Tree of Thoughts

What it is about:

Tree of Thoughts (ToT) is a framework designed to improve the problem-solving abilities of large language models (LLMs) for complex tasks. Traditional prompting techniques often struggle with tasks that require exploration, strategic planning, and considering multiple possibilities. ToT addresses this limitation by encouraging the LLM to explore a "tree" of potential solutions in a step-by-step manner.

How it works:

Building the Tree:

Imagine a tree structure where each branch represents a potential solution path. ToT guides the LLM to generate "thoughts" - coherent language sequences that act as intermediate steps towards solving the problem. These thoughts become nodes on the tree. Here, the LLM might leverage its existing knowledge and understanding of the task to generate initial thoughts.

Self-Evaluation:

The LLM doesn't blindly explore every branch. Instead, it evaluates each thought based on its likelihood of leading to the desired solution. This evaluation might involve:

• Relevance: Does the thought address the current problem or sub-problem?
• Feasibility: Is the thought a realistic or achievable step towards the solution?
• Promise: Does the thought seem like a promising path to explore further?

This self-evaluation process allows the LLM to prioritize its exploration and focus on the most relevant and promising branches.

Systematic Exploration:

ToT doesn't explore the tree randomly. It utilizes search algorithms like:

• Breadth-First Search (BFS): Explores all the nodes at a given level before moving to the next level. This ensures a comprehensive exploration of all possibilities at each stage.
• Depth-First Search (DFS): Chooses a branch and explores it as deeply as possible before backtracking and exploring other branches. This can be useful for finding solutions quickly if one branch leads to the desired outcome.

The choice of search algorithm depends on the specific task and the desired balance between comprehensiveness and efficiency.

Refining the Path:

Based on the self-evaluation and search algorithm, the LLM focuses on the most promising branches. It might explore sub-branches of these promising thoughts, generating additional intermediate steps to refine the solution path. This process of exploration and refinement continues until the LLM reaches a satisfactory solution or determines there are no viable options.

Examples:

Writing a Persuasive Essay:

Imagine you're using ToT to write an essay arguing for renewable energy sources.

• Thought 1: "The environmental impact of fossil fuels is a major concern." (Evaluated as promising)
  • Sub-thought 1.1: "Provide specific examples of environmental damage caused by fossil fuels." (Further explores thought 1)
• Thought 2: "Renewable energy sources are becoming more affordable." (Evaluated as promising)
  • Sub-thought 2.1: "Compare the cost of solar panels today to their cost 10 years ago." (Further explores thought 2)

By evaluating and exploring these intermediate thoughts, the LLM can develop a strong and well-supported argument for its essay.

Solving a Math Problem:

Let's say you're using ToT to solve a complex math equation.

• Thought 1: "Factor the equation to simplify it." (Evaluated as promising)
  • Sub-thought 1.1: "Identify common factors in the equation." (Further explores thought 1)
• Thought 2: "Apply a specific mathematical formula to solve the equation." (Evaluated as maybe)
  • Sub-thought 2.1: "Determine if the formula applies based on the equation's structure." (Refines thought 2)

When to use it:

Tree of Thoughts is particularly helpful for tasks that involve:

• Complex Problem Solving: When the task requires breaking down a problem into smaller steps and considering multiple approaches.
• Strategic Planning: When the solution requires looking ahead and anticipating potential consequences of different choices.
• Justification and Explanation: When the LLM needs to explain its reasoning or justify its solution steps.