Dynamic Prompts

Dynamic Prompts is a prompting technology that allows prompts to be dynamically adjusted based on specific tasks or instances in natural language processing (NLP) and other artificial intelligence applications. This technology can significantly improve the performance and adaptability of the model.

The concept and method of Dynamic Prompts were jointly proposed by researchers from the University of California, Santa Barbara and NEC Laboratories America. In their paper published in March 2023,Dynamic Prompting: A Unified Framework for Prompt Tuning"This method is described in detail in the paper. This paper proposes a unified dynamic prompting (DP) adjustment strategy that dynamically determines different factors of prompts based on different tasks and instances. By using a lightweight learning network with Gumbel-Softmax technology, it is able to learn instance-dependent guidance. Experimental results highlight the significant performance improvement of dynamic prompt adjustment in a wide range of tasks including NLP tasks, visual recognition tasks, and vision-language tasks.