When Should Dense Retrievers Be Updated in Evolving Corpora? Detecting Out-of-Distribution Corpora Using GradNormIR

This work addresses the challenge of maintaining dense retriever performance in evolving corpora. We propose GradNormIR, an unsupervised method that detects out-of-distribution shifts in document collections using gradient norms, enabling timely updates of dense retrievers without manual intervention. Experiments on BEIR benchmark demonstrate that our approach effectively identifies when retriever updates are necessary, maintaining and improving performance as corpus distributions change.

Recommended citation: Dayoon Ko, Jinyoung Kim, Sohyeon Kim, Jinhyuk Kim, Jaehoon Lee, Seonghak Song, Minyoung Lee, Gunhee Kim. (2025). "When Should Dense Retrievers Be Updated in Evolving Corpora? Detecting Out-of-Distribution Corpora Using GradNormIR." Findings of ACL 2025.
Download Paper