Learn PromptingZero Shot Chain of Thought
Zero Shot Chain of Thought (Zero-shot-CoT) prompting is a follow up to CoT prompting, which introduces an incredibly simple zero shot prompt. They find that by appending the words "Let's think step by step." to the end of a question, LLMs are able to generate a chain of thought that answers the question. From this chain of thought, they are able to extract more accurate answers.
Technically, the full Zero-shot-CoT process involves two separate prompts/completions. In the below image, the top bubble on the left generates a chain of thought, while the top bubble on the right takes in the output from the first prompt (including the first prompt itself), and extracts the answer from the chain of thought. This second prompt is a self augmented prompt.
Full Zero Shot CoT Process (Kojima et al.)
Example
Here are a few demos (which only perform reasoning extraction). This first demo shows GPT-3 (davinci-003) failing a simple math question, while the second demo uses a Zero-shot-CoT prompt and successfully solves the problem. Feel free to enter your OpenAI API key (Click Generate) and play around with the examples. Note how much simpler the Zero-shot-CoT prompt is compared to the CoT prompt.
Incorrect
Correct
Results
Zero-shot-CoT was also effective in improving results on arithmetic, commonsense, and symbolic reasoning tasks. However, unsurprisingly, it was usually not as effective as CoT prompting. An important use case for Zero-shot-CoT is when obtaining few shot examples for CoT prompting is difficult.
Ablations of Interest
Kojima et al. experiment with a number of different Zero-shot-CoT prompts (e.g. "Let’s solve this problem by splitting it into steps." or "Let’s think about this logically."), but they find that "Let's think step by step" is most effective for their chosen tasks.
Notes
The extraction step often must be task specific, making Zero-Shot-CoT less generalizable than it appears at first.
Anecdotally, I've found that Zero-shot-CoT style prompts are sometimes effective
in improving the length of completions for generative tasks. For example, consider
the standard prompt Write a story about a frog and a mushroom who become friends.
Appending the words Let's think step by step. to the end of this prompt leads to
a much longer completion.
Sander Schulhoff
Sander Schulhoff is the Founder of Learn Prompting and an ML Researcher at the University of Maryland. He created the first open-source Prompt Engineering guide, reaching 3M+ people and teaching them to use tools like ChatGPT. Sander also led a team behind Prompt Report, the most comprehensive study of prompting ever done, co-authored with researchers from the University of Maryland, OpenAI, Microsoft, Google, Princeton, Stanford, and other leading institutions. This 76-page survey analyzed 1,500+ academic papers and covered 200+ prompting techniques.