Determinants of Source-Language Response Variance

Identify and quantify the factors that contribute to response variance in source-language prompts for large language models on knowledge-intensive tasks, and characterize how these factors propagate to determine target-language variance and cross-lingual gaps.

Background

The authors show that cross-lingual gaps diminish when source confidence (the probability of the modal response) is high and present theoretical and empirical links between source and target variance, implying that source-side factors are pivotal.

However, they explicitly state that the drivers of source-language variance are unknown and provide related observations (e.g., entity-centric gaps) without establishing causal factors.