Why do LLMs hallucinate, and what actually reduces it?
- 1 min — why: plausibility objective, guess-rewarding training
- 3 min — levers in order: ground, permit abstention, constrain, verify
- 2 min — measurement: golden set with unanswerables, two metrics
- 1 min — ownership and the near-miss retrieval trap
Pretraining rewards plausible continuations, not verified ones. When a fact is weak or absent in parametric knowledge, the most probable continuation is still fluent text — an invented citation, a non-existent API flag, a wrong total. Post-training and most benchmarks historically scored a guess above "I don't know", so abstention is under-trained unless you ask for it, reward it and measure it.
Reduction levers, in order of leverage: put the fact in context (RAG / grounding); permit and instruct abstention; constrain fields with known valid values (structured decoding); verify after generation (schema checks, tool lookups, faithfulness scoring). None eliminates it.
A clinical-intake summariser occasionally lists a medication the patient never mentioned. Golden set: 300 intake notes with clinician-labelled medication lists plus 60 notes with no medications. Baseline: 4.1% of summaries contain an unsupported medication. Changes: (1) instruct "list only medications quoted in the note; output none if none"; (2) require a quoted span per medication; (3) a deterministic check that each span exists verbatim in the note, otherwise drop and flag. Result on the same set: 0.6% unsupported, correct-empty on 57/60 no-medication notes. Remaining failures route to clinician review.
- Track two numbers together: wrong-answer rate and correct-abstention rate. Pushing one down moves the other; pick the operating point per domain (clinical: abstain more; marketing copy: abstain less).
- LLM-as-judge faithfulness scoring scales review but must be calibrated — measure judge–human agreement on 200+ items and only trust it above ~0.8 Cohen's κ.
- Cheapest verification is deterministic: quote-and-check, schema enums, existence checks against an API or catalog.
- Org scope: domain teams own their golden sets and thresholds; a central evals team owns the judge, its calibration and the shared dashboard. Buy the model; build the verification, because it encodes your domain's ground truth.
- Where it breaks: assuming RAG fixes it. Models still fabricate around retrieved context, especially when retrieval returns near-miss passages. Test with deliberately near-miss contexts.
What the answer sounds like at each level
Explains plausibility vs truth and that RAG and 'say I don't know' instructions reduce hallucination.
Builds a golden set with unanswerable items, measures unsupported-claim rate and abstention together, and adds deterministic quote checks.
Sets per-domain operating points, owns judge calibration as a shared service, and makes verification part of the platform rather than per-team prompt tweaks.
Follow-ups the interviewer will push
1Can you just ask the model for a confidence score?
Verbalised confidence is poorly calibrated. Use it only after measuring its reliability on labelled data; prefer agreement across samples or verification against sources.
2Now the domain is regulated (insurance claims). What changes?
Every generated fact needs a source span (attribution), unsupported claims block auto-approval, and the audit log stores the context the model saw so a reviewer can reproduce the decision.
Red flags interviewers listen for
- Claims RAG or a bigger model eliminates hallucination.
- Measures only accuracy, never abstention on unanswerable inputs.
- Relies on an uncalibrated LLM judge as ground truth.
How it's scored
| Dimension | What strong looks like |
|---|---|
| Root cause | Ties hallucination to the training objective and incentive to guess. |
| Mitigation | Orders levers by leverage and includes deterministic verification. |
| Measurement | Paired metrics with an explicit operating point and judge calibration. |
Sources
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — Lewis et al., 2020paper
- Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena — Zheng et al., 2023paper