Pith. sign in

REVIEW 4 major objections 5 minor 48 references

SPARC compresses each historical interaction into a single token while preserving context-dependent side information, improving generative recommendation without increasing backbone input length.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 02:42 UTC pith:TWWATJBI

load-bearing objection A genuinely new compression architecture for generative recommendation with promising numbers; the causal attribution to context-conditioning (rather than added capacity) is real but unproven and can be fixed with capacity-matched ablations and error bars. the 4 major comments →

arxiv 2607.25339 v1 pith:TWWATJBI submitted 2026-07-28 cs.IR

SPARC: Sequence-aware Progressive Attribute Routing and Compression Framework for Generative Recommendation

Chang Liu , Changfa Wu , Hui Qian , Binbin Cao , Jian Wu , Yuliang Yan , Han Zhu , Bo Zheng This is my paper
classification cs.IR
keywords generative recommendationsemantic IDssequence-aware compressionmulti-field behavior modelingcontext-aware attribute routingtoken consolidationindustrial recommender system
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

SPARC is a compression front-end for generative recommender systems that turns each historical interaction, with its many heterogeneous fields, into a single token for the generative backbone while preserving the side information the current behavior context makes relevant. The paper argues that static field compression (mean pooling, MLP, or query-based transformers) discards context-dependent signals prematurely, and that modeling each field's evolution across the user history before routing it into a fixed number of slots retains more of what matters. If this holds, generative recommenders can absorb rich behavioral attributes — category, brand, price, behavior type, recency — without paying the quadratic cost of longer input sequences, and a user's most recent intent (e.g., shifting to higher-priced purchases) is reflected in the compressed history. Experiments on a large Taobao dataset and on Amazon Beauty and Toys report consistent gains over strong conventional and generative baselines, with ablations claiming the gain is due to context-conditioned retention.

Core claim

The central discovery is that a sequence-aware, progressively compressed representation of multi-field user history — produced by field-wise context encoding, attention-style routing into a small set of side slots, and a gated consolidation into a single item token — improves autoregressive target SID prediction without expanding the backbone input. The paper further claims that the improvement comes from the context conditioning itself, not merely from a larger or more expressive compression module, and backs this with ablations that replace the routing with static MLP, QFormer, and modulated compressors.

What carries the argument

Field-wise Context Modeling (FCM), Context-aware Attribute Routing (CAR), and Sequence-level Token Consolidation (STC). FCM runs a lightweight Transformer over each field's timeline to produce context-aware field representations; CAR concatenates original, contextual, and field-identity embeddings, projects them, and uses soft attention normalized over side fields to fill a fixed set of side slots, so the same item is compressed differently in different histories; STC flattens all intermediate tokens, applies a gated residual Transformer over them, and pools identity and side tokens into a single token. The routing weights are the mechanism that decides which fields survive the fixed-capacit

Load-bearing premise

The load-bearing premise is that the static-compression variants in Table 2 (w/ MLP, w/ QFormer, w/ Modulated) are matched to SPARC in model capacity and tuning effort; if those baselines are undertuned or have far fewer parameters, the central claim that context-conditioned retention, not added expressiveness, causes the improvement does not follow.

What would settle it

Train SPARC and the w/ QFormer variant with identical parameter budgets, training steps, and hyperparameter search on the Taobao dataset; if the matched static compressor matches or exceeds SPARC's HR@20 and HR@1000 click metrics, the claim that context-conditioned routing is the driver would be refuted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Each historical interaction can be represented by exactly one backbone token, so the input length of the generative model stays constant while the information content of each token grows, directly addressing the quadratic cost of full field expansion.
  • Because side-field routing is context-dependent, the same item gets different compressed representations in different user histories, allowing retrieval to emphasize the fields that best explain the current behavior pattern.
  • The improvement over static compressors on both an industrial Taobao dataset and public Amazon subsets suggests the principle generalizes to dense and sparse interaction regimes.
  • The framework leaves the target SID generation objective untouched, so it can be layered on any SID-based generative backbone without retraining the tokenizer.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same contextualize-then-compress principle could apply to other sequential inputs with heterogeneous features, such as event logs or medical timelines, wherever static pooling discards conditionally relevant information.
  • The routing patterns in Figures 3 and 4 suggest that slots become interpretable (e.g., one slot tracks behavior type and recency, another tracks seller and brand); one could test whether slot specializations are stable across domains or token budgets.
  • An untested extension is adaptive token-budget allocation: instead of a fixed number of side slots per interaction, the routing weights could be used to allocate more slots to interactions where context is more informative.
  • The slot-level division of labor could be exported as an interpretability tool for generative recommendation, explaining why a particular item was retrieved by pointing to the fields a slot emphasized in that user's context.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. SPARC proposes a front-end compression module for SID-based generative recommendation. It first models each field type along the user history (Field-wise Context Modeling, FCM), then routes original, contextual, and identity representations of side fields into a small number of slots (Context-aware Attribute Routing, CAR), and finally refines and consolidates the intermediate tokens into one token per historical interaction (Sequence-level Token Consolidation, STC). The compressed sequence maintains the same length as the original item-level history, so the generative backbone input length is unchanged. The paper evaluates SPARC on industrial Taobao and public Amazon Beauty/Toys data, reports consistent gains over conventional and generative baselines, and attributes the gains to context-conditioned information retention based on comparisons with static compression variants and routing analyses.

