Tree of Thought (ToT) is a reasoning technique that organizes language-model problem solving into multiple coherent paths. The model can explore alternatives, evaluate them, refine promising options, and backtrack instead of committing token by token. section 2.5.6
ToT is useful when a task requires deliberate planning rather than a single direct answer:
- Explore multiple reasoning paths before choosing an approach. section 2.5.6
- Evaluate and refine intermediate options as the solution develops. section 2.5.6
- Backtrack when an earlier choice leads to a poor result. section 2.5.6
- Apply the approach to strategic tasks such as Game of 24, creative writing, and mini crosswords. section 2.5.6
Game of 24
ToT helped GPT-4 solve 74% of Game of 24 problems, compared with 4% using Chain of Thought, illustrating the potential value of exploring and revisiting alternatives. section 2.5.6
Table 2.2 In-context learning prompts
|
In-context learning technique
|
Explanation
|
|---|---|
|
Zero-shot learning
|
No example is provided in the prompt.
|
|
One-shot learning
|
One example is provided in the prompt.
|
|
Two-shot learning
|
Two examples are provided in the prompt.
|
|
Few-shot learning
|
A number of examples are provided in the prompt.
|
|
Chain of Thought (CoT)
|
A number of examples are provided in the prompt, and for each example, all the logical steps to achieve an objective are clearly explained.
|
Compare ToT with Chain of Thought: the prompt reference defines CoT through examples whose logical steps are explicitly explained. section 2.5.6
