Hallucination mitigation
The collection of techniques for reducing an LLM making up things that sound plausible but are wrong. No method eliminates hallucinations entirely, but several layers together make output considerably more reliable.
Techniques: ground in sources (RAG — give the model actual material to answer from, the most effective), citation requirements (force the model to point to the source for each claim, so ungrounded statements show), post-hoc verification (a second model/run fact-checks the answer), let the model say "don't know" (calibrate it against guessing), low temperature for factual tasks, and self-consistency (several runs, look for agreement). Also: structured output and tool use (let a calculator/database answer what the model is bad at). Trade-off: more checks = more cost/latency, and too-aggressive mitigation makes the model overly cautious. A core question in building reliable AI systems. Related to RAG and grounding.