Significance. If the claims hold, the paper addresses a practical bottleneck in generative recommendation: heterogeneous behavioral fields are either expanded into the backbone context or prematurely collapsed. The staged 'contextualize before compress' design is plausible, and the fixed per-interaction token budget is attractive for industrial deployment. The industrial Taobao evaluation and the use of RankGR as a strong generative backbone are strengths. The paper gives a concrete architecture, and the slot-level and same-item routing analyses help interpret the mechanism. However, the central causal claim is currently supported mainly by a comparison that is not controlled for model capacity or tuning budget, and the headline metric gains on public datasets are reported without variance or significance information. The paper's significance will depend on whether these controls can be supplied.

major comments (4)
  1. [§4.3, Table 2; Abstract] The abstract's conclusion that SPARC's improvement comes from 'context-conditioned information retention rather than merely increasing the expressiveness of the compression module' is not established. The static variants w/ MLP, w/ QFormer, and w/ Modulated are not matched to SPARC in parameter count, FLOPs, or tuning budget. SPARC adds seven field-wise 2-layer Transformer encoders, a 2-layer Transformer over the flattened 4L intermediate sequence, routing MLPs, value projections, and fusion parameters, while the static alternatives are described as a single MLP, a lightweight cross-attention block, and a projection with scaling factors. If the variants are undertuned or have substantially fewer parameters, Table 2 cannot separate contextualization from raw capacity. Please report parameter counts and training budgets for all variants, and add capacity-matched static compressors to suppo
  2. [§4.2, Table 1; §4.1.3] The main performance results are reported as single trials without error bars, repeated-seed variance, or significance tests. The Amazon gains are unusually large (e.g., Beauty HR@20 from 0.0466 for RankGR to 0.0794 for SPARC), so optimization noise cannot be ruled out. Also, the evaluation protocol is incomplete: HR@K is defined over a retrieved item set, but the manuscript does not specify whether top-K items are decoded from the full item space or sampled. This makes comparability across conventional and generative baselines unclear. Please report multi-seed statistics and a precise retrieval protocol.
  3. [§4.3, Table 2; RQ2] The ablation study that underlies the paper's main interpretation is performed only on TaoBao, where SPARC's gains over RankGR are modest. The public Amazon datasets, where gains are far larger, do not include any static-compression ablation. Since the paper claims context-conditioned retention is generally beneficial, the absence of w/ MLP, w/ QFormer, w/ Modulated on Beauty/Toys weakens the generality. Please add the same ablations on public datasets or explicitly restrict the causal claim to the industrial setting.
  4. [§4.4, Figures 3-4] The routing analyses are descriptive rather than causal. They show context-dependent slot specialization but do not demonstrate that these routing patterns drive the metrics. A stronger test would intervene on capacity while removing context-conditioning (e.g., freezing routing weights to dataset-level averages or replacing context-aware routing representations with static field embeddings) and compare retrieval performance. Without such an experiment, the figures support the mechanism qualitatively but not the attribution in the abstract.
minor comments (5)
  1. [§4.2] The sentence before Table 1 is duplicated: 'We evaluate the overall performance of the compared We evaluate the overall performance of the compared approaches across all three datasets.'
  2. [§4, RQ3] RQ3 refers to 'the slots of CIR'; the module is CAR (Context-aware Attribute Routing), not CIR. Please correct the acronym.
  3. [§3.4, Eq. (20)] The text says the sequence-level residual gate is initialized with a logit value of -5, but Eq. (20) uses the gate directly as lambda. Clarify whether lambda is the raw logit, a sigmoid-transformed probability, or another bounded parameter, and specify how it is optimized.
  4. [§3.4, Eq. (22)] The fusion weights eta_id and eta_side are described only as learnable. Please state whether they are normalized (e.g., softmax over the two scales), their initialization, and whether they are shared across positions. This affects the interpretation of the final token scale.
  5. [Table 1] For conventional baselines, the number of parameters and the candidate retrieval set are not described. Please ensure these settings match across all methods, or explain any differences that could affect HR@K.

