Describe two automated hallucination detection techniques and their tradeoffs

**Hallucination detection** is the job of automatically catching claims an LLM states confidently but cannot support. Two techniques cover most production needs, and a third often joins them. The first is NLI-based detection. You split the answer into individual claims, then run a Natural Language Inference model that decides whether each claim is entailed by a reference document or the retrieved context. Claims marked contradicted or neutral get flagged. It is fast and cheap, but only as good as the NLI model, which can stumble on domain-specific text. The second is self-consistency. You sample the same answer five to ten times at non-zero temperature and watch which claims stay stable. Facts that wobble between samples are likely invented. The catch: it needs no reference, but it doubles or triples inference cost, and it cannot catch a fact the model gets wrong the same way every time. A third option is an LLM judge that scores groundedness and must cite supporting evidence. It reads naturally but inherits judge bias and cost. In practice teams layer these: cheap NLI first, then judge or sampling on the flagged subset.

**Hallucination detection** in production usually combines three techniques, each fitting a different situation. First is NLI entailment. Decompose the answer into atomic claims. For each, an NLI model classifies it against the reference context as entailed, neutral, or contradicted. Flag the non-entailed ones. It is fast and cheap. The tradeoff: accuracy is capped by the NLI model, which can fail on domain text. Second is self-consistency. Sample the answer N times at non-zero temperature and check which claims recur. Wobbling claims are likely hallucinated. It needs no reference, so it fits closed-book tasks. The tradeoffs are real. It costs N times the inference. It also misses confident systematic errors, since a repeated wrong fact looks perfectly consistent. Third is LLM-judge groundedness. Ask a strong judge to rate whether each claim is supported and to cite the evidence span. Citation forces grounding rather than vibes. The tradeoff is judge bias and per-call cost. When to use each. NLI when you have a reference and want recall at low cost. Self-consistency for closed-book generation. The judge when nuance matters and budget allows. The precision and recall trade runs throughout: loosen thresholds for recall, tighten for precision.