Organizations: Microsoft Research Asia · Sungkyunkwan University · Singapore University of Technology and Design · University of British Columbia · Institute of Science Tokyo · Renmin University of China
Cultural value alignment requires not only model optimization but also a well-specified target grounded in real-human evidence. Yet how such targets should be grounded in human data remains largely unexplored, raising three fundamental questions: (1) whose values should be represented? (2) where should cultural values be observed? and (3) what signals reveal underlying values? We investigate these questions through Re3Culture, a large-scale dataset linking (i) Representative communities: 700+ human participants across four cultures recruited through cultural-affiliation screening and demographic-aware sampling; (ii) Realistic scenario: culturally sensitive scenarios selected by community members, e.g., family and kinship; and (iii) Revealed value signals: 1,400+ value surveys, 20,000+ topic-grounded behavioral principles and explicit value priorities with responses for 6,000+ situated questions. Our analysis shows why these choices matter for defining the target, with three implications. First, estimating cultural values requires both community affiliation and broad internal coverage. Substituting other cultures or restricting subgroup coverage shifts value estimates. LLM simulations also disproportionately reflect particular subgroups. Second, values must be observed in concrete, community-guided situations. Abstract questionnaire profiles weakly recover situated priorities, and communities differ in which scenarios they identify as culturally value-revealing. Third, value priorities should be elicited explicitly, not learned from response preferences or bottom-line answers alone. Preferences conflate value agreement with other response-level qualities, and similar answers can be derived from different value priorities. By addressing these requirements, Re3Culture provides the first human-grounded resource that explicitly informs target definition for cultural value alignment.
Figures & tables
Figure 1: Overview of our research framework and Re 3 Culture dataset. Top: three fundamental questions for grounding cultural value alignment in human data; Bottom: Re 3 Culture supports investigation of these questions by linking demographically sampled community members, realistic and culturally sensitive scenarios, and explicit value reasoning with contextual responses.
Figure 2: Overview of Re 3 Culture: human participants, annotation tasks, and corpus statistics.
Figure 3: Analysis results for Whose value perspectives should be represented? , including (i) distances between candidate proxies and the full target community; (ii) visualization of human value data and LLM-simulated data; (iii) subgroup attraction of LLM simulations.
Figure 4: Abstract and situated value priorities. (a) Accuracy of a participant’s high-level profile and increasingly dissimilar controls in predicting contextual value orderings. (b) Topic-specific shifts in contextual Schwartz value priorities.
Figure 5: Community-grounded selection of contexts. (a) Cross-cultural deviations in broad-topic selection, with mean importance and distinctiveness ratings. (b) Associations between culture and the fine-grained subtopic selected within each broad topic.
Figure 6: Preference reasons across value-distance quartiles. Proportion of conversations mentioning each reason, grouped by quartile of the chosen–rejected responses’ value distance.
Figure 7: Same direct answers can express different underlying values.
Figure 8: Annotated value priorities better align with responses than swapped priorities. Dashed arrows mark decreases in priority alignment and overall value alignment (the bottom two categories combined), in percentage points.
Figure 9: Annotated value priorities show some generalizability within the same culture, same topics.
Appendix figures & tables35 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 10: Questionnaire results for the practical need for cultural value alignment in LLMs
Figure 11: Demographic distribution of China Participants.
Figure 12: Demographic distribution of Japan Participants.
Figure 13: Demographic distribution of South Korea Participants.
Figure 14: Demographic distribution of Singapore Participants.
Source
Native unit
Candidate pool
License / access
WildChat-1M
Real-world human–LLM conversation
286,770 conversations
ODC-By (official Hugging Face card)
LMSYS-Chat-1M
Real-world human–LLM conversation
120,787 conversations
Custom gated agreement; redistribution prohibited
PRISM
Survey-linked human–LLM conversation
8,011 conversations
Human-authored text: CC BY 4.0; model outputs: CC BY-NC 4.0
CultureBank
Cultural descriptions from social media
22,990 records
MIT (official Hugging Face card)
CAReDiO
Synthetic question–response pair
Not directly ingested
-
CIVICS
Value-laden statement from a public source
699 statements
CC BY 4.0 (official Hugging Face card)
Appendix
Table 1: Candidate question sources. “Candidate pool” is the number of source records after initial source-level sampling or filtering and before query-level value filtering and deduplication.
