💬 NLP

PLURAL: A Global Dataset for Value Alignment

The paper introduces PLURAL, a large-scale dataset derived from the Integrated Values Survey across 92 countries that generates synthetic preference triplets to improve large language model alignment with diverse, non-Western cultural values, demonstrating significant improvements in both automated metrics and human evaluations.

Original authors: Dhruv Agarwal, Anya Shukla, Tanya Goyal, Aditya Vashistha

Published 2026-07-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dhruv Agarwal, Anya Shukla, Tanya Goyal, Aditya Vashistha

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you've built a super-smart robot chef who can cook a million different recipes. But there's a catch: this chef was trained almost entirely by a small group of people from one specific neighborhood in the West. As a result, the robot thinks everyone wants their food served with a side of Western values, even when a customer from Tokyo, Mumbai, or Rio asks for something different. The robot might be fluent in Japanese or Portuguese, but its "soul" still feels very American. It's like a fluent but foreign guest who speaks your language perfectly but doesn't understand your family traditions.

Enter PLURAL, a new project from researchers at Cornell University that tries to fix this robot's cultural blind spots.

The Big Idea: A Global Taste Test

The researchers realized that to teach a robot to respect different cultures, they couldn't just ask random people on the internet what they like. Instead, they went to the source: a massive, scientifically rigorous survey called the Integrated Values Survey (IVS). This survey has asked over 156,658 people in 92 different countries about their deepest beliefs—what they think is right, wrong, important, or silly.

Think of the IVS as a giant, global "taste test" where people from every corner of the world rated their favorite flavors of life. But here's the problem: you can't feed a robot a spreadsheet of survey answers. Robots need stories and conversations, not just "Yes/No" checkboxes.

The Magic Pipeline: Turning Surveys into Stories

So, the team built a two-stage "translation machine" to turn those dry survey answers into 500,000 realistic conversation scenarios (called "preference triplets").

Stage One (The Translator): They took a real person's survey answers—say, a 66-year-old man from Japan who values family duty over leisure—and asked an AI to imagine a real-life situation where that person would face a tough choice. The AI generated a prompt like, "My friend wants to play golf, but my daughter needs a babysitter. What do I do?"
Stage Two (The Storyteller): The AI then wrote two responses. One response (the "preferred" one) reflected the Japanese man's actual values (prioritizing family duty). The other (the "dispreferred" one) reflected a different, plausible viewpoint (prioritizing leisure).

They did this for 20 diverse countries, creating a massive library of "what would a real person from this country actually say?" scenarios. They even made sure to pick people from the survey who represented the real mix of ages, genders, and education levels in those countries, so the data wasn't just about the same type of person over and over.

Did It Work? The Taste Test Results

The researchers didn't just hope it worked; they put it to the test in three ways:

The "Is it Real?" Check: They verified that the new stories actually kept the unique flavor of each country. They found that the data still held onto the distinct differences between, say, Brazil and Japan, and even kept the variety of opinions within each country. It wasn't just a bunch of stereotypes; it was a rich, messy reflection of real human diversity.
The Robot Training: They took a standard language model and trained it on this new PLURAL data. They tested the robot using a different set of cultural questions (the GLOBE framework) that the robot had never seen before. The result? The robot trained on PLURAL became 27.7% better at matching the cultural profile of countries like India, compared to other strong methods. It wasn't just guessing; it was learning the specific "vibe" of the target culture.
The Human Verdict: They asked 176 real people from India, Brazil, and Japan to read answers from the new robot and the old robot. The humans overwhelmingly picked the PLURAL-trained robot's answers as feeling more "typical" of their own national values. One person even called a response "quintessentially Indian," noting how it captured the specific way families use WhatsApp to stay connected.

What It's NOT (The "Fluent but Foreign" Trap)

The paper is very clear about what this doesn't do. It argues against the idea that simply making a robot speak a local language (like Hindi or Arabic) is enough. Previous attempts created "fluent but foreign" models that could chat in local tongues but still pushed Western values. PLURAL shows that you need to train the model on the values themselves, not just the words.

