Least To Most Prompting: Definition and Examples

Least To Most Prompting is an advanced prompt engineering strategy developed by researchers at Google Brain in 2022. It relies on a simple but powerful principle: when faced with a complex problem, start by identifying the simplest sub-problems, then gradually progress to the more difficult ones, using previous answers as context to solve each subsequent step.

This technique unfolds in two distinct phases. The first phase, called decomposition, consists of asking the model to identify the sub-problems needed to solve the main question. The second phase, called sequential resolution, consists of solving each sub-problem in increasing order of difficulty, injecting the solutions of previous sub-problems into the prompt at each step.

The major advantage of Least To Most Prompting over classic Chain of Thought is its ability to generalize to harder problems than those seen in examples. Where Chain of Thought may fail when faced with a significantly more complex problem than the provided demonstrations, Least To Most Prompting adapts by building its answer brick by brick, each intermediate solution serving as a springboard for the next.

This approach is directly inspired by pedagogical methods used in teaching, where one guides a student by starting with fundamental concepts before tackling advanced notions. It is particularly effective for mathematical reasoning tasks, multi-step problem solving, and questions requiring compositional understanding of language.