Evaluation
Metric
Result
Four-class classification
Accuracy
69.0% (69/100)
Macro-F1
0.589
Category C
Precision
86.0% (43/50)
Recall
78.2% (43/55)
F1
0.819
Binary filtering
Accuracy
81.0% (81/100)
Appendix
Table 2: Value-relevance classifier agreement with human labels after prompt calibration on a held-out test set of 100 questions (N: 8, A: 18, B: 19, C: 55). Binary filtering retains C and discards N, A, and B. F1 scores are reported on a 0–1 scale.
Level-1 Topic
Level-2 Topics
1. Family and Kinship Relationships 11 Level-2 topics
1.1 Views on marriage and partner relationships 1.2 Gender roles in family 1.3 Decision-making power and voice within the family 1.4 Respect for elders and obedience to parents 1.5 Parental approaches to child education 1.6 Parent–child relationships 1.7 Intergenerational relationships and generational differences 1.8 Obligations and responsibilities of family members 1.9 Non-traditional family structures and family diversity 1.10 Attitudes toward intra-family romance/incest 1.11 Loyalty and trust in intimate relationships
2. Personal Life Choices and Life Direction 8 Level-2 topics
2.1 Personal lifestyle choices 2.2 Life goals and definitions of success 2.3 Personal priorities and life direction 2.4 Attitudes toward social expectations 2.5 Time allocation and work–life balance 2.6 Sources of personal happiness 2.7 Stability vs. freedom 2.8 Risk-taking vs. conservatism
3. Work, Career, and Economy 15 Level-2 topics
3.1 Priorities when choosing a job 3.2 Work motivation & purpose 3.3 Fairness between effort and reward at work 3.4 Career development and path 3.5 Work–life balance and time allocation 3.6 Views on job stability and risk 3.7 Obedience and self-expression in the workplace 3.8 Workplace relationships and hierarchical structures 3.9 Entrepreneurship, self-employment, and risk attitudes 3.10 Views on income inequality 3.11 Workplace safety and compliance 3.12 Workers’ rights and employers’ responsibilities 3.13 Consumption habits and attitudes toward money 3.14 Saving, investment, and financial planning 3.15 Housing, residential stability, and views on assets
4. Education, Learning, and Self-Improvement 11 Level-2 topics
4.1 Purposes and goals of education 4.2 Motivations for learning 4.3 Educational equity and opportunity 4.4 Pursuit of prestigious schools 4.5 Attitudes toward exams and competition 4.6 Whether academic performance represents ability 4.7 Teacher authority 4.8 Attitudes toward lifelong learning 4.9 Student pressure and mental health 4.10 Views on student autonomy 4.11 Standardized vs. personalized education
5. Social Interaction and Interpersonal Relationships 7 Level-2 topics
5.1 Interactions with friends and colleagues 5.2 Attitudes toward strangers 5.3 Politeness, respect, and social norms 5.4 Social hierarchy, status, and respect 5.5 Authenticity vs. pretense in social interactions 5.6 Trust in interpersonal relationships 5.7 Views on social mutual assistance
6. Social Responsibility and Social Order 9 Level-2 topics
6.1 Public behavioral norms 6.2 Moral norms in social interactions 6.3 Social responsibility and civic duties 6.4 Attitudes toward social order and safety 6.5 Attitudes toward uncivil behavior 6.6 Collective interests vs. individual freedom 6.7 Racism and discrimination 6.8 Gender and LGBTQ+ norms in social interactions 6.9 Gender, sexual orientation, and identity
Appendix
Table 3: Overview of the two-level value-laden topic taxonomy.
