Semantic IDs (SIDs) are compact discrete representations derived from multimodal item features, serving as a unified abstraction for ID-based and generative recommendation. However, learning high-quality SIDs remains challenging due to two issues. (1) Collision problem: the quantized token space is prone to collisions, in which semantically distinct items are assigned identical or overly similar SID compositions, resulting in semantic entanglement. (2) Collision-signal heterogeneity: collisions are not uniformly harmful. Some reflect genuine conflicts between semantically unrelated items, while others stem from benign redundancy or systematic data effects. To address these challenges, we propose Qualification-Aware Semantic ID Learning (QuaSID), an end-to-end framework that learns collision-qualified SIDs by selectively repelling qualified conflict pairs and scaling the repulsion strength by collision severity. QuaSID consists of two mechanisms: Hamming-guided Margin Repulsion, which translates low-Hamming SID overlaps into explicit, severity-scaled geometric constraints on the encoder space; and Conflict-Aware Valid Pair Masking, which masks protocol-induced benign overlaps to denoise repulsion supervision. In addition, QuaSID incorporates a dual-tower contrastive objective to inject collaborative signals into tokenization. Experiments on public benchmarks and industrial data validate QuaSID. On public datasets, QuaSID consistently outperforms strong baselines, improving top-K ranking quality by 5.9% over the best baseline while increasing SID composition diversity. In an online A/B test on Kuaishou e-commerce with a 5% traffic split, QuaSID increases ranking GMV-S2 by 2.38% and improves completed orders on cold-start retrieval by up to 6.42%. Finally, we show that the proposed repulsion loss is plug-and-play and enhances a range of SID learning frameworks across datasets.
Figures & tables
Figure 1. SID learning in large-scale recommendation: multimodal features are encoded and quantized into SIDs as a unified discrete interface for traditional and generative paradigms. Challenges: token collisions cause semantic entanglement, and overlap signals are heterogeneous (harmful vs. benign), so one-size-fits-all suppression is problematic.
Figure 2. The overall framework of QuaSID.
Figure 3. Representative in-batch collision cases in industrial SID training. (a) Partial collision with 2/4 token overlap. (b) Complete collision with identical 4-token SIDs between semantically dissimilar items.
Figure 4. QuaSID deployment and application pipeline.
Dataset
#Users
#Items
#Interactions
Amazon-Beauty
22,363
12,101
1,048,296
Amazon-Toys
19,412
11,924
905,253
Table 1. Dataset statistics used in the experiments.
Amazon-Beauty
Amazon-Toys
Tokenizer
HR@5
HR@10
NDCG@5
NDCG@10
Entropy
HR@5
HR@10
NDCG@5
NDCG@10
Entropy
RQ-VAE
0.0225
0.0300
0.0171
0.0195
9.3075
0.0206
0.0256
0.0164
0.0180
9.3068
Improved VQGAN
0.0232
0.0304
0.0167
0.0189
9.3569
0.0196
0.0245
0.0157
0.0172
9.3313
GRVQ
0.0222
0.0302
0.0161
0.0187
9.2755
0.0195
0.0242
0.0158
0.0173
9.2101
RQ-OPQ
0.0205
0.0271
0.0159
0.0180
9.3368
0.0228
0.0278
0.0179
0.0195
9.3521
RQ-VAE-Rotation
0.0236
0.0308
0.0175
0.0198
9.3455
0.0203
0.0257
0.0167
0.0187
9.3290
Table 2. Performance comparison on Amazon-Beauty and Amazon-Toys. Bold denotes the best result, and underlined denotes the second-best result.
Amazon-Beauty
Amazon-Toys
Tokenizer
HR@5
HR@10
NDCG@5
NDCG@10
Entropy
HR@5
HR@10
NDCG@5
NDCG@10
Entropy
RQ-VAE
0.0225
0.0300
0.0171
0.0195
9.3075
0.0206
0.0256
0.0164
0.0180
9.3068
+ LHaMR
0.0246 ↑9.3%
0.0348 ↑16.0%
0.0182 ↑6.4%
0.0215 ↑10.3%
9.3971 ↑1.0%
0.0239 ↑16.0%
0.0290 ↑13.3%
0.0187 ↑14.0%
0.0203 ↑12.8%
9.3795 ↑0.8%
Improved VQGAN
0.0232
0.0304
0.0167
0.0189
9.3569
0.0196
0.0245
0.0157
0.0172
9.3313
+ LHaMR
0.0257 ↑10.8%
0.0340 ↑11.8%
0.0192 ↑15.0%
0.0218 ↑15.3%
9.3974 ↑0.4%
0.0239 ↑21.8%
0.0301 ↑22.8%
0.0187 ↑18.9%
0.0207 ↑20.1%
9.3797 ↑0.5%
GRVQ
0.0222
0.0302
0.0161
0.0187
9.2755
0.0195
0.0242
0.0158
0.0173
9.2101
Table 3. Effect of HaMR ( LHaMR ) on different semantic ID learning methods.
