Self-Play Fine-Tuning Quiz

Answers and explanations appear as you take the quiz.

1. Which famous system reached superhuman Go play using only self-play and no human game records?
   • AlphaGo Zero
   • Deep Blue
   • Watson
   • GPT-2
2. In SPIN, what plays the role of the 'opponent' that the model must beat?
   • A separate human-trained reward model
   • The model's own previous checkpoint and its generated responses
   • A random text generator
   • The tokenizer
3. What is the main data-efficiency advantage of self-play fine-tuning?
   • It requires a brand-new human dataset each round
   • It only works with infinite data
   • It can extract more gains from a fixed dataset without new human labels
   • It needs no model at all
4. In SPIN's training objective, what is the model pushed to do?
   • Prefer its own generated answers over human references
   • Ignore all human data
   • Memorize the test set
   • Assign higher likelihood to human reference responses than to its own prior outputs
5. Which pair of games did the AlphaZero paper demonstrate beyond its Go experiments?
   • Chess and shogi
   • Poker and bridge
   • Checkers and Othello
   • Atari and StarCraft
6. What risk did researchers report when language models are recursively trained on generated data?
   • Model collapse, with loss of diversity and representation of the original distribution
   • Guaranteed monotonic improvement at every iteration
   • Automatic prevention of distribution shift
   • A higher count of verified human labels
7. What search component does AlphaGo Zero use alongside self-play?
   • A nearest-neighbor lookup over human game records
   • A language-model tokenizer
   • Monte Carlo tree search guided by a policy and value network
   • A fixed list of opening moves
8. Why can automatic checkers be useful in self-play for math or code tasks?
   • They guarantee that a model never learns incorrect strategies
   • They replace the need for a task specification
   • They make all generated examples independent of the model
   • They can verify generated solutions without requiring a new human label for each attempt