Tree of Thoughts (ToT) explores multiple reasoning paths, enabling backtracking and strategic exploration. Mechanism: Generate multiple candidate "thoughts" at each step, evaluate/score each path, explore promising branches, backtrack from dead ends, use BFS/DFS search strategies. Comparison to CoT: Chain-of-thought follows single path, ToT maintains tree of possibilities, enables recovery from mistakes. Components: Thought generator (propose next steps), state evaluator (score partial solutions), search algorithm (BFS, DFS, or best-first). Use cases: Game playing (puzzles, chess), planning, creative tasks with multiple valid approaches, math problems with multiple solution paths. Implementation: Can use single model for generation and evaluation, or specialized evaluator model. Trade-offs: Much more expensive than CoT (many more LLM calls), slower, better for high-stakes decisions. Frameworks: LangChain has ToT components, research implementations available. When to use: Complex problems where backtracking matters, tasks with exploration/exploitation trade-off. Variants: Graph of Thoughts extends to arbitrary graph structures, not just trees.