Cross-Cultural Alignment for Large Language Models

This direction is merged into another direction: Western-centric training bias. Its last version stays readable and citable here.

Version 1 of 1102 papers111 conceptsfrozen 2026-10-04

Most modern Large Language Models (LLMs) are "multilingual" in the sense that they can translate text between hundreds of languages. However, researchers have begun to realize that speaking a language is not the same as understanding the culture of its speakers. Current models suffer from a "Western default" bias, where they apply the values, social norms, and commonsense of Western, Educated, Industrialized, Rich, and Democratic (WEIRD) societies to the rest of the world.

The emerging field of cross-cultural alignment seeks to move beyond linguistic translation to ensure that AI systems can reason within the diverse value systems and social contexts of global populations.

The Western Default and Algorithmic Monoculture

The root of cultural bias in LLMs lies in their training data. Because models are trained on massive, web-scale datasets like CommonCrawl and Wikipedia, they inherit the perspectives of the most digitally active populations. Daniel Hershcovich and colleagues describe this as a risk of "algorithmic monoculture," where dominant Western perspectives systematically marginalize or erase indigenous knowledge and local values Challenges and Strategies in Cross-Cultural NLP.

This bias isn't just about what the model knows; it's about how it behaves. Even when a model is developed outside the West, it often defaults to Western norms because the underlying English-centric training data is so pervasive. Yuhang Wang demonstrated this using CDEval, a benchmark based on Hofstede’s Cultural Dimensions Theory. The study found that most mainstream models align more closely with wealthy, Western cultural personalities than with the cultures of the regions they were intended to serve CDEval: A Benchmark for Measuring the Cultural Dimensions of Large Language Models.

From Trivia to Cultural Reasoning

Early attempts to measure cultural awareness focused on "trivia"—static facts like "What is the national dish of Egypt?" However, Inha Cha and others argue for intentionally cultural evaluation, which treats culture as a lens through which all tasks are filtered Culture is Everywhere: A Call for Intentionally Cultural Evaluation.

The difficulty lies in "procedural" or "tacit" knowledge—the unwritten rules of social etiquette, rituals, and traditions. For example, Amir Hossein Yari introduced CAPTex to evaluate how models understand complex cultural procedures, finding that even top-tier models like GPT-4o have significant "blind spots" when it comes to non-Western social etiquette Unveiling Cultural Blind Spots: Analyzing the Limitations of mLLMs in Procedural Text Comprehension.

To address this, Lingzhi Shen proposed CALM (Culturally self-Aware Language Models). Instead of treating culture as a database of facts, CALM models it as a dynamic reasoning process, using a "reflective" loop to disentangle cultural signals from the logic of a task CALM: Culturally Self-Aware Language Models.

The Challenge of Value Probing

A major hurdle in this research is that a model's "tongue" (its language) does not always match its "thought" (its underlying values). Arnav Arora and Isabelle Augenstein developed a "Value Probing" framework to see if multilingual models inherit the values of the languages they speak. They found a "value gap": while models show different profiles for different languages, these profiles rarely align with real-world human survey data from those cultures Probing Pre-Trained Language Models for Cross-Cultural Differences in Values.

This suggests that models might develop language-agnostic "concept representations" that default to the most prevalent cultural interpretation in the training data. Clément Dumas used activation patching to show that while these universal semantic spaces exist, they risk prioritizing dominant cultural perspectives at the expense of diversity Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in Transformers.

Strategies for Multicultural Alignment

How do we fix the Western default? Researchers are exploring several data-efficient strategies:

  1. Semantic Augmentation: CHENG LI developed CultureLLM, which uses a tiny seed dataset (just 50 questions from the World Value Survey) and semantic augmentation to fine-tune models for specific cultures at a very low cost CultureLLM: Incorporating Cultural Differences into Large Language Models.
  2. Cross-Cultural Transfer: Saeed Almheiri and Preslav Nakov discovered that cultural alignment is highly transferable. Aligning a model with a few examples from one Arab nation (e.g., Egypt) significantly improves its performance across other Arab nations, suggesting that models can learn broad cultural clusters efficiently Cross-Cultural Transfer of Commonsense Reasoning in LLMs: Evidence from the Arab World.
  3. Geo-Diverse Knowledge Bases: Models like GD-COMET, developed by Mehar Bhatia and Vered Shwartz, augment standard reasoning models with massive knowledge bases of global cultural assertions to break the Western-centric bias in commonsense inference GD-COMET: A Geo-Diverse Commonsense Inference Model.

The ultimate goal of this research is to move AI from being a "Western-centric assistant" to a truly global partner capable of navigating the nuanced, pluralistic reality of human culture.

Papers in this version

Showing 60 of 102 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cross_cultural_alignment_for_large_langu_ad93fc31,
  title = {Cross-Cultural Alignment for Large Language Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-cultural-alignment-for-large-language-models/rdr_ad93fc31f9f6425191933fe2760d1976},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_cross_cultural_alignment_for_large_langu_ad93fc31_v1,
  title = {Cross-Cultural Alignment for Large Language Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-cultural-alignment-for-large-language-models/rdr_ad93fc31f9f6425191933fe2760d1976/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}