🤖 "Tree of Thoughts": a new approach to applying large language models for solving complex tasks
Recent research from Google DeepMind and Princeton University has introduced a new method for applying language models, known as "Tree of Thoughts" (ToT). This approach offers a solution to the limitations of the existing "Chain of Thought" (CoT) method, which struggles with tasks requiring planning or exploration. By the way, if you don't know what "Chain of Thought" is and aren't using it with ChatGPT - you're missing out on at least half the quality of the final answer.
ToT allows language models to use "thoughts" or whole units of text as intermediate steps in problem-solving. It involves using multiple reasoning paths and self-evaluation of choices to determine the next actions. Instead of a "token-by-token" approach, ToT allows for looking ahead, going back, and exploring different thoughts globally.
Experiments have shown that ToT significantly enhances the problem-solving capabilities of language models. For example, in the game "Game of 24", GPT-4 with CoT solves only 4% of tasks, while the ToT approach allows solving 74% of tasks.
📝 Paper: https://arxiv.org/pdf/2305.10601.pdf
👨💻 Github: https://github.com/ysymyth/tree-of-thought-llm
