Reasoning Strategies › The Relationship to Human Problem-Solving

It is worth stepping back to consider why these strategies work from a cognitive science perspective. They are not random tricks; they reflect specific insights about how intelligent problem-solving proceeds.

Self-consistency mirrors the statistical intuition behind peer review and expert consensus. A single expert can be wrong, but the convergent judgment of multiple independent experts is more reliable. The key word is "independent": the value comes from diverse perspectives that fail in different ways, not from redundant agreement between identically-trained reviewers. Language model sampling at temperature $T > 0$ produces this independence by exploring different token choices at each step.

Tree of Thought mirrors the deliberate search process described by cognitive scientists studying expert problem-solvers. Kahneman's "thinking, fast and slow" framework distinguishes rapid intuitive responses (System 1) from slow deliberate reasoning (System 2). Standard language model generation is pure System 1: fast, fluent, and unreliable on hard problems. ToT adds a System 2 layer: slow, effortful evaluation that checks whether the fast responses are good. The language model's language understanding provides the intuitive generation; the evaluator provides the deliberate scrutiny.

Least-to-most prompting mirrors the cognitive strategy of problem reduction: converting a hard problem into a sequence of easier problems. This is the foundation of dynamic programming in computer science and of structured proof-writing in mathematics. The key insight is that problem difficulty is not an inherent property but a function of available resources. A 10-step problem is hard when you must hold all 10 steps in working memory simultaneously; it becomes easy when you can offload solved steps and focus on one at a time.

Decomposition strategies mirror the broader principle of modular problem-solving: breaking a complex task into components that can be handled by appropriate specialists. Human organizations work this way: accountants handle finance, engineers handle implementation, lawyers handle contracts. The coordination overhead is real, but the quality gain from specialist handling exceeds the cost. Language model decomposition strategies apply this organizational principle to single-model inference.