Prefix-Tuning: Optimizing Continuous Prompts for Generation

TL;DR

Prefix-tuning is a lightweight method that optimizes a small continuous vector to adapt large language models for generation tasks, achieving comparable or better results than fine-tuning with fewer parameters.

Contribution

The paper introduces prefix-tuning, a novel approach that keeps model parameters frozen and optimizes a small prefix, reducing storage and improving low-data and unseen-topic performance.

Findings

Achieves comparable performance to fine-tuning with only 0.1% of parameters.

Outperforms fine-tuning in low-data scenarios.

Generalizes better to unseen topics.

Abstract

Fine-tuning is the de facto way to leverage large pretrained language models to perform downstream tasks. However, it modifies all the language model parameters and therefore necessitates storing a full copy for each task. In this paper, we propose prefix-tuning, a lightweight alternative to fine-tuning for natural language generation tasks, which keeps language model parameters frozen, but optimizes a small continuous task-specific vector (called the prefix). Prefix-tuning draws inspiration from prompting, allowing subsequent tokens to attend to this prefix as if it were "virtual tokens". We apply prefix-tuning to GPT-2 for table-to-text generation and to BART for summarization. We find that by learning only 0.1\% of the parameters, prefix-tuning obtains comparable performance in the full data setting, outperforms fine-tuning in low-data settings, and extrapolates better to examples with topics unseen during training.