Self-Refine Iterative Output Improvement Quiz

Answers and explanations appear as you take the quiz.

1. What three roles does a single model play in the Self-Refine loop?
   • Generator, critic, and reviser
   • Trainer, tester, deployer
   • Encoder, decoder, classifier
   • Planner, actor, scorer
2. What does Self-Refine require beyond prompting to work?
   • A separately trained reward model
   • Nothing extra; it needs no additional training or reward model
   • Human reviewers in the loop
   • A fine-tuned critic network
3. Why is generating feedback in a separate prompt from generating the draft helpful?
   • It hides the draft from the model
   • It reduces the token count
   • It lets the model evaluate from a fresh framing instead of defending its first attempt
   • It disables the model's memory
4. What kind of feedback makes the refinement step most effective?
   • Vague encouragement like 'make it better'
   • A simple numeric score with no detail
   • Random suggestions
   • Specific, actionable feedback pointing at concrete flaws
5. When does Self-Refine tend to fail to improve an output?
   • When the model cannot detect the flaw it would need to fix
   • When the task is short
   • When the prompt is in English
   • When temperature is set to zero
6. How does the loop typically decide to stop?
   • It always runs exactly 100 times
   • It stops when the model judges the output good enough or hits a step limit
   • It stops only when a human approves
   • It never stops on its own
7. What information does Self-Refine feed into later refinement rounds?
   • Only the final token of the original prompt
   • A reward model's parameters
   • The draft together with the feedback history
   • A new human-written answer with no prior context
8. In the guide's coding example, what might the critic flag before a revision?
   • An incorrect microphone sample rate
   • A ship arriving outside its port window
   • A mismatch between image and audio codecs
   • Missing edge-case handling