Circularity Check

0 steps flagged

No significant circularity: the paper's claims are empirical and its compression property holds by construction rather than by fit-to-input.

full rationale

SPARC does not derive a prediction from fitted parameters in a way that reduces to its own inputs. The only mathematical claim is Eq. (22), where each historical interaction is compressed into a single token z_t by pooling identity and side tokens; this is a stated architectural design property, not an empirical result derived from the data. The central empirical claim, that SPARC outperforms strong baselines, is tested on held-out Taobao and Amazon data against conventional and generative baselines, including methods outside the author group. The attribution of the gains to 'context-conditioned information retention rather than merely increasing the expressiveness of the compression module' in Section 4.3 is an interpretation of ablations (w/ MLP, w/ QFormer, w/ Modulated). That interpretation is somewhat underdetermined because the paper does not report parameter counts or tuning budgets for the static variants, but underdetermination is an evidential weakness, not circular reasoning. The use of RankGR as the backbone is a self-affiliated citation, but it is not load-bearing for the method's derivation: FCM, CAR, and STC are defined independently, and SPARC is evaluated against RankGR itself as well as external baselines. No fitted value is renamed as a prediction, no uniqueness theorem is imported to force the design, and no equation reduces by construction to the claim being tested. The slot-level and context-dependent routing analyses are descriptive and do not constitute circular evidence. Overall, the paper is self-contained as an empirical study, and no circular step can be exhibited.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The ledger is dominated by hand-chosen architectural hyperparameters (slot count, field count, history length, gate init) and domain assumptions about the field encodings and evaluation objective. There are no new physical entities, but the model introduces learnable side-slot embeddings and routing parameters that are internal representations, not externally falsifiable inventions.

free parameters (4)
  • Number of side slots R = 2
    Chosen by hand as the fixed token budget for side information; no sensitivity analysis over R is reported.
  • Number of side fields = 7
    Dataset-specific field selection (category, brand, seller, price, tagging, decay, action type); no analysis of which fields contribute most.
  • Maximum history length L = 50
    Standard truncation choice; no sensitivity analysis is reported.
  • STC residual gate initialization logit = -5
    Hand-chosen to stabilize early training; effect on final results is not reported.
axioms (5)
  • standard math Self-attention cost grows approximately quadratically with sequence length, making L×F field tokens prohibitive (§1, Eq. 5).
    This motivates why compression to one token per interaction is necessary; if attention were subquadratic or the backbone had different scaling, the cost argument would weaken.
  • domain assumption The autoregressive SID generation objective in Eq. (3) is a faithful objective for recommendation quality.
    The paper optimizes negative log-likelihood of target SID sequences and evaluates HR@K; the alignment between this token-level objective and retrieval quality is inherited from prior generative recommendation work.
  • domain assumption Each side field can be sufficiently encoded by token embedding plus mean pooling or another lightweight field encoder (Eq. 4).
    No details are given for how price, timestamp, or decay are embedded; if these encodings are weak or lossy, the reported gains may not generalize.
  • ad hoc to paper The chosen 9-field schema (2 SID fields + 7 side fields) is a complete enough representation of user behavior.
    The field list is fixed per dataset and unmodeled fields could change the conclusions; no sensitivity analysis is provided.
  • ad hoc to paper Two side slots (R=2) suffice to capture complementary side information under the fixed budget.
    R=2 is chosen without an ablation over the number of slots, so the adequacy of this capacity is not demonstrated.

pith-pipeline@v1.3.0-alltime-deepseek · 15061 in / 12635 out tokens · 131783 ms · 2026-08-01T02:42:07.478063+00:00 · methodology

