Least-to-most prompting is a technique where you break a complex problem into a sequence of smaller sub-problems, then ask an AI model to solve them one at a time, feeding each answer into the next step — instead of asking it to solve the whole thing in one shot. It was introduced by Google researchers as a way to help language models handle tasks that are harder than any single example they’ve seen, by decomposing the task first and solving it second.
How Does Least-to-Most Prompting Work?
The technique runs in two stages. First, you prompt the model to decompose the original problem into a list of simpler sub-questions, ordered from least to most complex. Second, you prompt the model to answer each sub-question in sequence, and critically, each new prompt includes the original question plus every sub-question and answer that came before it. By the time the model reaches the final, hardest sub-question, it already has a chain of correct intermediate answers to build on, rather than having to solve everything from scratch.
The original research, published by Denny Zhou and colleagues at Google, tested this on compositional generalization tasks — problems where the test cases are structurally more complex than the training examples — and found it substantially outperformed standard few-shot prompting on tasks like symbolic manipulation and math word problems.
Why Does Breaking a Problem Into Steps Improve Accuracy?
Language models generate answers token by token, and errors compound when a model tries to jump straight to a complex conclusion without working through the intermediate logic. When you force the decomposition step first, each sub-answer acts as a checkpoint — if the model gets an early step wrong, that’s visible and correctable before it poisons the final answer. This is similar in spirit to how chain-of-verification prompting catches mistakes, except least-to-most focuses on structuring the problem itself rather than checking the output afterward.
How Is It Different From Chain-of-Thought or Tree-of-Thought Prompting?
Chain-of-thought prompting asks a model to “think step by step” within a single response, which helps with reasoning but doesn’t guarantee the steps are built in a useful order. Tree-of-thought prompting goes further by letting the model explore and evaluate multiple reasoning paths before picking one. Least-to-most prompting is narrower and more structured: it explicitly separates “figure out the sub-problems” from “solve them in order,” and it reuses each prior answer as context for the next step rather than branching into alternatives. In practice, it’s most useful when a task has a clear hierarchy of difficulty — where solving the easy part genuinely unlocks the hard part — rather than when a task has many equally plausible approaches.
How Do You Write a Least-to-Most Prompt?
You don’t need the model to label anything as “least-to-most” — you just need to guide it through the two stages explicitly:
- Decomposition prompt: “Break this problem into a series of simpler sub-questions, ordered from easiest to hardest, where each one builds on the last.”
- Sequential solving: Ask the model to answer the first sub-question, then feed the original question plus that answer back in and ask for the next one, repeating until you reach the original question.
- Final synthesis: Once every sub-question has an answer, ask the model to use the full chain to answer the original, complex question.
For a one-shot version that doesn’t require multiple round trips, you can also just ask the model to “first list the sub-steps this problem requires, in order of difficulty, then solve them one at a time before giving a final answer” inside a single prompt — it’s less rigorous than the original multi-call method but captures most of the benefit for everyday use.
When Should You Use Least-to-Most Prompting?
It earns its extra steps on tasks that are genuinely compositional: multi-part math problems, code that depends on several smaller functions working correctly, long-form content that needs an outline fleshed out section by section, or any analysis where getting an early fact wrong would wreck everything after it. For short, simple requests, it’s overkill — the decomposition step costs you time and tokens for no real accuracy gain. It also pairs naturally with prompt chaining, since both techniques are about structuring a workflow as a sequence of smaller, verifiable calls instead of one large, opaque one.
Frequently Asked Questions
Is least-to-most prompting the same as chain-of-thought prompting?
No. Chain-of-thought asks a model to reason step by step within one answer. Least-to-most explicitly decomposes the problem into ordered sub-questions first, then solves each one in sequence using the prior answers as context — it’s a more structured, multi-stage process rather than a single instruction to “think it through.”
Does least-to-most prompting work with any AI model?
It works best with models capable of multi-step reasoning, which includes most current general-purpose chat models. The technique is model-agnostic since it’s really a prompting strategy, not a model feature, though newer reasoning-focused models may already perform some decomposition internally, which narrows the gap it closes.
Do I need multiple separate prompts, or can it be done in one message?
The original technique uses multiple prompts, each building on the last. A simplified single-prompt version — asking the model to list sub-steps in order of difficulty and solve them sequentially before answering — is less rigorous but still captures much of the benefit for everyday tasks that don’t need the full multi-call setup.
Least-to-most prompting was first described in the paper “Least-to-Most Prompting Enables Complex Reasoning in Large Language Models” by Denny Zhou and colleagues at Google Research. For a deeper look at related reasoning techniques, see our guide to reducing AI hallucinations with better prompts.
