From a Static Multi-Level Small Semantic Codebook to a Dynamic Single-Level Large Semantic Codebook for Generative Recommendation

Tianlu Xie; Xin Ku; Mingjie Sun; Yunhao Sha; Lixiang Wang; Peng Wang; Yiyu Wang; Wenjin Wu

cs.IRAug 21, 2026

From a Static Multi-Level Small Semantic Codebook to a Dynamic Single-Level Large Semantic Codebook for Generative Recommendation

Authors: Tianlu Xie, Xin Ku, Mingjie Sun, Yunhao Sha, Lixiang Wang, Peng Wang, Yiyu Wang, Wenjin Wu,

Organizations: Kuaishou Technology, Beijing, China

Abstract

Generative recommendation represents each item with a sequence of discrete Semantic IDs (SIDs) and predicts the sequence to retrieve the next item. Typical systems use multi-level residual quantization, which increases autoregressive decoding cost and creates a large hierarchical space that may be sparsely occupied. Static codebooks also become misaligned with current traffic as new items arrive and exposure distributions change. We propose a single-level large semantic codebook that replaces multiple residual semantic codes with one semantic token while retaining a separate collaborative disambiguation token to reduce item collisions. We further introduce an exposure-aware dynamic update mechanism based on temporal weight decay, exponential moving-average center updates, and an exposure-weighted penalty on SID changes. We also develop an offline evaluation framework covering representation quality, code utilization, cluster load, full-SID collision, and temporal stability. On two public datasets, the two-level SID improves mean Recall@10 by 5.0%-8.8% and mean NDCG@10 by 4.1%-5.1% for OneRec-V1, and by 7.1%-8.7% and 3.8%-8.5%, respectively, for OneRec-V2. Dynamic updating provides further gains on KuaiRec. Across three serving architectures, the shorter SID reduces estimated autoregressive-decoding FLOPs by 47.93%-48.70% and increases single-card QPS by 28.57%-47.0%. A five-day online A/B test serving 2.5% of production traffic improves the primary consumption metric by 0.792%.

Figures & tables

Figure 1: Global and conditional utilization of the two semantic levels in the deployed three-level SID. The histogram shows, for each active SID1, the fraction of the SID2 vocabulary observed under that code. The table reports global utilization and summarizes the conditional utilization distribution.

Figure 2: Overview of the proposed single-level large semantic codebook. Initial construction combines exposure-weighted semantic quantization (SID1) with deterministic collaborative disambiguation (SID2). Online updates maintain temporal exposure weights, penalize exposure-weighted SID1 switches, and update active centers while keeping SID2 unchanged.

Dataset	Users	Items	Interactions	Temporal coverage
Amazon Beauty (5-core)	22,363	12,101	198,502	May 1996–July 2014
KuaiRec 2.0 Big Matrix	7,176	10,728	12,530,806	28 observed days

Table 1: Statistics of the public datasets used in our experiments.

Figure 3: OneRec-V1/V2 scaling results on Amazon Reviews 2014 (Beauty). Curves and shaded regions denote the mean and sample standard deviation over seeds

\{7,42,2024\}

, respectively.

Model	SID	Recall@5	Recall@10	NDCG@5	NDCG@10
TIGER	S3	$0.0369\pm 0.0078$	$0.0520\pm 0.0136$	$0.0253\pm 0.0051$	$0.0302\pm 0.0070$
TIGER	S2	$0.0328\pm 0.0009$	$0.0463\pm 0.0008$	$0.0230\pm 0.0010$	$0.0273\pm 0.0010$
SEATER	S3	$0.0380\pm 0.0014$	$0.0562\pm 0.0014$	$0.0256\pm 0.0009$	$0.0314\pm 0.0010$
SEATER	S2	$0.0385\pm 0.0002$	$0.0584\pm 0.0008$	$0.0258\pm 0.0005$	$0.0323\pm 0.0006$
RPG	S3	$0.0402\pm 0.0004$	$0.0569\pm 0.0008$	$0.0284\pm 0.0003$	$0.0337\pm 0.0005$
RPG	S2	$0.0441\pm 0.0009$	$0.0646\pm 0.0006$	$0.0310\pm 0.0007$	$0.0375\pm 0.0006$