Level-1 Taxonomy
Level-2 Task
Definition and Example
1. Open-Ended Expression or Reasoning Tasks 4 Level-2 tasks
1.1 Everyday Questions with Implicit Value Orientation
Definition Ask questions grounded in everyday life situations that involve what to do or how to view something, implicitly reflecting value priorities or life choices. Example How do you balance work and family life in your daily routine?
1.2 Opinion Expression on Social Issues
Definition Ask for opinions on social phenomena or public issues, where the response reflects values, moral views, or life perspectives. Example What do you think about the 996 work schedule and the pressure it puts on young people today?
1.3 Value Comments on Controversial Issues
Definition Ask for opinions or evaluations on controversial topics involving value judgments. Example A family member is severely ill and in long-term pain and says they want euthanasia. Would you support it?
1.4 Philosophical or Conceptual Questions
Definition Raising philosophical questions about life, meaning, or values, or asking about the meaning/essence of an abstract value concept, requiring an opinion or interpretation. Example People often say being respectable/dignified matters. What does dignity really mean, and how should we think about it when it conflicts with real needs or personal freedom?
2. Decision Making and Problem-Solving Tasks 5 Level-2 tasks
2.1 Value-Oriented Decision Support
Definition Present a situation that requires trade-offs or decisions and ask how one should choose, with the response reflecting value priorities. Example I have two job offers: A pays more but often requires overtime; B pays less but is more stable and lets me spend time with family. Can you weigh them using long-term growth / family relationships / physical & mental health / financial pressure and give me a recommendation?
2.2 Value-Oriented Policy and Strategy Making
Definition Ask questions about designing or evaluating rules, policies, or systems, where responses should reflect value priorities. Example Should governments limit the amount of time people can spend on short-video platforms?
Appendix
Table 4: Overview of the two-level value-laden task taxonomy, including definitions and examples.
Figure 15: Topic selection and topic-grounded principle elicitation. Participants select a level-1 and level-2 topic and formulate three to five ranked behavioral principles for the target culture.
Figure 16: Task and candidate-question selection. Participants select a task type, choose a candidate question, and validate its cultural value importance, cultural distinctiveness, and real-life plausibility.
Figure 17: Situated annotation from the perceived-community perspective. Participants select and rank the relevant values, explain their priority relationships, construct a response consistent with those priorities, and specify inappropriate behaviors or responses.
Figure 18: Situated annotation from the personal perspective. Participants repeat the same annotation structure based on their personally endorsed values and preferred responses.
Instrument
Version and format
Constructs
Paraphrased example items
PVQ-21
Portrait Values Questionnaire, 21-item short form
Schwartz’s ten basic human values
- “Being creative and developing original ideas is important to this person.”; - “Having money and expensive possessions is important to this person.”
VSM 2013
Values Survey Module, 2013 edition
Hofstede’s six cultural dimensions; interpreted at the aggregate cultural-group level
- “How important is having sufficient time for your personal or home life?”; - “How often do you feel nervous or tense?”
Appendix
Table 5: Abstract questionnaires administered before the situated annotation task. Example items are concise paraphrases included to illustrate item form rather than reproduce the official instruments. Scoring follows the corresponding official procedures.
Quality item
Rubric
1. Survey Questionnaires
1.1 Complete all three questionnaires and ensure that the username entered in each questionnaire matches the system username.
Pass: All three questionnaires are completed and the username matches. Needs correction: All three are completed, but the username does not match; note the discrepancy. Fail: Any questionnaire is missing and must be completed.
1.2 In the Cultural Values Questionnaire, answer both the “Personal Values Statement” and the “Cultural Values Statement” from at least three distinct perspectives.
Pass: Both statements cover at least three perspectives. Fail: Either statement covers fewer than three perspectives.
