Least-to-Most Prompting: Decompose, Then Solve in Order

Hard, multi-step problems break LLMs — they leap to the answer and slip on a middle step. Least-to-Most Prompting fixes it: make the model decompose the problem into sub-problems first, then solve them in order, each building on the last.

🧩 Watch one-shot vs least-to-most: https://dev48v.infy.uk/prompt/day18-least-to-most.html

Two stages

1. Decompose — ask the model to break the problem into simpler sub-problems, ordered easiest → hardest.
2. Solve sequentially — answer sub-problem 1, then feed its answer into sub-problem 2, and so on, until the final question falls out.

The demo races a direct answer (fumbles a middle step) against the decomposed path that builds up the correct result.

Why it beats chain-of-thought sometimes

CoT reasons forward in a single pass. Least-to-Most makes the structure explicit — it literally lists the steps, then solves them with the earlier answers in context. That's powerful for compositional problems (think: last-letter concatenation, multi-hop math) where you must reuse intermediate results.

Cost & cousins

It's multiple calls (decompose + N solves). Pair it mentally with Step-Back (abstract up) and Tree-of-Thoughts (search branches) — three different ways to give a model room to think.

🔨 Full pattern (decompose prompt → sequential solve loop → final answer) on the page: https://dev48v.infy.uk/prompt/day18-least-to-most.html

Part of PromptFromZero. 🌐 https://dev48v.infy.uk