Least-to-Most Prompting

Least-to-most prompting tackles a hard problem by explicitly breaking it into a chain of simpler sub-problems, then solving them in order so each answer feeds the next. Introduced by Zhou et al. (2022), it differs from plain chain-of-thought in two ways: decomposition is a distinct first stage, and the sub-problems are solved sequentially with earlier solutions inserted into later prompts. The payoff shows up on tasks that require "compositional generalization" — where the test cases are harder or longer than anything in your examples, like multi-step word problems, symbolic manipulation, or applying a policy with several nested conditions. For SaaS builders, it's a good pattern when a single prompt keeps skipping steps or collapsing under complexity: instead of one mega-instruction, you have the model list the sub-questions, then answer them one at a time. The cost is more tokens and more calls, so reserve it for genuinely multi-step logic rather than simple lookups where zero-shot already works.