ModelRefs / Least-to-Most Prompting — AI Glossary

Least-to-Most Prompting — AI Glossary

A prompting strategy decomposing a complex problem into subproblems of increasing difficulty, solved sequentially to build toward the answer.

Overview

Least-to-most prompting (Zhou et al. 2022) first asks the model to identify constituent subproblems, then solves them in order, feeding each answer into the next subproblem. Outperforms standard chain-of-thought on compositional generalization tasks (SCAN, CFQ). Particularly effective when subtasks have clear dependencies.

Reference details

Topic	prompting
Last reviewed	2026-06-24

How it works

A first call decomposes the problem into sub-problems ordered from simplest to most dependent, rather than attempting a solution.
Each sub-problem is then solved in its own call, in that order.
The answer to each is appended to the context of the next, so later sub-problems reason over concrete earlier results rather than re-deriving them.
That explicit carrying is the mechanism: a single chain-of-thought must hold intermediate results in its own generated text and can lose or corrupt them.
Because each step is a separate call with a visible answer, a wrong intermediate is inspectable before it propagates.

Example: Where the gain over one long chain comes from

A problem with four dependent stages solved in one chain-of-thought pass: if each stage is right 85% of the time the whole is 0.85⁴ ≈ 52%, and a slip at stage two silently poisons the rest with nothing marking where. Solved as four calls, the per-stage rate is similar but each intermediate is a separate, checkable output, so a validator or a human can catch stage two before stage three consumes it. The reliability gain is the inspection point, not a better model.

Commonly confused with

Grouped with chain-of-thought and with decomposition, and it is a specific combination of both. Chain-of-thought asks for reasoning within one generation; this makes each sub-problem its own call and feeds answers forward. Plain task decomposition produces a list of sub-tasks and may run them in any order or in parallel; the defining feature here is the ordering by dependency and the explicit forwarding of each result. Against step-back prompting the difference is direction: step-back generalises upward once, this descends through a dependency chain.

When to use it

Reach for it when:

Problems with clear dependency ordering where later steps need earlier results.
Intermediate answers should be inspectable rather than buried in one generation.
A single chain-of-thought loses or corrupts intermediate results.

Reach for something else when:

Sub-problems are independent, where parallel decomposition is faster.
The problem is a single step, where the extra calls buy nothing.
Latency cannot absorb several sequential calls.

Primary source

Zhou et al., Least-to-Most Prompting Enables Complex Reasoning in Large Language Models (arXiv)

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Least-to-Most Prompting — AI Glossary.

Frequently asked questions

What is Least-to-Most Prompting?