SelfCheckGPT
Sample multiple responses. Compare via NLI or LLM judge. Divergence → hallucination. Free but slow (5-10x cost).
Advertisement
Retrieval verification
For each claim, retrieve supporting doc. NLI classifier checks 'doc entails claim.' Requires KB.
Advertisement
Chain-of-verification
Model self-generates verification questions + answers. Discussed in prompt engineering series.
Uncertainty estimation
Semantic entropy: sample N outputs, measure semantic diversity. High entropy → uncertain → possibly hallucinated.