Multi-Decoder OneRec: Controllable Generative Retrieval for Multi-Objective Industrial Recommendation

2026-07-29

AI Summary

This work addresses the challenges of fragmented multi-objective retrieval pathways in industrial recommendation systems, which complicate modeling and serving, and the inability of existing single-decoder generative approaches to simultaneously achieve objective disentanglement and candidate diversity. To this end, we propose Multi-Decoder OneRec, a unified generative framework featuring multiple decoders that share a common user context representation while incorporating objective-specific LoRA experts and a coordinated decoding mechanism for controllable multi-objective retrieval. Our approach innovatively combines shared modeling with objective-specific control, enhancing candidate complementarity and reducing cross-objective overlap through gradient routing, stop-gradient references, KL regularization, and constrained beam search. Evaluated on the Kwai26 benchmark, our method improves Recall@512 by 1.69%–5.62%; online A/B tests further demonstrate a 0.37% increase in per-user watch time, 0.19% gains in both seven-day retention and shared device count, and a 2.09% improvement in cold-start performance for new content.

Abstract

Industrial recommender systems build candidate pools by assigning explicit quotas to objective-specific retrieval routes. This design offers quota control but increasingly fragments modeling, training, and serving as the route set grows. Semantic-ID-based generative retrieval provides a unified alternative, yet a single decoder entangles objective policies and limits candidate complementarity. We propose Multi-Decoder OneRec, a controllable framework that combines shared representations, isolated objective adaptation, and coordinated decoding. All objectives share a user-context module and the General Decoder, while each objective adds an isolated, parameter-efficient LoRA expert. During training, exposure-sample next-token prediction (NTP) updates the shared base, target-filtered NTP updates the event-based experts, and Kullback-Leibler (KL)-regularized policy optimization updates the Watch-time expert; gradient routing isolates these updates, and the General Decoder supplies a stop-gradient reference. At inference, explicit route quotas allocate the fixed budget and Multi-Decoder Constrained Beam Search reduces cross-route overlap. We publicly release Kwai26, a large-scale multi-objective benchmark with 1.31 billion raw item-level records, 31.85 million Item-ID entries, and 25.03 million items with valid Semantic IDs, together with predefined splits and an evaluation protocol. Under the same 512-item retrieval budget, Multi-Decoder OneRec improves over the single-decoder OneRec baseline by 1.69%-5.62% across four Recall@512 metrics. In a production A/B test, it yields relative gains of 0.37% in app usage time per device, 0.19% in Day-7 retained users, 0.19% in devices with at least one share, and 2.09% in new-content Cold-Start. These results show that generative retrieval can combine shared modeling with objective-specific control and complementary candidate generation.

Problem

• multi-objective recommendation
• generative retrieval
• candidate complementarity
• quota control
• industrial recommender systems

Innovation

• Multi-Decoder
• Generative Retrieval
• LoRA
• Controllable Recommendation
• Multi-Objective Optimization