I think it hallucinates in scenarios where it either doesn't "remember" the answer or there is no answer. For example: since the text generation is sequential,

I think it hallucinates in scenarios where it either doesn't "remember" the answer or there is no answer. For example: since the text generation is sequential, based on context the highest probability continuation in a situation might be: "A famous paper that discussed this phenomenon is" However, when the network gets to this point it doesn't know the name of such a paper. Maybe it already knew that it doesn't know the paper's name but the above sentence was generated despite that, since everything else was lower probability or it got sampled due to higher temperature. In any case, it can't go back and delete the sentence so it must make up something. So it ends up going with the most plausible sounding (highest sample probability) continuation, which ends up being a "hallucination". Basically hallucinations are produced when the network "writes itself into a corner" due to the sequential nature of text generation.