Tag
The paper introduces a bias-variance theory for dense retrieval, comparing shared and dual projections, and proposes the CARS method to select optimal geometry based on directional signal estimation from training data.