📚 Methodology Chapter ​

Overview ​

This chapter is divided into 9 major sections, including Design Principles, Framework, Basic Prompt, Advanced Prompt, Automated Prompt, Chain-Of-Thought(COT), In Context Learning(ICL), Knowledge Augumented Prompt, and Evaluation and Reliability. Among them, Design Principles section mainly introduce how readers can improve the quality and accuracy of large language models' answers and enhance the overall user experience by following design principles; for the Basic Prompt section, several case studies and scenarios are used to quickly cultivate writing prompt habits and techniques; for the Advanced Prompt section, advanced and complex prompts are designed using relevant cutting-edge technologies and taking into account many factors; for the Automated Prompt section, related machine learning methods are used to automate the design of prompts and minimize manual design; for the COT section, several papers are introduced on the use of thinking chains in large models and their application scenarios; for the ICL section, related papers are also used to introduce this section; for the Knowledge Augumented Prompt section, this column is explained in conjunction with corresponding papers; for the Evaluation and Reliability section, explanations are provided based on relevant cutting-edge papers.

Design Principles ​

This section presents several design principles for prompts in natural language processing, which provide input text to models and guide them to generate appropriate responses. Some of the design principles for prompts include:

• Simplicity principle: Clearly state it.
• Diversity principle: Use multiple different prompts.
• Length principle: Try to use short prompts as much as possible.
• Structure principle: Use simple sentence structures.

Based on these principles, we have condensed more intuitive design principles without distinguishing them into categories.

Framework ​

This section proposes a prompt design framework consisting of five key components: context, instruction, relevance, constraints, and demonstrations. This standardized structure can help simplify prompt design, ensure adherence to consistent procedures, and produce more effective prompts.

Basic Prompt ​

The design of basic prompt is designed to facilitate readers in quickly establishing a writing prompt design habit.

Advanced Prompt ​

• Batch Prompting
• Successive Prompting
• PAL Prompting
• ReAct Prompting
• Self-ask
• Context-faithful Prompting
• REFINER: Reasoning Feedback on Intermediate Representations
• Reflexion
• Progressive-Hint Prompt
• Self-Refine
• RecurrentGPT

Automatic Prompt ​

This section focuses on using advanced machine learning and other technologies to generate prompts for language models (LM/LLM). It trains the model to generate high-quality, context-relevant and accurate prompts, guiding LM/LLM to produce more coherent, precise, semantically rich, and context-relevant results. This column will introduce this section through selected papers.

CoT ​

This section mainly introduces the thinking chain from the following three aspects:

1. What is COT?
2. What are the forms of CoT prompts?
3. How to design innovative CoT prompts?

In-Context Learning ​

This section mainly introduces context learning from the following four aspects:

1. What is in context learning (ICL)?
2. What are the characteristics of ICL?
3. What are the advantages of ICL?
4. What are the components of research on ICL?

Knowledge Augumented Prompt ​

Evaluation and Reliability ​

This section introduces the latest cutting-edge methods for evaluating and verifying the reliability of large models through selected papers.