Adaptive Language Sampling Method Targets Cross-Lingual Transfer for Low-Resource Languages

A new arXiv paper proposes an online adaptive sampling strategy for realigning multilingual language models, aiming to improve cross-lingual transfer to extremely low-resource languages. The authors note that existing realignment approaches typically use uniform or random sampling, which may underuse informative language pairs. Their method adjusts sampling dynamically as training proceeds within a distributed setup.