Pith. sign in

REVIEW 3 major objections 5 minor 57 references

SITA claims that by organizing compressed user interests into semantic groups aligned with item codes, a recommender can be target-aware and globally informed while keeping online cost O(BNd) and per-user storage O(NK).

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 ↗

SITA learns semantic interest tokens per user and selects them with per-item semantic codes, giving target-aware long-sequence modeling at O(N) inference cost.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection SITA's architecture is clever and the deployment results are credible, but the 'semantic' alignment claim is under-supported and needs an ablation with random SIDs before I'd trust it. the 3 major comments →

arxiv 2608.03692 v1 pith:3RJZEWRA submitted 2026-08-04 cs.IR

SITA: Semantic Interest Tokens for Target-Aware Compression in Long-Sequence Recommendation

Rui Zhou , Bo Chen , Qinglin Jia , Jiezhou Ji , Chaoyi Ma , Ruiming Tang , Hao Wang , Enhong Chen This is my paper
classification cs.IR
keywords long-sequence recommendationtarget-aware modelinginterest compressionsemantic identifiersvector quantizationclick-through rate predictionstructured interest tokensmultimodal recommendation
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.

The reading

Long-sequence recommendation has faced a trade-off: retrieval-based methods adapt to the target item but pay target-dependent inference cost and may discard global interests, while compression-based methods keep global interests and cheap inference but produce the same representation for every candidate. SITA is a proposal to break that trade-off. The paper claims that by first assigning every item a compositional semantic identifier (N codebooks of K codes), then compressing each user's full behavior history into N×K structured interest tokens organized by those same codebooks, a recommender can store one compact token set per user and still compute a target-specific user representation by looking up the tokens indexed by the candidate item's identifier. If correct, target-aware global modeling is achievable at the same order of online cost as pure compression, and the paper reports consistent gains over retrieval- and compression-based baselines on two public datasets and meaningful AUC/GAUC lifts in an industrial deployment.

Core claim

The paper's central claim is that semantic identifiers can serve as the indexing structure that makes compressed interests target-selectable without per-user-item storage. BPQ learns N parallel codebooks of K codewords from item multimodal representations, giving each item a structured semantic identifier while representing K^N possible codes from only NK units. SIC compresses the whole behavior sequence into NK interest tokens grouped by the same N codebooks, with intra-group SwiGLU refinement and group-level self-attention for cross-group exchange; these tokens are stored offline per user. At serving time, SGS reads the target item's semantic identifier and selects one token per group, the

What carries the argument

Balanced Parallel Quantization (BPQ) — N parallel semantic codebooks of K codewords that turn each item into a compact semantic identifier (SID); it reduces hypothetical per-user-item storage O(|U||V|) to O(|U|NK) while retaining K^N combinatorial patterns. Structured Interest Compression (SIC) — stacked blocks that turn the full behavior sequence into NK structured interest tokens via cross-attention, group-specific SwiGLU intra-group modeling, and group-level self-attention inter-group interaction. SID-Guided Selection (SGS) — at serving time, one index lookup per codebook selects N tokens from the stored token set for the candidate item. Together they make target-aware selection a cheap i

Load-bearing premise

The framework assumes the user-behavior embeddings fed into the compression blocks already live in the same semantic space as the item codebooks, so that selecting tokens by the candidate item's semantic code is meaningful; the paper never states how those user embeddings are constructed.

What would settle it

Train SITA on XLong twice, once with user sequence embeddings formed from BPQ codebook embeddings and once from plain item-ID embeddings while keeping everything else fixed; if AUC does not drop, the claimed semantic correspondence between identifiers and interest groups is not doing the work. Alternatively, permute the mapping from learned semantic identifiers to interest-token groups after training: a large performance drop confirms the semantic structure, while no drop shows the selection is incidental.

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

If this is right

  • Target-aware modeling no longer requires target-dependent computation over the raw sequence; online cost stays O(BNd), independent of sequence length.
  • Stored user representations can remain compact (NK tokens per user) while supporting K^N distinct target-specific activation patterns.
  • Compression-based recommenders can add target adaptation without abandoning offline sequence encoding, so long histories remain practical at scale.
  • Item semantic coding and user interest modeling become one structured space: semantic identifiers index both items and interest groups.
  • The reported industrial gains suggest the design transfers from public benchmarks to deployed click-through rate ranking.

