario and paired responses shown to evaluators for cultural value [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗

**Figure 8.** Figure 8: Survey questions used to elicit value judgments from evaluators. [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗

**Figure 9.** Figure 9: Similar participants yield diverse scenarios. No significant template collapse. Since the data generation relies on LLMs, we check whether similar demographic groups collapse into narrow templatized questions (e.g., always mapping “elderly Japanese male who values family” to the same scenario). To do so, we first group participants by demographic and value profiles, yielding “buckets” of similar partici… view at source ↗

**Figure 10.** Figure 10: Automated evaluation on additional base models for Brazil (lower is better). [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗

**Figure 11.** Figure 11: PCA embedding of ground-truth country profiles, the vanilla model, Community Alignment, and our country-specific DPO models in the 9-dimensional GLOBE space. The target profiles are more widely separated than the adapted models, indicating that DPO retains only part of the cross-country spread. Method Retained DPO (1 epoch) 18% SFT (1 epoch) 23% SFT (2 epochs) 30% [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · 3 internal anchors

[1]

URLhttps://arxiv.org/abs/2406.15951. Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, 11 Preprint. Under review. Aviya Skowron, Lintang Sutawika, Eric Ta...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1017/cbo9780511790881 2024
[2]

ISBN 9781605589299

Association for Computing Machinery. ISBN 9781605589299. doi: 10.1145/1753326. 1753522. URLhttps://doi.org/10.1145/1753326.1753522. Jiaming Ji, Donghai Hong, Borong Zhang, Boyuan Chen, Juntao Dai, Boren Zheng, Tianyi Qiu, Jiayi Zhou, Kaile Wang, Boxuan Li, Sirui Han, Yike Guo, and Yaodong Yang. Pku- saferlhf: Towards multi-level safety alignment for llms ...

work page doi:10.1145/1753326 2025
[3]

NormAd: A Framework for Measuring the Cultural Adaptability of Large Language Models

URLhttps://arxiv.org/abs/2404.12464. Milton Rokeach.The Nature of Human Values. Free Press, New York, 1973. 13 Preprint. Under review. Shalom H. Schwartz.Universals in the Content and Structure of Values: Theoretical Advances and Empirical Tests in 20 Countries, pp. 1–65. Elsevier, 1992. ISBN 9780120152254. doi: 10.1016/ s0065-2601(08)60281-6. URLhttp://d...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/s0065-2601(08)60281-6 1973
[4]

A Roadmap to Pluralistic Alignment

ISSN 1932-6203. doi: 10.1371/journal.pone.0329179. URL http://dx.doi.org/10. 1371/journal.pone.0329179. Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi. A roadmap to pluralistic alignment, 2024. URL https://arxiv. ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1371/journal.pone.0329179 1932
[5]

prescriptive

URLhttps://arxiv.org/abs/2412.11167. Pardis Sadat Zahraei and Ehsaneddin Asgari. I am aligned, but with whom? mena values benchmark for evaluating cultural alignment and multilingual bias in llms, 2025. URL https://arxiv.org/abs/2510.13154. Lily Hong Zhang, Smitha Milli, Karen Jusko, Jonathan Smith, Brandon Amos, Wassim Bouaziz, Manon Revel, Jack Kussman,...

work page arXiv 2025
[6]

A question a user might ask an AI

work page
[7]

# Methodology We have access to the user’s response to a bunch of value-related survey questions

A dispreferred answer Such that, if the user is presented with the dispreferred and preferred answers as a response to the question, they would prefer the preferred response. # Methodology We have access to the user’s response to a bunch of value-related survey questions. We are trying to infer their value preferences based on their answers. To do so, we ...

work page
[8]

Demographic details about the user

work page
[9]

One or more value-related questions they were asked

work page
[10]

Their response to the question(s) # Output You are required to transpose them into a preference dataset style. Output:

work page
[11]

A question the user might ask an AI

work page
[12]

A preferred response, based on their response to the survey question

work page
[13]

- The scenario should be realistic with detailed descriptions

A dispreferred response, contrary to their response to the survey question # Tips - The question should be in first person, framed as if the user is asking for advice in a given scenario. - The scenario should be realistic with detailed descriptions. Instead of generic scenarios, present *specific* scenarios that a user might encounter in real life. - Uti...

work page 2024
[14]

These values may or may not match your own views.)

Which response is more typical of values commonly held in Japan? (Think about Japan overall. These values may or may not match your own views.)

work page
[15]

Which response better aligns with your own values? (Think about your own personal views.)

work page
[16]

the traditional Japanese virtue of ‘Ganbari’ (doing one’s best)

Briefly explain your choices (optional). Figure 8: Survey questions used to elicit value judgments from evaluators. At the end of each survey (10 questions), we performed an attention check and collected demographic information. All evaluators passed the attention check. Personal vs. Country Values.Evaluators answered two separate questions: one about whi...

work page 2021
[17]

Performance is measured on MMLU (Hendrycks et al., 2021) and TinyMMLU (Polo et al.,

by comparing llama-3.1-8b-instruct before and after DPO fine-tuning on PLURAL. Performance is measured on MMLU (Hendrycks et al., 2021) and TinyMMLU (Polo et al.,

work page 2021
[18]

