Process Supervision for Math Reasoning Quiz

Answers and explanations appear as you take the quiz.

1. Where does process supervision provide feedback compared with outcome supervision?
   • Only at the physical GPU temperature
   • At intermediate reasoning steps rather than only the final result
   • Only at the document filename
   • Only at an increased vocabulary size
2. Why can outcome-only supervision be misleading for math problems?
   • It is too slow to compute
   • It cannot handle decimals
   • It requires human graders for every problem
   • A model can reach the right answer through flawed steps that happen to cancel out
3. What did OpenAI release alongside the 'Let's Verify Step by Step' work?
   • PRM800K, a dataset of ~800,000 step-level labels
   • A new GPU architecture
   • The GSM8K test set
   • A chess engine
4. In best-of-N math selection with a process reward model, what does the verifier do?
   • Guarantees that a correct candidate was generated
   • Changes every tokenizer token into a proof axiom
   • Scores generated candidate solutions under the chosen aggregation rule
   • Replaces the original problem with a translation
5. What approach reduces the need for expensive human step labels?
   • Disabling the reward model entirely
   • Monte Carlo rollouts that estimate each step's value automatically (e.g., Math-Shepherd)
   • Using only the final answer
   • Training on more languages
6. What can low step-level verifier scores usefully indicate to someone reviewing a solution?
   • Steps that merit closer inspection
   • An infallible location of the first actual error
   • A mathematical proof that all other steps are correct
   • The exact temperature used to sample every token
7. For step scores 0.9 and 0.8, what is the solution score under a product aggregation rule?
   • 0.8
   • 0.85
   • 1.7
   • 0.72
8. Which data-selection strategy did Let’s Verify Step by Step study to improve labeling efficiency?
   • Choosing examples only by alphabetical filename order
   • Removing every incorrect-answer candidate from the labeling pool
   • Active learning that selected informative candidate solutions
   • Using an unlabeled test answer as the final human label