Relevance posterior boosts neural IR by 0.054 nDCG on TREC DL-2020

Modern information retrieval systems overwhelmingly model relevance as a monolithic query-conditional score computed at inference time. In a new paper accepted to SIGIR 2026, Andrew Parry and colleagues challenge this assumption, arguing that neural ranking systems are better understood as performing approximate posterior inference. They extend classical probabilistic retrieval formalisms to modern learned systems, decomposing relevance into two components: a query-dependent likelihood (e.g., from cross-encoders or bi-encoders) and a query-independent prior (the inherent utility of a document). This decomposition allows practitioners to precompute document priors offline, then fuse them with lightweight query-time scores via simple score fusion.

The empirical results are striking. On the TREC DL-2019 benchmark, a learned prior improves first-stage retrieval by up to 0.046 nDCG@10 through rank fusion alone. The gains persist downstream: re-rankers see consistent improvements, with the largest observed for the LLM-based RankZephyr model—0.054 nDCG@10 on TREC DL-2020. These results highlight how allocating modeling capacity between offline priors and online interaction can boost effectiveness with minimal additional query-time computation. The paper also discusses connections to retrieval-augmented generation and multi-stage pipelines, outlining how this perspective could lead to more efficient, scalable search systems.