Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Overview

Tree of Thoughts (ToT) is a framework from Google and Princeton that extends chain-of-thought prompting by enabling language models to explore multiple reasoning paths in a tree structure. It generates intermediate thought steps and uses search algorithms like breadth-first or depth-first search to evaluate and select the most promising branches. This approach allows deliberate problem solving for tasks that require planning, exploration, or backtracking.

Best for

Best for
Researchers and developers tackling hard reasoning tasks that require exploration of multiple solution paths

Use cases

• Solving complex math or logic puzzles that benefit from multiple reasoning paths
• Planning tasks like game moves or itinerary generation where sequential decisions matter
• Decision-making scenarios that require evaluating alternatives before committing to an answer

Notes

Tree of Thoughts (ToT) is a framework from Google and Princeton that extends chain-of-thought prompting by enabling language models to explore multiple reasoning paths in a tree structure. It generates intermediate thought steps and uses search algorithms like breadth-first or depth-first search to evaluate and select the most promising branches. This approach allows deliberate problem solving for tasks that require planning, exploration, or backtracking.

Use cases

• Solving complex math or logic puzzles that benefit from multiple reasoning paths
• Planning tasks like game moves or itinerary generation where sequential decisions matter
• Decision-making scenarios that require evaluating alternatives before committing to an answer

Pros

• Significantly improves reasoning accuracy on tasks where chain-of-thought fails
• Provides a structured way to explore and backtrack, mimicking human deliberate thinking
• Works with existing LLMs without fine-tuning, only requiring prompt engineering

Cons

• Higher token and compute cost due to generating and evaluating multiple thought branches
• More complex to implement than standard prompting, needs search logic and evaluation metrics
• Limited by the language model’s ability to generate useful intermediate thoughts and evaluate them correctly

Indexed from awesome-llm and enriched against its public facts.