It does not guess the next word, the sampler chooses subword tokens. Your explanation can't even explain why it generates coherent words.