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

Industrial recommender systems traditionally rely on multiple specialized retrieval routes, each assigned a specific quota to ensure a balanced candidate pool. While this provides control, it leads to fragmented model development and high maintenance costs. Generative retrieval offers a unified alternative, but existing single-decoder models often suffer from policy entanglement, where optimizing for one objective negatively impacts others. The authors propose Multi-Decoder OneRec, a framework that balances shared modeling with objective-specific control. It uses a shared user-context module and a General Decoder, while attaching parameter-efficient LoRA experts to each objective. This architecture allows for isolated, gradient-separated updates, ensuring that specialized tasks (e.g., watch-time, cold-start) can adapt without perturbing the base model.

Multi-Decoder OneRec introduces three key innovations: a shared-yet-isolated architecture, feedback-adaptive optimization with gradient isolation, and Multi-Decoder Constrained Beam Search (MD-CBS). During training, the General Decoder learns from broad exposure data, while objective-specific experts are updated via filtered samples or policy optimization with KL regularization to anchor them to the shared prior. At inference, MD-CBS enforces explicit quotas and prevents cross-route overlap by masking claimed Semantic ID prefixes. The authors demonstrate the effectiveness of this approach using the newly released Kwai26 benchmark, showing 1.69%–5.62% improvements in Recall@512 across four metrics. A production A/B test confirmed these gains, yielding significant increases in app usage time, user retention, and cold-start content performance.

This research bridges the gap between the flexibility of traditional multi-route systems and the efficiency of unified generative retrieval. By enabling independent objective adaptation within a single model, Multi-Decoder OneRec allows practitioners to scale the number of objectives without the overhead of managing separate models or the performance degradation associated with coupled single-decoder architectures. The release of the Kwai26 benchmark also provides a critical resource for future research in large-scale, multi-objective generative recommendation.