I haven't got systematic measurements that I can present yet, but informal measurements on a number of XMark queries show similar results to those found by Marian and Simeon. That is, a pretty dramatic improvement. The question really is, how typical is this scenario: it works only for relatively simple queries in situations where you are parsing and building the document in order to run a single query, and then only when the query is extracting a small amount of information from a large document.