Prefix Tuning Quiz

Answers and explanations appear as you take the quiz.

1. What does prefix tuning actually train?
   • All of the model's transformer weights
   • Only the model's output vocabulary layer
   • A separate copy of the model for each task
   • A small set of continuous prefix vectors while the model stays frozen
2. Where are the prefix vectors injected?
   • Into the keys and values at every attention layer
   • Only at the final output layer
   • Into the loss function only
   • Into the tokenizer before embedding
3. Who introduced prefix tuning and roughly when?
   • OpenAI in 2018
   • Li and Liang at Stanford in 2021
   • Google in 2015
   • Meta in 2023
4. Why is prefix tuning attractive for serving many tasks?
   • It requires no training data at all
   • It makes inference run on no hardware
   • Each task needs only a tiny prefix instead of a full model copy
   • It permanently rewrites the base model for all tasks at once
5. What trick is often used to stabilize prefix training?
   • Doubling the learning rate every step
   • Freezing the prefix and training the model
   • Removing all attention layers
   • Reparameterizing the prefix through a small feed-forward network that is later discarded
6. On which kind of task did prefix tuning especially shine?
   • Natural language generation like table-to-text and summarization
   • Image classification
   • Audio waveform synthesis
   • Database indexing
7. What proportion of model parameters did the original prefix-tuning paper report learning in its experiments?
   • About 10%
   • About 0.1%
   • About 50%
   • All model parameters
8. How did prefix-tuning perform in the paper’s low-data settings?
   • It always matched full fine-tuning on every task
   • It could not train with limited examples
   • It outperformed full fine-tuning in some evaluated low-data cases
   • It changed every backbone parameter