Cross-Lingual Transfer

Cross-lingual transfer studies how models use data, representations, or capabilities learned in one language to perform tasks in another, especially where labeled data are scarce. Common approaches include multilingual pretraining and joint training, translation and annotation projection, source-language selection, parameter-efficient fine-tuning, and inference-time prompting or activation steering; applications range from classification and retrieval to reasoning, speech, and sign language. Benchmarks such as XNLI, MGSM, and MIRACL test zero-shot and few-shot transfer across languages, scripts, domains, and resource levels. Transfer can reduce annotation needs and help low-resource languages, but results are often asymmetric and task-dependent: script, tokenization, data quality, and cultural fit matter, while gains may come with interference or fail to preserve safety and factual accuracy.