Analysis of Numerical Localisation in LLM Translations

The authors extend Tang et al.'s research on numerical translation by focusing on localization—adapting times, numbers, and dates to the target culture—rather than direct translation. Five LLMs were selected specifically for their ability to run on commodity hardware, making the study accessible and reproducible. A baseline quality metric was computed for each model, followed by testing three distinct improvement strategies. The paper reports findings that contrast with Tang et al.'s original conclusions, suggesting that localization poses unique challenges separate from general translation. These results are relevant for developers building multilingual applications where cultural formatting accuracy matters.