Where Pith is reading between the lines

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

  • A direct test the paper does not run: train SITA with the user sequence embeddings S_u constructed from plain item-ID embeddings rather than BPQ-derived embeddings; if the gains survive, target-awareness is not actually semantically grounded in the codebook.
  • The same BPQ/SIC/SGS pattern could be extended beyond long behavior sequences to other large user-side memories, such as short-term session buffers or multimodal action logs, with the same indexing trick.
  • Because NK tokens are all that is stored per user, the framework implies that per-user memory can be held in embedding tables, opening the door to richer token budgets or more frequent user updates at fixed storage.
  • One can test whether the combinatorial capacity is real by holding NK fixed while varying N and K; the paper's own hyperparameter study indicates a balanced configuration outperforms lopsided ones.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper proposes SITA, a target-aware compression framework for long-sequence CTR prediction. BPQ learns N parallel codebooks over item-side multimodal representations, producing structured semantic identifiers (SIDs). SIC compresses the user's full behavior sequence into N×K learnable interest tokens via cross-attention, group-specific SwiGLU blocks, and group-level self-attention. At serving time, SGS selects one token per group using the target item's SID, yielding a target-aware user representation with online complexity O(BNd) and per-user storage O(NK). Experiments on Taobao-MM, XLong, and two industrial scenarios report consistent gains over retrieval- and compression-based baselines, with relative AUC/GAUC improvements of about 0.05–0.08% in the industrial study.

Significance. If the central design claim holds, SITA is a meaningful step: it offers target-dependent user representations without per-user-item storage and with a small, fixed token set per user. The complexity analysis in Section 4.5 is internally consistent, and the offline-computable, online-accessible token storage is practical. The paper also reports an industrial deployment study, which is a strength. However, the load-bearing 'semantic correspondence' between BPQ SIDs and SIC interest tokens is not established: the construction of S_u is unspecified, no parameter tie or loss couples token groups to codebooks, and the ablations do not test the semantic-content hypothesis. The empirical support is additionally weakened by missing variance information and by a very small margin over the strongest baseline on Taobao-MM.

major comments (3)
  1. [Section 4.3, Eq. (8)] The embedded user behavior sequence S_u is never defined. Eq. (8) uses S_u as K/V in cross-attention, and Eq. (13) selects tokens indexed by the BPQ SID; the paper's claim of 'semantic correspondence' between interest tokens and SIDs is therefore load-bearing. The text in Section 5.3 asserts that SID-guided selection outperforms SimGS because of semantic correspondence, but no parameter tie, initialization, or loss explicitly connects token group n to codebook n, or token (n,k) to codeword (n,k). If S_u is item-ID embeddings, or if any fixed index assignment works, SITA reduces to a hashed token memory. Please specify S_u, add an alignment signal (e.g., initialize or regularize interest tokens toward the BPQ codeword embeddings), and test with randomized or permuted SID assignments.
  2. [Table 3, Section 5.2] No error bars or standard deviations are reported. SITA's best Taobao-MM AUC is 0.6550 versus 0.6543 for C-Former, an absolute margin of 0.0007. The dagger only states a paired t-test with p<0.05, without reporting the number of runs, the variance, or exact p-values. Given the tiny margin on this dataset, 'consistently outperforms' is not adequately supported. Please report means ± std over multiple seeds, confidence intervals, and test details for both public datasets.
  3. [Section 5.3, Figure 3] The SimGS ablation compares SID-guided selection with similarity-based Top-N selection, but it does not compare against a random or permuted SID assignment. End-to-end training can exploit any fixed index mapping as a regularizer, independent of whether the SIDs carry semantic content. A permutation test—e.g., randomly permuting the mapping from SIDs to token slots, or replacing BPQ SIDs with random codes—is needed to show that the semantic content of the identifiers, not merely the indexing structure, drives the observed gain. Without such a test, the claimed semantic contribution remains unsupported.
minor comments (5)
  1. [Section 4.2, Eq. (6)] The usage-balance loss uses 'soft assignment probabilities' p̄_n, but Eq. (4) only defines hard assignments. Please define the soft assignment distribution used to compute p̄_n.
  2. [Section 4.5] The complexity statement O(BNd) for 'target attention' should clarify whether this is a single-query cross-attention from the target representation to the N selected tokens. If it is self-attention over the N tokens, the cost would be O(BN²d), not O(BNd).
  3. [Section 5.6] The text says 'Figure 4 presents the relative improvements' but the industrial results are in Table 4. Please correct the cross-reference.
  4. [Figure 4(a)] The left subplot is described as showing 'the balanced configuration' best, but the axis labels and the mapping of 'Setting' to (N,K) values are unclear. Please make the configurations explicit and indicate which one is considered balanced.
  5. [General] The ACM Reference Format and conference year contain placeholder text ('Conference acronym ’XX, June 03–05, 2018'); the final manuscript must be updated.

