Rethinking Cross-lingual Gaps from a Statistical Viewpoint

This paper presents a novel approach to understanding cross-lingual gaps in Large Language Models (LLMs) by focusing on response variance as a key factor contributing to accuracy drops when querying in target languages. The authors formalize the cross-lingual gap in terms of biased and unbiased errors and empirically validate their hypothesis through inference-time interventions, demonstrating that ensemble methods can reduce response variance and improve source-target transfer scores by up to 12 absolute points, translating to relative gains of 8% to over 50% across various LLMs. This research is significant for practitioners as it offers new strategies to enhance cross-lingual performance in AI applications.