Google DeepMind Tests Unified Ranking Model for Search
Google DeepMind researchers have proposed replacing the two-stage ranking architecture behind search with a single large language model. The paper, written with academics from the University of Massachusetts Amherst and the University of Texas at Austin, sets out a method called Autoregressive Ranking, or ARR.
Today's ranking systems pair a fast Dual Encoder for retrieval with a more precise Cross Encoder for ordering the candidates. The researchers argue Dual Encoders run into trouble as a corpus grows, because the embedding dimension has to grow with it to represent every possible ordering. They prove that an ARR model with a constant hidden dimension can in theory rank any number of documents.
To train the model they developed SToICaL, a loss function that reweights documents by their ground-truth relevance and marginalises over prefix trees, teaching the LLM which results belong higher and suppressing irrelevant ones. Tests on WordNet and ESCI Shopping Queries showed gains on several ranking metrics, though one version of the method got worse at putting the single most relevant result first in the shopping test.
The theoretical result does not guarantee the same behaviour in a live search system, and the authors call for more work on the shopping-search weakness. For brands chasing citations in AI answers, a single learned ranker would change how documents earn their place, though nothing here is in production.
Reported by Search Engine Journal ↗
Topics: Google Search