First Published 22 June 2023

Background: Introduction to Tree of Thought (ToT)

First, we have to understand some limitations of llms. Language models, such as ChatGPT, have proven their prowess in generating text with impressive speed and accuracy. However, their effectiveness has historically been limited when it comes to complex, contemplative thinking that characterizes human “System 2” cognition. Tasks requiring exploration, strategic lookahead, or nuanced decision-making often challenge these models. That’s where the Tree of Thoughts comes into play. However, by incorporating the principles of deliberate, logical thinking, ToT offers a new approach that expands the horizons of AI capabilities. In this blog post, we will delve into the concept of Tree of Thoughts and explore its implications in enhancing AI models like ChatGPT.

Tree of Thoughts

Development:

Developed by researchers from Princeton University and Google DeepMind, the Tree of Thoughts framework offers a departure from traditional approaches in language model inference. It builds upon the popular “Chain-of-Thought” (CoT) strategy by introducing a tree structure that allows models to navigate through coherent units of text or thoughts. This tree-based organization enables exploration, self-evaluation, and a more deliberate decision-making process. Also, for more info, see a paper that I will be referring to here: link.

Benefits of ToT:

The Tree of Thoughts technique allows for a hot, new form of AI problem-solving. By incorporating this approach, AI models can overcome the limitations of quick, instinctive responses associated with “System 1” thinking. Here are some key benefits of the Tree of Thoughts approach:

First, preliminary results have demonstrated remarkable success. In tasks like the mathematical card game “Game of 24,” ChatGPT using Tree of Thoughts achieved a success rate of 74%, a significant performance improvement(1) over traditional CoT prompting. Next, ToTs helps AI models tackle challenges that demand high-level planning and creative thinking(2). When writing creatively, ChatGPT generated coherent passages preferred by humans, showcasing its ability to engage in sophisticated reasoning. Lastly, AI models like ChatGPT can evolve into general-purpose problem solvers(3) using ToT. Through the infusion of conscious, deliberative thinking of humans, they become more adept at exploring multiple options and making informed decisions.

References:

https://arxiv.org/abs/2305.10601