0 comments
read the original abstract

Generative recommendation tokenizes items as discrete Semantic IDs (SIDs) and autoregressively generates target items from users' historical SID sequences. Although existing SIDs incorporate multimodal and structured information, they are typically statically assigned and independent of the current interaction context. In industrial scenarios, each behavior also contains heterogeneous attributes, such as category, brand, price, behavior type, and timestamp. Fully expanding these features greatly increases the input length, while directly compressing them into a single representation may prematurely discard context-relevant information. We propose \textbf{SPARC}, \uline{\textbf{S}}equence-aware \uline{\textbf{P}}rogressive \uline{\textbf{A}}ttribute \uline{\textbf{R}}outing and \uline{\textbf{C}}ompression Framework for Generative recommendation. SPARC first models the sequential dependencies of each field type to obtain context-aware field representations. It then routes the original, contextual, and identity representations of different fields into multiple slots to preserve complementary information under a fixed capacity. Finally, lightweight cross-item interaction integrates the intermediate tokens and compresses each historical item into a single token. Following the principle of contextualizing before compression, SPARC enriches user-history representations without increasing the input length of the generative backbone. Experiments on industrial Taobao and public Amazon datasets demonstrate that SPARC outperforms strong conventional and generative baselines. Further comparisons with static compression variants show that the improvement of SPARC comes from context-conditioned information retention rather than merely increasing the expressiveness of the compression module.

Figures

Figures reproduced from arXiv: 2607.25339 by Binbin Cao, Bo Zheng, Changfa Wu, Chang Liu, Han Zhu, Hui Qian, Jian Wu, Yuliang Yan.

Figure 1
Figure 1. Figure 1: Illustration of context-dependent field importance [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of SPARC. SPARC compresses each multi-field historical interaction into a single backbone [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Slot-level routing patterns learned by CAR. The [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 7 linked inside Pith

  1. [1]

    Yimeng Bai, Chang Liu, Yang Zhang, Dingxian Wang, Frank Yang, Andrew Rabinovich, Wenge Rong, and Fuli Feng. 2025. Bi-Level Optimization for Gener- ative Recommendation: Bridging Tokenization and Generation.arXiv preprint arXiv:2510.21242(2025)

    Pith/arXiv arXiv 2025

  2. [2]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. InProceedings of the 10th ACM conference on recommender systems. 191–198

    2016

  3. [3]

    Xinyan Fan, Zheng Liu, Jianxun Lian, Wayne Xin Zhao, Xing Xie, and Ji-Rong Wen. 2021. Lighter and better: low-rank decomposed self-attention networks for next-item recommendation. InProceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 1733–1737

    2021

  4. [4]

    Kairui Fu, Changfa Wu, Kun Yuan, Binbin Cao, Dunxian Huang, Yuliang Yan, Junjun Zheng, Jianning Zhang, Silu Zhou, Jian Wu, et al. 2026. RankGR: Rank- Enhanced Generative Retrieval with Listwise Direct Preference Optimization in Recommendation.arXiv preprint arXiv:2602.08575(2026)

    arXiv 2026

  5. [5]

    Kairui Fu, Tao Zhang, Shuwen Xiao, Ziyang Wang, Xinming Zhang, Chenchi Zhang, Yuliang Yan, Junjun Zheng, Yu Li, Zhihong Chen, et al . 2025. Forge: Forming semantic identifiers for generative retrieval in industrial datasets.arXiv preprint arXiv:2509.20904(2025)

    Pith/arXiv arXiv 2025

  6. [6]

    Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. 2023. Learning vector-quantized item representation for transferable sequential recommenders. InProceedings of the ACM Web Conference 2023. 1162–1171

    2023

  7. [7]

    Yupeng Hou, Binbin Hu, Zhiqiang Zhang, and Wayne Xin Zhao. 2022. Core: simple and effective session-based recommendation within consistent represen- tation space. InProceedings of the 45th international ACM SIGIR conference on research and development in information retrieval. 1796–1801

    2022

  8. [8]

    Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating long semantic ids in parallel for recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 956–966

    2025

  9. [9]

    Wenyue Hua, Shuyuan Xu, Yingqiang Ge, and Yongfeng Zhang. 2023. How to index item ids for recommendation foundation models. InProceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region. 195–204

    2023

  10. [10]

    Chengkai Huang, Tianqi Gao, Hongtao Huang, Quan Z Sheng, and Lina Yao. 2026. Beyond Item Order: Temporal Gap Tokenization for Generative Recommendation with Semantic IDs.arXiv preprint arXiv:2607.03918(2026)

