LLM Reranking

LLM reranking studies how to reorder a shortlist of documents, products, answers, or tools so that the most useful candidates appear first, often improving retrieval-augmented generation and recommendations. Common approaches ask models to score items individually, compare them in pairs or lists, or reason over evidence; other methods fine-tune or distill smaller rerankers, calibrate attention scores, or use reinforcement learning. Evaluations include TREC and BEIR retrieval benchmarks as well as scientific, conversational-memory, recommendation, and domain-specific tasks, increasingly measuring downstream answer quality and efficiency alongside ranking metrics. Key tensions are inference cost and sensitivity to candidate order, prompts, and bias, while reranking cannot recover relevant items missing from the initial retrieval pool and topical relevance may not match downstream usefulness.