Cross-Cultural Transfer of Commonsense Reasoning in LLMs: Evidence from the Arab World

Saeed Almheiri, Rania Elbadry, Mohamed Attia, Chenxi Wang, Preslav Nakov, Timothy Baldwin, Fajri Koto2025 · DOI

Introduction

Large Language Models (LLMs) are frequently criticized for their "Western-centric" bias, a byproduct of training datasets dominated by English-language internet content. While recent efforts have attempted to align models with broad regional values (e.g., "Eastern" vs. "Western"), these approaches often overlook the rich intra-regional diversity within specific cultural blocs. The Arab world, comprising 22 nations with shared linguistic roots but distinct social norms, provides a unique case study for this challenge.

This research investigates whether aligning an LLM with the cultural nuances of a single "high-resource" country can improve its performance across other culturally similar "low-resource" nations. The findings suggest that cultural knowledge is not merely a collection of isolated facts but a transferable "cultural logic."

Methodology and Data

The study utilizes the ArabCulture dataset, a comprehensive benchmark featuring approximately 3,200 handcrafted cultural statements and Multiple Choice Questions (MCQs). These cover 13 Arab countries across 12 distinct domains, including food, weddings, family relationships, and traditional games.

To evaluate cultural alignment, the researchers compared several strategies:

  1. In-Context Learning (ICL): Few-shot prompting without weight updates.
  2. DITTO (Demonstration-based Iterative Task Tuning Optimization): A lightweight reinforcement learning method that uses high-quality demonstrations to iteratively refine model preferences.
  3. Baselines: Standard Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO).

Experiments were conducted on four models: two multilingual models (Qwen2.5-7B and Gemma-2-9B) and two Arabic-centric models (ALLaM-7B and SILMA-9B).

The Power of Cross-Cultural Transfer

A central finding of the study is the high degree of transferability between cultures. Aligning a model with as few as 12 culture-specific examples from one country (e.g., Egypt) can boost its reasoning performance across the entire Arab world by an average of 10%. In some instances, specific models saw accuracy gains of up to 34%.

The "Teacher" Country Effect

The research identifies that certain countries act as more effective "teachers" than others. Training on data from Syria and Jordan yielded the highest average performance gains across the region. This suggests that some cultures may occupy a more "central" position in the regional cultural landscape, or their social norms share more commonalities with the broader Arab bloc.

Model Plasticity: Multilingual vs. Arabic-Centric

Surprisingly, multilingual models like Gemma and Qwen exhibited significantly larger gains from cultural alignment than native Arabic models like ALLaM and SILMA. While Arabic-centric models maintain a higher baseline of cultural knowledge, multilingual models appear more "plastic"—they adapt more dramatically to new cultural tuning. This suggests that general-purpose models possess latent capabilities that can be unlocked with minimal, targeted data.

Reasoning vs. Fact Memorization

One of the most intriguing discoveries was that training a model on US or Indonesian cultural examples (using the IndoCulture dataset) could occasionally match or outperform in-culture training for MCQ tasks. This implies that the alignment process is not just feeding the model specific facts about a country; rather, it is teaching the model how to reason about culture as a category. The model learns to identify the types of social structures and norms that define cultural identity, which it then applies to the Arab context.

Probing the Latent Space

By analyzing the internal layers of the models, the researchers found that LLMs already represent different cultures as distinct clusters in their latent space. Alignment does not create these clusters from scratch; instead, it makes them more "linearly separable." This mathematical refinement helps the model distinguish between similar but distinct traditions, such as the nuances of an Emirati wedding versus an Egyptian one.

Conclusion and Future Directions

The study demonstrates that cross-cultural transfer is a robust and efficient way to improve LLM performance in low-resource cultural settings. While "structured" domains like family relationships transfer well, highly localized topics like idioms still require specific data.

Future work should expand this research to all 22 Arab League nations and explore whether these gains hold for open-ended dialogue. The ultimate goal is the creation of "curated" cross-cultural alignment, where models are trained on a strategic mix of "teacher" countries to develop a more globally aware and culturally sensitive reasoning engine.

Summary updated

Abstract

Large language models (LLMs) often reflect Western-centric biases, which limits their effectiveness in diverse cultural contexts.Although some previous work has explored cultural alignment, the potential for cross-cultural transfer, using alignment in one culture to improve performance in others, remains underexplored.Here we investigate cross-cultural transfer of commonsense reasoning in the Arab world, where linguistic and historical similarities coexist with local cultural differences.Using a culturally grounded commonsense reasoning dataset covering thirteen Arab countries, we evaluate lightweight alignment methods such as in-context learning and demonstration-based reinforcement, alongside baselines such as supervised fine-tuning and direct preference optimization.Our results show that merely twelve culture-specific examples from one country can improve performance in others by 10% on average, within multilingual models.We further demonstrate that out-of-culture demonstrations from Indonesian and US contexts can match or surpass in-culture alignment, highlighting cultural commonsense transferability beyond the Arab world.These findings demonstrate that efficient cross-cultural alignment is possible and offer a promising approach to adapt LLMs to low-resource cultural settings.* Equal contribution Before After Abeer prepares fesikh for lunch.A. Abeer serves feskih with rice.B. Abeer serves fesikh with bread.C. Abeer serves fesikh with couscous.Egypt UAE culturally align LLM using only Egyptian data Bakhita performed Al-Na'eesh at her wedding.A. Al-Na'eesh is a traditional women's dance involving dancing with sticks.B. Al-Na'eesh is a traditional women's dance involving coordinated leg movements.C. Al-Na'eesh is a traditional women's dance involving coordinated hair movements.Abeer prepares fesikh for lunch.A. Abeer serves feskih with rice.B. Abeer serves fesikh with bread.C. Abeer serves fesikh with couscous.Egypt UAE Bakhita performed Al-Na'eesh at her wedding.A. Al-Na'eesh is a traditional women's dance involving dancing with sticks.B. Al-Na'eesh is a traditional women's dance involving coordinated leg movements.C. Al-Na'eesh is a traditional women's dance involving coordinated hair movements. Models
BibTeX
@misc{almheiri2025cross,
  author = {Saeed Almheiri and Rania Elbadry and Mohamed Attia and Chenxi Wang and Preslav Nakov and Timothy Baldwin and Fajri Koto},
  title = {{Cross-Cultural Transfer of Commonsense Reasoning in LLMs: Evidence from the Arab World}},
  year = {2025},
  doi = {10.18653/v1/2025.findings-emnlp.247},
}
Figure