Grokking and Delayed Generalization Quiz

Answers and explanations appear as you take the quiz.

1. Which performance pattern characterizes the studied grokking phenomenon?
   • Held-out accuracy must improve before any training example is learned
   • Training accuracy can never become high
   • Every model loses all training information at once
   • Held-out generalization improves substantially after the training examples have already been fit
2. Which tasks did the original grokking paper use as a controlled setting?
   • Only large-scale photorealistic image generation
   • Small algorithmically generated tasks such as modular arithmetic
   • Only real-time online advertising auctions
   • Only unlabelled random filenames
3. Which regularization intervention was among those examined in the original grokking experiments?
   • Weight decay
   • Mandatory deletion of every activation function
   • A proof that all test labels must be randomized
   • Automatic enlargement of the vocabulary on every step
4. What computation did the cited modular-addition mechanistic study identify in its trained small transformers?
   • A universal rule for every large deployed model
   • Only a table containing every held-out answer
   • A Fourier-based algorithm using trigonometric relationships
   • No identifiable structured computation
5. Which phase ordering does Progress Measures for Grokking describe in the studied training dynamics?
   • Cleanup, deletion of all data, then no computation
   • Memorization, circuit formation, then cleanup
   • Perfect held-out performance before any circuit formation
   • A mandatory switch to image-only inputs
6. What practical caution does grokking suggest for practitioners?
   • Always stop training the moment training accuracy hits 100%
   • Validation accuracy is irrelevant
   • An early validation plateau does not always mean the model has failed to learn
   • Regularization should never be used
7. For the modular-addition task, what is (6+5) modulo 7?
   • 11
   • 7
   • 1
   • 4
8. What dataset-size trend did the original grokking paper report in its studied setups?
   • Smaller training datasets required more optimization to reach generalization
   • Smaller datasets always generalized immediately with no optimization
   • Dataset size was never varied
   • Every larger dataset prevented all training accuracy