Hallucinations: detection and mitigation
- Evals
- Asked in most interview loops
- About 1 h 30 min to learn
- 3 vetted resources
What this topic covers
A hallucination is output the model states with confidence but that isn’t supported. Before you can fix one, decide what counts in your product: a claim with no support in the context, a claim that is false about the world, or a made-up citation. Interviewers want a layered plan, not one trick. Before generation, ground the model in good context and let it abstain. During generation, constrain the output and ask for quotes. After generation, check each claim against the sources, sample several answers to spot disagreement, and show citations in the UX. A strong answer also explains how you measure faithfulness offline and keep tracking it in production, and why models guess rather than say “I don’t know”.
You’re ready when you can
What interviewers check, from the resources below.
- Explain why fine-tuning on new knowledge can increase hallucination
- Describe a claim-level check that breaks an answer into facts and verifies each against sources
- Explain sampling-based detection (several samples that disagree signal a guess)
- Order fixes into before, during and after generation for a RAG product
- Explain why giving the model permission to abstain reduces made-up answers
- Describe the extract-quotes-then-answer pattern for long documents
- Describe a retract-if-no-supporting-quote step after drafting
- Explain how an NLI model scores whether a summary is entailed by the source
- Describe question-generation-and-answering checks for faithfulness
- Say why reference-based overlap metrics miss unfaithful output
What to study, in order
Each link opened and checked by hand; the date says when.
Start here
Extrinsic Hallucinations in LLMs (opens in a new tab)
Lil'Log · Lilian Weng · 75 min · Free · checked 21 Oct 2026
- Why
- It is the most complete free map of why models hallucinate, how to detect it and what reduces it. That lets you give a structured answer instead of a list of tricks.
- What to read
- Read “What Causes Hallucinations?” (both parts) and “Hallucination Detection” (retrieval-augmented evaluation, sampling-based detection, calibration of unknown knowledge, indirect query). Then read “Anti-Hallucination Methods” up to and including “Sampling Methods”.
- What to skip
- Skim “Fine-tuning for Factuality” and “Fine-tuning for Attribution”. Skip the benchmarks appendix, citation and references.
Free alternative
Reduce hallucinations (opens in a new tab)
Anthropic · 15 min · Free · checked 21 Oct 2026
- Why
- It is a short, practical checklist of prompt-level fixes. Interviewers expect you to try these before reaching for training or extra models.
- What to read
- Read “Basic hallucination minimization strategies” (permission to say “I don’t know”, quote-first grounding, verifying with citations) and “Advanced techniques”. Rewrite one of the example prompts for your own use case.
- What to skip
- You don’t need to open every example; the first two show the pattern.
Go deeper
Evaluation & Hallucination Detection for Abstractive Summaries (opens in a new tab)
eugeneyan.com · Eugene Yan · 45 min · Free · checked 21 Oct 2026
- Why
- When asked “how would you measure hallucination?”, this gives you concrete detectors and their trade-offs. It goes beyond “use an LLM judge”.
- What to read
- Read “Context-based (reference-free) metrics”, including the part on detecting inconsistencies with the source, then “Sampling-based metrics”. Skim the opening on fluency, coherence, relevance and consistency for vocabulary.
- What to skip
- Skip “Reference-based metrics” (ROUGE, BERTScore). They correlate poorly with faithfulness, and the post says as much.
Companies that test Evals most
Put Hallucinations in your plan
Tell us your interview date and hours a day; it lands on the right day, in the right order.