Table 2: Recommendation performance on Amazon Reviews 2014 (Beauty). Results are mean

\pm

sample standard deviation over seeds

\{7,42,2024\}

Metric	S3	S2	PV-S2
Mean reconstruction cosine similarity	0.9958	0.9962	0.9939
Reconstruction P10 / median	0.9879 / 0.9991	0.9879 / 0.9996	0.9823 / 0.9995
Semantic-code utilization	L1: 100.00% L2: 100.00%	L1: 97.56%	L1: 94.92%
Semantic-group load P95 / maximum / CV	9.0 / 256 / 3.1685	26.1 / 376 / 1.9938	34.4 / 248 / 1.6390

Table 3: Offline codebook quality on KuaiRec.

Figure 4: OneRec-V1/V2 scaling results on KuaiRec under S3, S2, and PV-S2. Curves and shaded regions denote the mean and sample standard deviation over seeds

\{7,42,2024\}

, respectively.

Model	SID	Recall@5	Recall@10	NDCG@5	NDCG@10
TIGER	S3	$0.1057\pm 0.0032$	$0.1429\pm 0.0019$	$0.0740\pm 0.0010$	$0.0860\pm 0.0006$
TIGER	S2	$0.1094\pm 0.0022$	$0.1483\pm 0.0012$	$0.0761\pm 0.0027$	$0.0886\pm 0.0023$
TIGER	PV-S2	$0.1087\pm 0.0028$	$0.1492\pm 0.0024$	$0.0742\pm 0.0030$	$0.0873\pm 0.0028$
SEATER	S3	$0.1032\pm 0.0019$	$0.1374\pm 0.0007$	$0.0697\pm 0.0028$	$0.0808\pm 0.0022$
SEATER	S2	$0.1077\pm 0.0018$	$0.1473\pm 0.0040$	$0.0729\pm 0.0019$	$0.0858\pm 0.0002$
SEATER	PV-S2	$0.1069\pm 0.0022$	$0.1478\pm 0.0028$	$0.0717\pm 0.0022$	$0.0849\pm 0.0024$

Table 4: Recommendation performance of the completed SID comparisons on KuaiRec. Results are mean

\pm

sample standard deviation over seeds

\{7,42,2024\}

Figure 5: Fixed-date evaluation setup for the static–dynamic comparison on KuaiRec. Bars indicate the data windows used for codebook fitting or updating and recommender training. Both arms use the first positive interaction per user on September 4 for validation and on September 5 for testing.

Figure 6: OneRec-V1/V2 results under the fixed-date KuaiRec comparison. Curves and shaded regions denote the mean and sample standard deviation over seeds

\{7,42,2024\}

, respectively. Static C1 uses the first-stage PV-S2 codebook, while Dynamic C2 uses the updated codebook and matched warm-start fine-tuning.

Model	Codebook	Recall@5	Recall@10	NDCG@5	NDCG@10
TIGER	Static C1	$0.0075\pm 0.0025$	$0.0092\pm 0.0011$	$0.0048\pm 0.0009$	$0.0053\pm 0.0004$
TIGER	Dynamic C2	$0.0496\pm 0.0775$	$0.0532\pm 0.0812$	$0.0372\pm 0.0584$	$0.0383\pm 0.0596$
SEATER	Static C1	$0.1484\pm 0.0014$	$0.1788\pm 0.0033$	$0.1161\pm 0.0034$	$0.1260\pm 0.0031$
SEATER	Dynamic C2	$0.1570\pm 0.0106$	$0.1890\pm 0.0090$	$0.1215\pm 0.0066$	$0.1331\pm 0.0068$
RPG	Static C1	$0.0356\pm 0.0238$	$0.0669\pm 0.0312$	$0.0206\pm 0.0151$	$0.0307\pm 0.0164$
RPG	Dynamic C2	$0.0465\pm 0.0136$	$0.0688\pm 0.0105$	$0.0335\pm 0.0073$	$0.0392\pm 0.0073$

