LLM-based rerankers typically score documents by making repeated local comparisons—listwise, pairwise, or pointwise. Each comparison requires a separate model call, so the total number of sequential calls grows quickly with the size of the candidate set.
The new paper explores an alternative: feeding the entire pool of documents to a long-context language model at once, or in larger sets. This 'whole-pool' or 'setwise' strategy lets the model compare many documents in a single pass, potentially eliminating the need for many sequential calls.
The abstract says the authors study how long-context LLMs can 'drastically reduce' this computation, but the text cuts off before giving specific results. As a single-source digest, we can only report the proposed direction, not the measured outcome.