Pith. sign in

REVIEW 4 major objections 5 minor 55 references

GenCDSR claims that generative cross-domain sequential recommenders can cut inference latency by 85.1% and slightly improve accuracy by tokenizing items into shared-specific SIDs and decoding them in two passes.

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

2026-08-03 00:44 UTC pith:P6SAUSY4

load-bearing objection GenCDSR is a genuine engineering win for fast generative cross-domain recommendation, but the serial-parallel decoding speedup needs a recomputation baseline to confirm it isn't quietly losing accuracy. the 4 major comments →

arxiv 2607.28659 v1 pith:P6SAUSY4 submitted 2026-07-21 cs.AI

Empowering Cross-Domain Sequential Recommendation with Hybrid Tokenization and Serial-Parallel Decoding

Yuxuan Hu , Yuhao Wang , Tianbo Huang , Chao Zhang , Ziwei Liu , Lihua Zhang , Xiangyu Zhao This is my paper
classification cs.AI
keywords cross-domain sequential recommendationgenerative recommendationhybrid tokenizationresidual quantizationsemantic identifiersserial-parallel decodinginference latencyRQ-VAE
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 paper tries to establish that generative cross-domain sequential recommendation can be both accurate and fast if the item-identifier construction and the decoding strategy are designed together. It argues that existing tokenizers either ignore cross-domain collaborative correlations or wash out domain-specific distinctions, and that standard autoregressive decoding such as beam search is too slow for real-time serving while fully parallel decoding loses accuracy. GenCDSR's answer is a two-stage hybrid tokenization—shared codes for cross-domain commonality plus fine-grained domain-specific codes—matched to a serial-parallel decoder that generates the first identifier positions serially and the rest in parallel. If the claims hold, generative recommenders can approach beam-search accuracy at a fraction of the decoding cost, making real-time generative recommendation practical.

Core claim

The central discovery is that the hierarchical structure of semantic identifiers (SIDs) can be exploited twice: once during tokenization, by making the first stage of the identifier capture cross-domain shared semantics and the second stage capture domain-specific fine detail, and once during decoding, by generating the shared-level tokens serially while predicting the specific-level tokens in parallel. GenCDSR reports that this joint design outperforms state-of-the-art baselines by 1.5% on average in accuracy while reducing inference latency by 85.1% relative to beam search, evaluated on Clothing-Sports, Electronics-Phone, and Book-Movie datasets. The authors argue that the two-stage struct

What carries the argument

The load-bearing mechanism is the cross-domain serial-parallel decoding strategy over a two-stage SID. Items are tokenized by a multi-tower RQ-VAE: Stage 1 quantizes shared and domain-specific latent representations, fused by a Gumbel-Softmax router, into L1 tokens, and Stage 2 quantizes the residual into L2 domain-specific tokens. At inference, the first LLM call precomputes hidden states for L1 placeholder positions; Step 1 decodes the L1 tokens serially while a carried context state s_l is updated only by adding the embedding of each selected token; a second LLM call fills the L1 positions and Step 2 predicts all L2 positions in parallel conditioned on the carried state. This two-pass sta

Load-bearing premise

The fast decoder assumes that hidden states computed with placeholder tokens in a single forward pass, combined with a context state that only accumulates token embeddings, capture enough of the true autoregressive context to predict the later identifier positions in parallel; if that approximation is lossy, decoding quality falls toward that of naive parallel decoding.

What would settle it

Run the exact GenCDSR model with beam search over all L positions and compare against its serial-parallel decoding: if beam search achieves materially higher Hit@10 on the same test set, the state-carry parallel step is losing information. A cleaner test: fix the tokenizer and the Step 1 serial output, then decode the remaining L2 positions with beam search instead of the parallel Step 2 head; if accuracy rises, the parallel approximation is the bottleneck.

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

If this is right

  • If the central claim holds, generative cross-domain recommendation becomes viable in latency-sensitive settings, since the 85.1% latency reduction relative to beam search moves it toward real-time deployment.
  • The hybrid tokenizer's shared-specific design implies cross-domain collaborative signals can be injected into SIDs at tokenization time, not only at training time, improving downstream generative modeling.
  • Serial-parallel decoding preserves per-position accuracy better than fully parallel decoding, suggesting the two-stage SID structure is a natural fit for partial parallelization.
  • A balanced allocation of codebook levels (2 shared-specific plus 2 fine-grained) is claimed to be optimal, indicating the shared/specific capacity trade-off is tunable.
  • The framework is instantiated on two different backbone models with consistent results, suggesting the improvements transfer across backbone sizes.

