Comment by jmalicki

Of course that can happen!

But they are actual tokens produced, that are then read by the answer generation as part of the prompt, nonetheless. And the hidden state of course has a ton of logic that may not be apparent by the tokens produced as well!

Unlike humans, this thinking cannot possibly be retrograde, since causal masking means it is strictly generated before the answer and cannot be affected by it (though the model may have some concept of an answer by the time it starts generating the thinking tokens, and there is no guarantee the thoughts generated by thinking are actually attended to by the text generation).