# Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4

**Authors:** Sondos Mahmoud Bsharat, Aidar Myrzakhan, Zhiqiang Shen  
**arXiv:** [2312.16171](https://arxiv.org/abs/2312.16171)  
**Date:** 2023-12-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.16171v1/

## Key points

- Introduces 26 guiding principles for querying and prompting large language models
- Simplify formulation of questions for different scales of models
- Verify effectiveness through experiments on LLaMA-1/2 and GPT-3.5/4
- Provide researchers with a better guide for prompting large language models
- Use phrases like "Your task is" and "You MUST" to enhance user engagement
- Incorporate consequences for incorrect responses using phrases like "You will be penalized"
- Assign a role to the language model to shape its behavior
- Break down complex tasks into simpler prompts within an interactive conversation
- Use specific prompts like "Explain [specific topic] in simple terms" or "Explain to me like I'm 11 years old"
- Example-driven prompting using few-shot prompts is effective
- Format prompts with clear instructions, examples, questions, context, and input data
- Use affirmative language and avoid negative language in prompts
- Guide the model's response with leading words like "think step by step"
- Allow the model to ask questions for precise details before providing output
- Test understanding by asking the model to teach a theorem or topic without giving away answers but providing feedback on correctness
- Assign a role to large language models and use delimiters to enhance performance

## Summary

, , , , 

This paper introduces 26 guiding principles aimed at streamlining the process of querying and prompting large language models. The goal is to simplify the formulation of questions for different scales of these models, understand their capabilities, and enhance user comprehension regarding their behaviors when fed with various prompts. The effectiveness of these principles on instructions and prompt design is verified through extensive experiments conducted on LLaMA-1/2 (7B, 13B, and 70B) as well as GPT-3.5/4. The objective is to provide researchers working on the prompting of large language models with a better guide. The project page for this work can be found at https://github.com/VILA-Lab/ATLAS. The expanded summary also provides additional insights into user interaction and engagement strategies. It suggests incorporating phrases like "Your task is" and "You MUST" to enhance user engagement. Additionally, it recommends using phrases such as "You will be penalized" to establish consequences for incorrect responses. In terms of content and language style, assigning a role to the language model helps in shaping its behavior. Phrases like "Answer a question given in natural language form" can be used in prompts to specify the desired response format. Instead, it is advised to get straight to the point. For complex tasks and coding prompts that involve multiple files, generating a programming script that automates file creation or modification can be helpful. Breaking down complex tasks into simpler prompts within an interactive conversation is also recommended. To ensure clarity or deeper understanding of a topic, specific prompts like "Explain [insert specific topic] in simple terms" or "Explain to me like I'm 11 years old" can be utilized. Example-driven prompting using few-shot prompts is another effective approach. Formatting prompts with clear instructions, examples, questions, context, and input data can improve comprehension. The refined summary also suggests incorporating phrases like "Your task is" and "You MUST" to provide clear directives to the model. It emphasizes the use of affirmative language while avoiding negative language. Leading words like "think step by step" can guide the model's response. Allowing the model to ask questions for precise details and requirements before providing output is encouraged. A specific prompt example for testing understanding includes asking the model to teach a theorem or topic without giving away answers but providing feedback on correctness. Assigning a role to large language models and using delimiters can further enhance their performance. Overall, this expanded summary provides detailed insights into user interaction strategies, content and language style considerations, handling complex tasks and coding prompts effectively, as well as various techniques for improved engagement with large language models.