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

Scientific Impact Assessment: Multi-Decoder OneRec

1. Core Contribution

This paper addresses a real and well-articulated tension in industrial recommender systems: traditional multi-route retrieval offers explicit quota control over candidate composition but fragments modeling/training/serving, while single-decoder generative retrieval (Semantic-ID-based) unifies modeling but couples objective policies and produces overlapping candidates. The proposed Multi-Decoder OneRec resolves this trade-off with three coordinated mechanisms: (1) a shared user-context encoder and "General Decoder" backbone augmented by per-objective LoRA experts (plus objective-specific BOS embeddings and SID embedding residuals); (2) gradient-isolated feedback-adaptive training, where the shared base learns from exposure NTP, event-based experts use target-filtered SFT, and the watch-time expert uses KL-regularized relative-reward policy optimization (L-GBPO) with the General Decoder as a stop-gradient reference; and (3) Multi-Decoder Constrained Beam Search (MD-CBS), which assigns explicit route quotas and masks already-claimed SID prefixes to reduce cross-route overlap. A secondary but valuable contribution is the public release of Kwai26, a billion-scale multi-objective generative retrieval benchmark.

The framing is precise and the design maps cleanly onto the stated three requirements (shared, isolated, coordinated). This is an integration of known components (LoRA, gradient stop, constrained beam search, GRPO-style rewards) rather than a fundamentally new primitive, but the specific combination targeting quota-controllable generative retrieval is genuinely novel and non-obvious.

2. Methodological Rigor

The experimental design is solid for an industrial systems paper. Baselines span discriminative (DSSM, SASRec, HSTU) and generative (TIGER, OneRec) paradigms, all trained under matched conditions (same sessions, chronological split, identical history truncation, same budget of 512). The most important baseline — single-decoder OneRec sharing the same encoder and General Decoder — is well-chosen and isolates the multi-decoder contribution. Ablations are thorough: single-vs-multi-decoder decomposition, Multi-BOS baseline (to isolate LoRA's contribution beyond token conditioning), MD-CBS granularity (L2/L3/No CBS), KL weight sensitivity, reward-history size, SFT-vs-RL for continuous targets, LoRA rank, and efficiency (params/FLOPs). The leakage-safe data pipeline (strict time truncation, last-session held-out) is carefully described.

Notable gaps: statistical significance is claimed (p<0.05) but no error bars or variance across seeds are reported for offline results. The production A/B test has an asymmetric and modest traffic allocation (14.64% control vs 7.32% treatment), and gains are small in absolute terms (+0.37% usage time), though these magnitudes are meaningful at industrial scale. The RL formulation (L-GBPO) borrows heavily from GRPO/DeepSeekMath-style normalization and is not deeply theoretically justified.

3. Potential Impact

The practical relevance is high. Multi-objective retrieval under fixed budgets is a universal problem in large-scale recommenders (Kuaishou, YouTube, Pinterest, etc.). The paper offers a concrete architecture that preserves the operational lever practitioners care about — explicit quota control — while capturing generative retrieval's modeling efficiency. The parameter-efficient LoRA approach (only +20% params for 3 experts) makes it deployable. The A/B test on Kwai Brazil, where generative retrieval already drives 57% of impressions, demonstrates real production integration.

The Kwai26 dataset is potentially the most durable contribution: billion-scale public benchmarks with valid Semantic IDs and predefined multi-objective splits are scarce, and this could become a standard evaluation resource for the fast-growing generative retrieval subfield. This addresses a genuine reproducibility bottleneck in the OneRec/TIGER line of work.

4. Timeliness & Relevance

Highly timely. Generative retrieval (TIGER, OneRec, OneRec-V2, DiffGRM) is one of the most active areas in recommender systems (2023–2026), and the coupling/controllability problem for multi-objective settings is a natural next bottleneck that many industrial teams are hitting. The paper is essentially a first mover in equipping shared SID retrievers with objective-specific decoders and coordinated quota-aware decoding.

5. Strengths & Limitations

Strengths: Clear problem framing; clean architectural decomposition matching stated requirements; comprehensive ablations that isolate each component's contribution (especially the Multi-BOS baseline confirming LoRA's value); both offline and online validation; a substantial public dataset with detailed reproducibility appendices (algorithms, hyperparameters, vocabulary sizes); honest reporting of efficiency costs (2.23× inference FLOPs).

Limitations: Novelty is combinatorial rather than paradigmatic — every ingredient is borrowed. Online gains are modest and the A/B setup is somewhat asymmetric. No variance/seed analysis offline. The method's decoding cost grows with the number of routes (2.23× for 4 routes), which partially reinstates the scalability concern it critiques for multi-route systems, though at lower cost. Results are entirely on short-video (Kuaishou) data; generalization to e-commerce or other domains is untested. The paper carries a future dated venue (KDD '27) and arXiv ID (2607/2026), suggesting a preprint whose peer-review status is uncertain.

Reproducibility: Strong for an industrial paper — the Kwai26 pipeline is released on GitHub, hyperparameters and data construction are documented in detail. The proprietary production stack and A/B test are inherently not reproducible, but the offline benchmark is.

Overall

A well-executed, practically motivated industrial systems paper that makes a clean contribution to a timely problem and, importantly, releases a large-scale benchmark. Its impact will likely be moderate-to-strong within the generative recommendation subfield — cited and built upon by teams working on multi-objective generative retrieval — but it is an engineering synthesis rather than a conceptual breakthrough. The dataset may prove to be its most cited legacy.