Tree of Thoughts: the LLM that reasons by exploring alternative branches
In one sentence Princeton and DeepMind propose Tree of Thoughts: the LLM generates and evaluates multiple reasoning paths as a search tree, clearly outperforming Chain-of-Thought.
Chain-of-Thought (CoT) helps models reason step by step, like a student showing their work. But it has a flaw: it always picks the first path that seems right, never backtracking if it's wrong.
Tree of Thoughts fixes this by letting the model explore multiple paths simultaneously, like a tree with many branches. At each step it evaluates which branch is most promising — pruning included — and can backtrack if a branch hits a dead end.
On problems requiring planning and backtracking — like Game of 24 or creative writing with precise constraints — ToT clearly beats standard CoT, jumping from 4% to 74% success on Game of 24.
Companies
Princeton University, Google DeepMind
Tools
Tree of Thoughts, GPT-4
Tags
Sources