Also, the paper explicitly rules out the idea that you can just "prompt" a robot to be culturally aware by telling it, "You are from Brazil." While that helps a tiny bit, it's not nearly as effective as actually training the robot on thousands of real, value-grounded examples.

The Catch: The Robot Still Gets a Bit "Squished"

Here is the part where the researchers are honest about the limits. While the robot got much better at understanding different cultures, the training process itself seemed to "squish" the differences a little bit.

Imagine the cultural profiles of different countries as points scattered across a wide field. The real countries are far apart from each other. After training, the robot's new "cultural points" moved in the right direction, but they ended up clustered closer together than the real countries are. The robot captured about 18% of the original diversity. The researchers suggest this isn't because the data was bad, but because the current training methods (a technique called DPO) tend to smooth things out too much. It's like a photo filter that makes everyone look a bit more alike, even if they are still clearly different.

The Bottom Line

PLURAL is a huge, open library of 500,000 value-focused stories grounded in real surveys from 20 countries (and ready to expand to 92). It proves that we can teach robots to respect a wider range of human values, moving them away from a single Western perspective. It's a major step toward a future where our digital assistants don't just speak our language, but truly understand our hearts and homes.

Technical Summary: PLURAL: A Global Dataset for Value Alignment

Problem Statement

Large Language Models (LLMs) currently disproportionately reflect Western, English-speaking values, leading to representational harms for international users, such as the homogenization of cultural values in generated text. While recent efforts have focused on "sovereign" language models for specific languages, these models often remain "fluent but foreign," mastering local syntax while retaining American cultural value systems. Existing alignment datasets (e.g., PRISM, Community Alignment) suffer from two primary limitations: (1) Reach: they lack nationally representative data across a broad range of countries, often skewing heavily toward Western or highly educated demographics; and (2) Usability: they capture preferences over generic assistant responses rather than isolating the underlying normative cultural values that shape those preferences. The core challenge is to create a scalable resource that captures diverse human values across cultures while producing data suitable for post-training alignment.

Methodology

The authors introduce PLURAL (Preference Library for Multi-Region Alignment), a large-scale synthetic preference dataset grounded in the Integrated Values Survey (IVS), a nationally representative survey covering 92 countries with over 156,000 participants.

Data Construction Pipeline

The construction of PLURAL involves a two-stage generation pipeline designed to transform terse, fixed-choice survey responses into realistic preference triplets $(prompt, preferred\_response, dispreferred\_response)$ :

Grounding and Filtering:
- The authors utilize the IVS (combining World Values Survey and European Values Study data).
- They reconstruct 59 original survey question groups to preserve survey structure.
- Using Milton Rokeach's hierarchy of beliefs, they filter these groups to retain only prescriptive beliefs (normative values regarding what people ought to do), excluding descriptive (perceptions of reality) and primitive (faith-based truths) beliefs. This results in 41 value-focused question groups.
Stratified Sampling:
- To ensure representativeness, the authors employ stratified proportionate sampling based on sex, age (four bands), and education (three levels).
- An initial release covers 20 culturally diverse countries with 100 representative participants per country ( $N=2,000$ ), yielding approximately 500,000 preference triplets.
Synthetic Generation (Two-Stage):
- Stage 1 (Triplet Generation): For each participant and question group, an LLM is provided with the participant's demographic profile and survey responses. The model generates a realistic scenario (prompt) and two responses: one aligned with the participant's values (preferred) and one reflecting a contrasting stance (dispreferred). Five triplets are generated per group to ensure scenario diversity.
- Stage 2 (Response Expansion): A second LLM pass expands the terse responses from Stage 1 into natural, conversational assistant-like answers while strictly maintaining the underlying value stance.
- Quality Control: Generation is performed in English to isolate values from linguistic variation. The pipeline uses a pool of frontier LLMs (GPT-5 Mini, Mistral Large 3, DeepSeek-v3.2) with stochastic decoding to reduce model-specific bias. Manual verification confirms the data is faithful to IVS responses and avoids generic stereotypes.

