Concept

Prefix Tuning Architecture

Prefix tuning is a parameter-efficient fine-tuning method where a sequence of trainable vectors, known as prefixes, are prepended to the hidden states at each layer of a Transformer model. For any given layer l, the input consists of the prefixes for that layer (e.g., p0l,p1lp_0^l, p_1^l) followed by the hidden states (h0l1,h1l1,h_0^{l-1}, h_1^{l-1}, \ldots) computed by the previous layer. The core LLM parameters remain frozen, while only these layer-specific prefix vectors are optimized during training to steer the model's output for a downstream task.

0

1

Updated 2026-02-06

Contributors are:

Who are from:

Tags

Ch.3 Prompting - Foundations of Large Language Models

Foundations of Large Language Models

Foundations of Large Language Models Course

Computing Sciences