gating

This paper challenges the assumption that reranking always improves few-shot selection for LLMs, proposing a training-free gated reranking approach that uses model uncertainty to decide when to rerank, reducing computational costs by 15-80% while slightly improving performance.