Hallucination Mitigation

hallucination mitigation is a toolbox of techniques for cutting fabricated content, on three main lines: (1) RAG — grounding generation on retrieved documents, the conditional becomes $P(y \mid x, z)$ with $z$ the retrieved evidence, plus requiring citations; (2) honest refusal — letting the model answer "I don't know" when uncertain, e.g. abstaining when the top-token confidence $\max_t p_t < \tau$; (3) uncertainty prompts — system prompts demanding the model state uncertainty, combined with self-consistency and verification-style generation to filter contradictions.

factual-critical products (customer support, QA, report generation); preventing RAG fabrication when retrieval is irrelevant; preventing agents from making up tool results; a favorite interview combo — how RAG + honest refusal + uncertainty prompts work together.

internalized knowledge can't cover long-tail facts and models can't express ignorance on their own. RAG externalizes knowledge to verifiable evidence, honest refusal makes uncertainty explicit, and uncertainty prompts push models to abstain at low confidence — the three complement each other across "no evidence, no confidence, ambiguous instruction" scenarios.

Why does RAG reduce hallucination, and how do you defend against irrelevant retrieval making it worse?

Implementing honest refusal: confidence thresholds, answerability classifiers, and abstain training data — how do you design each?

Why does self-consistency reduce reasoning hallucination, and what is the majority-vote formula?

What post-hoc verification methods exist, and how do you automatically check factual errors?

Evaluating mitigation: how do you measure the refusal-vs-helpfulness trade-off and tune false refusals?