Input Composition Formula for Prompt Tuning

An engineer is adapting a large, pre-trained language model for a new task. To do this efficiently, they keep all the original model's parameters frozen. Their adaptation strategy involves modifying the input sequence before it is processed by the model. For any given text, they first convert the text into its standard sequence of numerical token representations. Then, they prepend a separate, short sequence of newly initialized, trainable numerical vectors to the beginning of that sequence. Only these new vectors are updated during training on the new task. Which statement best distinguishes the nature of these prepended, trainable vectors from the standard token representations?

You are examining the input layer of a large language model adapted using a parameter-efficient technique. The input is formed by combining two distinct types of numerical vectors. Match each vector type with its correct description.

Prefix Tuning Architecture

Parameter-Efficient Model Adaptation