Circularity Check

0 steps flagged

No significant circularity: SITA's held-out CTR gains and ablations are genuine measurements; SID-guided selection is target-aware by construction but the semantic-correspondence claim is empirical, not a derived tautology.

full rationale

After walking the derivation chain, I find no circular step that reduces a claimed prediction or result to its inputs by construction. SITA's components (BPQ, SIC, SGS) are trained end-to-end on the CTR objective and the reported AUC/GAUC improvements are measured on temporally split held-out data (Section 5.1.1, Table 3) and on an online industrial evaluation (Section 5.6, Table 4). The SID-guided selection in Eq. (13) is target-aware by construction, but the paper does not derive the semantic-correspondence claim from the index alignment alone; that claim is an empirical interpretation supported by ablations (Section 5.3) and qualitative case studies (Section 5.5). The self-citations (e.g., [50]) are background taxonomy references and are not load-bearing for the method's derivation. The main weakness—the unspecified relationship between S_u and the BPQ codebook space—is a completeness/validity concern, not a circularity: nothing in the equations defines the claimed semantic alignment as an input, so the results are not forced by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The central claims rest on a few hand-chosen hyperparameters, the availability of meaningful multimodal item embeddings, and the unstated compatibility between the sequence embedding and the BPQ codebook space. These are reasonable engineering assumptions but are not independently verified.

free parameters (5)
  • N (number of parallel codebooks / semantic groups) = 16 (Taobao-MM), 8 (XLong)
    Hyperparameter chosen by search; controls the number of interest token groups and the size of selection. Varying N changes AUC in Figure 4.
  • K (codebook size / tokens per group) = 16 (both datasets)
    Hyperparameter chosen by search; controls the number of codewords per codebook and tokens per group.
  • L (number of SIC blocks) = 4
    Chosen based on Figure 4(b), where performance peaks at L=4.
  • d (embedding dimension) = 16
    Fixed for all models; not tuned.
  • BPQ loss weights (reconstruction, codebook, commitment, usage balance) = not specified
    Influence the learned semantic space but the values are not reported.
axioms (4)
  • domain assumption Item multimodal embeddings are semantically meaningful and available for all items in both public and industrial datasets.
    BPQ learns semantic identifiers by reconstructing these embeddings (Section 4.2). If they are noisy or misaligned, the SIDs used for selection are not semantically coherent.
  • domain assumption The user behavior sequence embeddings S_u are embedded in a space compatible with the BPQ semantic space (or can be made compatible via the cross-attention).
    SIC cross-attention (Eq. 8) mixes S_u with interest tokens; the paper does not specify the embedding used for S_u, yet the semantic alignment of tokens depends on it.
  • domain assumption Offline AUC/GAUC on Taobao-MM and XLong are reliable proxies for online recommendation quality, and the reported small AUC gaps (as low as 0.0007 over the best baseline) are practically meaningful.
    The paper uses these to claim consistent outperformance; a paired t-test is reported but no effect sizes or confidence intervals are shown.
  • domain assumption Baselines are compared fairly under identical training protocols and tuned sufficiently.
    The paper states all models share the same training protocol for the same number of epochs, but does not show per-baseline hyperparameter tuning results.
invented entities (1)
  • Semantic Interest Tokens (Z_u) no independent evidence
    purpose: A per-user set of N*K learnable vectors, organized into N groups of K, that compress the full behavior sequence and support target-aware selection via SID lookup.
    Internal model parameters with no falsifiable handle outside the trained system; they are introduced as the paper's compression unit.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SITA: Semantic Interest Tokens for Target-Aware Compression in Long-Sequence Recommendation." pith.science (2026). https://pith.science/paper/3RJZEWRA

