Domain-Specific Decomposition Prompts

Decomposition few-shots rarely transfer across domains — invest in within-domain demos because a correct split usually decides success.

Least-to-most prompting needs a decomposition context tuned to the domain. A math-word decomposition demo does not teach the model to break down common-sense questions like “Did Aristotle use a laptop?” — a new decomposition prompt is required for optimal performance. Even within a domain, decomposition generalization is hard: nearly all GSM8K L2M failures become solvable given a correct manual split.

Implication for harness design: treat decompose-stage exemplars as first-class artifacts under offline prompt evaluation, not as one universal “how to split problems” preamble. Symbolic wins (SCAN, last-letter) partly reflect easy decomposition; do not assume those rates transfer to domains where inventing subproblems is the hard part. Pair with compositional mapping exemplars for the solve stage once the split is reliable.

---

Sources

• Least-to-Most Prompting (Zhou et al.), pp. 1–15