[anon291]

That is one way of sampling tokens. It is not the only way. Logits do not map neatly to belief, although it is convenient to behave as if they do