Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Paper-reading notes: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

October 20, 2025 · 314 words

Table of Contents

Two simple methods to unlock the reasoning ability of LLMs
What this paper do
Why this method is important
Result
Related work

Two simple methods to unlock the reasoning ability of LLMs

Thinking in steps helps:
When the model explains each step before the answer, it understands the problem better.
Learning from examples:
When we show a few examples with step-by-step answers, the model learns to do the same.

What this paper do

Combine these two ideas
- help the language models think step by step to generate a clear and logical chain of ideas that shows how they reach the final answer.
Given a prompt that consists of triples: <input, chain of thought, output>

Why this method is important

It doesn’t need a big training dataset.
One model can do many different tasks without extra training.

Result

It only works well for giant models, not smaller ones.
It only works well for more-complicated problems, not the simple ones.
Chain-of-thought prompting with big models gives results as good as or better than older methods that needed finetune for each task.

Some methods make the input part of the prompt better — for example, adding clearer instructions before the question.
But this paper does something different (orthogonal): it improves the output part, by making the model generate reasoning steps (chain of thought) before giving the final answer.