Researchers explained how to improve the quality of AI responses - ForkLog: cryptocurrencies, AI, singularity, future

# Researchers explained how to improve the quality of AI responses

Carefully crafted contextual prompts increase the accuracy of AI model responses. This conclusion is found in an article by the Shanghai Artificial Intelligence Laboratory.

Despite the ability of neural networks to understand natural speech, they still require additional information and clearly formulated requests to provide quality results. For example, if you ask the AI to “plan a trip,” it may suggest a luxury cruise without knowing about the limited budget.

Good questions help avoid “entropy” — confusion due to excessive uncertainty.

How to Create Quality Prompts

The article suggests ways to improve communication efficiency with artificial intelligence. They are based on prompt design (prompt engineering).

Some tips:

1. It is necessary to start with the basics: who, what, why. It is important to always include background information to create context. Instead of the prompt “Write a poem”, it is worth trying the request: “You are a romantic poet writing in honor of my anniversary. The theme is eternal love. Let the poem be short and sweet.”
2. It is worth structuring information “layer by layer”, like a cake — from general to specific**.** Start with the general, then add details. For the programming task: “I am a beginner programmer. First, explain the basics of Python. Then help me debug this code [insert code]. Context: this is for a simple game application.” This will help the AI handle complex requests without overwhelming.
3. Using tags and structure. Prompts should be organized using labels. Example: “Goal: to plan a budget vacation. Constraints: $500. Suitable for family. Preferences: beach destination.” This is similar to providing artificial intelligence with a roadmap.
4. Inclusion of multimodal elements. If the request involves the use of visual elements or previous chats, a description needs to be made. Example: “Based on this image [description or link], suggest outfit options. Previous context: I prefer a casual style.” For long tasks, it is necessary to briefly summarize the history.
5. Noise Filtering. The prompt should only include what is absolutely necessary. If the AI “gets off track,” it is important to add clarification. For example: “Ignore irrelevant topics — focus only on health benefits.”
6. Accounting for past mistakes. It's important to think ahead, for example: “Last time you suggested X, but it didn't work due to Y — adjust accordingly.”

Recall that in October, a study by the University of Pennsylvania showed that large language models respond more accurately when addressed roughly.