Hallucination

Hallucinations range from subtle factual errors (wrong date, wrong number) to entirely invented citations or events. Every LLM does it, at some rate. Mitigations include: RAG (ground responses in retrieved documents), RLHF (train to prefer honest answers), reasoning models (extended thinking catches errors), and verification layers (secondary model or tool checks the output).

Hallucination root cause: LLMs predict next-token probability, not truth. When the model lacks information, it predicts plausible continuations based on pattern-matching, producing output that looks right but isn't. Older models: 10-20%+. Mitigations rank: RAG > reasoning + verification > RLHF > prompt engineering. Hallucination is hardest in: obscure facts, numeric details, citations, legal/medical advice, code that uses deprecated APIs. See the related terms and live BenchGecko data for current examples.

Measured on TruthfulQA (817 adversarial questions designed to elicit common misconceptions), HaluBench, and custom evals. RAG reduces hallucination 3-10× when retrieved context is relevant; near zero reduction when retrieval fails. Self-consistency (sample k outputs, pick majority) reduces some types. Reasoning models with extended thinking catch more errors via internal self-critique. Production systems stack: RAG retriever → verifier → reasoning model → citation extractor. Total hallucination rate for well-engineered pipelines: <0.5% on factual queries.