# MetaPrompting: Learning to Learn Better Prompts

**Authors:** Yutai Hou, Hongyuan Dong, Xinghao Wang, Bohan Li, Wanxiang Che  
**arXiv:** [2209.11486](https://arxiv.org/abs/2209.11486)  
**Date:** 2022-09-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.11486v1/

## Key points

- Few-shot natural language processing has made progress with prompting methods
- Recent research uses continuous "soft prompts" instead of discrete token-based "hard prompts"
- Soft-prompting methods rely heavily on good initialization to be effective
- Generalized soft prompting method called MetaPrompting has been proposed
- MetaPrompting adopts model-agnostic meta-learning algorithm to find better prompt initialization for fast adaptation to new tasks
- Extensive experiments show that MetaPrompting improves performance on four different datasets, achieving new state-of-the-art performance
- Over 6 points improvement in accuracy for the 1-shot setting was observed
- Authors of this paper are Yutai Hou, Hongyuan Dong, Xinghao Wang, Bohan Li, and Wanxiang Che.
- MetaPrompting can learn to learn better prompts and overcome challenges associated with initializing soft prompts effectively.
- This research represents an important step forward in few-shot natural language processing by providing a more efficient and effective way to initialize soft prompts for improved performance on various tasks.

## Summary

The field of few-shot natural language processing has made significant progress with the development of prompting methods. Recent research has shifted from using discrete token-based "hard prompts" to continuous "soft prompts," which employ learnable vectors as pseudo prompt tokens and have shown improved performance. However, these soft-prompting methods heavily rely on good initialization to be effective. Obtaining a perfect initialization for soft prompts requires an understanding of inner language models and elaborate design, which is a challenging task that needs to be repeated for each new task. To address this issue, a generalized soft prompting method called MetaPrompting has been proposed. This method adopts the model-agnostic meta-learning algorithm, which automatically finds better prompt initialization that facilitates fast adaptation to new prompting tasks. Extensive experiments show that MetaPrompting tackles the soft prompt initialization problem and brings significant improvement on four different datasets, achieving new state-of-the-art performance. In particular, there was over 6 points improvement in accuracy for the 1-shot setting. The authors of this paper are Yutai Hou, Hongyuan Dong, Xinghao Wang, Bohan Li, and Wanxiang Che. Their work demonstrates how MetaPrompting can learn to learn better prompts and overcome the challenges associated with initializing soft prompts effectively. Overall, this research represents an important step forward in few-shot natural language processing by providing a more efficient and effective way to initialize soft prompts for improved performance on various tasks.