Graph of Thoughts Explained!

After Chain-of-Thought, Tree-of-Thought we now have Graph-of-Thoughts. Prompting LLMs is an essential technique that helps us get the most out of LLMs without changing the weights.

Graph-of-thoughts Illustration created by StabilityAI

A gentle reminder for people who don’t know what a prompt is, LLM like ChatGPT, Claude, LLama and you name it are decoder-only based models.

They learn to model statistically a language representation, this means that given one word or two, the model gives you next most probable word based on it’s training data and techniques.

A prompt is simply the prefix that you give to a language model, when you open https://chat.openai.com/ and write something that is a prompt.

Now referring to what we said above, what you specify in the prompt has a great impact on how the model chooses the next tokens.

If you think about it, we humans also behave the same. For example if you ask me to write a blog, there are many ways to do it.

1. “WRITE A BLOG FOR ME”
2. “Can you please write a blog for me”
3. “Can you please write a step by step blog for me and ask for my approval after every paragraph”

The answer to every possibility is different. This has resulted in Prompt Engineering.

People have started playing around with LLMs and discovered patterns and ways to improve these prompts. We’ll go over the famous ones.

Chain-of-Thoughts: The idea is to show the model how to think step by step in examples, this makes the model use more tokens and more probable to give the correct answer. To do this there are two ways, you either provide examples or you add “Let’s think step by step”

Chain-of-Thoughts-Self-Consistency: Same as Chain-of-Thoughts, but you generate many chain of thoughts and you finally pick the best one.

Tree-of-Thoughts: Tree of thoughts allows the possibility to back track from one thought to another or explore that thought even further.

Graph-of-Thoughts: Expands on Tree of Thoughts by representing thoughts as Direct Acyclic Graphs, where you could loop over one thought or aggregate two thoughts or more together to form a better one.