Tree Of Thought: Definition and Examples

Tree of Thought (ToT) is a prompting method inspired by human problem-solving processes. Unlike classic Chain of Thought, which follows a linear step-by-step reasoning, ToT allows the model to simultaneously explore multiple lines of thought, evaluate them, and backtrack if a branch proves unproductive.

Specifically, the model generates multiple intermediate 'thoughts' at each reasoning step, forming a tree of possibilities. Each node represents a partial state of reasoning, and the model can evaluate the promise of each branch before deciding which to pursue. This exploration and self-evaluation mechanism mirrors the human ability to consider different approaches to a complex problem.

This technique was formalized in 2023 by researchers at Princeton and Google DeepMind (Yao et al.) in their paper "Tree of Thoughts: Deliberate Problem Solving with Large Language Models." Their work demonstrated significant improvements on tasks requiring planning, strategic search, and decision-making, such as the 24 game, creative writing, and logic puzzles.

ToT can be implemented in two main ways: through carefully crafted prompts that explicitly ask the model to explore multiple branches, or via an external framework that orchestrates model calls using search algorithms like BFS (breadth-first search) or DFS (depth-first search). This flexibility makes it a powerful tool for tasks where a direct answer would be insufficient.