least-to-most prompting
**Least-to-Most Prompting** is the **structured prompt engineering technique that teaches language models to solve complex problems by first decomposing them into progressively simpler sub-problems, then solving from easiest to hardest** — developed by Google Research as a systematic approach that significantly outperforms standard chain-of-thought prompting on tasks requiring compositional generalization, mathematical reasoning, and multi-step problem solving.
**What Is Least-to-Most Prompting?**
- **Definition**: A two-stage prompting strategy where the model first decomposes a problem into sub-problems ordered from simplest to most complex, then solves each sequentially.
- **Core Innovation**: Explicitly separates the decomposition step from the solving step, ensuring systematic coverage of all reasoning components.
- **Key Difference from CoT**: Chain-of-thought generates reasoning inline; least-to-most structures reasoning as an explicit ordered sequence of sub-problems.
- **Origin**: Introduced by Zhou et al. (2023) at Google Research.
**Why Least-to-Most Prompting Matters**
- **Compositional Generalization**: Enables models to solve problems more complex than any seen in few-shot examples.
- **Systematic Reasoning**: The ordered decomposition ensures no reasoning steps are skipped or duplicated.
- **Transfer Learning**: Solutions to simpler sub-problems directly inform solutions to harder ones.
- **Reliability**: More consistent than free-form chain-of-thought on structured problems.
- **Interpretability**: The explicit sub-problem chain makes reasoning fully transparent.
**How It Works**
**Stage 1 — Decomposition**:
- Present the complex problem to the model.
- Prompt the model to list sub-problems from simplest to most complex.
- Each sub-problem builds on solutions to previous simpler ones.
**Stage 2 — Sequential Solving**:
- Solve the simplest sub-problem first.
- Feed the solution as context for the next sub-problem.
- Continue until the most complex (original) problem is solved.
**Comparison with Other Prompting Strategies**
| Strategy | Decomposition | Solving Order | Context Passing |
|----------|--------------|---------------|-----------------|
| **Standard Prompting** | None | Direct answer | None |
| **Chain-of-Thought** | Implicit | Left-to-right inline | Implicit |
| **Least-to-Most** | Explicit, ordered | Simplest first | Explicit sub-answers |
| **Tree-of-Thought** | Branching | Parallel exploration | Branch-specific |
**Applications & Results**
- **Math Word Problems**: 16.2% improvement over CoT on GSM8K-style problems.
- **Symbolic Reasoning**: Near-perfect accuracy on last-letter concatenation tasks where CoT fails.
- **Code Generation**: Effective for breaking complex programming tasks into incremental steps.
- **Multi-Step Planning**: Natural fit for tasks requiring ordered action sequences.
Least-to-Most Prompting is **a foundational advance in structured reasoning for LLMs** — demonstrating that explicitly ordering sub-problems from simple to complex enables compositional generalization impossible with standard prompting approaches.