I have a hunch they made their decision to train off that particular type of A* traces to avoid an exponential number of embeddings.