🌟 Prompting Techniques That Outperform Cutting-Edge Reasoning AI Models by 10x

Hello everyone!

Reasoning models are very popular lately, and it's quite impressive, isn't it? Like OpenAI's o1 or o3, or Claude's extended thinking mode.

These reasoning models have a special type of prompting called CoT (Chain of Thought) built into them.

However, what would you think if there were a prompting technique that could outperform this by about 10 times?

That is what I will talk about this time.

If you master this prompting method, you will be able to elicit surprisingly high-quality answers from the AI you are currently using, without needing expensive, state-of-the-art models!

This is also an important perspective from the viewpoint of the democratization of AI.

[Related Video 1]

[Related Video 2]

🚀 Astonishing 10x Power! What is the Prompting That Overwhelms CoT?

First, to state the conclusion, the prompting technique that outperforms CoT (Chain of Thought) by 10 times is ToT (Tree of Thought). This ToT is a relatively newly developed technique. Therefore, I believe its potential is immeasurable.

Unbelievable! Demonstrating 2 to 24 times the strength in search tasks

Around May 2023, the Tree of Thought (ToT) prompting technique, announced by a joint research team from Google DeepMind and Princeton University, showed performance that significantly exceeded the conventional Chain of Thought (CoT).

What was particularly shocking was the result that for a problem that GPT-4 could only solve 4% of the time with CoT, it achieved a 74% success rate with ToT. That is a performance improvement of over 18 times by simple calculation! On average, ToT is considered to have about a 10-fold advantage over CoT in tasks that require searching.

"Wait, is it really that different?"

You might be thinking that. But this is not an exaggeration; it is an actual research result!

I have experimented with it many times myself, and there were problems that I could not reach the answer to no matter how hard I tried with CoT, but could easily solve using ToT. I couldn't help but think, 'Why didn't I know about this method sooner?'

When I say 10 times, it is just me personally stating that because the difference is about 2 to 18 times, so I took the middle ground, but I think there is roughly that much difference in results.

Tree of Thought recognized by the world's top research team

This revolutionary technique was born from a research team at Google DeepMind and Stanford University, which are at the forefront of AI. They observed human thought processes and focused on the obvious fact that 'people don't think in a straight line when solving difficult problems, do they?'

Humans think about multiple possibilities simultaneously and, when they hit a dead end, rethink from a different angle. ToT is an implementation of this natural thinking mechanism in AI. This can truly be called a bridge between AI and human thought!

🤔 "Single Path" vs. "Exploring the Forest"—Why ToT is a Game Changer

The Pitfall of CoT: Once You Go Off the Rails, You Can't Go Back!

The conventional Chain of Thought (CoT) prompting method, or "think step-by-step," is certainly convenient, but it has a major weakness. That is... once you head in the wrong direction, you can't get out of it!

For example, when thinking about the difference between "sawayaka" (refreshing) and "sugasugashii" (refreshing/pure), if you establish the false premise in CoT that "both are words for the same season," no matter how much you think afterward, you will never reach the correct answer. Even if you ask "Anything else?", the AI will only look for other angles based on that false premise and will not notice the fundamental error.

This is similar to when we get lost and think, "This must be the right way!" and keep going deeper, isn't it?

The Appeal of ToT (Tree of Thoughts): The Flexibility to Explore Freely Anywhere

On the other hand, in Tree of Thoughts (ToT), thinking "branches out," so if one path is a dead end, you can try another route. Furthermore, with a technique called "backtracking," if you realize "this path is wrong," you can turn back and explore a different path.

This is similar to the freedom of spreading out a map when exploring a forest, considering multiple routes, and trying a different path if you hit a dead end. When thinking about "sawayaka," you can explore from multiple perspectives simultaneously, such as "seasonal feeling," "physical sensation," and "linguistic usage," so the possibility of finding more essential differences increases.

This flexible exploration capability is the secret to why ToT overwhelms CoT!

🧠 The Foundation of Thought Branches—Not Getting Lost with Higher-Level Concepts

Thinking Under a Common Umbrella—The Magic of Lower-Level Concepts and Partial Vector Spaces

To master ToT (Tree of Thoughts), you first need a map of "higher-level concepts." For example, if you are comparing an "apple" and a "strawberry," the higher-level concept is "fruit."