    Pith/arXiv arXiv 2026

  11. [11]

    Jianchao Ji, Zelong Li, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Juntao Tan, and Yongfeng Zhang. 2024. Genrec: Large language model for generative recommen- dation. InEuropean Conference on Information Retrieval. Springer, 494–502

    2024

  12. [12]

    Clark Mingxuan Ju, Liam Collins, Leonardo Neves, Bhuvesh Kumar, Louis Yufeng Wang, Tong Zhao, and Neil Shah. 2025. Generative Recommendation with Seman- tic IDs: A Practitioner’s Handbook. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 6420–6425

    2025

  13. [13]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE international conference on data mining (ICDM). IEEE, 197–206

    2018

  14. [14]

    Xinyu Lin, Haihan Shi, Wenjie Wang, Fuli Feng, Qifan Wang, See-Kiong Ng, and Tat-Seng Chua. 2025. Order-agnostic identifier for large language model-based generative recommendation. InProceedings of the 48th international ACM SIGIR conference on research and development in information retrieval. 1923–1933

    2025

  15. [15]

    Chang Liu, Yimeng Bai, Xiaoyan Zhao, Yang Zhang, Fuli Feng, and Wenge Rong

  16. [16]

    Chang Liu, Yimeng Bai, Xiaoyan Zhao, Yang Zhang, Qifan Wang, Fuli Feng, and Wenge Rong. 2026. Intuition-Guided Latent Reasoning for LLM-Based Recommendation. arXiv:2606.27684 [cs.IR] https://arxiv.org/abs/2606.27684

    Pith/arXiv arXiv 2026

  17. [17]

    Chang Liu, Xiaoguang Li, Guohao Cai, Zhenhua Dong, Hong Zhu, and Lifeng Shang. 2021. Noninvasive self-attention for side information fusion in sequential recommendation. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 4249–4256

    2021

  18. [18]

    Enze Liu, Bowen Zheng, Cheng Ling, Lantao Hu, Han Li, and Wayne Xin Zhao

  19. [19]

    Langming Liu, Liu Cai, Chi Zhang, Xiangyu Zhao, Jingtong Gao, Wanyu Wang, Yifu Lv, Wenqi Fan, Yiqi Wang, Ming He, et al. 2023. Linrec: Linear attention mechanism for long-term sequential recommender systems. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 289–299

    2023

  20. [20]

    InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Generative recommender with end-to-end learnable item tokenization. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 729–739

  21. [21]

    Luyi Ma, Wanjia Zhang, Kai Zhao, Abhishek Kulkarni, Lalitesh Morishetti, Anjana Ganesh, Ashish Ranjan, Aashika Padmanabhan, Jianpeng Xu, Jason HD Cho, et al

  22. [22]

    Zihan Liu, Yupeng Hou, and Julian McAuley. 2024. Multi-behavior generative recommendation. InProceedings of the 33rd ACM international conference on information and knowledge management. 1575–1585

    2024

  23. [23]

    Anima Singh, Trung Vu, Nikhil Mehta, Raghunandan Keshavan, Maheswaran Sathiamoorthy, Yilin Zheng, Lichan Hong, Lukasz Heldt, Li Wei, Devansh Tandon, et al. 2024. Better generalization with semantic ids: A case study in ranking for recommendations. InProceedings of the 18th ACM Conference on Recommender Systems. 1039–1044

    2024

  24. [24]

    InProceedings of the Nineteenth ACM Conference on Recommender Systems

    Grace: Generative recommendation via journey-aware sparse attention on chain-of-thought tokenization. InProceedings of the Nineteenth ACM Conference on Recommender Systems. 135–144

  25. [25]

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al

  26. [26]

    Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommenda- tion via convolutional sequence embedding. InProceedings of the eleventh ACM international conference on web search and data mining. 565–573

    2018

  27. [27]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Chang Liu et al. Table 3: Statistical details of the evaluation datasets. Dataset #Users #Items #Interactions Sparsity TaoBao21 million 0.27 billion 26 b...

