Reranking

Reranking is the secondary scoring of a set of candidate items produced by an initial retrieval or generation stage, with the aim of improving the quality of the top results. It is used to refine decisions in information retrieval, question answering, machine translation, and related tasks.

In practice, a fast first-stage system provides a candidate list, which is then re-scored by a more expensive model that can use richer features or cross-item interactions. Methods include cross-encoder rerankers that jointly encode query and candidate, and bi-encoder or feature-based models that score using precomputed representations.

Training data may come from relevance judgments, click logs, or synthetic labels, and objectives can be pointwise, pairwise, or listwise. Reranking is often deployed in a two-stage pipeline to balance speed and accuracy.

Common applications include document or passage retrieval, answer selection in question answering, and reranking translation hypotheses or recommendations. Evaluation uses ranking metrics such as mean average precision, NDCG, or precision@k on held-out data.

Practical considerations include latency and computational cost, potential overfitting or calibration issues, and sensitivity to the quality of the first-stage results. Limitations arise from biased training data and diminishing returns when the initial ranking is already strong. Reranking remains a flexible technique for improving end-task performance without redesigning the entire retrieval or generation system.