Researchers have developed a new scheduling strategy called "DealRank" for listwise Large Language Model (LLM) rerankers, which are used to sort large numbers of documents. Traditional methods often cut off relevant documents due to fixed quotas, necessitating costly repair calls. DealRank addresses this by distributing document candidates in a round-robin fashion based on their initial retriever rank, significantly reducing the need for these repair calls. Experiments show DealRank performs within 0.1 nDCG@10 points of optimal schedules across six LLMs and is substantially faster than existing methods like BlitzRank, improving latency without sacrificing quality. AI
IMPACT Improves efficiency and latency in document retrieval systems that leverage LLMs for reranking.
RANK_REASON Academic paper detailing a new method for LLM reranking. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →