Tree of Thoughts

Tree-of-Thoughts (ToT) makes the model explore multiple reasoning paths and pick the best — like a chess engine searching a move tree. The cost: for 3 branches × 3 depth = 9 inference calls per response, a 9× increase in compute and latency. It's impractical for real-time apps but valuable for offline decision support where correctness matters more than speed.

Use ToT for high-stakes decisions: (1) model generates 3 candidate solutions, (2) for each, generate 3 next steps, (3) evaluate all 9 paths with a scoring prompt ("Rate this solution on correctness, completeness, safety 1–10"), (4) keep best-scoring path, repeat from step 2 until depth limit, (5) output best complete path. Best suited for math proofs, code architecture, or legal analysis where 30s latency is acceptable.