hillbig LLM causes Hallucination even if the training data is correct. This is because a correctly calibrated predictive distribution, like Good-Turing estimation, assigns the same probability to an unknown fact as to a fact observed only once during training. Showing that hallucination reduction requires prior training followed by another Calibrated Language Models Must Hallucinate Recent language models have a mysterious tendency to generate false but plausible-sounding text. Such “hallucinations” are an obstacle to the usability of language-based AI systems and can harm…
It could also be interpreted that Homo sapiens who find meaning in things they have only experienced once are just not calibrating enough.
This page is auto-translated from /nishio/一度だけ見たものは未知 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.