How do you detect and mitigate hallucinations?
LLM & AILevel: IntermediateHallucinations
How do you detect and mitigate hallucinations?
Hints
- Clarify the threat: an unsupported fact, an untrusted instruction or an unsafe action?
- Separate input checks, authority limits and output validation; consider false positives.
Answer
Detection:
- Grounding check: for RAG systems, verify every claim in the response appears in retrieved context (RAGAS faithfulness metric)
- Self-consistency: sample the same query with high temperature N times; if answers disagree significantly, flag as uncertain
- Factual verification pipeline: extract claims from the response; verify each against a knowledge base or search engine
- Confidence calibration: chain-of-thought prompting where the model explicitly states its confidence and reasoning
Mitigation:
- RAG + source citation: ground responses in retrieved documents; require citations for factual claims
- Explicit "I don't know" training: SFT or RLHF reward model trained on responses that appropriately express uncertainty
- Temperature reduction: lower temperature (0.0–0.3) for factual queries
- Constitutional self-check: after generation, prompt model: "Review this response for factual errors or unsupported claims"
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0