Evaluation Framework

The authors evaluate PLURAL using three complementary methods across five target countries (India, Brazil, Japan, Malaysia, Zimbabwe):

Dataset-Level Validation: Verifying that the synthetic data preserves cross-country value differences and within-country diversity observed in the original IVS. This involves training classifiers to predict country from value embeddings and comparing pairwise distance distributions.
Automated Evaluation: Fine-tuning instruction-tuned models (specifically llama-3.1-8b-instruct) on PLURAL using Direct Preference Optimization (DPO). The resulting models are evaluated against the LLM-GLOBE benchmark, which probes cultural alignment across nine dimensions (e.g., Uncertainty Avoidance, Performance Orientation) using open-ended dilemmas. Performance is measured by Mean Absolute Error (MAE) against ground-truth country GLOBE scores.
Human Evaluation: A blind pairwise evaluation with 176 evaluators in India, Brazil, and Japan, asking them to judge which model response is more typical of their national values.

Key Results

1. Dataset Validity

Cross-Country Differences: A classifier trained on PLURAL embeddings achieved 78.0% accuracy in predicting a respondent's country (chance = 20%), demonstrating that significant country-specific value signals are preserved after synthetic generation.
Within-Country Diversity: The distribution of pairwise cosine distances between respondents in PLURAL closely matches the original IVS data (average Wasserstein distance of 0.051), indicating no "template collapse" where diverse individuals are mapped to identical scenarios.

2. Automated Alignment Performance

Improved Alignment: Models fine-tuned on PLURAL consistently outperformed strong baselines (Vanilla, Demographic Prompting, Aggregate Prompting, and Community Alignment DPO).
Error Reduction: Training on PLURAL reduced the Mean Absolute Error (MAE) relative to target country GLOBE profiles by 15.7% (Japan) to 27.7% (India).
Dimensional Gains: Improvements were observed across 7 of the 9 GLOBE dimensions, particularly in Hierarchy, Ambition, and Social Coordination. Dimensions like Gender Egalitarianism showed minimal change, likely due to conflicts with existing safety norms in base models.
Generalization: The value signal transferred across model families, showing improvements on qwen-2.5-7b and qwen-2.5-14b models.

3. Human Evaluation

Preference: Evaluators preferred the PLURAL-aligned model over the Vanilla baseline in 66% of cases overall. Significant gains were observed in India (72%) and Japan (71%).
Qualitative Feedback: Evaluators described the PLURAL-aligned responses as "quintessentially Indian" or aligned with "traditional Japanese virtues," noting the capture of specific cultural artifacts (e.g., WhatsApp usage for family coordination in India).

4. Baseline Comparisons

Vs. Prompting: Demographic and Aggregate prompting yielded negligible or inconsistent gains compared to fine-tuning on grounded data.
Vs. Existing Datasets: PLURAL outperformed the Community Alignment dataset (which covers fewer countries) and showed superior results compared to a synthetic baseline grounded in diverse personas (PersonaHub) rather than real survey data.
Capability Preservation: Fine-tuning on PLURAL did not degrade general capabilities; MMLU scores remained stable or slightly improved.

Significance and Claims

The paper claims that PLURAL provides a scalable, learnable signal for value steering, demonstrating that it is possible to align LLMs with the value distributions of diverse populations using nationally representative survey data.

Pluralistic Alignment: The work advances the agenda of pluralistic alignment by moving beyond "fluent but foreign" models to those that can engage with the plurality of human values.
Grounding in Reality: By grounding synthetic data in the IVS, the authors show that real-world survey responses provide a stronger value-alignment signal than purely persona-conditioned generation.
Limitations and Future Work: The authors modestly note that while PLURAL contains the signal, current post-training methods (specifically DPO) compress cross-country diversity, retaining only ~18% of the variation present in ground-truth profiles. They highlight that the tension between cultural representation and safety norms (e.g., Gender Egalitarianism) remains an open challenge. They position PLURAL not as a final solution, but as a resource enabling researchers to empirically study how to navigate conflicting values and preserve plurality in future alignment systems.