Under this umbrella of "fruit," there are characteristics such as "color," "shape," "size," and "growing environment." These are called "partial vector spaces," and they are like a coordinate system for comparison. If you want to know the differences, you won't get an answer unless you compare these.

When you firmly set higher-level concepts, lower-level concepts are relativized, making the objects of comparison clear and helping you avoid irrelevant comparisons like "apple" and "car." Our thinking also becomes organized, and we can make more meaningful comparisons.

"Understanding" by "Dividing"—The Power of Comparing Partial Vector Spaces

I have long said that "to understand is to divide." True understanding is born from grasping the commonalities and differences between things.

In ToT (Tree of Thoughts), we clarify subtle differences between concepts through the comparison of partial vector spaces. For example, when thinking about the difference between "ozanari" (perfunctory) and "naozari" (neglect), if you compare them from the perspective of "presence of action," you can see the essential difference: "ozanari" involves action even if it is formal, whereas "naozari" means the action itself is abandoned.

AI processes this as a vector representation and explains it to us in an easy-to-understand way. You should be able to have an "Aha! So that was the difference!" moment. For AI, the partial vector spaces where "ozanari" and "naozari" exist are perceived as completely different things.

I think this kind of AI co-creation learning will become very important in the future. Perhaps ToT might even develop in the field of education.

💡 Eye-opening examples—A new world revealed by ToT (Tree of Thoughts)

Understanding through familiar examples—A new perspective on strawberries and apples

To understand the basics of ToT (Tree of Thoughts), it is recommended to start with familiar examples. You know the difference between a "strawberry" and an "apple," right?

However, when displayed in a ToT tree structure, concepts like color, shape, and size branch out under the higher-level concept of "fruitiness," leading to a more systematic understanding. In terms of "redness," strawberries are more vivid, while in terms of "roundness," apples are uniformly round, whereas strawberries are pointed at the bottom—the differences become clear.

Don't you feel like, "Wow, even with things I take for granted, organizing them like this leads to new discoveries"?

Unraveling subtle nuances of language—The world of 'ozanari' and 'naozari'

As an example of the depth of the Japanese language, let's look at the difference between "ozanari" and "naozari."

It is difficult to explain these subtle differences with ordinary prompts or CoT, but using ToT, the essential difference becomes visible from the perspective of "how one faces a subject."

"Ozanari" refers to a state of doing something just for the sake of form, while "naozari" means abandoning the action itself, containing a serious message of neglect or indifference. "Ah, so that's why the impression is so different when using these words," it's useful for choosing the right words, isn't it?

Exploring the secrets of seasonal words—The cultural difference between 'sugasugashii' and 'sawayaka'

An even more interesting example is the difference between "sugasugashii" (refreshing) and "sawayaka" (fresh/breezy). They are often used with similar meanings in daily conversation. In fact, the vectors of both in daily conversation are very close, but they actually have quite different partial vector spaces.

ToT unraveled the difference that normal prompts or CoT couldn't notice from the perspective of "seasonal symbols." It turns out that "sugasugashii" is a summer seasonal word, while "sawayaka" is an autumn seasonal word!

This might be common knowledge in the world of haiku and literature, but it's a fresh discovery for the rest of us. With ToT, even such cultural and literary differences can be brought to light.

🛠️ You can use this starting today! How to create ToT prompts

Growing a tree of thought—The power of tree-format display

The most important thing in practicing ToT (Tree of Thoughts) is "displaying in a tree format." In this format, a lot of information can be organized compactly, making exploration and comparison much easier.

Explicitly state "output in a tree format" in your prompt, and adding "everything in a single tree format" makes it easier to organize your thoughts. Since I switched to this format, the quality of my dialogue with AI has improved dramatically!

An explorer's mindset—Magic words to switch AI to exploration mode

ToT (Tree of Thoughts) shows its true value in tasks that require "exploration." You can turn on the AI's exploration mode with a simple word: "explore."

