Reasoning Strategies › Reasoning Strategies

Chain of thought prompting, which we explored in the previous chapter, teaches language models to work through problems step by step. But chain of thought is just one approach in a broader set of reasoning strategies. Researchers have developed a family of techniques that address different failure modes: when a single reasoning path leads to the wrong answer, when problems need to be broken into sub-problems, or when complex tasks require systematic search through possible approaches.

This chapter covers four major reasoning strategies: self-consistency, tree of thought, least-to-most prompting, and decomposition strategies more broadly. Each addresses a distinct limitation of naive prompting and offers a different intuition about what "good reasoning" means for language models. Together, they represent the current state of the art in eliciting reliable, accurate reasoning from large language models without any additional training.

Understanding why these strategies work requires revisiting a fundamental property of language models: they are stochastic text generators. Given the same prompt, a model can produce different outputs depending on its sampling temperature. This stochasticity, which looks like a bug, turns out to be a feature. If you sample multiple completions from a model, you get diverse reasoning paths, and those paths contain different errors. The strategies in this chapter exploit that diversity in complementary ways.