Researchers from MBZUAI, University of Washington, and other institutions presented studies at EMNLP 2024 exploring how LLMs represent cultures. A survey analyzed dozens of recent studies on LLMs and culture and proposes a new framework for future research. The survey found that there is no widely accepted definition of 'culture' in NLP, making it challenging to interpret how models represent culture through language. Why it matters: This highlights a key gap in the field and emphasizes the need for a more rigorous and consistent understanding of culture in AI, especially as LLMs become more globally integrated.
A new paper from MBZUAI introduces JEEM, a benchmark dataset for evaluating vision-language models on their understanding of images grounded in four Arabic-speaking societies (Jordan, UAE, Egypt, and Morocco) and their ability to use local dialects. The dataset comprises 2,178 images and 10,890 question-answer pairs reflecting everyday life and culturally specific scenes. Evaluation of several Arabic-capable models (Maya, PALO, Peacock, AIN, AyaV) and GPT-4o revealed that while models can generate fluent language, they struggle with genuine understanding, consistency, and relevance, especially when cultural context is important. Why it matters: This research highlights the challenges of building AI systems that can truly understand and interact with diverse cultures, emphasizing the need for culturally grounded datasets and evaluation metrics.
MBZUAI researchers release JEEM, a new benchmark dataset for evaluating vision-language models on Arabic dialects. The dataset covers image captioning and visual question answering tasks using images from Jordan, UAE, Egypt, and Morocco. Results show models struggle with cultural understanding and relevance despite fluent language generation.
This study introduces a Probabilistic Graphical Model (PGM) framework utilizing Pearl's do-operator to causally audit LLM safety mechanisms, specifically isolating the effect of injecting cultural demographics into prompts. A large-scale empirical analysis was conducted across seven instruction-tuned models from diverse origins, including the UAE's Falcon3-7B, as well as models from the US, Europe, China, and India, using ToxiGen and BOLD datasets. The findings revealed a disparity between observational and interventional bias, demonstrating that standard fairness metrics can overestimate demographic bias. Western models exhibited higher causal refusal rates for specific demographic groups, while Eastern models showed low overall intervention rates with targeted sensitivities toward regional demographics. Why it matters: This research highlights the geopolitical nuances of LLM safety alignment and the potential for demographic-sensitive over-triggering to restrict benign discourse, which is particularly relevant for diverse regions like the Middle East in developing culturally-aware AI.
MBZUAI researchers found that only 5.7% of music in existing datasets used to train generative music systems comes from non-Western genres. They discovered that 94% of the music represented Western music, while Africa, the Middle East, and South Asia accounted for only 0.3%, 0.4%, and 0.9% respectively. The team also tested whether parameter-efficient fine-tuning with adapters could improve generative music systems on underrepresented styles, presenting their findings at NAACL. Why it matters: This research highlights the critical need for more diverse datasets in AI music generation to better serve global musical traditions and audiences.
MBZUAI researchers, in collaboration with over 70 researchers, have created the Culturally diverse Visual Question Answering (CVQA) benchmark to evaluate cultural understanding in multimodal LLMs. The CVQA dataset includes over 10,000 questions in 31 languages and 13 scripts, testing models on images of local dishes, personalities, and monuments. Testing of several multimodal LLMs on the CVQA benchmark revealed significant challenges, even for top models. Why it matters: This benchmark highlights the need for AI models to better understand diverse cultures, promoting fairness and relevance across different languages and regions.
MBZUAI researchers presented two studies at NAACL 2025 concerning how LLMs understand cultural differences, with one study winning the SAC award. One study, titled "Reading between the lines: Can LLMs identify cross-cultural communication gaps," assesses GPT-4o's ability to identify cultural references in Goodreads book reviews. The researchers created a benchmark dataset using annotations from 50 evaluators across different cultures to measure the LLM's ability to identify culture-specific items (CSIs). Why it matters: Improving LLMs' cross-cultural understanding is crucial for ensuring these models can be used effectively and equitably across diverse global contexts.