This is similar to trying various possibilities when solving a maze, such as "What happens if I go right from here?" or "What about the left?" With this simple but effective instruction, the AI will think from more diverse perspectives.

Compare to Deepen—Enhancing the Quality of Thought with Evaluation Criteria

To evaluate search results, the prompt "compare and evaluate" is effective. Furthermore, specifying concrete criteria such as "pros and cons," "strengths and weaknesses," or "advantages and disadvantages" allows for a more detailed comparison.

My favorite is the prompt "rate the degree of difference from 1 to 10." When you include this, the AI quantifies the extent of the differences, making it easy to see at a glance what is truly important. By extracting only those with a "difference score of 9 or higher," you can focus solely on the essential differences.

The Five Magics—Optimizing Thought with Magic Numbers

The magic number to remember when implementing ToT (Tree of Thoughts) is "5." Research suggests that 5 branches of thought are the most efficient. This is based on prior studies showing that a beam width of "5" optimizes both accuracy and efficiency.

Furthermore, as for why it is 5: if there are too few branches, possibilities are limited, and if there are too many, quality drops and computational costs increase. The number 5 strikes the perfect balance between exploration and exploitation.

By specifying "create 5 branches" in your prompt and setting the number of steps to "3," you complete the magic combination of a 5x3 tree of thought. Just remembering this number will significantly improve the quality of your dialogue with AI.

This magic number of 5x3 is very important for tasks like brainstorming, so I have covered it in a past article. Please take a look if you'd like.

🎯 Toward a New Era of AI Utilization—Maximizing the Potential of ToT (Tree of Thoughts)

Paving the Way for the Future—Essential Exploration Capabilities for AI Agents

Tree of Thoughts (ToT) is a technique that significantly enhances "exploration capability," which is a weakness of current reasoning models. It will likely have a major impact on future AI agent development as well.

For tasks known to AI agents, I think current CoT (Chain of Thought) can handle them, but it might be difficult for unknown tasks. That is why I believe ToT will eventually be incorporated into AI agent prompting.

If future AI agents are integrated ToT x CoT systems that cost hundreds of thousands per month, you might be able to compete by incorporating this method without paying for those services. Well, we won't know until the future arrives, but it's a technique worth remembering!

Wisdom in Usage—Prompt Selection Techniques Tailored to Tasks

That said, ToT is not optimal for every situation. For tasks that require digging deep into a single point or problems that can be solved with a single path of thought, a focused, breakthrough-style CoT may be more suitable.

The important thing is to choose your method of thinking according to the nature of the problem. By using ToT when exploration is needed and CoT when exploitation or deep diving is required, you will be able to elicit the best answers from AI. Tools truly show their worth depending on how you use them.

Combining Two Powers—Infinite Possibilities through the Fusion of ToT and CoT

In the future, an approach that combines ToT and CoT might become the strongest. First, explore broadly with ToT, and once a promising path is found, dig deeper with CoT... If such a combination becomes possible, we will be able to solve even more complex problems. I have introduced this kind of prompting before as a breakthrough prompt, so please take a look if you are interested.

Dialogue with AI is still in its infancy. Our individual ingenuity and experiments are what expand the possibilities of AI. Let's open the new door of ToT (Tree of Thoughts) and make our interactions with AI even richer!

See you later!

[Profile]
Wonder Motohiko Sato
Organizer of MBBS & AI Co-Creation Innovation.
After working at medical and psychological research institutes, I became independent and am currently researching AI and the mind-body connection.
Author of "Oriental Medicine and the Latent Motor System," I have been developing AI co-creation writing while working on writing projects such as a two-year series in a professional journal and academic papers.
I am developing AI Co-Creation Studies by applying psychology, counseling, and coaching techniques to AI.
I conduct AI schools, corporate AI training, and AI application development.

✅ note articles (including articles linked to YouTube)
https://note.com/mbbs

✅ AI Co-Creation Innovation
https://mbbs-ai.jimdofree.com/

✅ Facebook
https://www.facebook.com/motohiko1977

✅ Motohiko Sato Comprehensive Links
https://linktr.ee/motohiko.sato