LLMs produce much more consistent XML than JSON because JSON is a horrible language that can be formatted in 30 different ways with tons of useless spaces everywhere, making for terrible next token prediction.