Disentangling Linguistic Relatedness from Task Alignment in Cross-Lingual Transfer

This study investigates cross-lingual transfer by fine-tuning seven large language models ranging from 4B to 671B parameters on Arabic, assessing their zero-shot performance on Semitic and non-Semitic languages. The results indicate that models with weak baselines significantly improve across all languages, while those with strong baselines show minimal gains, suggesting that enhancements derive more from task-format alignment than from true cross-lingual knowledge transfer. This insight is crucial for practitioners as it highlights the importance of model baseline performance and task alignment strategies in multilingual applications.