Paired Demonstration Format

Repeated input-output pairs act as a structural trigger that makes the task signals in few-shot demonstrations usable by the model.

The paired sequence structure of in-context learning examples can preserve most few-shot gains even when only one side of each pair contains task-relevant information. Removing one side and concatenating only inputs or only outputs performs near or below a zero-demonstration baseline. The pair pattern appears to trigger continuation of the demonstrated task structure, allowing demonstration input distribution or demonstration output space to guide the completion.

Format includes diversity and repetition, not just delimiters. Constant labels or repeated identical inputs may behave like separators and change the pattern the model perceives. Preserve varied pairs when testing a compact or partially synthetic demonstration suite.

---

Sources

• Min et al. (2022), full paper, pp. 1–19