I guess the biggest limitation of this approach is that the max output length is fixed before generation starts. Unlike autoregressive LLM, which can keep generating forever.