Revision Prompting: Diff your inputs for consistency and efficiency (revisionprompting.info)

A new technique called Revision Prompting has been introduced to enhance the efficiency and consistency of interacting with Large Language Models (LLMs). This approach addresses two major issues in industrial prompting: the non-deterministic nature of LLMs, which can lead to inconsistent outputs when inputs are updated, and the inefficient processing of entire inputs when only minor changes occur. By focusing on the differences between the original and updated inputs—termed "diff"—Revision Prompting allows users to apply small patches to previous outputs instead of generating entirely new ones. This method significantly reduces both processing time and costs. For instance, when updating a translation of an e-bike product page due to an increase in range from 80 km to 100 km, instead of re-translating the entire text, Revision Prompting identifies the change and generates a concise output patch. As a result, this technique can lead to approximately 80% reductions in processing time and 65% in costs, making it a promising advancement in the AI/ML community for automating and optimizing language model applications across various industrial processes.