We use the Brazil-adapted LoRA checkpoint described in Section3.2 for this comparison

using lm-eval (Gao et al., 2024). We use the Brazil-adapted LoRA checkpoint described in Section3.2 for this comparison. 23 Preprint. Under review. Growing up in a rural area, I’ve seen how con- flicts can devastate farming communities. Now, with rumors of unrest, my elderly parents rely on me to manage our family land. If called to fight, I’d have to lea...

work page 2024

[1] [1]

URLhttps://arxiv.org/abs/2406.15951. Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, 11 Preprint. Under review. Aviya Skowron, Lintang Sutawika, Eric Ta...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1017/cbo9780511790881 2024

[2] [2]

ISBN 9781605589299

Association for Computing Machinery. ISBN 9781605589299. doi: 10.1145/1753326. 1753522. URLhttps://doi.org/10.1145/1753326.1753522. Jiaming Ji, Donghai Hong, Borong Zhang, Boyuan Chen, Juntao Dai, Boren Zheng, Tianyi Qiu, Jiayi Zhou, Kaile Wang, Boxuan Li, Sirui Han, Yike Guo, and Yaodong Yang. Pku- saferlhf: Towards multi-level safety alignment for llms ...

work page doi:10.1145/1753326 2025

[3] [3]

NormAd: A Framework for Measuring the Cultural Adaptability of Large Language Models

URLhttps://arxiv.org/abs/2404.12464. Milton Rokeach.The Nature of Human Values. Free Press, New York, 1973. 13 Preprint. Under review. Shalom H. Schwartz.Universals in the Content and Structure of Values: Theoretical Advances and Empirical Tests in 20 Countries, pp. 1–65. Elsevier, 1992. ISBN 9780120152254. doi: 10.1016/ s0065-2601(08)60281-6. URLhttp://d...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/s0065-2601(08)60281-6 1973

[4] [4]

A Roadmap to Pluralistic Alignment

ISSN 1932-6203. doi: 10.1371/journal.pone.0329179. URL http://dx.doi.org/10. 1371/journal.pone.0329179. Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi. A roadmap to pluralistic alignment, 2024. URL https://arxiv. ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1371/journal.pone.0329179 1932

[5] [5]

prescriptive

URLhttps://arxiv.org/abs/2412.11167. Pardis Sadat Zahraei and Ehsaneddin Asgari. I am aligned, but with whom? mena values benchmark for evaluating cultural alignment and multilingual bias in llms, 2025. URL https://arxiv.org/abs/2510.13154. Lily Hong Zhang, Smitha Milli, Karen Jusko, Jonathan Smith, Brandon Amos, Wassim Bouaziz, Manon Revel, Jack Kussman,...

work page arXiv 2025

[6] [6]

A question a user might ask an AI

work page

[7] [7]

# Methodology We have access to the user’s response to a bunch of value-related survey questions

A dispreferred answer Such that, if the user is presented with the dispreferred and preferred answers as a response to the question, they would prefer the preferred response. # Methodology We have access to the user’s response to a bunch of value-related survey questions. We are trying to infer their value preferences based on their answers. To do so, we ...

work page

[8] [8]

Demographic details about the user

work page

[9] [9]

One or more value-related questions they were asked

work page

[10] [10]

Their response to the question(s) # Output You are required to transpose them into a preference dataset style. Output:

work page

[11] [11]

A question the user might ask an AI

work page

[12] [12]

A preferred response, based on their response to the survey question

work page

[13] [13]

- The scenario should be realistic with detailed descriptions

A dispreferred response, contrary to their response to the survey question # Tips - The question should be in first person, framed as if the user is asking for advice in a given scenario. - The scenario should be realistic with detailed descriptions. Instead of generic scenarios, present *specific* scenarios that a user might encounter in real life. - Uti...

work page 2024

[14] [14]

These values may or may not match your own views.)

Which response is more typical of values commonly held in Japan? (Think about Japan overall. These values may or may not match your own views.)

work page

[15] [15]

Which response better aligns with your own values? (Think about your own personal views.)

work page

[16] [16]

the traditional Japanese virtue of ‘Ganbari’ (doing one’s best)

Briefly explain your choices (optional). Figure 8: Survey questions used to elicit value judgments from evaluators. At the end of each survey (10 questions), we performed an attention check and collected demographic information. All evaluators passed the attention check. Personal vs. Country Values.Evaluators answered two separate questions: one about whi...

work page 2021

[17] [17]

Performance is measured on MMLU (Hendrycks et al., 2021) and TinyMMLU (Polo et al.,

by comparing llama-3.1-8b-instruct before and after DPO fine-tuning on PLURAL. Performance is measured on MMLU (Hendrycks et al., 2021) and TinyMMLU (Polo et al.,

work page 2021

[18] [18]

We use the Brazil-adapted LoRA checkpoint described in Section3.2 for this comparison

using lm-eval (Gao et al., 2024). We use the Brazil-adapted LoRA checkpoint described in Section3.2 for this comparison. 23 Preprint. Under review. Growing up in a rural area, I’ve seen how con- flicts can devastate farming communities. Now, with rumors of unrest, my elderly parents rely on me to manage our family land. If called to fight, I’d have to lea...

work page 2024