UniGD: A Unified Generative-Discriminative Framework for Industrial Retrieval

2026-08-04

AI Summary

This work addresses the suboptimal performance and excessive inference overhead in industrial search advertising caused by the decoupled optimization between generative retrieval and standalone relevance scoring models. To bridge this gap, the authors propose UniGD, a unified generative-discriminative framework that jointly optimizes retrieval and relevance scoring within a single end-to-end model. Key innovations include Conflict-Aware Gradient Enhancement (CAGE) to mitigate interference among competing objectives, Codebook-Anchored Representation (CAM) that injects semantic priors via frozen multi-granularity codebooks, and Heterogeneous Ad Material modeling (HAM) to capture both shared and ad-specific characteristics. Deployed on Kuaishou, UniGD increases ad revenue by 5.78% and reduces inference latency by 33%, while achieving Recall@10 improvements of 8.44% and 3.19% on NQ320K and MS300K benchmarks, respectively.

Abstract

Generative retrieval (GR) is a promising paradigm for industrial search advertising, yet its deployment is constrained by strict relevance and latency requirements. Existing systems cascade GR with an independent relevance model, decoupling the generative likelihood objective from query-ad relevance discrimination, which compromises effectiveness and increases serving costs. We propose a Unified Generative-Discriminative framework (UniGD) that integrates retrieval and relevance scoring within a single model. To mitigate gradient interference in joint optimization, UniGD introduces Conflict-Aware Gradient Enhancement (CAGE) to adaptively coordinate the two objectives. UniGD further designs a Codebook-Anchored Representation Module (CAM) that anchors item representations to frozen hierarchical codebooks distilled from a multimodal pretrained model, thereby endowing them with rich and generalizable semantic priors. For heterogeneous short-video, product, and live-stream ads, UniGD proposes Heterogeneous Ad-material Modeling (HAM), which captures cross-type semantic commonality over a shared backbone while preserving type-specific modeling capacity. Online AB tests on Kuaishou search advertising platform show that UniGD raises ad revenue by 5.78%, reduces inference latency by 33%, and improves discriminative relevance estimation. On NQ320K and MS300K, UniGD improves Recall@10 over the strongest reproduced GR baseline by 8.44% and 3.19%, respectively.

Problem

• Generative Retrieval
• Search Advertising
• Relevance Modeling
• Latency Constraint
• Industrial Retrieval

Innovation

• Unified Generative-Discriminative
• Conflict-Aware Gradient Enhancement
• Codebook-Anchored Representation
• Heterogeneous Ad-material Modeling
• Industrial Retrieval