    2017

  28. [28]

    Jiajie Su, Chaochao Chen, Zibin Lin, Xi Li, Weiming Liu, and Xiaolin Zheng

  29. [29]

    InProceedings of the 31st ACM international conference on multimedia

    Personalized behavior-aware transformer for multi-behavior sequential recommendation. InProceedings of the 31st ACM international conference on multimedia. 6321–6331

  30. [30]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  31. [31]

    Ming Xia, Zhiqin Zhou, Guoxin Ma, and Dongmin Huang. 2026. Unleash the Potential of Long Semantic IDs for Generative Recommendation.arXiv preprint arXiv:2602.13573(2026)

    arXiv 2026

  32. [32]

    Yueqi Xie, Peilin Zhou, and Sunghun Kim. 2022. Decoupled side information fusion for sequential recommendation. InProceedings of the 45th international ACM SIGIR conference on research and development in information retrieval. 1611– 1621

    2022

  33. [33]

    Yi Xu, Moyu Zhang, Chenxuan Li, Zhihao Liao, Haibo Xing, Hao Deng, Jinxin Hu, Yu Zhang, Xiaoyi Zeng, and Jing Zhang. 2026. Mmq: Multimodal mixture- of-quantization tokenization for semantic id generation and user behavioral adaptation. InProceedings of the Nineteenth ACM International Conference on Web Search and Data Mining. 788–797

    2026

  34. [34]

    Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2024. Learnable item tokenization for generative recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2400–2409

    2024

  35. [35]

    Yuhao Wang, Junwei Pan, Xinhang Li, Maolin Wang, Yuan Wang, Yue Liu, Dapeng Liu, Jie Jiang, and Xiangyu Zhao. 2025. Empowering large language model for sequential recommendation via multimodal embeddings and semantic ids. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 3209–3219

    2025

  36. [36]

    Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, et al . 2024. Eager: Two-stream generative recommender with behavior-semantic collaboration. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3245–3254

    2024

  37. [37]

    Tingting Zhang, Pengpeng Zhao, Yanchi Liu, Victor S Sheng, Jiajie Xu, Deqing Wang, Guanfeng Liu, Xiaofang Zhou, et al . 2019. Feature-level deeper self- attention network for sequential recommendation.. InIJCAI. 4320–4326

    2019

  38. [38]

    Bowen Zheng, Hongyu Lu, Yu Chen, Wayne Xin Zhao, and Ji-Rong Wen. 2025. Universal item tokenization for transferable generative recommendation.arXiv preprint arXiv:2504.04405(2025)

    Pith/arXiv arXiv 2025

  39. [39]

    Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S3-rec: Self-supervised learning for se- quential recommendation with mutual information maximization. InProceedings of the 29th ACM international conference on information & knowledge management. 1893–1902

    2020

  40. [40]

    Wencai Ye, Mingjie Sun, Shaoyun Shi, Peng Wang, Wenjin Wu, and Peng Jiang

  41. [41]

    InProceedings of the 34th ACM International Conference on Information and Knowledge Management

    DAS: Dual-Aligned Semantic IDs Empowered Industrial Recommender System. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 6217–6224

  42. [42]

    Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis, Joemon M Jose, and Xi- angnan He. 2019. A simple convolutional generative network for next item recommendation. InProceedings of the twelfth ACM international conference on web search and data mining. 582–590

    2019

  43. [43]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152(2024)

    Pith/arXiv arXiv 2024

  44. [47]

    Jieming Zhu, Mengqun Jin, Qijiong Liu, Zexuan Qiu, Zhenhua Dong, and Xiu Li

  45. [2019]

    InProceedings of the 28th ACM international conference on information and knowledge management

    BERT4Rec: Sequential recommendation with bidirectional encoder rep- resentations from transformer. InProceedings of the 28th ACM international conference on information and knowledge management. 1441–1450

  46. [2023]

    Recommender systems with generative retrieval.Advances in Neural Information Processing Systems36 (2023), 10299–10315

    2023

  47. [2024]

    InProceedings of the 18th ACM Conference on Recommender Systems

    Cost: Contrastive quantization based semantic tokenization for generative recommendation. InProceedings of the 18th ACM Conference on Recommender Systems. 969–974. A Dataset We evaluate the effectiveness of SPARC on both industrial and public datasets. For the industrial dataset, denoted asTaoBao, we collect one day of real-world interaction data from Tao...

    2007

  48. [2025]

    arXiv:2506.15576 [cs.IR] https://arxiv.org/abs/2506.15576

    DiscRec: Disentangled Semantic-Collaborative Modeling for Generative Recommendation. arXiv:2506.15576 [cs.IR] https://arxiv.org/abs/2506.15576

    Pith/arXiv arXiv