Tree of Thoughts: Deliberate Problem Solving with Large Language Models
- Thesis: Inspired by cognitive science and AI in terms of how humans and animals approach problem solving, Tree of Thoughts (ToT) make use of system 1 & 2 by using LM to generate different possible paths (thoughts) and evaluate each path in making progress towards solving the problem in the form of of tree. The search for the best possible answer is done through either breadth-first search (BFS) or depth-first search (DFS), which helps in lookahead and backtracking
- Methods:
- Thought decomposition: Decompose intermediate thought steps. Each step must be big enough and coherent
- Thought generator: Use CoT prompt to generate
ksamples and then propose thoughts sequentially using propose prompt - State evaluator: Value each state (node in a tree) independently using value prompt via lookahead simulations and commonsense. Finally vote across states which one would lead to the most promising solution using vote prompt
- Search: Use any of the search algorithms depending on the nature of the tree: BFS or DFS
- Contribution:
- ToT framework
- Importance: ToT framework has many benefits such as 1) generality to any problem solving, 2) modularity where each stage of the framework can be changed independently, 3) interpretability/debugging, 4) no need to fine-tuning for different tasks
- Takeaways: ToT significantly improves performance of LLMs over CoT or self-consistency with CoT due to the fast that it mimics how problem-solving is carried in humans/animals using insights from cognitive science. Utilizing exploration and lookahead, ToT can generate and evaluate different paths at any given step.
- Improvements:
- It is tested on three simple tasks, so we need more diverse set of tasks to check for applicability and generalizability
- It is not needed and an overkill for tasks that current LLMs excel at
- It is much more compute intensive than other sampling approaches
#nlp #llm #agents