Optimal balance between few-shot prompting and guideline-based prompting

Determine whether an optimal hybrid prompting strategy exists that combines few-shot in-context examples with clinically curated entity-annotation guidelines for small language models performing question-answering-based information extraction from paediatric renal biopsy histopathology reports, or establish that few-shot prompting and guideline-based prompting should be treated as alternative, non-complementary strategies.

Background

The study evaluated prompting strategies for small LLMs using clinician-curated annotation guidelines and few-shot examples on paediatric renal biopsy reports. Guidelines and few-shot examples each improved accuracy substantially over zero-shot baselines, but combining them did not yield additional benefits in accuracy, speed, or error rates.

This unexpected lack of complementarity raises a methodological uncertainty about how best to integrate—or whether to integrate—few-shot examples with structured domain guidelines in clinical information extraction tasks under resource constraints.