The Dense Retriever is the core optimization component of Revela, a novel self-supervised training framework. This framework was proposed by a joint team from Darmstadt University of Technology, the University of Washington, Carnegie Mellon University, Microsoft, and Tencent AI Lab, and the related research results have been published in a paper. Revela: Dense Retriever Learning via Language ModelingIt has been accepted by ICLR 2026.

The core mechanism of dense search engines is to map queries and documents into a high-dimensional vector space, and then determine content relevance by calculating vector similarity, thereby helping language models acquire external expertise. Traditionally, training high-quality dense search engines relies heavily on costly manually labeled query-document data, making it difficult to apply on a large scale in complex professional domains such as coding. In the latest Revela research, dense search engines have completely overcome this bottleneck: they are cleverly integrated into the "predict the next word" task of language models, and jointly optimized by introducing a cross-document attention mechanism. Experiments show that this dense search engine, trained without labeled data, not only outperforms supervised models with larger parameter sizes in specific domains and complex reasoning tasks, but also achieves state-of-the-art (SOTA) performance in general domains with extremely low data and computational costs.