Escrever uma avaliação PREreview

In this paper, we introduce an innovative perspective, proposing that prompts in large language models (LLMs) can be viewed as hypernetworks. From this viewpoint, we further suggest that prompt engineering acts as a form of post-training for LLMs. Building upon this foundation, we present a novel training-free approach to transform system prompts into model parameters, serving as a sleep mechanism within LLMs. Our method effectively enables the conversion of knowledge and memory contained in system prompts into model parameters through the sleep mechanism, enhancing the adaptability and efficiency of language models without traditional training processes.