Thought as Evaluable Unit
In Tree-of-Thoughts prompting, a “thought” is defined as a coherent language sequence that represents progress toward solving a problem.
The key innovation: thoughts can be evaluated for their contribution to the solution. The model assesses how much progress each thought represents, enabling deliberate search through the space of possible reasoning paths.
This makes reasoning decomposable and searchable:
- Generate candidate thoughts
- Evaluate each thought’s progress toward solution
- Expand promising directions
- Backtrack from unpromising ones
The thought becomes the atomic unit of a search algorithm, not just a step in a sequence, but a node in a decision tree that can be scored, compared, and abandoned.