#prompt-tuning

Question 1: Select all that apply

Soft prompts (prompt tuning) prepend trainable embedding vectors to the input of a frozen model, instead of hand-written text prompts. What are their main drawbacks? Select all that apply.

They must be learned by gradient descent on task data, unlike written prompts1
The learned vectors match no words, so users cannot read or edit them2
A forward pass is much slower than with a text prompt of the same length3
They need a transformer architecture different from the standard one4

Timed test