Google Researchers Explore a New LLM-Based Approach to Search Ranking

Google researchers and academic collaborators have published new research exploring whether large language models (LLMs) could change how search engines rank documents.

The paper, “Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders,” introduces a method called Autoregressive Ranking (ARR), which uses an LLM-based approach to rank search results rather than relying solely on the traditional multi-stage retrieval architecture.

The researchers also propose SToICaL, a training method designed to make the model more aware of document relevance and ranking order. Their experiments on WordNet and ESCI shopping-search datasets found that the approach improved several ranking metrics and was able to better suppress irrelevant document results.

The research also presents a theoretical argument that autoregressive ranking can represent rankings without the same embedding-dimension limitations associated with dual encoders.

However, the findings are still research-stage results, rather than evidence that Google has replaced its production search-ranking systems with LLMs. The experiments were conducted on specific research datasets, so their implications for real-world search remain to be established.