Hallucination
Fluent output that is not true.
When a model produces confident, well-formed text that is factually wrong or unsupported. It is a natural consequence of predicting plausible continuations rather than consulting a source.
Two flavours matter in practice. Factuality hallucinations come from missing knowledge — fix them with retrieval. Faithfulness failures drift from material already in the prompt — fix them by clearing noise or re-anchoring the instruction.
The framing that leads to better fixes: hallucination is partly an incentive problem. If a task is graded only on correctness, and a guess sometimes scores and abstaining never does, then guessing is the higher-scoring strategy. Changing the scoring is often more effective than adding warnings to the prompt.
Rewarding the admission
An extraction task where the correct answer is often "not stated in the document".
If abstention is never counted as correct, the model fills the field with something plausible. Making "not stated" a first-class correct answer removes a large share of the fabricated values without touching the model.
The citation that proves it
A confidently worded claim with no source.
Require a quotable span from the retrieved text for every factual sentence, and display it. Unsupported sentences become visible instead of fluent, which shifts the work from "detect hallucination" to "check the citation", which is much cheaper.
Treating every error as "the model lying". Most hallucinations are a retrieval or grounding gap, not deception.
- paperWhy language models hallucinate ↗
The statistical argument that training and grading regimes reward guessing over admitting uncertainty — so hallucination is an incentive problem, not a gap in knowledge.
- standardLLM09: Misinformation ↗
Treats overreliance as the risk: the failure is users accepting confident output without verification, so the mitigation is validation and visible limitations, not just better models.
- paperRetrieval-augmented generation for knowledge-intensive NLP tasks ↗
The original RAG formulation: combine a parametric model with a non-parametric dense index, so answers can be grounded in retrieved text.
Parametric Knowledge
What the model stores inside its weights.
RAG · Retrieval-Augmented Generation
Look things up, then put what you found in the prompt.
Faithfulness · Grounding
Sticking to the material it was given.
Knowledge Cutoff
The date the model’s knowledge stops.
Grounding
Tying outputs to evidence the user can check.