@misc{pith2026260803692,
  author       = {Pith},
  title        = {Pith review of: SITA: Semantic Interest Tokens for Target-Aware Compression in Long-Sequence Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3RJZEWRA}},
  note         = {Machine review of arXiv:2608.03692}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

As user behavior histories continue to grow on modern Internet platforms, effectively modeling long behavior sequences has become crucial for predicting user interests in candidate items. Existing methods have evolved along two directions. One line dynamically retrieves target-relevant behaviors from long histories, enabling target-aware modeling but requiring target-dependent computation during inference. The other line compresses entire behavior sequences into compact user representations, achieving high efficiency and scalability but sacrificing target-specific adaptation due to target-independent encoding. The key challenge is therefore to enable target-aware modeling while preserving the efficiency and scalability of compressed user representations. To address this challenge, we propose \textbf{SITA}, a target-aware compression framework for long-sequence recommendation. SITA enables target-aware compression by organizing compressed interests into semantic structures through semantic identifiers learned via parallel semantic quantization. Conditioned on the semantic identifier of the target item, SITA adaptively aggregates the corresponding structured interests to construct the target-specific user representation. Extensive experiments on public datasets and a large-scale industrial dataset demonstrate that SITA consistently outperforms representative baselines while maintaining strong scalability, highlighting its strong potential for real-world recommender systems.

Figures

Figures reproduced from arXiv: 2608.03692 by Bo Chen, Chaoyi Ma, Enhong Chen, Hao Wang, Jiezhou Ji, Qinglin Jia, Ruiming Tang, Rui Zhou.

Figure 1
Figure 1. Figure 1: Comparison of representative long-sequence mod [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SITA. BPQ maps each item into a semantic identifier (SID). SIC compresses the original behavior sequence [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation study of SITA. The left panel shows the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of target-aware selection in the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Examples of items sharing the same SID. Each row shows three representative items associated with one SID, where [PITH_FULL_IMAGE:figures/full_fig_p009_6.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

57 extracted references · 29 canonical work pages · 3 internal anchors

  1. [1]

    Yong Bai, Rui Xiang, Kaiyuan Li, Yongxiang Tang, Yanhua Cheng, Xialong Liu, Peng Jiang, and Kun Gai. 2025. Chime: A compressive framework for holistic interest modeling.arXiv preprint arXiv:2504.06780(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    Zheng Chai, Qin Ren, Xijun Xiao, Huizhi Yang, Bo Han, Sijun Zhang, Di Chen, Hui Lu, Wenlin Zhao, Lele Yu, et al . 2025. Longer: Scaling up long sequence modeling in industrial recommenders. InProceedings of the Nineteenth ACM Conference on Recommender Systems. 247–256

    2025

  3. [3]

    Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, et al. 2023. TWIN: TWo-stage interest network for lifelong user behavior modeling in CTR prediction at kuaishou. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3785–3794

    2023

  4. [4]

    Bo Chen, Jinlong Jiao, Tijian Hu, Ruihao Zhang, Yanzhi Liu, Chenghou Jin, Qinglin Jia, Baixuan He, Hechang Pan, Yiwu Liu, et al. 2026. UniFormer: Efficient and Unified Model-Centric Scaling for Industrial Recommendation.arXiv preprint arXiv:2606.27058(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  5. [5]

    Bo Chen, Yichao Wang, Zhirong Liu, Ruiming Tang, Wei Guo, Hongkun Zheng, Weiwei Yao, Muyu Zhang, and Xiuqiang He. 2021. Enhancing explicit and implicit feature interactions via information sharing for parallel deep CTR models. In Proceedings of the 30th ACM international conference on information & knowledge management. 3757–3766

    work page 2021

  6. [6]

