UniGD: Unified Generative-Discriminative Framework for Industrial Retrieval
August 5, 2026
UniGD integrates generative retrieval and relevance scoring into a single model to reduce latency and costs. It uses Conflict-Aware Gradient Enhancement (CAGE) to manage joint optimization and a codebook-anchored module for semantic priors.
HOW THIS AFFECTS YOU
●
builderThis framework offers a more efficient way to deploy generative retrieval in high-throughput production environments.
●
researcherYou can use CAGE to coordinate competing objectives in joint generative-discriminative training.