Structure
Root = problem. Each level = one 'thought'. Model generates K next thoughts per node. LLM-based evaluator scores nodes. BFS/DFS with pruning.
Advertisement
Where it wins
Problems needing backtracking: game of 24, creative writing, crosswords, planning. Not for straightforward tasks.
Advertisement
Cost
10-1000x single-pass. Worth it for problems where accuracy matters + linear CoT fails.
Alternatives
ReAct: interleave reasoning + tool use. AlphaGo-style MCTS + LLM (AlphaProof). Modern 'reasoning models' internalize this.