http://arxiv.org/pdf/2201.11903

• This was back when models couldn’t do 0 shot simple arithmetic, good times 😔
• Goal is to mimic human reasoning steps when solving multi-step problems
• The paper also uses few shot prompting in experiments
  • Include reasoning steps in the answer part of the QnA format as below

• CoT seems to be an emergent capability in larger models, just didn’t work with smaller ones, gets better with scale
• The improvement of CoT is relatively robust to different linguistic styles

• NOT IN THE PAPER but: when without few shot examples, the phrase “A: let’s solve this step by step to make sure we have the correct answer” at the end helps.
  • This was mid 2023 with GPT 3-4