1.3 Pass the attention checks in Schwartz Q6 and Q19 and Cultural Values Q7 and Q24.
Pass: All four attention checks are answered correctly. Fail: Any attention check is answered incorrectly; the relevant questionnaire must be completed again.
1.4 Responses to oppositely worded items should differ in direction: Schwartz (Q15, Q4), (Q7, Q10), and Cultural Values (Q17, Q23).
Pass: Responses to all three pairs are directionally reasonable. Fail: Any response is directionally unreasonable; the relevant questionnaire must be completed again.
1.5 Do not tell annotators that the questionnaires contain test questions; evaluate their performance directly.
Reviewer instruction: Keep the embedded validation items undisclosed during administration.
Appendix
Table 6: Complete human quality-assurance scorecard. Items and rubrics follow the English version of the operational review sheet.
Diagnostic
Observed value
Interpretation
Codebook size
396 codes
Final number of induced value codes after ten refinement iterations.
Code frequency
min 6; median 757.5; max 27,996
Number of value-expression instances assigned to each code.
Assignment coverage
mean cosine 0.7206; 2.47% below 0.55
Most expressions are close to their assigned centroid; the reported tail remains below the extension threshold after the capped refinement run.
Assignment margin
mean 0.0647; median 0.0550
Assigned-code similarity minus the highest alternative-code similarity; 46.86% are below 0.05, indicating limited separation for many expressions.
Name-fit dispersion Dk
mean 0.1037; max 0.1716
Lightweight diagnostic used to identify potentially conflated codes.
Cross-source breadth
388 codes in ≥2 role cultures; all 396 in ≥2 models
Most codes recur across cultural role conditions, and none is supported by only one culture/model source.
Appendix
Table 7: Diagnostics for the constructed value codebook. Similarities are cosine similarities in the normalized embedding space.
Rank
Value code
Frequency
Rank
Value code
Frequency
1
maintaining social harmony and stability
27,996
26
respecting people’s freedom to choose their lives
3,803
2
protecting people’s dignity and saving face
13,108
27
taking responsibility for one’s actions
3,737
3
communicating openly and honestly
12,652
28
maintaining strong social bonds and unity
3,688
4
respecting people’s choices about their own life
11,903
29
making your own major life choices
3,652
5
showing respect to authority figures
11,093
30
holding people in power accountable
3,550
6
respecting and maintaining personal boundaries
9,387
31
putting the greater good first
3,545
Appendix
Table 8: The 50 most frequent induced value codes, ordered down the left half and then down the right half.
Culture
# Participants
# Schwartz Questionnaires
# Hofstede Questionnaires
# Principles
# Situated Annotations
China
210
210
210
7,055
2,047
Japan
186
186
186
4,317
1,314
South Korea
186
186
186
3,904
1,167
Singapore
342
342
342
6,355
1,874
Total
924
924
924
21,631
6,402
Appendix
Table 9: Dataset sizes by culture. Participants are counted as the union of identifiers across the four data files; questionnaire counts refer to non-empty survey profiles. Each situated annotation contains both personal and perceived-community perspectives and is counted once.
Culture
# Unique Topic-1
# Unique Topic-2
# Unique Task-1
# Unique Task-2
# Unique Questions
China
15
145
6
17
1,500
Japan
15
136
6
17
880
South Korea
15
127
6
17
706
Singapore
15
138
6
17
1,133
Total
15
146
6
17
4,219
Appendix
Table 10: Question and taxonomy coverage by culture. Questions are deduplicated by exact text after trimming leading and trailing whitespace within each culture; the total sums these culture-specific counts. Taxonomy totals count the union of canonical English labels across cultures.
