Generative recommendation has been widely adopted in large-scale industrial recommendation systems, where multimodal item embeddings are quantized into offline semantic identifiers (SIDs) and items are retrieved by autoregressively generating their SIDs. However, conventional SID construction often underrepresents item utility reflected in post-exposure behaviors such as shares, long views, and video completion, making semantically similar items with different utility insufficiently distinguishable while struggling to accommodate rapidly changing utility signals. Moreover, generation likelihood may be misaligned with downstream utility, causing high-utility candidates to be prematurely pruned during beam search before post-generation reranking or training-time reward alignment can effectively compensate. To address these limitations, we propose UAGR, which aligns both SID construction and decoding with downstream multi-objective utility. Our Utility-Enriched Semantic Codes preserve the original semantic information while incorporating item-level utility without increasing SID length or vocabulary size, supporting incremental utility updates with no additional tokenization or sequence-decoding cost. Meanwhile, our Utility-Guided Decoding estimates candidate utility by combining predictions of multiple post-exposure outcomes and incorporates these estimates into beam scoring, reducing the risk of prematurely pruning high-utility paths while enabling flexible utility control without interfering with generative backbone training. Extensive offline experiments and online A/B tests under real-world traffic demonstrate the effectiveness and deployability of UAGR, which has been fully deployed on Kuaishou and increases video watch time by 0.206% online.
Figures & tables
Figure 1. Overview of UAGR. (a) Utility-Enriched Semantic Codes encode item-level utility in the final SID token without increasing SID length or vocabulary size. (b) The utility estimation module, with stop-gradient isolation, predicts post-exposure feedback from decoder hidden states. (c) Utility-Guided Decoding combines cumulative generation log-likelihood with utility estimates to select candidate prefixes during beam search.
Figure 2. Illustration of utility-guided beam search. Expanded SID prefixes are scored by cumulative generation log-likelihood and current-step utility estimates before pruning to retain the top- K prefixes. At each decoding step, candidate SID prefixes are expanded. Their cumulative generation log-likelihood and current-step utility estimates jointly determine which prefixes survive pruning and continue to the next step.
Type
Method
KuaiRand
Recall@3 ↑
NDCG@3 ↑
Recall@5 ↑
NDCG@5 ↑
Discriminative
Caser ( Tang and Wang, 2018 )
2.0462
1.5378
3.0924
1.9753
SASRec ( Kang and McAuley, 2018 )
1.8460
1.4063
2.8637
1.8436
BERT4Rec ( Sun et al., 2019 )
1.6482
1.3563
2.6452
1.7629
S3-Rec ( Zhou et al., 2020 )
2.3681
1.8630
3.4612
2.2975
HSTU ( Zhai et al., 2024 )
2.4831
1.8742
3.6223
2.3593
Table 1. Overall offline performance on KuaiRand (all metrics in %). The best result is bold, and the second-best is underlined.
Method
KuaiRand
Recall@3 ↑
NDCG@3 ↑
Recall@5 ↑
NDCG@5 ↑
ICR ↑
Gini ↓
Entropy ↑
RQ-KMeans ( Chen et al., 2010 )
2.1114
1.6432
3.1493
2.0460
0.9571
0.02660
5.9732
OPQ ( Ge et al., 2013 )
2.0095
1.6174
3.0695
2.0323
0.9582
0.00941
7.9522
RQ-OPQ ( Chen et al., 2025 )
2.2364
1.8171
3.2123
2.2171
0.9843
0.01030
7.8756
UIST-RQ ( Liu et al., 2024 )
2.2137
1.7982
3.1641
2.1943
0.9842
0.00990
7.8642
LCRec ( Zheng et al., 2024 )
2.1473
1.7463
3.3651
2.2579
0.9997
0.01250
7.5144
Table 2. Comparison of different SID construction methods on KuaiRand (Recall and NDCG in %). The best result is highlighted in bold, and the second-best result is underlined.
Variant
Recall@ K
NDCG@ K
ICR
K=3
K=5
K=3
K=5
Full UAGR
2.774
3.963
2.274
2.776
0.9902
Base
2.111
3.149
1.643
2.046
0.9571
w. LFQ
2.246
3.288
1.654
2.064
0.9789
w. UE-SID
2.409
3.585
1.967
2.436
0.9902
w. UGD
2.578
3.712
2.047
2.499
–
Table 3. Component ablations and comparisons of other utility alignment strategies on KuaiRand (Recall and NDCG in %). Each individual variant is independently added to Base.
App Usage Time
Video Watch Time
Video Views
QPS
Δ
+0.074%
+0.206%
+0.164%
−0.03%
Table 4. Online A/B test results on the Kuaishou platform. All reported user-consumption gains are statistically significant under a two-sided t -test with p<0.05 .
Figure 3. High-utility prefix retention during beam search for Base and Base with UGD at the same beam width. Annotated gaps show improvements in percentage points. The figure compares high-utility prefix retention during beam search between Base and Base with UGD at the same beam width. Gap annotations show differences in retention rates in percentage points.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Policy
Users
Videos
Interactions
KuaiRand-27K
Standard
27,285
32,038,725
322,278,385
Random
27,285
7,583
1,186,059
KuaiRand-1K
Standard
1,000
4,369,953
11,713,045
Random
1,000
7,388
43,028
KuaiRand-Pure
Standard
27,285
7,551
1,436,609
Random
27,285
7,583
1,186,059
Appendix
Table 5. Statistics of KuaiRand before preprocessing. Standard and Random denote standard recommendation and randomized exposure, respectively.