Where Pith is reading between the lines

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

  • The state-carry approximation could be stress-tested by extending it to longer SID sequences or more than two domains; the same two-pass idea might apply to any hierarchical generative retrieval system where identifiers have a coarse-to-fine structure.
  • The Gumbel-Softmax router in tokenization could be trained jointly with the recommender rather than separately, potentially yielding SIDs even better aligned with the downstream decoding objective.
  • The reported latency numbers depend on the L1:L2 split and on placeholder-conditioned hidden states; rebalancing the split or using a stronger backbone might shift the accuracy-latency Pareto frontier.
  • A testable extension is to apply the serial-parallel decoder to single-domain generative recommendation, where the shared-specific stage would degenerate but the two-pass scheme might still cut latency.

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. GenCDSR is a generative cross-domain sequential recommendation (CDSR) framework that pairs a cross-domain hybrid tokenization mechanism (shared-specific multi-tower RQ-VAE with fine-grained specific codebooks) with a serial-parallel decoding strategy. The tokenizer is designed to capture both shared and domain-specific item semantics, and the decoder leverages the hierarchical SID structure to predict the first L1 tokens serially and the remaining L2 tokens in parallel, thereby reducing inference latency. The paper reports experiments on three public cross-domain datasets (Clothing–Sports, Electronics–Phone, Book–Movie) with T5 and Qwen3-0.6B backbones, claiming an average accuracy gain of 1.5% over SOTA baselines and an average latency reduction of 85.1% compared with beam search.

Significance. If the claimed gains hold, the paper makes a useful practical contribution: it directly addresses two known limitations of generative recommendation in CDSR—cross-domain tokenization and decoding efficiency. The paper includes several strengths: open-sourced code and datasets, comparisons against a wide range of single-domain and cross-domain baselines, an efficiency study against MTP/NEZHA/Beam Search, ablations of the two tokenization stages, and a feature-fidelity analysis. The serial-parallel decoding idea is interesting and could be applicable beyond CDSR. However, the central accuracy-latency trade-off rests on an under-specified train/inference protocol for the serial-parallel decoder, and several experimental-reporting choices weaken the empirical claims. These issues are addressable, but they need to be resolved before the results can be fully relied upon.