Per Participant
Personal Perspective
Perceived-Community Perspective
Culture
Principles
Situated Annotations
Value Labels
Response Characters
Inappropriate Items
Value Labels
Response Characters
Inappropriate Items
China
52.65
15.28
2.84
305.13
2.70
2.88
304.35
2.69
Japan
43.17
13.14
2.88
176.54
2.49
2.89
178.67
2.49
South Korea
40.25
12.03
2.72
163.85
1.99
2.69
159.07
1.91
Singapore
36.73
10.83
2.95
694.74
2.41
2.96
722.66
2.42
Total
42.92
12.70
2.86
367.03
2.40
2.87
374.52
2.39
Appendix
Table 11: Mean annotation quantities by culture. Per-participant means use only participants appearing in the corresponding principle or question–answer file. Perspective-specific means are computed over situated annotation records. Response length counts non-whitespace characters in the submitted language; inappropriate items count non-empty lines as an approximation to formatted entries, not semantically distinct behaviors. The total is computed from pooled records and participants, not an unweighted average of cultures.
Figure 19: Numbers of questions annotated by one, two, or three distinct participants in each culture.
Figure 20: Distributions of situated annotation records over (a) level-1 topics, (b) level-1 tasks, and (c) level-2 tasks. Percentages are normalized within each culture; questions annotated by multiple participants contribute multiple records.
ID
Variable
Explanation
Category
0
country
Country dataset to which the participant belongs.
Question text: Not applicable.
China
Japan
South_Korea
Singapore
Appendix
Table 12: Codebook for participant_basic_information.jsonl .
ID
Variable
Explanation
0
country
Country to which the participant belongs.
1
annotator_id
Pseudonymized participant identifier.
2
schwartz
Ordered list of 21 released Schwartz Value Survey items.
3
schwartz[].question_index
Consecutive released-item index from 1 to 21.
4
schwartz[].question
Canonical English question text.
5
schwartz[].personal_response
Response from the participant’s personal.
Appendix
Table 13: Codebook for value_survey_results.jsonl . Array notation denotes fields inside each survey item.
ID
Variable
Explanation
0
country
Country to which the participant belongs.
1
annotator_id
Pseudonymized participant identifier.
2
annotation_id
Pseudonymized annotation identifier.
3
topic_1
Level-1 topic label in the country’s annotation language.
4
topic_2
Level-2 topic label in the country’s annotation language.
5
topic_1_en
Canonical English level-1 topic label.
Appendix
Table 14: Codebook for principles.jsonl .
ID
Variable
Explanation
0
country
Country to which the participant belongs.
1
annotator_id
Pseudonymized participant identifier.
2
annotation_id
Pseudonymized annotation identifier.
3
topic_1
Level-1 topic label in the country’s annotation language.
4
topic_2
Level-2 topic label in the country’s annotation language.
5
topic_1_en
Canonical English level-1 topic label.
Appendix
Table 15: Codebook for question_answers.jsonl . Fields without the _person suffix refer to the cultural perspective.
Table 16: Models included in RQ1, grouped by model category and family.
Category
Training Resource
China
Japan
South Korea
Singapore
Survey-derived judgments
WVS
222
258
256
254
GlobalOpinionQA
655
891
790
271
Synthesized cultural question–answer data
CultureLLM
980
1,000
1,000
1,000
CulturePark
686
719
675
626
CAReDiO
2,048
2,060
936
1,836
Human-rated model responses
CARE
1,293
1,053
–
–
Appendix
Table 17: Cultural supervision categories and available SFT examples before the validation split. Each supported culture is trained separately; a dash denotes no checkpoint in this RQ1 comparison.
Figure 21: Hofstede supplementary analysis.
Figure 22: Schwartz supplementary analysis.
Figure 23: Relative demographic attraction under cultural role-playing. Schwartz-based attraction to human personal-value subgroups, centered within each model and demographic attribute. Points represent available models and diamonds their unweighted means; zero is each model’s attribute-specific reference, not perfect alignment. Region is shown for China, Japan, and South Korea, and ethnicity for Singapore.
Figure 24: Country-specific contextual value shifts. Rows are level-1 topics and columns are Schwartz dimensions. Red and blue indicate positive and negative deviations from each country’s equal-topic mean, respectively; all panels share one color scale. Their equal-country average yields Figure 4 (b).
