Disentangle demonstration-count and demonstration-type effects

Determine whether changes in model behavior arise from adding a third in-context demonstration, adding an unanswerable demonstration specifically, or using a particular unanswerable example and rationale in extractive radiology question answering.

Background

The few-shot conditions are cumulative: the final condition adds an unanswerable example after two answerable examples. Consequently, the observed changes cannot be uniquely attributed to the number of demonstrations, the presence of an unanswerable example, or the specific near-miss example and rationale used.

This unresolved attribution is important because the central behavioral effect of the final condition was a shift toward abstention. Controlled experiments with count-matched demonstrations and varied unanswerable examples would be needed to isolate the relevant cause.