Table 5: Fixed-date static–dynamic recommendation performance on KuaiRec. Results are mean

\pm

sample standard deviation over seeds

\{7,42,2024\}

Codebook representation	Day 1	Day 7	Change
Static S3, first semantic level	0.758357	0.754451	$-0.003906$
Static S3, two semantic levels	0.802241	0.798868	$-0.003373$
Static S2	0.805676	0.800899	$-0.004777$
Static PV-S2	0.821336	0.814508	$-0.006828$
Dynamic PV-S2	0.821336	0.826723	$+0.005387$

Table 6: Mean reconstruction similarity on the first and seventh industrial evaluation days.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Scale	$d_{\mathrm{model}}$	Enc. layers	Dec. layers	Heads	Enc. FFN	Expert FFN
0.1B	512	6	6	8	2,048	256
0.2B	640	8	8	10	2,560	320
0.3B	768	8	10	12	3,072	320
0.4B	896	8	8	14	3,584	448
0.5B	1,024	8	8	16	4,096	512
0.6B	1,024	10	10	16	4,096	512

Appendix

Table 7: OneRec-V1 public-data configurations. All variants use 24 experts and activate two experts per token.

Scale	$d_{\mathrm{model}}$	Layers	Attention heads	KV heads	FFN
0.1B	768	12	12	4	2,048
0.2B	1,024	16	16	4	2,048
0.3B	1,024	20	16	4	3,072
0.4B	1,280	18	20	4	3,072
0.5B	1,280	20	20	4	4,096
0.6B	1,280	24	20	4	4,096

Appendix

Table 8: OneRec-V2 public-data configurations.

Scale	SID	Recall@5	Recall@10	NDCG@5	NDCG@10
0.1B	S3	$0.0243\pm 0.0004$	$0.0411\pm 0.0016$	$0.0152\pm 0.0002$	$0.0206\pm 0.0006$
0.1B	S2	$0.0253\pm 0.0005$	$0.0429\pm 0.0006$	$0.0156\pm 0.0003$	$0.0212\pm 0.0004$
0.2B	S3	$0.0241\pm 0.0008$	$0.0397\pm 0.0011$	$0.0150\pm 0.0006$	$0.0200\pm 0.0007$
0.2B	S2	$0.0252\pm 0.0017$	$0.0415\pm 0.0013$	$0.0155\pm 0.0011$	$0.0208\pm 0.0010$
0.3B	S3	$0.0219\pm 0.0014$	$0.0372\pm 0.0008$	$0.0138\pm 0.0010$	$0.0187\pm 0.0009$
0.3B	S2	$0.0241\pm 0.0002$	$0.0402\pm 0.0015$	$0.0148\pm 0.0001$	$0.0199\pm 0.0007$

Appendix

Table 9: OneRec-V1 results on Amazon Reviews 2014 (Beauty).

Scale	SID	Recall@5	Recall@10	NDCG@5	NDCG@10
0.1B	S3	$0.0308\pm 0.0013$	$0.0476\pm 0.0022$	$0.0196\pm 0.0006$	$0.0250\pm 0.0008$
0.1B	S2	$0.0315\pm 0.0014$	$0.0509\pm 0.0008$	$0.0203\pm 0.0010$	$0.0266\pm 0.0008$
0.2B	S3	$0.0280\pm 0.0003$	$0.0444\pm 0.0014$	$0.0178\pm 0.0004$	$0.0231\pm 0.0004$
0.2B	S2	$0.0301\pm 0.0014$	$0.0492\pm 0.0009$	$0.0193\pm 0.0010$	$0.0255\pm 0.0008$
0.3B	S3	$0.0272\pm 0.0007$	$0.0437\pm 0.0014$	$0.0175\pm 0.0003$	$0.0228\pm 0.0006$
0.3B	S2	$0.0316\pm 0.0008$	$0.0503\pm 0.0017$	$0.0202\pm 0.0004$	$0.0263\pm 0.0007$

Appendix

Table 10: OneRec-V2 results on Amazon Reviews 2014 (Beauty).

