What is hallucination, and why do LLMs hallucinate?

Main causes

1. Objective mismatch: next-token prediction rewards plausibility. Nothing in pretraining checks truth.
2. Lossy, compressed knowledge: rare facts (a small company's founding year) are poorly memorised, and the model fills gaps with plausible patterns.
3. Knowledge cutoff: events after training aren't known, but the model may still answer.
4. Training incentives: evaluations and feedback often reward a confident answer over abstaining, which encourages guessing.
5. Prompt pressure: leading questions ("Why did X win the Nobel Prize?" when X didn't) and requests for specifics (citations, numbers).
6. Decoding: high temperature, or long generations where an early error snowballs.

Types: factual errors, fabricated references, wrong reasoning steps, unfaithfulness to provided context (in RAG), invented code APIs.

Mitigations

• Ground with retrieval (RAG) and tools (search, calculators, code execution).
• Ask for citations and verify them; allow and encourage "I don't know".
• Lower temperature for factual tasks.
• Post-hoc verification (a second model, NLI checks, self-consistency across samples).
• Use stronger models and models trained to abstain.

Common mistakes

• Claiming RAG or a bigger model eliminates hallucination. They reduce it, so you still need verification for high-stakes outputs.

Related

• Previous: Q10. Walk through the LLM training lifecycle: pretraining, SFT, and preference alignment.
• Next: Q12. What are zero-shot, few-shot, and chain-of-thought prompting?
• LLM Fundamentals
• RAGAS
• DeepEval