What is it?
A normal answer follows one path. Search-based reasoning tries several paths, rates how promising each partial attempt looks, drops weak ones and continues with strong ones — like exploring a maze rather than walking one corridor.
It can solve harder puzzles but needs a trustworthy way to judge half-finished work, and it uses a lot of compute.
Also known as: tree of thoughts, ToT, MCTS for LLMs, Monte Carlo tree search reasoning, beam search reasoning, step-level search, graph of thoughts, deliberate search.
Technical definition
Methods include Tree of Thoughts (Yao et al., 2023), beam search over reasoning steps, and Monte Carlo tree search guided by value or process reward estimates (Process Reward Model (PRM)). Components: a proposal policy (the LLM), a state evaluator (LLM self-evaluation, learned value/PRM, or programmatic check), and a search controller (BFS/DFS/beam/MCTS). Cost depends on branching and depth; benefit depends heavily on evaluator accuracy.
Entity type: technique · Freshness class: evolving.
When should I use it?
Use for puzzles, planning, theorem-like or combinatorial problems with checkable partial progress and a tolerance for latency. Not for chat or retrieval Q&A.
Implementation patterns
- Propose-evaluate-select loop with beam width k.
- Programmatic evaluator for puzzles and games (valid move, constraint satisfied).
- Value-guided MCTS where a learned model estimates how promising a partial solution is.
- Hybrid: shallow search for planning, single chain for execution.
Technologies and frameworks
Commonly associated: Tree of Thoughts reference code, process reward models, programmatic game/puzzle checkers. Names are examples of what exists, not recommendations, and projects change quickly.
Practical guide
- Check that the task has checkable intermediate states.
- Implement or choose an evaluator and test it alone.
- Start with small beam (2–4) and shallow depth.
- Compare against single-pass and best-of-N at equal compute.
- Add hard budgets.
Common failure modes
- Unreliable partial-step evaluator, so good branches are pruned and bad ones kept.
- Exponential cost from deep or wide trees.
- Latency unsuitable for interactive use.
- Complexity without gain on problems a single chain already solves.
Troubleshooting
- Search no better than single chain — Validate the evaluator on labelled partial solutions first.
- Cost spiralling — Reduce beam width/depth; add budget caps and early termination.
- Evaluator disagrees with ground truth — Prefer programmatic checks; otherwise train a PRM on step labels.
Contrasted with
- Best-of-N Sampling — Best-of-N scores completed answers; search scores partial steps and prunes early.
- Chain-of-Thought Prompting — CoT follows one path; search explores multiple.
Questions this page answers
- What is Tree of Thoughts?
- How is MCTS used with language models?
- When is search over reasoning better than chain-of-thought?
- What do I need to run tree search with an LLM?
Typical goals: solve problems needing exploration and backtracking; understand how search is combined with LLMs; decide whether tree search is worth the cost.
Sources and verification
Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.
Freshness class: evolving — Established techniques whose best practice shifts as research and tooling move. Re-review every 6 months.
Canonical sources (unverified leads):
- Yao et al., Tree of Thoughts: Deliberate Problem Solving with Large Language Models (2023) — arXiv:2305.10601
- Silver et al., Mastering the game of Go with deep neural networks and tree search, Nature (2016) — doi:10.1038/nature16961