Prompt Tuning Quiz

Answers and explanations appear as you take the quiz.

1. What is a 'soft prompt' in prompt tuning?
   • Learned continuous embedding vectors that need not map to real words
   • A politely worded text instruction
   • A smaller distilled copy of the model
   • A set of frozen output logits
2. Where does prompt tuning add its trainable vectors?
   • At every transformer layer
   • Only at the input embedding layer
   • Only at the output layer
   • Inside the optimizer state
3. What surprising effect did the original paper highlight about model scale?
   • Prompt tuning gets worse as models grow
   • Scale has no effect on prompt tuning
   • As models grow larger, prompt tuning approaches full fine-tuning quality
   • Larger models can no longer be prompt-tuned
4. Who introduced prompt tuning?
   • DeepMind in 2016
   • Anthropic in 2023
   • Microsoft in 2019
   • Researchers at Google (Lester, Al-Rfou, Constant) in 2021
5. How does prompt tuning enable serving many tasks from one model?
   • Each task uses its own tiny soft prompt while the model stays frozen
   • It retrains the whole model for each request
   • It deletes unused tasks from the weights
   • It requires a separate GPU per task
6. How does Lester and colleagues’ paper describe prompt tuning in relation to prefix tuning?
   • As a replacement for frozen language models
   • As a simpler related method that the paper compares with prefix tuning
   • As a tokenizer with no trainable parameters
   • As a method that updates every model weight
7. Which model family did Lester and colleagues use to study prompt tuning across scales?
   • WaveNet
   • AlphaFold2
   • T5
   • Isolation Forest
8. How are soft prompts optimized in prompt tuning?
   • By changing all model weights for each task
   • By replacing embeddings with hard-coded words only
   • By training a new tokenizer for every example
   • By backpropagation through a frozen language model to update the prompt vectors