Inject retrieved passages into the prompt and instruct: "Answer only using the context; if not present, say you don't know." This converts a factuality task into a faithfulness task, which is far easier to enforce and audit (citations). See RAG pipeline.
Model drafts → generates verification questions → answers them independently → revises. Independence prevents it from just re-justifying its own error.
Difference between factuality and faithfulness hallucination?
Why does high confidence NOT imply correctness in LLMs?
What does raising the abstention threshold τ do to risk and coverage?
Why decompose answers into atomic claims?
Recall Feynman: explain to a 12-year-old
Imagine a super-fast storyteller who never wants to look stuck. Ask anything and it instantly gives a smooth answer — even when it doesn't actually know, it just makes up something that sounds right. That made-up part is a "hallucination." To stop it we do three things: (1) hand it the answer book first (retrieval) and say "only use this"; (2) make it show its sources so we can catch the invented bits; (3) let it say "I'm not sure" instead of guessing. We also ask it the same question a few times — the real answer stays the same, the made-up one keeps changing, so we spot the fake.
A fluent, confident output that is factually wrong or unsupported by the given context, produced because the model predicts likely tokens rather than true ones.
Factuality vs faithfulness hallucination?
Factuality contradicts real-world truth; faithfulness contradicts the provided source/context.
Formula for hallucination rate over N atomic claims with C supported?
(N − C)/N.
Why does self-consistency reduce hallucinations?
Correct facts are stable across samples while fabrications vary, so majority voting filters the unstable false answers.
Effect of raising abstention threshold τ?
Coverage decreases (fewer answered) and selective risk decreases (fewer errors among answered).
Why doesn't RAG fully eliminate hallucinations?
The model can ignore/misread context (faithfulness failure) or the retriever can fetch wrong passages.
Why is temperature=0 insufficient alone?
It removes sampling-based errors but not knowledge-gap errors — it confidently returns the single most likely wrong answer.
Chain-of-Verification (CoVe) core idea?
Draft → generate verification questions → answer them independently → revise, so the model doesn't just re-justify its own mistake.
Why decompose answers into atomic claims for scoring?
So one false claim doesn't hide among true ones; you get a granular faithfulness fraction (FActScore).
Selective risk definition?
Error rate computed only over the questions the model chose to answer (confidence ≥ τ).
Dekho, ek LLM basically ek next-token predictor hai — woh sochta nahi ki "yeh sach hai ya nahi", woh sirf predict karta hai ki agla shabd kya aana chahiye jo natural lage. Isliye jab usko koi fact nahi pata, tab bhi woh confidently kuch bana ke bol deta hai — usko hi hallucination kehte hain. Problem yeh hai ki galat answer bhi utna hi smooth aur confident lagta hai jitna sahi answer, toh sirf sun ke pakadna mushkil hai.
Isko kam karne ka sabse bada trick hai RAG (grounding) — pehle relevant documents retrieve karo, model ko do, aur bolo "sirf isi context se answer do, warna keh do 'mujhe nahi pata'". Isse factuality problem ek faithfulness problem ban jaati hai jo check karna easy hai (citations se). Dusra, factual kaam ke liye temperature kam rakho taaki random low-probability shabd na aaye. Teesra, self-consistency — same question 5 baar poocho, majority answer lo, kyunki sach har baar same aata hai par jhoot badal jaata hai.
Ek zaroori baat: confidence ka matlab correctness nahi hai. Model miscalibrated hota hai, poori confidence se galti bol sakta hai. Isliye hum abstention thresholdτ use karte hain — jab confidence kam ho, model answer na de. τ badhao toh galtiyan kam hoti hain par kam sawaalon ka jawab milta hai (yeh risk–coverage trade-off hai, diagram mein dekho). Aur answer ko chhote-chhote atomic claims mein todo, har claim alag se verify karo — tabhi pata chalega exactly kitna hissa hallucinate hua.
Yaad rakhne ke liye mnemonic: G-C-A-V — Ground, Cite, Abstain, Verify. Bas yeh chaar walls laga do aur hallucination kaafi kam ho jaayega.
