AIMentor — AI Expertise Learning Platform

13 questions · pass mark 80% · 20 minute limit. The timer auto-submits at zero.

Time remaining: 20:00

Question 1easy
Why do LLMs hallucinate?
They deliberately lieThey predict the most probable next token, not the true one, with no built-in model of truthOnly poorly trained models do itBecause of the temperature setting alone
Question 2medium
What is the difference between intrinsic and extrinsic hallucination?
Intrinsic contradicts the provided source (unfaithful); extrinsic is wrong about the world with no sourceIntrinsic is always worseThey are the sameExtrinsic only happens with RAG
Question 3easy
What is the single most important reliability instruction in a grounded prompt?
Be conciseAnswer only from the context, and say 'I don't know' when it's insufficient (abstention)Use a higher temperatureAlways give three options
Question 4medium
How does requiring source citations improve reliability?
It makes answers longerIt makes each claim checkable ('supported by chunk 2') and exposes unsupported statementsIt speeds up generationIt removes the need for retrieval
Question 5hard
What is chain-of-verification?
Encrypting the model's outputGenerate → verify each claim against the source with a second pass → revise to remove unsupported claimsRunning the model on a blockchainChaining multiple embeddings
Question 6medium
How does constraining output to a schema/enum prevent a class of hallucination?
It makes the model fasterThe model physically can't emit a value outside the allowed set (e.g. no fourth category), eliminating made-up structured valuesIt improves prose qualityIt disables the model's memory
Question 7hard
In a reliability layer, what are the three typical actions?
Retry, cache, logAbstain (can't ground), escalate to human (verification failed), answer (grounded + verified)Train, evaluate, deployEmbed, retrieve, generate
Question 8hard
For irreversible high-stakes actions (medical, legal, financial, sending, deleting), what additional control is needed?
A higher temperatureA human-in-the-loop approval gate (e.g. LangGraph interrupt) before anything shipsMore few-shot examplesA bigger context window
Question 9medium
Can a defense stack make an LLM 100% reliable?
Yes, with enough guardrailsNo — the goal is to reduce and contain hallucination and fail safe (abstain/escalate) when it happensYes, by fine-tuningOnly with GPT-4
Question 10hard
Which are layered defenses against hallucination? (Select all that apply)
Grounding (retrieve facts) + forced abstentionSource citation and a verification passStructured-output constraints (schema/enum)Raising temperature to add variety
Select all that apply.
Question 11medium
Why set temperature to 0 for grounded factual answering?
It makes the model creativeIt reduces improvisation on facts, keeping the answer closer to the grounded contextIt is required by the APIIt increases token limits
Question 12medium
What one-word behavior — saying 'I don't know' when context is insufficient — is the key first-line reliability defense?
Question 13hard
An answer that contradicts its own provided source is a failure of which quality (one word) from RAG evaluation?

0 / 13 answered

Hallucination & Reliability — Final Exam