When English Isn't the Best Teacher: Source Language Effects in Cross-Lingual In-Context Learning

This study investigates the effects of source language selection on cross-lingual In-Context Learning (ICL) across seven tasks and six models, revealing that traditional fine-tuning assumptions do not hold in ICL scenarios. The research identifies language confusion as a significant challenge for generative tasks and proposes new heuristics for selecting source languages to improve transfer quality. This has implications for practitioners aiming to optimize multilingual models and enhance performance in cross-lingual applications.