Compositional Mapping Exemplars

Teach solve-stage few-shots to name component outputs and how they concatenate or repeat, so composition reuses simpler meanings instead of flat I/O pairs.

Under least-to-most prompting, the mapping / solving context is not the same as standard few-shot I/O. Standard SCAN-style demos pair a command with a finished action string. Compositional mapping demos instead:

  1. Name which simpler commands’ outputs are concatenated or repeated.
  2. Restate each component’s output.
  3. Conclude the composed action (or other IR).

That framing trains reuse of previously demonstrated (or already-solved) meanings — primitives, directional actions, repetition, conjunction, “after”, opposite/around — rather than memorizing flat input→output rows. It is the exemplar counterpart of feeding prior sub-answers into the next prompt: both make intermediate results explicit working memory for composition.

When engineering L2M for symbolic or instruction-following domains, invest in this solve-stage rhetoric separately from the decomposition context; compact IRs (+, *, small expressions) keep composed answers short enough for context engineering as nesting grows.

Sources

  • Least-to-Most Prompting (Zhou et al.), pp. 16–30