PLURAL: A Global Dataset for Value Alignment

Most large language models (LLMs) are trained on data that disproportionately reflects Western, Educated, Industrialized, Rich, and Democratic (WEIRD) perspectives. While these models are used globally, they often act as "fluent but foreign" agents. They can speak a local language but lack the cultural nuance to respect local value systems. This creates a representational harm. AI homogenization flattens diverse moral landscapes into a singular, Western-centric viewpoint.

Current attempts to fix this usually fall into two camps. Some developers build regional models optimized for local syntax. Others use prompting (providing instructions to a model at inference time) to tell a model which culture to emulate. Neither solves the underlying problem of value alignment. Regional models often retain American cultural biases despite their linguistic fluency. Prompting is a fragile, shallow mechanism. It struggles to steer a model's deeper reasoning. A major question remains: how can we create a scalable dataset that captures the actual normative values of diverse global populations?

The limitations of linguistic fluency

The challenge of pluralistic alignment—the ability of an AI to engage with diverse human values—is currently bottlenecked by a scarcity of high-quality data. Existing datasets used to steer models via Direct Preference Optimization (DPO, a method to optimize models based on preferred vs. dispreferred pairs) are notoriously skewed. For example, the PRISM dataset is heavily dominated by white respondents from the US and UK. The US and UK account for over 40% of its participants. Similarly, the Community Alignment (CA) dataset covers only five countries.

Furthermore, even when datasets attempt to capture cultural differences, they often focus on generic assistant responses. They do not isolate the underlying normative values—the "oughts" of human behavior—that shape those preferences. Without a way to move from terse survey answers to rich, naturalistic conversational scenarios, researchers cannot effectively teach a model to navigate cultural tensions.

Transforming survey data into synthetic dialogue

To bridge this gap, the authors introduce PLURAL (Preference Library for Multi-Region Alignment). This pipeline transforms massive social science surveys into actionable training data. The architecture of the PLURAL pipeline rests on three pillars:

1. Grounding in the IVS: Instead of inventing personas, the researchers use the Integrated Values Survey (IVS). This is a rigorous dataset containing responses from 156,658 participants across 92 countries. This ensures the data is nationally representative rather than stereotypical.
2. Normative Filtering: To ensure the data is useful for steering behavior, the authors apply Milton Rokeach’s hierarchy of beliefs. They discard "descriptive" beliefs (how the world is perceived) and "primitive" beliefs (faith-based truths). They retain only "prescriptive" beliefs—standards regarding how people ought to behave. This leaves 41 question groups targeted at value-laden dilemmas.
3. Two-Stage Synthetic Generation: Raw survey responses are too terse for model training. The authors use a two-stage LLM pipeline. In Stage 1, an LLM takes demographics and survey answers to generate a "preference triplet" .

This consists of a user prompt, a preferred response, and a dispreferred response. In Stage 2, these responses are expanded into natural, conversational assistant-style text.

Evidence of learnable cultural signals

The authors demonstrate that PLURAL carries a genuine cultural signal. They first validated the dataset by testing if a classifier could predict a respondent's country based on their PLURAL-derived value representations. The classifier achieved 78.0% accuracy. This is significantly above chance and proves that the synthetic generation process preserved cross-country value differences. They also confirmed the pipeline avoids "template collapse" .

This means the model does not simply repeat the same few scenarios for similar people.

The most critical finding involves the effectiveness of the data for model steering. By performing country-specific DPO fine-tuning on a Llama-3.1-8b-instruct base model, the authors found that PLURAL reduced the Mean Absolute Error (MAE) relative to the GLOBE framework (a sociological measure of cultural dimensions). For example, in India, PLURAL-aligned models reduced MAE by up to 27.7% compared to the vanilla baseline .

This means the model's responses became much closer to the target country's actual cultural profile. This improvement was consistent across dimensions like Performance Orientation and Power Distance .

Finally, blind human evaluations in India, Brazil, and Japan corroborated these results. Evaluators judged the PLURAL-aligned responses as being more typical of their national values. The overall preference probability against the vanilla model was 0.66.

The compression of diversity

Despite these successes, the paper identifies a significant hurdle: the "compression" of diversity during post-training. While the PLURAL dataset contains a wide array of distinct cultural signals, fine-tuning models using DPO pulls them toward a more homogenous center.

The authors report that ground-truth country profiles are widely separated in the 9-dimensional GLOBE space. However, the resulting adapted models cluster much more tightly together. Quantitatively, the adapted models preserve only approximately 18% of the cross-country variation found in the original profiles . This suggests that while the signal in PLURAL is strong enough to steer a model, current optimization techniques are reductive. They may erase the very nuances we seek to represent. Furthermore, the authors note a tension between cultural alignment and safety. Improving alignment with certain cultural dimensions may conflict with safety guardrails already embedded in modern LLMs.

A scalable resource for pluralism

The verdict is clear: PLURAL provides a robust, scalable proof-of-concept for pluralistic alignment. By grounding synthetic data in established social science, the researchers have bypassed the demographic bottlenecks of current datasets. The results prove that cultural values are "learnable" signals that can be injected into models through post-training.

However, the utility of PLURAL for practitioners depends on solving the compression problem. If the goal is to build a truly pluralistic ecosystem, we cannot rely solely on DPO. We may need new training objectives that prioritize the retention of variance. For those looking to implement this today, the initial version of the dataset is available at huggingface.co/datasets/agdhruv/plural-alignment.

Figures from the paper