XML is pretty good for information extraction with LLMs. It will parse your input text into a tree structure. JSON and YAML will conjure slightly different skills from the LLM. Maybe it all comes from the slightly different applications of these formats in the training corpus.