reCAPTCHA WAF Session Token

Unraveling the Power of Chain-of-Thought Prompting in Large Language Models

Image created by author with Midjourney


Large Language Models (LLMs) have revolutionized the field of artificial intelligence, offering unprecedented capabilities in natural language understanding and generation. However, their ability to perform complex reasoning tasks has been a subject of intense research. One technique that has shown promise in this regard is Chain-of-Thought (CoT) prompting. This article explores the intricacies of CoT prompting and its implications for the future of LLMs.

CoT prompting, as introduced in a recent paper, is a method that encourages LLMs to explain their reasoning process. This is achieved by providing the model with a few-shot exemplars where the reasoning process is explicitly outlined. The LLM is then expected to follow a similar reasoning process when answering the prompt. This technique has been found to significantly improve the model’s performance on tasks that require complex reasoning.

One of the key advantages of CoT prompting is its ability to enhance the performance of LLMs on tasks that require arithmetic, commonsense, and symbolic reasoning. The technique has been shown to yield performance gains, particularly with models of approximately 100 billion parameters. Smaller models, however, have been found to produce illogical chains of thought, leading to lower accuracy than standard prompting.


At its core, CoT prompting is about guiding the LLM to think step by step. This is achieved by providing the model with a few-shot exemplar that outlines the reasoning process. The model is then expected to follow a similar chain of thought when answering the prompt. This approach is particularly effective for complex tasks that require a series of reasoning steps before a response can be generated.

Here is a sample CoT prompt (taken from this article on Machine Learning Mastery), using a few-shot strategy:

Q: Joe has 20 eggs. He buys 2 more cartons of eggs. Each carton contains 12 eggs. How many eggs does Joe have now?
A: Joe started with 20 eggs. 2 cartons of 12 eggs is 24 eggs. 20 + 24 = 44. Therefore, Joe has 44 eggs, and the answer is 44.

Q: Farmer John had 93 chickens. If he sold 20 to Farmer Bill and bought twice that number more, how many chickens does Farmer John have now?


ChatGPT’s response:

A: Farmer John sold 20 chickens to Farmer Bill, so he was left with 93 – 20 = 73 chickens. He then bought twice the number of chickens he sold, so he bought 2 * 20 = 40 chickens. Adding these newly bought chickens to his remaining stick, Farmer John now has 73 + 40 = 113 chickens.


CoT prompting can also be used in a zero-shot setting. This involves adding a phrase like “Let’s think step by step” to the original prompt, which can also be made use of alongside few-shot prompting. This simple addition has been found to be effective at improving the model’s performance on tasks where there are not many examples to use in the prompt.


While CoT prompting can be effective, it often involves hand-crafting examples, which can be time-consuming and may lead to suboptimal solutions. To address this, researchers have proposed an approach known as Automatic Chain-of-Thought (Auto-CoT). This method leverages LLMs to generate reasoning chains for demonstrations automatically, thereby eliminating the need for manual effort.

Auto-CoT consists of two main stages: question clustering and demonstration sampling. In the question clustering stage, questions from a given dataset are partitioned into a few clusters. In the demonstration sampling stage, a representative question is selected from each cluster, and its reasoning chain is generated using Zero-Shot-CoT with simple heuristics. This process encourages the model to use simple and accurate demonstrations.

The Auto-CoT process:

  1. Question clustering: Partition questions of a given dataset into a few clusters
  2. Demonstration sampling: Select a representative question from each cluster and generate its reasoning chain using Zero-Shot-CoT with simple heuristics


While CoT prompting has shown promise, it is not without its limitations. For one, it only yields performance gains when used with models of approximately 100 billion parameters. Smaller models tend to produce illogical chains of thought, leading to lower accuracy than standard prompting. Furthermore, the performance boosts from CoT prompting are generally proportional to the size of the model.

Despite these limitations, CoT prompting represents a significant step forward in the quest to enhance the reasoning capabilities of LLMs. Future research will likely focus on refining this technique and exploring ways to make it more effective across a wider range of tasks and model sizes.


Chain-of-Thought prompting represents a significant advancement in the field of artificial intelligence, particularly in enhancing the reasoning capabilities of Large Language Models. By encouraging these models to explain their reasoning process, CoT prompting has shown promise in improving performance on complex tasks that require arithmetic, commonsense, and symbolic reasoning. While the technique has its limitations, it opens up exciting possibilities for the future of LLMs.

As we continue to push the boundaries of what LLMs can achieve, techniques like CoT prompting will play a crucial role. By enabling these models to think step by step and explain their reasoning, we can not only improve their performance on complex tasks but also gain valuable insights into their inner workings. The journey towards fully reasoning LLMs is still long, but with techniques like CoT prompting, we are certainly on the right path.

Matthew Mayo (@mattmayo13) is a Data Scientist and the Editor-in-Chief of KDnuggets, the seminal online Data Science and Machine Learning resource. His interests lie in natural language processing, algorithm design and optimization, unsupervised learning, neural networks, and automated approaches to machine learning. Matthew holds a Master’s degree in computer science and a graduate diploma in data mining. He can be reached at editor1 at kdnuggets[dot]com.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *

WP Twitter Auto Publish Powered By :