    Qiwei Chen, Changhua Pei, Shanshan Lv, Chao Li, Junfeng Ge, and Wenwu Ou

  7. [7]

    Zhimin Chen, Chenyu Zhao, Ka Chun Mo, Yunjiang Jiang, Jane H Lee, Khush- hall Chandra Mahajan, Ning Jiang, Kai Ren, Jinhui Li, and Wen-Yun Yang. 2025. Massive memorization with hundreds of trillions of parameters for sequential transducer generative recommenders.arXiv preprint arXiv:2510.22049(2025)

    arXiv 2025

  8. [8]

    Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment.arXiv preprint arXiv:2502.18965 (2025)

    Pith/arXiv arXiv 2025

  9. [9]

    Ningya Feng, Junwei Pan, Jialong Wu, Baixu Chen, Ximei Wang, Xian Hu, Jie Jiang, and Mingsheng Long. 2025. Long-sequence recommendation models need decoupled embeddings. InInternational Conference on Learning Representations, Vol. 2025. 20123–20144

    work page 2025

  10. [10]

    Jingtong Gao, Bo Chen, Menghui Zhu, Xiangyu Zhao, Xiaopeng Li, Yuhao Wang, Yichao Wang, Huifeng Guo, and Ruiming Tang. 2024. Hierrec: Scenario-aware hierarchical modeling for multi-scenario recommendations. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 653–662

    2024

  11. [11]

    Lin Guan, Jia-Qi Yang, Zhishan Zhao, Beichuan Zhang, Bo Sun, Xuanyuan Luo, Jinan Ni, Xiaowen Li, Yuhang Qi, Zhifang Fan, et al. 2026. Make it long, keep it fast: End-to-end 10k-sequence modeling at billion scale on Douyin. InProceedings of the ACM Web Conference 2026. 7989–7998

    work page 2026

  12. [12]

    Huifeng Guo, Bo Chen, Ruiming Tang, Weinan Zhang, Zhenguo Li, and Xiuqiang He. 2021. An embedding learning framework for numerical features in ctr prediction. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2910–2918

    work page 2021

  13. [13]

    Ruijie Hou, Zhaoyang Yang, Yu Ming, Hongyu Lu, Zhuobin Zheng, Yu Chen, Qinsong Zeng, and Ming Chen. 2024. Cross-domain lifelong sequential modeling for online click-through rate prediction. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5116–5125

    work page 2024

  14. [14]

    Zheng Hu, Yuxin Chen, Yongsen Pan, Xu Yuan, Yuting Yin, Daoyuan Wang, Boyang Xia, Zefei Luo, Hongyang Wang, Songhao Ni, et al. 2026. Stop Treating Collisions Equally: Qualification-Aware Semantic ID Learning for Recommenda- tion at Industrial Scale.arXiv preprint arXiv:2603.00632(2026)

    arXiv 2026

  15. [15]

    Guowen Li, Yuepeng Zhang, Shunyu Zhang, Yi Zhang, Xiaoze Jiang, Yi Wang, and Jingwei Zhuo. 2026. SID-Coord: Coordinating Semantic IDs for ID-based Ranking in Short-Video Search.arXiv preprint arXiv:2604.10471(2026)

    Pith/arXiv arXiv 2026

  16. [16]

    Xinchen Luo, Jiangxia Cao, Tianyu Sun, Jinkai Yu, Rui Huang, Wei Yuan, Hezheng Lin, Yichen Zheng, Shiyao Wang, Qigen Hu, et al . 2025. Qarm: Quantitative alignment multi-modal recommendation at kuaishou. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 5915– 5922

    2025

  17. [17]

    Qiushi Pan, Hao Wang, Guoyuan An, Luankang Zhang, Wei Guo, and Yong Liu

    work page

  18. [18]

    Qi Pi, Weijie Bian, Guorui Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Practice on long sequential user behavior modeling for click-through rate prediction. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. 2671–2679

    work page 2019

  19. [19]

    Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. InProceedings of the 29th ACM International Conference on Information & Knowledge Management. 2685–2692

    2020

  20. [20]

    Jiarui Qin, Weinan Zhang, Xin Wu, Jiarui Jin, Yuchen Fang, and Yong Yu. 2020. User behavior retrieval for click-through rate prediction. InProceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval. 2347–2356

    work page 2020

  21. [21]

    Kan Ren, Jiarui Qin, Yuchen Fang, Weinan Zhang, Lei Zheng, Weijie Bian, Guorui Zhou, Jian Xu, Yong Yu, Xiaoqiang Zhu, et al. 2019. Lifelong sequential modeling with personalized memorization for user response prediction. InProceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval. 565–574

    2019

  22. [22]

    Rong Shan, Jianghao Lin, Chenxu Zhu, Bo Chen, Menghui Zhu, Kangning Zhang, Jieming Zhu, Ruiming Tang, Yong Yu, and Weinan Zhang. 2025. An automatic graph construction framework based on large language models for recommenda- tion. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 4806–4817

    work page 2025

  23. [23]

    Tingjia Shen, Hao Wang, Chuhan Wu, Jin Yao Chin, Wei Guo, Yong Liu, Huifeng Guo, Defu Lian, Ruiming Tang, and Enhong Chen. 2026. P-Law: predicting quantitative scaling law with entropy guidance in large recommendation models. Advances in Neural Information Processing Systems38 (2026), 12911–12940

    work page 2026

  24. [24]

    Zihua Si, Lin Guan, ZhongXiang Sun, Xiaoxue Zang, Jing Lu, Yiqun Hui, Xingchao Cao, Zeyu Yang, Yichen Zheng, Dewei Leng, et al. 2024. Twin v2: Scaling ultra- long user behavior sequence modeling for enhanced ctr prediction at kuaishou. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 4890–4897

    2024

  25. [25]

    Xing Tang, Chaohua Yang, Yuwen Fu, Dongyang Ao, Shiwei Li, Fuyuan Lyu, Dugang Liu, and Xiuqiang He. 2025. Retrieval Augmented Cross-Domain Life- Long Behavior Modeling for Enhancing Click-through Rate Prediction. InPro- ceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 4891–4900

    2025

  26. [26]

    Huimu Wang, Xingzhi Yao, Yiming Qiu, Qinghong Zhang, Haotian Wang, Yufan Cui, Songlin Wang, Sulong Xu, and Mingming Li. 2026. Towards Efficient and Generalizable Retrieval: Adaptive Semantic Quantization and Residual Knowl- edge Transfer.arXiv preprint arXiv:2602.23978(2026)

    Pith/arXiv arXiv 2026

  27. [27]

    Jiancheng Wang, Mingjia Yin, Hao Wang, and Enhong Chen. 2025. En- hancing CTR prediction with de-correlated expert networks.arXiv preprint arXiv:2505.17925(2025)

    arXiv 2025

  28. [28]

    Kefan Wang, Hao Wang, Wei Guo, Yong Liu, Jianghao Lin, Defu Lian, and Enhong Chen. 2025. DLF: Enhancing explicit-implicit interaction via dynamic low-order- aware fusion for CTR prediction. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2213– 2223

    work page 2025

  29. [29]

    Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. Dcn v2: Improved deep & cross network and practical lessons for web-scale learning to rank systems. InProceedings of the web conference 2021. 1785–1797

    2021

  30. [30]

    Xingmei Wang, Shiyao Wang, Wuchao Li, Jiaxin Deng, Song Lu, Defu Lian, and Guorui Zhou. 2025. Transformers are Good Clusterers for Lifelong User Behavior Sequence Modeling. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 3123–3132

    work page 2025

  31. [31]

    Zhuoxing Wei, Qi Liu, and Qingchen Xie. 2025. Deep Multiple Quantization Network on Long Behavior Sequence for Click-Through Rate Prediction. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 3090–3094

    2025

  32. [32]

    Zhuoxing Wei, Qingchen Xie, Qi Liu, and Jingsong Yu. 2026. DMGIN: How Multimodal LLMs Enhance Large Recommendation Models for Lifelong User Post- click Behaviors. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 15933–15940

    work page 2026

  33. [33]

