Progressing beyond Art Masterpieces or Touristic Clichés: how to assess your LLMs for cultural alignment?

2026-04-28

AI Summary

This study addresses the scarcity of high-quality evaluation datasets for cultural alignment in large language models, which hinders the effective assessment of whether models genuinely adapt to specific cultural contexts. The authors systematically analyze the limitations of existing datasets and, for the first time, propose a set of principled guidelines for constructing culturally aligned data from the perspective of annotators. Following these principles, they introduce a new benchmark dataset. Through controlled comparative experiments, they demonstrate that this dataset exhibits significantly stronger discriminative power in distinguishing between culturally specialized models and general-purpose models, thereby establishing a reliable benchmark for evaluating cultural alignment in language models.

Abstract

Although the cultural (mis)alignment of Large Language Models (LLMs) has attracted increasing attention -- often framed in terms of cultural bias -- until recently there has been limited work on the design and development of datasets for cultural assessment. Here, we review existing approaches to such datasets and identify their main limitations. To address these issues, we propose design guidelines for annotators and report on the construction of a dataset built according to these principles. We further present a series of contrastive experiments conducted with this dataset. The results demonstrate that our design yields test sets with greater discriminative power, effectively distinguishing between models specialized for a given culture and those that are not, ceteris paribus.

Problem

• cultural alignment
• Large Language Models
• dataset design
• cultural bias
• model evaluation

Innovation

• cultural alignment
• dataset design
• large language models
• contrastive evaluation
• annotation guidelines