EP059: Tree of Thoughts Unlocks System 2 Thinking

The paper introduces "Tree of Thoughts" (ToT), a novel framework designed to enhance the general problem-solving capabilities of Large Language Models (LLMs).Traditionally, LLMs generate text through a left-to-right, token-level decision-making process, which is similar to fast, automatic "System 1" human cognition. This standard mechanism often falls short in complex tasks that require strategic lookahead, exploration of alternatives, or backtracking. To overcome these limitations, the ToT framework augments LLMs with deliberate, conscious "System 2" planning.ToT generalizes the popular "Chain of Thought" prompting approach by framing problem-solving as a search over a combinatorial problem space represented as a tree. The core mechanics of this framework include:Thoughts as Nodes: The problem-solving process is decomposed into coherent units of text called "thoughts," which serve as intermediate steps toward a final solution.Generation and Evaluation: The LM generates multiple potential thoughts from a given state and self-evaluates the progress each thought makes toward solving the problem.Search Algorithms: ToT incorporates classical search algorithms, such as Breadth-First Search (BFS) and Depth-First Search (DFS), allowing the model to systematically explore different reasoning paths, look ahead, and backtrack to previous states when necessary.The authors demonstrated the effectiveness of ToT through experiments on three challenging tasks that require non-trivial planning, search, and reasoning: Game of 24, Creative Writing, and Mini Crosswords. The results show that ToT significantly outperforms existing prompting methods. For instance, in the Game of 24 mathematical reasoning challenge, GPT-4 with standard chain-of-thought prompting only achieved a 4% success rate, whereas the ToT method achieved a 74% success rate.