    Bin Wu, Feifan Yang, Zhangming Chan, Yu-Ran Gu, Jiawei Feng, Chao Yi, Xiang- Rong Sheng, Han Zhu, Jian Xu, Mang Ye, et al. 2025. MUSE: A Simple Yet Effective Multimodal Search-Based Framework for Lifelong User Interest Modeling.arXiv preprint arXiv:2512.07216(2025)

    arXiv 2025

  34. [34]

    Tian Xia, Jiaqi Zhang, Yueyang Liu, Hongjian Dou, Tingya Yin, Jiangxia Cao, Xulei Liang, Tianlu Xie, Lihao Liu, Xiang Chen, et al. 2026. QARM V2: Quantitative Alignment Multi-Modal Recommendation for Reasoning User Sequence Modeling. arXiv preprint arXiv:2602.08559(2026)

    arXiv 2026

  35. [35]

    Xue Xia, Saurabh Joshi, Kousik Rajesh, Kangnan Li, Yangyi Lu, Nikil Pancha, Dhruvil Badani, Jiajing Xu, and Pong Eksombatchai. 2025. TransAct V2: Lifelong User Action Sequence Modeling on Pinterest Recommendation. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 6881–6882

    2025

  36. [36]

    Wenjia Xie, Hao Wang, Minghao Fang, Ruize Yu, Wei Guo, Yong Liu, Defu Lian, and Enhong Chen. 2025. Breaking the bottleneck: User-specific optimization and real-time inference integration for sequential recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 3333–3343

    work page 2025

  37. [37]

    Wenjia Xie, Hao Wang, Luankang Zhang, Rui Zhou, Defu Lian, and Enhong Chen

    work page

  38. [38]

    Xiang Xu, Hao Wang, Wei Guo, Luankang Zhang, Wanshan Yang, Runlong Yu, Yong Liu, Defu Lian, and Enhong Chen. 2025. Multi-granularity interest retrieval and refinement network for long-term user behavior modeling in ctr prediction. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 2745–2755

    2025

  39. [39]

    Yi Xu, Chaofan Fan, Jinxin Hu, Yu Zhang, Zeng Xiaoyi, and Jing Zhang. 2025. STORE: Semantic Tokenization, Orthogonal Rotation and Efficient Attention for Scaling Up Ranking Models.arXiv preprint arXiv:2511.18805(2025)

    arXiv 2025

  40. [40]

    Jing Yan, Liu Jiang, Jianfei Cui, Zhichen Zhao, Xingyan Bin, Feng Zhang, and Zuotao Liu. 2024. Trinity: Syncretizing multi-/long-tail/long-term interests all in one. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6095–6104

    work page 2024

  41. [41]

    Yufei Ye, Wei Guo, Jin Yao Chin, Hao Wang, Hong Zhu, Xi Lin, Yuyang Ye, Yong Liu, Ruiming Tang, Defu Lian, et al. 2025. Fuxi-𝛼: Scaling recommendation model with feature interaction enhanced transformer. InCompanion Proceedings of the ACM on Web Conference 2025. 557–566

    work page 2025

  42. [42]

    Yufei Ye, Wei Guo, Hao Wang, Luankang Zhang, Heng Chang, Hong Zhu, Yuyang Ye, Yong Liu, Defu Lian, and Enhong Chen. 2026. Fuxi-linear: Unleashing the power of linear attention in long-term time-aware sequential recommendation. arXiv preprint arXiv:2602.23671(2026)

    arXiv 2026

  43. [43]

    Mingjia Yin, Junwei Pan, Hao Wang, Ximei Wang, Shangyu Zhang, Jie Jiang, Defu Lian, and Enhong Chen. 2025. From feature interaction to feature generation: A generative paradigm of ctr prediction models.arXiv preprint arXiv:2512.14041 (2025)

    arXiv 2025

  44. [44]

    Hongwei Zhang, Qiqiang Zhong, Jiangxia Cao, Yiyang Lv, Huanjie Wang, Li- wei Guan, Jing Yao, Yiyu Wang, Junfeng Shu, Zhaojie Liu, et al . 2026. UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence.arXiv preprint arXiv:2605.09040(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  45. [45]

