By "low-resource" they seem to mean "not much training data". I wonder if the fix for this was generic generalization ability, or just making a lot of training data for the specific languages involved. This distinction matters a lot for the question of whether new programming languages can be developed without having LLMs be bad at them.