Last time I researched AST parsing was quite rudimentary for LMs. The problem was preserving their structural properties, whereby a flattening approach via node traversal tended to remove. But you needed to do to put it into a format that language models could parse.