Chain-of-Verification for Hallucination Reduction Quiz

Answers and explanations appear as you take the quiz.

1. What is the first step in Chain-of-Verification?
   • Writing a baseline draft answer
   • Searching the web
   • Deleting the prompt
   • Retraining the model
2. Why do CoVe variants answer checking questions separately from the draft?
   • To prove that every model error becomes independent
   • To remove all need for a final revision
   • To guarantee that the model's stored knowledge is correct
   • To reduce bias toward repeating or confirming an earlier mistaken claim
3. What kind of questions does the planning step generate?
   • Random trivia unrelated to the answer
   • Targeted questions probing the draft's factual claims
   • Questions about the user's mood
   • Programming syntax questions
4. What does the final step of CoVe produce?
   • A new training dataset
   • A list of hyperparameters
   • A revised answer that fixes or drops unsupported claims
   • An image of the answer
5. Which problem is CoVe specifically designed to reduce?
   • Slow tokenization
   • High GPU temperature
   • Short context windows
   • Confident factual hallucinations
6. How could a tool-assisted CoVe extension check a draft claim beyond the model's stored knowledge?
   • Consult relevant external records or a suitable calculation for the targeted checking question
   • Treat repetition of the original claim as external evidence
   • Delete every verification answer before revision
   • Assume that every retrieved page is necessarily reliable
7. What cost should be included when comparing CoVe with one direct model response?
   • Only the number of letters in the final answer's title
   • A mandatory full retraining of the model for every claim
   • The computation and latency of planning, checking, and revising in addition to drafting
   • A guarantee that every additional checking stage has zero cost
8. Why might a CoVe checking answer still repeat a false claim even when prompted in a separate context?
   • Separate contexts prove the model has obtained new external evidence
   • The model can retain the same knowledge error or misconception across calls
   • Every factual error is caused only by seeing the original draft
   • A checking question automatically replaces all model parameters