Scale	SID	Recall@5	Recall@10	NDCG@5	NDCG@10
0.1B	S3	$0.0826\pm 0.0025$	$0.1098\pm 0.0037$	$0.0539\pm 0.0036$	$0.0628\pm 0.0040$
0.1B	S2	$0.0854\pm 0.0037$	$0.1196\pm 0.0025$	$0.0553\pm 0.0021$	$0.0664\pm 0.0017$
0.1B	PV-S2	$0.0846\pm 0.0047$	$0.1179\pm 0.0057$	$0.0556\pm 0.0030$	$0.0664\pm 0.0033$
0.2B	S3	$0.0733\pm 0.0057$	$0.1001\pm 0.0062$	$0.0483\pm 0.0029$	$0.0569\pm 0.0031$
0.2B	S2	$0.0735\pm 0.0122$	$0.1053\pm 0.0062$	$0.0456\pm 0.0065$	$0.0559\pm 0.0045$
0.2B	PV-S2	$0.0719\pm 0.0033$	$0.1039\pm 0.0033$	$0.0466\pm 0.0021$	$0.0570\pm 0.0021$

Appendix

Table 11: OneRec-V1 results on KuaiRec.

Scale	SID	Recall@5	Recall@10	NDCG@5	NDCG@10
0.1B	S3	$0.0931\pm 0.0024$	$0.1244\pm 0.0020$	$0.0656\pm 0.0013$	$0.0757\pm 0.0013$
0.1B	S2	$0.1014\pm 0.0067$	$0.1351\pm 0.0058$	$0.0713\pm 0.0049$	$0.0823\pm 0.0046$
0.1B	PV-S2	$0.0977\pm 0.0055$	$0.1335\pm 0.0058$	$0.0692\pm 0.0038$	$0.0808\pm 0.0040$
0.2B	S3	$0.0964\pm 0.0036$	$0.1271\pm 0.0025$	$0.0663\pm 0.0037$	$0.0762\pm 0.0029$
0.2B	S2	$0.0968\pm 0.0069$	$0.1342\pm 0.0056$	$0.0659\pm 0.0020$	$0.0781\pm 0.0017$
0.2B	PV-S2	$0.0966\pm 0.0036$	$0.1322\pm 0.0020$	$0.0676\pm 0.0032$	$0.0791\pm 0.0027$

Appendix

Table 12: OneRec-V2 results on KuaiRec.

Scale	Codebook	Recall@5	Recall@10	NDCG@5	NDCG@10
0.1B	Static C1	$0.1268\pm 0.0076$	$0.1467\pm 0.0082$	$0.0972\pm 0.0089$	$0.1046\pm 0.0106$
0.1B	Dynamic C2	$0.1328\pm 0.0104$	$0.1567\pm 0.0084$	$0.1028\pm 0.0056$	$0.1115\pm 0.0044$
0.2B	Static C1	$0.1138\pm 0.0170$	$0.1388\pm 0.0118$	$0.0832\pm 0.0208$	$0.0913\pm 0.0189$
0.2B	Dynamic C2	$0.1102\pm 0.0206$	$0.1374\pm 0.0197$	$0.0832\pm 0.0197$	$0.0921\pm 0.0194$
0.3B	Static C1	$0.1198\pm 0.0069$	$0.1445\pm 0.0053$	$0.0875\pm 0.0078$	$0.0956\pm 0.0074$
0.3B	Dynamic C2	$0.1212\pm 0.0111$	$0.1515\pm 0.0128$	$0.0959\pm 0.0126$	$0.1061\pm 0.0132$

Appendix

Table 13: OneRec-V1 results in the fixed-date KuaiRec comparison.

