Hallucinations mitigation
Why do LLMs hallucinate, and what product techniques reduce harm?
Answers use simple, clear English.
Audio N/AQuick interview answer
Training optimizes next-token likelihood, not truthfulness; models fill gaps fluently. Reduce harm with RAG+citations, tool use for facts/math, constrained decoding/JSON schemas, abstain when confidence/retrieval weak, human review for high-risk domains, and evals for faithfulness. Never present ungrounded chat as authoritative policy.
Detailed answer
Training optimizes next-token likelihood, not truthfulness; models fill gaps fluently. Reduce harm with RAG+citations, tool use for facts/math, constrained decoding/JSON schemas, abstain when confidence/retrieval weak, human review for high-risk domains, and evals for faithfulness. Never present ungrounded chat as authoritative policy.
Real example & use case
Medical triage bot refuses diagnosis; routes to clinician with retrieved guidelines.
Pros & cons
Pros: safer UX. Cons: more abstains may feel less 'smart'.