Tree of thought
Also known as: ToT, tree-of-thoughts prompting, multi-path reasoning
Chain-of-thought gets the model to reason step by step in a straight line. Tree of thought (ToT) goes further: it asks the model to generate multiple candidate next steps, evaluate which ones are promising, explore those branches, and prune the ones that don't work out. It's more like a deliberate search process than a single reasoning trace.
The pattern was designed for problems where the right answer isn't obvious from a single direction, such as creative planning, logic puzzles, or multi-constraint optimization. By explicitly branching and backtracking, the model can recover from wrong initial assumptions rather than committing to them.
In practice, tree-of-thought is expensive and slow because it requires multiple model calls per step. It's most useful when you're willing to trade cost for quality on a hard task and have no better alternative. With the rise of reasoning models (models that have internalized extended thinking into their pre-training), the explicit tree-of-thought pattern has become less necessary for many use cases, but the underlying idea, exploring multiple paths before committing, is still alive inside those models.