They use a separate ngram model to generate the proposed sequence instead of extra heads on top of the main model. The process of verifying the proposed sequence appears to be the same.