Tree of Thoughts: Deliberate Problem Solving


Paper: Yao, Yu, Zhao, Shafran, Griffiths, Cao, Narasimhan, Tree of Thoughts: Deliberate Problem Solving with Large Language Models, NeurIPS 2023. arXiv:2305.10601

Why this paper matters

Chain-of-Thought is a single left-to-right path. If it takes a wrong step early, it can never recover — there is no lookahead and no backtracking. Tree of Thoughts (ToT) reframes reasoning as a search over a tree of partial solutions, so the model can explore, evaluate, and backtrack. This is the difference between “guessing the next word” and “actually solving.” It is a key ingredient in the “agentic reasoning” trend (and echoes the search behind o1-style models).

The core idea: thoughts as search nodes

The final answer is read from the best leaf found.

BFS variant in one line

At each depth, keep only the b highest-valued states (e.g., b = 5); expand those; repeat. This systematically explores before committing — unlike CoT, which commits instantly.

Key results

Why it matters today

ToT is deliberate problem solving: it adds lookahead and backtracking on top of the ReAct loop (acting) and Reflexion (learning from failure). When an agent faces a task with a verifiable goal — math, code, puzzles, multi-step planning — a search over thoughts beats a straight shot. Modern “reasoning models” are, in spirit, ToT at scale with a learned verifier.

References

© 2026 Jiawei    粤ICP备18035774号