Figure 5. Hyperparameter sensitivity analysis of QuaSID on the Amazon-Beauty dataset. The bars indicate NDCG@K (left y-axis), while the curves denote HR@K (right y-axis). In each subfigure, one hyperparameter is varied while all others are fixed to their default values.
Setting
Completed Orders
GMV-S1
GMV-S2
Generative Retrieval
+0.21%
+1.03%
+0.55%
Ranking
+0.20%
+1.44%
+2.38%
Completed Orders
GMV
GPM
Retrieval
+1.09%
+1.69%
+3.25%
Retrieval 100vv
+6.42%
+4.67%
+0.21%
Retrieval 600vv
+4.69%
+3.11%
+0.53%
Table 4. Online A/B test results on Kuaishou e-commerce. Relative uplifts (%) brought by integrating QuaSID-learned SIDs into retrieval and ranking.
Tokenizer
Amazon-Beauty
Amazon-Toys
HR@5
NDCG@5
HR@5
NDCG@5
QuaSID
0.0277
0.0193
0.0266
0.0193
w/o CVPM
0.0263
0.0181
0.0264
0.0190
w/o HaMR
0.0254
0.0170
0.0261
0.0189
Table 5. Ablation study results on Amazon-Beauty and Amazon-Toys datasets.
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.
Song-Li Wu, Weinan Gan, Zhaocheng Du +2
Tsinghua University · Huawei Noah’s Ark Lab · University of Science and Technology of China
Existing work on Semantic IDs (SIDs) for generative recommendation treats SID construction as a representation learning problem: encode items into a quantised latent space and read off codes. We argue this view is incidental. SID construction is, at heart, a recursive clustering problem, and once stated this way the natural object to cluster is a graph whose nodes carry semantic content and whose edges carry collaborative signal; SID assignment becomes a hierarchical graph partition. This reframing yields a unified framework, Graph-Informed Semantic IDs (GrIS), that subsumes prior approaches rather than displacing them. RQ-VAE and RQ-KMeans are recovered as the special case where the graph is empty, exposing content-only quantisation as one corner of a larger design space along two so-far-collapsed axes: graph construction and recursive partition algorithm. We explore two contrasting instantiations: RecDMoN, which performs hierarchical assignment via differentiable graph pooling, and RQ-GAE, which extends RQ-VAE with graph-aware item representations and a graph reconstruction objective. On multiple real-world datasets, GrIS consistently improves over CF-aware SOTA, with gains of up to +52% Hit@10. Because graph construction and partition are explicit, separately configurable components, improvements on either axis can be combined and evaluated systematically.
Aleksei Medvedev, Alejandro Ariza-Casabona, Steven Derby +3
Semantic IDs (SIDs) are now a central component of generative recommendation. Current SID-based systems assign three roles to the same token sequence. Shared prefixes are intended to organize related items, the complete SID identifies an individual item, and each generated token narrows the items that can still be returned. We systematically investigate SIDs from item encoding and SID construction to autoregressive generation and final recommendation. We examine how SID construction changes item representations and how those changes affect generation. Across three Amazon domains and eight SID constructions, SID neighborhoods recover only 32.2% of the encoder's ten nearest neighbors on average. Alternative item descriptions still retrieve the corresponding item first in 99.57% of controlled cases, yet change 38.4% of exact SIDs. These results show that SIDs retain broad organization but lose much of the encoder's fine local structure, while their exact tokens are not determined by item meaning alone. This loss becomes consequential during generation. After the final semantic token, TIGER retains only 29.9% of held-out targets that were plausible recommendations before SID filtering. Motivated by these findings, we propose Item-Supported Decoding (ISD), a lightweight inference-time method that allows a user-specific item ranking to support corresponding SID prefixes before beam search discards them. The same ranking then orders the generated items. ISD requires no additional parameters or retraining of the SID constructor or decoder. We empirically show that ISD improves NDCG@10 over the corresponding SID backbone in every evaluated setting, with relative gains of up to 31.2%. Our results show that SIDs provide useful coarse item organization, but their fine boundaries should not alone determine which items remain available during generation.
Junting Wang, Xinrui He, Yunzhe Li +1
University of Illinois Urbana-Champaign Urbana, Illinois, USA