Prompted IR Expansion

Expand compact intermediate expressions into flat outputs with a few-shot rewrite prompt when you lack — or prefer not to run — a dedicated executor.

Least-to-most pipelines often emit a compact intermediate representation (e.g. SCAN’s Python-style + / * action expressions) to save context. Expanding that IR to a flat sequence can be a script or another LM prompt: few-shots that optionally Rewrite (fold products) then emit enumerated actions. Zhou et al. report ~99.7% expansion accuracy on 1000 random expressions under code-davinci-002, supporting near-ceiling combined accuracy without a Python executor.

Use when IR is simple algebra over strings and executor sandboxing is costly; prefer a real interpreter when IR can embed arbitrary code or when expansion errors would silently poison compositional mapping results. Keep expand-stage prompts separate from decompose/solve contexts under domain-specific decomposition prompts.

---

Sources

• Least-to-Most Prompting (Zhou et al.), pp. 31–45