Google DeepMind proposes autoregressive ranking for search
A paper from Google DeepMind with co-authors at the University of Massachusetts Amherst and the University of Texas at Austin proposes Autoregressive Ranking (ARR), a single model meant to replace the two-stage pipeline most search systems use today — dual encoders for initial retrieval, cross encoders for final scoring. In tests on WordNet and Amazon’s ESCI shopping-search dataset, the authors report ARR, trained with a new loss function called SToICaL, matched cross-encoder accuracy while keeping dual-encoder efficiency, though they note top-1 ranking of the single most relevant result sometimes worsened even as overall metrics improved. Per Search Engine Journal, which reported the preprint Sept. 10, this is a research proposal only — Google has not announced any change to production search ranking.
Why it matters: It's an early academic signal that search's two-stage retrieve-then-rerank pipeline could eventually collapse into a single model — nothing to act on yet, but a preview of where relevance-scoring research is headed.
The record: Google
Via Search Engine Journal ↗ · arXiv preprint (Rozonoyer et al.) ↗
Posted to the wire September 10, 2026. Edited by Joe Balewski.