The paper tries several different ways to translate five low-resource Turkic languages, instead of forcing one method to fit all.
Different programming languages scale differently when training code AI models, so treating them all the same wastes compute and lowers performance.