RE2+CoT: The Prompting Technique That Improves AI Accuracy(+ demo you can try)

Today I found out about a new prompting technique. Asking LLM to re-read the questions boosts the quality of its answers by ~5% on top of the Chain-of-Thought technique. I even made a CustomGPT to demo this technique for you!

AI use disclosure: ChatGPT was used to help structure and format my thoughts and materials

The Research Behind the Technique

Figure from original paper

Do not have a Medium Account? Read here!

A recent research paper introduced this innovative approach. The idea is straightforward: by prompting the AI to re-read the question, it enhances its understanding and reasoning capabilities. Here’s the prompt structure they used:

💡 Prompt structure RE2+CoT:
Q: {Input Query}
Q: Read the question again: {Input Query}
Q: Thought-eliciting prompt (“Let’s think step by step”)

This method was tested on several language models, including davinci-003, ChatGPT, and LLaMA-2 (both 13B & 70B parameters). The experiments spanned across 14 different datasets covering: