New AI Breakthrough Makes Search Results Way Better

Embedding-based retrieval models have known expressive limits, and multi-vector embeddings consistently outperform single-vector ones—but the exact gap was unclear. Following Jayaram's work, researchers provided the first explicit family of query and document sets with relevance matrices where single-vector embeddings need exponential size to rank relevant documents above irrelevant ones, while polynomial-size multi-vector embeddings suffice. This proves an exponential separation for ranking documents, not just approximating numerical scores. They also introduce ANDOR, a benchmark built from these hard cases. State-of-the-art single-vector models perform poorly on ANDOR in the zero-shot setting and show only marginal gains after fine-tuning; multi-vector models consistently outperform them and improve substantially with fine-tuning, closely matching the theoretical predictions.