fact verification

Fact verification checks LLM-generated claims against retrieved documents or knowledge bases for accuracy. **Motivation**: LLMs hallucinate - generating plausible but false statements. Verification provides factual grounding. **Approaches**: **Retrieval-based**: For each claim, retrieve relevant documents, check if claim supported. **Entailment models**: NLI classifier determines if retrieved text entails claim (entailment/neutral/contradiction). **LLM-as-judge**: Use model to compare claim with retrieved evidence. **Pipeline**: Extract claims from response → retrieve evidence per claim → verify each → flag unsupported claims. **Granularity**: Sentence-level, entity-level, or full-response verification. **Response options**: Reject unsupported claims, add caveats, regenerate with constraints. **Tools**: FactScore, TRUE benchmark, Minicheck, custom NLI pipelines. **Challenges**: Defining ground truth, handling opinions vs facts, partial support, temporal validity. **Production use**: Critical for high-stakes domains (medical, legal, financial), news/content generation. **Trade-offs**: Adds latency and cost, may over-reject valid claims without retrieved support. Essential for trustworthy AI systems.