Correct answers do not necessarily reflect cultural understanding. We introduce CRaFT, an explanation-based multilingual evaluation framework designed to assess how large language models (LLMs) reason across cultural contexts. Rather than scoring outputs solely based on accuracy, CRaFT evaluates model explanations using four interpretable metrics: Cultural Fluency, Deviation, Consistency, and Linguistic Adaptation. We apply the framework to 50 culturally grounded questions from the World Values Survey, translated into Arabic, Bengali, and Spanish, and evaluate three models (GPT-4o, DeepSeek, FANAR) across over 2,100 answer–explanation pairs. Results reveal significant cross-lingual variation in reasoning: Arabic reduces fluency, Bengali enhances it, and Spanish remains largely stable. While GPT-4o adapts more effectively across languages, it exhibits lower consistency; FANAR shows stable but rigid reasoning. These findings suggest that cultural awareness in LLMs is not intrinsic but emerges through linguistic framing. CRaFT offers a new lens for evaluating cross-cultural reasoning in multilingual settings, providing actionable insights for building culturally adaptive language models.

LLM-as-a-Judge has emerged as a scalable alternative to human evaluation in multilingual cultural reasoning, yet its reliability for culturally grounded judgments remains poorly understood. We present a human-aligned validation study using the CRaFT framework across Arabic, Bengali, and Spanish, evaluating six judge models, including five open-weight models and GPT-4.1, against paired human annotations across two structurally observable dimensions (Answer Consistency and Explanation Consistency) and three culturally interpretive dimensions (Cultural Fluency, Linguistic Adaptation, and Deviation). Our results reveal a consistent reliability divide: automated judges align substantially better with humans on structural dimensions than on culturally interpretive ones requiring sociocultural reasoning. Reliability further varies across languages, judge architectures, and response-generation models, with several failures reflecting systematic calibration limits rather than prompt underspecification alone. These findings suggest that current automated evaluators are best treated as complementary diagnostic tools rather than replacements for human judgment in culturally sensitive multilingual settings.

Addressing the challenges of Arabic intent detection amid extensive dialectal variation, this study presents a crossdialtectal, multilingual approach for classifying intents in banking and migration contexts. By augmenting dialectal inputs with Modern Standard Arabic (MSA) and English translations, our method leverages cross-lingual context to improve classification accuracy. We evaluate single-input (dialect-only), dual-input (dialect + MSA), and triple-input (dialect + MSA + English) models, applying language-specific tokenization for each. Results demonstrate that, in the migration dataset, our model achieved an accuracy gain of over 50% on Tunisian dialect, increasing from 43.3% with dialect-only input to 94% with the full multilingual setup. Similarly, in the PAL (Palestinian dialect) dataset, accuracy improved from 87.7% to 93.5% with translation augmentation, reflecting a gain of 5.8 percentage points. These findings underscore the effectiveness of our approach for intent detection across various Arabic dialects.