Cross-Encoder Reranking

Cross-encoder reranking studies how to improve search results by scoring a query and each candidate document together, usually after a faster sparse or dense retrieval stage. Common approaches use Transformer cross-attention models trained with relevance labels or ranking losses; newer work also distills these models into smaller rerankers or adapts instruction-tuned language models. The methods are used in retrieval-augmented generation, scientific and biomedical search, and multi-hop question answering, with ranking metrics such as nDCG, MRR, and Recall used alongside downstream answer quality. A central tension is that reranking can improve which evidence appears near the top, but adds latency and depends on candidate recall, training data, and fit to the target domain.