Method
Collision (%) ↓
Recall@3 ↑
NDCG@3 ↑
Recall@5 ↑
NDCG@5 ↑
Base (RQ-KMeans)
4.2915
2.1114
1.6432
3.1493
2.0460
LFQ-13
2.1122
2.2461
1.6538
3.2875
2.0643
LFQ-13 + Mask
2.6793
2.3074
1.7147
3.2964
2.1117
LFQ-13 + Mask + Utility (UE-SID)
0.9801
2.4089
1.9669
3.5851
2.4357
LFQ-10 + Utility
0.9967
2.1675
1.6419
3.1972
2.0626
Appendix
Table 6. Detailed ablation and comparative studies of UE-SID on KuaiRand (Recall and NDCG in %). The best results are highlighted in bold, and the second-best results are underlined.
Semantic IDs (SIDs) provide the discrete item vocabulary used by generative recommendation, but their quality depends on what item evidence is preserved before quantization. In product recommendation, surface metadata often misses latent usage intent, visual evidence may be only weakly reflected in text, and downstream policy learning provides sparse feedback about whether a generated SID corresponds to a semantically useful item. We introduce DeepInterestGR, an intent-enriched SID framework for generative recommendation. Before SID quantization, CMSA enriches item representations through two complementary evidence paths: recommendation-oriented VLM captions and projected image embeddings. DCIM then uses an LLM to mine item-side intent descriptors -- latent usage motivations implied by product content rather than personalized user states. During policy training over the constructed SIDs, QARM adds a relevance-gated semantic-quality bonus on top of standard SID rewards, applying the bonus only when the generated SID decodes to the target item. Thus, semantic quality cannot reward a fluent but irrelevant item prediction. Experiments on three Amazon Product Review categories (Beauty, Sports, and Instruments) show that DeepInterestGR improves over competitive generative and RL-based baselines, with relative gains of up to 15.1% in NDCG@5 and 13.9% in NDCG@10 over the strongest per-metric baseline. Component ablations, CMSA branch analyses, reward variants, and SID-level case studies support a bounded claim: enriching pre-quantization item evidence with visual cues and item-side intent descriptors, together with relevance-gated semantic rewards, improves SID-based generative recommendation under the evaluated settings.
Personalized recommendation depends on learning semantic and collaborative relations among items from online content and user interactions. Generative recommendation (GR) offers a promising paradigm by autoregressively generating discrete semantic IDs (SIDs) of items. However, randomly initialized SID token embeddings and GR models trained from scratch do not directly inherit the semantic relations inherent in items, while inter-item collaborative relations are learned largely implicitly through token-level attention and next-token supervision. To address this limitation, we propose REGER, a unified framework that explicitly incorporates inter-item relations into semantic pretraining, history encoding, and SID generation. We pretrain the GR model through generative semantic retrieval, enabling it to inherit inter-item semantic similarity by learning to generate semantic neighbors' SIDs. We combine global attention with local attention aligned with SID boundaries to capture historical context and directed transitions between adjacent items. We formalize item-level preference distillation on the SID prefix tree, deriving an exact factorization of item preference distributions into prefix conditionals and a weighted decomposition of item-level KL divergence into conditional KL terms. Guided by this decomposition, we renormalize the generator's probabilities over legal next tokens and match the resulting distributions to the teacher's conditional preferences along the ground-truth SID path, jointly optimizing the distillation objective with cross-entropy. Extensive experiments on real-world datasets validate the effectiveness of the proposed preference distillation and demonstrate that REGER consistently outperforms existing GR baselines while retaining efficient inference.
Yuxi Huang, Silin Zhou, Chenhao Wang +6
University of Electronic Science and Technology of China Chengdu, China
Generative recommendation (GR) of social media platforms commonly retrieves items by autoregressively decoding discrete Semantic IDs (SIDs) with finite-width beam search. This work examines how this output interface constrains whether held-out targets enter finite returned lists, while distinguishing catalog-wide scoreability from empirical top-K support. We introduce two continuous-query paradigms: CQG-Single generates one retrieval query in a single pass, and CQG-AR autoregressively generates continuous residual updates, retaining step-wise decoding while removing the SID item-token vocabulary, code lookup, and discrete beam expansion. These designs thereby eliminate the SID-specific reachability constraint induced by finite-beam decoding and avoid the technical challenge of aligning a discrete SID vocabulary with the pretrained backbone's vector space. Both methods make every encoded item directly scoreable through dense or approximate-nearest-neighbor retrieval. Under a unified protocol with valid-prefix constrained decoding, CQG-Single exceeds the TIGER-style baseline in all seven runs while using 55.3% as many trainable parameters. CQG-AR remains comparable on ML-1M and is numerically stronger on Beauty and Sports datasets. A controlled depth ablation further shows that additional continuous refinement does not monotonically improve shallow accuracy. These findings motivate decoupling output representation, autoregressive modeling, and retrieval budget when evaluating generative recommendation systems.