IR models degrade with collection growth; independent scoring more robust

_reachsumit · x · 2026-08-25

The paper studies robustness of IR models to collection growth, finding that adding unrelated documents hurts retrieval quality, and scoring documents independently is more robust than conditioning on other documents. Experiments show all models degrade, but Multi-Document-Agnostic (MDA) models are more effective for retrieval, while MDD and MDA rerankers are equally effective. Accepted to CIKM 2026.

Original post →

More from Research

Research channel →