Figure 25: Level-2 subtopics with the largest cross-cultural enrichment differences. The 30 rows are ranked by the range of enrichment across the four culture columns. Enrichment is the within-parent selection share divided by a uniform share over all observed subtopics under that parent. Red indicates enrichment above 1 and blue below 1; white denotes the uniform baseline. Selection by cross-cultural range is descriptive, not a significance criterion.
Field
Participant A
Participant B
Case 1: personal perspective, within Singapore
Question
During a heatwave, should neighbors pressure residents to follow requests to set air-conditioning to 25∘ C, or should this remain a personal choice?
Direct answer
Treat it as a personal choice; neighbors should not pressure residents to comply.
Treat it as a personal choice; neighbors should not pressure people to comply.
Ranked values
1. Educating people with facts and figures; 2. Using resources wisely and responsibly.
1. Respecting people’s freedom to choose their lives; 2. Keeping participation voluntary; 3. Adapting support to individual needs.
Metrics
Response cosine: 0.969; value OT: 0.276.
Case 2: personal perspective, Singapore (A) and China (B)
Appendix
Table 18: Similar answers, different annotated priorities. Questions and extracted answers are shortened or translated for display; value labels are translated into English and listed in their original priority order. Similarities and OT distances are computed from the unabridged extracted answers and mapped value profiles. A and B denote different anonymous participants.
Figure 26: Value-priority consistency across four cultures. Each panel compares annotated and top-two-swapped rankings using the same annotation–perspective pairs). Dashed arrows show percentage-point changes in priority alignment and overall value alignment (the bottom two categories combined).
Large Language Models (LLMs) are deployed across cultural contexts but often reflect homogenized values inherited from training data. Evaluations of cultural alignment typically rely on direct prompting with survey-style questions, which frequently elicit neutral or safety-aligned responses and fail to capture underlying model preferences. We propose a framework for probing and steering latent cultural representations in LLMs along the two Inglehart--Welzel axes of the World Values Survey (WVS). By translating social value questions into scenario-based behavioral dilemmas, we extract token-level probabilities to measure implicit values and apply activation steering, optionally combined with country-conditioned prompting, to shift model behavior without retraining. Across three open-source LLMs and four target cultures, we find substantial variation in steerability and identify latent entanglement, where interventions along one cultural dimension induce shifts along another. This coupling mirrors correlations in human WVS data and persists across activation, prompt, and hybrid steering. It constrains axis-independent alignment, though general task performance is largely preserved.
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
Researchers increasingly treat LLM survey responses as a proxy for human cultural values. This includes projecting model outputs onto instruments like the Inglehart-Welzel Cultural Map and drawing conclusions about which cultures a model resembles. While a model's answer to a value-laden questions may be interpreted as a cultural signal, it also carries sampling noise and, can be quite sensitive to question framing. In this paper, we separate survey responses, sampling noise and question framing for multiple LLMs. We decompose response variance from these models into variation across random seeds, prompt rewordings. We employ noise-to-signal ratio (NSR) to test whether a model's apparent cultural position is distinguishable from noise. When applied across a dozen models from four geographic origins, calibrated against 88 Integrated Values Survey countries, the answer is often no. NSR exceeds 1.0 on 49 of 117 valid model-question pairs (42%), reaching 5.56 in the worst case. Two models even refuse to answer sufficient number of survey questions outright. Our results corroborate previous findings that LLMs cluster toward Western, English-speaking cultural positions. However, what does not hold up in this study is the precision with which anyone can currently interpret a specific model's coordinates: prompt tone alone can shift a model by 2.4 map units, comparable to the distance between actual countries in the Inglehart-Welzel Cultural Map. These findings suggest that cultural attribution from LLM survey responses requires establishing the reliability of the underlying measurements before interpreting model coordinates as evidence of cultural representation.
An Duy Nguyen, Muhammad Aurangzeb Ahmad
University of Washington · University of Washington Bothell