A research team from the Department of Automation at Tsinghua University has put forward a Local Demo Editing approach, termed LDE, and developed a compact editing model called Jev-LDE. Utilizing an existing set of retrieved ICL examples, this method makes a single, localized modification—be it retaining, deleting, or replacing elements—without the necessity of repeatedly calling upon the target large model. Experimental results indicate that this method surpasses the initial retrieval outcomes in 46 out of 48 classification scenarios. During 1-shot testing across three classification datasets and four large target models, the average accuracy rose from 81.2% to 88.1%, with only an approximate 11% uptick in end-to-end inference time. The trained editing strategy is adaptable to unseen target models and retrievers. Presently, this method exhibits consistent performance in classification tasks, yet it has not delivered similar advantages in tasks like mathematical reasoning.