    Jiaqing Zhang, Hao Wang, Mingjia Yin, Bo Chen, Qinglin Jia, Rui Zhou, Ruiming Tang, ChaoYi Ma, and Enhong Chen. 2026. DIET: Learning to Distill Dataset Continually for Recommender Systems.arXiv preprint arXiv:2603.24958(2026)

    work page arXiv 2026

  46. [46]

    Jiaqing Zhang, Mingjia Yin, Hao Wang, Yawen Li, Yuyang Ye, Xingyu Lou, Jun- ping Du, and Enhong Chen. 2025. Td3: Tucker decomposition based dataset distillation method for sequential recommendation. InProceedings of the ACM on Web Conference 2025. 3994–4003

    work page 2025

  47. [47]

    Luankang Zhang, Kenan Song, Yi Quan Lee, Wei Guo, Hao Wang, Yawen Li, Huifeng Guo, Yong Liu, Defu Lian, and Enhong Chen. 2025. Killing two birds with one stone: Unifying retrieval and ranking with a single generative recom- mendation model. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2224–2234

    2025

  48. [48]

    Qihang Zhao, Zhongbo Sun, Xiaoyang Zheng, Xian Guo, Siyuan Wang, Zihan Liang, Mingcan Peng, Ben Chen, and Chenyi Lei. 2026. COINS: Semantic Ids Enhanced Cold Item Representation for Click-through Rate Prediction in E- commerce Search. InProceedings of the ACM Web Conference 2026. 8529–8532

    work page 2026

  49. [49]

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1059–1068

    2018

  50. [50]

    Rui Zhou, Qinglin Jia, Bo Chen, Peng Xu, Yijia Sun, Siyuan Lou, Chaoxin Fu, Mengyuan Fu, Guoming Shen, Zheli Zhou, et al. 2026. A Survey of User Lifelong Behavior Modeling: Perspectives on Efficiency and Effectiveness. (2026)

    work page 2026

  51. [51]

    Rui Zhou, Hao Wang, Wei Guo, Qinglin Jia, Wenjia Xie, Xiang Xu, Yong Liu, Defu Lian, and Enhong Chen. 2025. Mit: A multi-tower information transfer framework based on hierarchical task relationship modeling. InCompanion Proceedings of the ACM on Web Conference 2025. 651–660

    work page 2025

  52. [52]

    Wen-Ji Zhou, Yuhang Zheng, Yinfu Feng, Yunan Ye, Rong Xiao, Long Chen, Xiaosong Yang, and Jun Xiao. 2024. ENCODE: Breaking the trade-off between performance and efficiency in long-term user behavior modeling.IEEE Transac- tions on Knowledge and Data Engineering37, 1 (2024), 265–277

    2024

  53. [53]

    Yongchun Zhu, Guanyu Jiang, Jingwu Chen, Feng Zhang, Qi Wu, and Zuotao Liu

    work page

  54. [57]

    InCompanion Proceedings of the ACM on Web Conference

    Long-Term Interest Clock: Fine-Grained Time Perception in Streaming Rec- ommendation System. InCompanion Proceedings of the ACM on Web Conference

    work page

  55. [2021]

    arXiv preprint arXiv:2108.04468(2021)

    End-to-end user behavior retrieval in click-through rateprediction model. arXiv preprint arXiv:2108.04468(2021)

    Pith/arXiv arXiv 2021

  56. [2024]

    SITA: Semantic Interest Tokens for Target-Aware Compression in Long-Sequence Recommendation Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

    Breaking determinism: Fuzzy modeling of sequential recommendation using discrete state space diffusion model.Advances in neural information processing systems37 (2024), 22720–22744. SITA: Semantic Interest Tokens for Target-Aware Compression in Long-Sequence Recommendation Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

    work page 2024

  57. [2025]

    Revisiting scalable sequential recommendation with Multi-Embedding Approach and Mixture-of-Experts.arXiv preprint arXiv:2510.25285(2025)

    work page arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.