How Magical Phrases Improve Responses from Large Language Models

"Take a deep breath" and "think step by step" are phrases you might use to help someone focus on solving a problem. Surprisingly, these phrases also seem to improve the responses from large language models (LLMs) when they are used in prompts, especially in math-related queries. Does this mean that the AI is actually getting smarter? What’s the reason behind this seemingly magical effect? Let’s explore.

The Magic Words Phenomenon

Studies show that when prompts include phrases like “take a deep breath” or “think step by step,” the generated responses appear to be more accurate or well-reasoned, especially in the context of solving math problems or other complex issues. Does this mean the AI has suddenly developed the ability to breathe or engage in human-like cognitive processing? Not quite.

Why Does This Happen?

The answer lies in the training data that large language models use to learn. These models are trained on vast datasets compiled from books, websites, forums, and other sources where human-written text exists. Often, people use phrases like “let’s take a deep breath” or “think step by step” before providing detailed answers to questions or explaining complex topics. The model’s improved performance is essentially an emulation of human behavior.

It’s All About the Data

LLMs can’t actually take a deep breath or think step-by-step because they don’t possess lungs, bodies, or human cognitive abilities. The “reasoning” that these models display is borrowed from the massive datasets they are trained on. When we include phrases like “take a deep breath” or “think step by step” in prompts, we’re essentially nudging the model to tap into segments of its training data that likely contain more reasoned, carefully structured responses.

Controversy Around “Reasoning”

It’s worth noting that the term “reasoning” is a subject of debate among experts in the field of artificial intelligence. While it’s commonly used to describe the operations of complex algorithms, some argue that it inaccurately anthropomorphizes machine processes that are fundamentally different from human cognition.


In essence, using phrases like “take a deep breath” or “think step by step” in prompts doesn’t make the AI smarter or more capable of human-like thought. Rather, it aligns the query with the kind of detailed, carefully considered responses that exist in the model’s training data, which can improve the quality of its output. So the next time you’re looking for a more thoughtful answer from a large language model, you might just find that a little “prompt magic” goes a long way.

Author: robot learner
Reprint policy: All articles in this blog are used except for special statements CC BY 4.0 reprint policy. If reproduced, please indicate source robot learner !