Scheduled Sampling and Exposure Bias Quiz

Answers and explanations appear as you take the quiz.

1. Exposure bias arises because models trained with teacher forcing only condition on:
   • Ground-truth tokens, but must use their own at inference
   • Their own predicted tokens
   • Random noise tokens
   • Future tokens
2. What does scheduled sampling do during training?
   • Always feeds the true token
   • Randomly chooses between the true token and the model's own prediction
   • Disables the loss function periodically
   • Doubles the learning rate over time
3. How does the probability of using the ground-truth token change over training in scheduled sampling?
   • It stays constant
   • It starts low and increases
   • It starts high and decays toward zero
   • It oscillates randomly
4. Who introduced scheduled sampling, and roughly when?
   • Vaswani et al., 2017
   • Hinton et al., 2006
   • LeCun et al., 1998
   • Bengio et al., 2015
5. A recognized technical drawback of feeding sampled tokens back during training is that:
   • The discrete sampling step is non-differentiable, complicating gradients
   • It makes training fully deterministic
   • It requires no validation data
   • It eliminates the need for embeddings
6. Which schedule shape is commonly cited for decaying the teacher-forcing probability?
   • Constant
   • Inverse-sigmoid decay
   • Step-up linear increase
   • Random walk
7. At a decoding step where scheduled sampling chooses the model-output branch, what is fed as the next context token?
   • The next ground-truth token every time
   • A token from a separate validation sequence
   • A token sampled from the model’s predicted distribution
   • The average of all vocabulary embeddings
8. Why does scheduled sampling usually begin with a high probability of using ground truth?
   • It guarantees unbiased gradients
   • It removes the need for a training schedule
   • It makes inference use ground-truth tokens
   • Early predictions are poor, so a correct prefix can stabilize initial learning