Paper “Prompt Reporting: A Systematic Overview of Prompt Engineering Technology”

By /

This paper provides a comprehensive survey report on hinting engineering techniques in generating artificial intelligence systems. It aims to establish a structured understanding of hinting engineering. It provides a detailed vocabulary of 33 terms, a classification of hinting technologies for 58 large language models, and 40 technologies in other modalities, and provides best practices and guidelines for hinting engineering.

Paper title: The Prompt Report: A Systematic Survey of Prompt Engineering Techniques(Chinese can be translated as “Prompt Report: A Systematic Summary of Prompt Engineering Technology”)
arXiv number: 2406.06608, first submitted June 6, 2024, latest revision (v6) released February 26, 2025

Overview of core content

This article focuses on discussing “Prompt EngineeringIn this field, through a systematic literature survey, the terms, techniques, methods and practices in this field have been clearly organized.

1. Terminology and technical classification

  • establishedGlossary of 33 related terms, helps clearly remind concepts commonly used in engineering

2. Technical system and list

  • proposed58 text-only prompting technologies;
  • Also compiled40 multimodal prompting technologies (such as images, audio, video, etc.)

3. methods and strategies

  • Contains “meta-analysis” to systematically evaluate literature related to prefix-prompting in natural language.
  • Various text prompt strategies were analyzed, such as:
    • In-Context Learning（ICL）
    • Zero-Shot
    • Thought Generation
    • Decomposition
    • Ensembling
    • Self-Criticism such

4. expand the field

  • Not limited to English text, but also covers multiple prompts, prompt template language selection, machine translation, image/audio/video/3D multimodal prompts, etc.
  • Explore high-level application directions, such as:Tools use Agent, Code Generation Agent, RAG (Retrieval Augmented Generation) and other content

5. Practical guides and safety assessments

  • Provide tips on how toBest practices and guidance, especially suitable for current mainstream models (such as ChatGPT)
  • Propose methods to evaluate alert technology and discuss security issues (such as alert attack risks)

6. Summary and positioning

  • The author’s team comes from many top research institutions such as OpenAI, Stanford, and Microsoft. The article is considered to be the most comprehensive and systematic review of prompt engineering to date.

Summary (concise version)

This paper creates a complete reminder project through a systematic literature reviewterminology systemwithtechnical classification, covering text and multimodal prompt methods, and providing operational suggestions and safety considerations based on actual application requirements, can be regarded as a “programmatic resource” in this field.

Directions that you may be interested in

  • Break down in detail the definition and application of certain types of prompting technologies (such as Self-Criticism, Ensembling, etc.)?
  • Compare the similarities and differences between text prompts and multimodal prompts and application scenarios?
  • Summary tips on engineering best practices and common pitfalls?
  • Explore the potential and challenges of prompting technology in multilingual or tool-use agent models?

Original text:https://arxiv.org/abs/2406.06608

Oil tubing:

Scroll to Top