Defines hallucination narrowly as fabricated, ungrounded model output
Distinguishes in-context vs extrinsic hallucination clearly
Frames the two requirements: be factual, admit ignorance
Key quotes
“To avoid hallucination, LLMs need to be (1) factual and (2) acknowledge not knowing the answer when applicable.”
“Equally importantly, when the model does not know about a fact, it should say so.”
Transcript
Hallucination in large language models usually refers to the model generating unfaithful, fabricated, inconsistent, or nonsensical content. As a term, hallucination has been somewhat generalized to cases when the model makes mistakes. Here, I would like to narrow down the problem of hallucination to cases where the model output is fabricated and not grounded by either the provided context or world knowledge. There are two types of hallucination: In-context hallucination: The model output should be consistent with the source content in context. Extrinsic hallucination: The model output should be grounded by the pre-training dataset. However, given the size of the pre-training dataset, it is too expensive to retrieve and identify conflicts per generation. If we consider the pre-training data corpus as a proxy for world knowledge, we essentially try to ensure the model output is factual and verifiable by external world knowledge. Equally importantly, when the model does not know about a fact, it should say so. This post focuses on extrinsic hallucination. To avoid hallucination, LLMs need to be (1) factual and (2) acknowledge not knowing the answer when applicable.
