Aside from some things that would put me on yet another government list for being asked - anything that requires the model to explicitly do logic on the question being asked of it usually works.