Glance lets a generator rerank its own retrieved passages
Adilet Metinov and Chingiz Arziev introduce Glance, a generator-based reranking method that adds a lightweight decision adapter and pointer head to score candidate passages from a shared prefill. The paper reports nDCG@10 of 0.741 on SciFact and 0.345 on NFCorpus after 39 minutes of MS MARCO fine-tuning, plus a 0.045 answer-F1 improvement on 300 HotpotQA questions, but evaluation remains limited to two retrieval datasets, one QA dataset and one backbone family.
Glance demonstrates a way to combine passage reranking and answer generation within one resident 9B backbone, avoiding the separate reranker model and repeated query-passage prefilling used in conventional RAG systems.
Computer Science > Computation and Language
arXiv:2610.40454 (cs)
[Submitted on 29 Oct 2026]
Title:Glance: The Generator Is Its Own Reranker
Authors:Adilet Metinov, Chingiz Arziev
View a PDF of the paper titled Glance: The Generator Is Its Own Reranker, by Adilet Metinov and 1 other authors