Scale	Codebook	Recall@5	Recall@10	NDCG@5	NDCG@10
0.1B	Static C1	$0.1406\pm 0.0284$	$0.1718\pm 0.0334$	$0.1095\pm 0.0235$	$0.1197\pm 0.0249$
0.1B	Dynamic C2	$0.1432\pm 0.0066$	$0.1804\pm 0.0066$	$0.1078\pm 0.0080$	$0.1200\pm 0.0077$
0.2B	Static C1	$0.1366\pm 0.0131$	$0.1659\pm 0.0099$	$0.1090\pm 0.0131$	$0.1186\pm 0.0122$
0.2B	Dynamic C2	$0.1386\pm 0.0176$	$0.1728\pm 0.0140$	$0.1067\pm 0.0160$	$0.1179\pm 0.0149$
0.3B	Static C1	$0.1391\pm 0.0061$	$0.1684\pm 0.0060$	$0.1072\pm 0.0123$	$0.1168\pm 0.0124$
0.3B	Dynamic C2	$0.1465\pm 0.0074$	$0.1838\pm 0.0140$	$0.1137\pm 0.0026$	$0.1259\pm 0.0050$

Appendix

Table 14: OneRec-V2 results in the fixed-date KuaiRec comparison.

Semantic ID Recommendation

Semantic identifier recommendation uses discrete, content-aware item codes to retrieve and rank items, especially in generative systems.

See all 38 papers

FineSID: Scalable and Efficient Semantic Identifier Learning for Generative Recommendation
Sep 29, 2026Song-Li Wu, Weinan Gan, Zhaocheng Du +2Semantic ID Recommendation Vector Quantization
Tsinghua University · Huawei Noah’s Ark Lab · University of Science and Technology of China
Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation
Jul 27, 2026Junting Wang, Xinrui He, Yunzhe Li +1Semantic ID Recommendation Language Model Decoding
University of Illinois Urbana-Champaign Urbana, Illinois, USA
Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation
Aug 10, 2026Chenxi Li, Yuchen Lu, Xu YangSemantic ID Recommendation Generative Recommendation
University of the Chinese Academy of Sciences · Institute of Automation, Chinese Academy of Sciences

Sep 29, 2026J/K move · Enter open · S save

FineSID: Scalable and Efficient Semantic Identifier Learning for Generative Recommendation

Song-Li Wu, Weinan Gan, Zhaocheng Du, Xianquan Wang+1

Tsinghua University · Huawei Noah’s Ark Lab · University of Science and Technology of China

Semantic ID Recommendation Vector Quantization Generative Recommendation

Open paper AskPDF arXiv

A critical prerequisite of generative recommendation is designing semantic identifiers (SIDs) that are both scalable to large item sets and efficiently learnable. Existing SID learning methods fundamentally rely on Top-1 hard assignment during vector quantization. While heuristic strategies -- such as clustering-based initialization or forced post-hoc collision resolution -- can artificially inflate codebook coverage, they often disrupt end-to-end semantic alignment and fail to address the underlying optimization bottleneck: sparse gradient propagation. In standard Top-1 assignment, gradients concentrate on a narrow subset of frequently selected codewords, leaving the majority inherently under-trained and causing severe SID collisions. To overcome this limitation natively without relying on complex initialization priors, we propose FineSID, a unified quantization framework that moves beyond Top-1 assignment by enabling fine-grained gradient propagation across the entire codebook. Instead of updating only a single selected codeword, FineSID distributes learning signals to all codewords in a soft, differentiable manner. This design promotes globally balanced codebook optimization while strictly preserving semantic consistency, effectively alleviating SID collisions and stabilizing training in large, high-dimensional codebooks. Extensive experiments on multiple public benchmarks demonstrate that FineSID is robust to initialization configurations and consistently improves both codebook utilization and recommendation accuracy. Our work provides a principled, initialization-agnostic solution for semantic identifier learning, advancing the practicality of generative recommendation.

Abstract

Semantic ID Recommendation

Explore similar work

FineSID: Scalable and Efficient Semantic Identifier Learning for Generative Recommendation

Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation

Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation