UniGD: Unified Generative-Discriminative Framework for Industrial Retrieval

A novel framework named UniGD (Unified Generative-Discriminative) has been introduced to tackle issues in industrial search advertising, particularly in generative retrieval (GR). While GR shows great potential, it struggles with stringent relevance and latency demands. Current systems typically use a cascade approach, combining GR with a separate relevance model, which separates the generative likelihood from query-ad relevance evaluation, resulting in reduced effectiveness and higher operational costs. UniGD merges retrieval and relevance scoring into one model. To address gradient interference during joint optimization, it employs Conflict-Aware Gradient Enhancement (CAGE) for adaptive coordination of the two goals. Furthermore, UniGD features a Codebook-Anchored Representation Module (CAM) that ties item representations to fixed hierarchical codebooks derived from a multimodal pretrained model, providing them with robust and transferable semantic priors. The comprehensive framework is discussed in a paper accessible on arXiv (arXiv:2608.03150).