major comments (4)
  1. [§3.3, Eqs. (19)–(21)] The serial-parallel decoder computes all Step-1 logits p^(1)_l from hidden states h^1_l obtained in a single forward pass in which every L1 position is a placeholder. The context state s_l is updated only by adding the embedding e_l of the previously selected token. Thus, for l>1, h^1_l does not contain the selected tokens at earlier positions, and the only information carried from those tokens is an additive embedding. The paper never states whether Unified Recommender Training (Eq. 22) uses the same placeholder-conditioned forward passes or standard teacher forcing with ground-truth token embeddings. If training uses teacher forcing, the inference-time state-carry is off-distribution and Eq. (24) is not the objective optimized in Eq. (22); if training uses placeholders, the approximation is self-consistent but still may lose accuracy. Please clarify the training protocol and add a reco
  2. [§4.6, Figure 3] The L1:L2 codebook allocation is selected by sweeping over {0:4, 1:3, 2:2, 3:1, 4:0} and choosing 2:2 on the basis of test-set HR@10/NDCG@10. No validation-based selection or multiple-comparison correction is reported. This means the headline numbers in Table 2 are partially tuned on the test data, which can inflate the reported gains. Please select hyperparameters on a validation split, or report results for all configurations with an explicit statement of the selection procedure, and identify which Table 2 numbers correspond to the selected configuration.
  3. [§4.1.3, Table 2] Experiments are averaged over 3 runs, but no standard deviations, confidence intervals, or per-run results are reported. The significance stars in Table 2 are therefore not interpretable: many reported margins are very small (e.g., Clothing N@5: 0.5373 vs 0.5361; Sports H@5: 0.5015 vs 0.4979), and a two-sided t-test on 3 runs is sensitive to variance that is not disclosed. Please report per-run standard deviations, exact p-values or effect sizes, and the number of test users used in the significance tests.
  4. [§4.1.3, Table 3] Table 3 shows that in several settings the serial-parallel decoder is less accurate than Beam Search (e.g., T5 Clothing H@5: 0.5869 vs 0.6013; Qwen Book H@10: 0.1519 vs 0.1658), so the statement that it 'largely preserves generation fidelity' is not uniformly supported. In addition, §4.1.3 states that beam search with size 20 is used during inference, but §3.3 does not describe how beam search operates together with the serial-parallel state-carry procedure. Please specify the exact inference algorithm, including how beams are scored and expanded, and report the average accuracy gap vs Beam Search across all datasets and backbones, as well as any settings where the gap exceeds a practically meaningful threshold.
minor comments (5)
  1. [Abstract and §1] The abstract and introduction say '85.1% reduction in inference latency compared with SOTA baselines,' but the 85.1% figure is computed only against Beam Search, not against the best accuracy baseline GenCDR or against MTP/NEZHA. Please disambiguate the comparison baseline for the latency claim.
  2. [§3.3, Eq. (20)] The notation Head1(h^1_l, s_l) is not formally defined: is the head a function of the hidden state and the state vector through concatenation, addition, or some other operation? Also, the update s_{l+1}=s_l+e_l in Eq. (21) inside the parallel loop cannot affect p^(2)_l for later positions because all Step-2 logits are computed with s_{L1+1}; please clarify whether the sequential update is meant to describe training, inference, or both.
  3. [§4.3, Table 3] Latency is reported in milliseconds without standard deviations, batch size, warm-up details, or a specification of how many test samples were measured. Since the latency comparison is a central contribution, these details are needed for reproducibility.
  4. [§4.5, Eq. (25)] The feature-fidelity metric is defined on concatenated reconstructed embeddings, but the paper does not specify which tokenization output is used as \hat{x}_A, \hat{x}_B—the Stage-1-only reconstruction, the full two-stage reconstruction, or the router-fused output. Please clarify.
  5. [§4.6, Figure 3] The caption says 'across two datasets' but the figure contains four subfigures (Clothing, Sports, Book, Movie), and no results are shown for Electronics–Phone. Please either correct the caption or provide the missing results.

Circularity Check

0 steps flagged

No significant circularity: the accuracy and latency claims are empirical comparisons against external baselines; the serial-parallel decoder is an approximation, not a circular derivation.

full rationale

The paper's central claims are measured, not derived: the 1.5% average accuracy improvement and 85.1% average latency reduction are reported from comparisons with external baselines on three public datasets (Tables 2 and 3) using leave-one-out validation/testing. The tokenization and decoding designs are trained and then evaluated on held-out interactions, so no fitted parameter is renamed as a prediction. The serial-parallel decoding in Section 3.3, Eqs. (19)-(21), does define an approximation: hidden states h^1_l are computed with placeholder tokens while a context state s_l is updated by adding selected-token embeddings. This is a modeling assumption that may carry a train/inference mismatch and is under-specified—the paper does not state whether Eq. (22)'s training objective uses the same placeholder-conditioned forward passes or teacher forcing. However, that is a correctness/evaluation risk, not circularity: the reported accuracy and latency are measured outcomes, not consequences entailed by the equations. The paper also cites some prior work by overlapping authors (e.g., LLM4CDSR [29], LLM-EDT [30], NEZHA [43]), but these serve as baselines or related work and are not load-bearing justification for the central result. There is no uniqueness theorem, no ansatz smuggled in solely by citation, and no known result merely renamed as a new contribution. The feature-fidelity analysis (Eq. 25) is aligned with the reconstruction objective, but it is used as a diagnostic comparison rather than as a fitted prediction. Overall, the derivation chain is self-contained and externally benchmarked; no circular step meets the evidentiary bar for a 6+ finding.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard generative-recommendation machinery (RQ-VAE, LLM backbones, LoRA) plus a purpose-built router and placeholder-based parallel decoding. The most consequential free parameter is the L1:L2 split, which is tuned on the same test metrics used for final comparison. No new physical entities are introduced.

free parameters (5)
  • L1:L2 codebook allocation = 2:2
    Total code levels L=4; the split between shared-specific Stage 1 and fine-grained Stage 2 is chosen by sweeping on test metrics in RQ5 (Figure 3).
  • per-level codebook size/dimension = K=256, dim=128
    Set in Section 4.1.3 without ablation; determines tokenization capacity and affects downstream accuracy.
  • Gumbel-Softmax temperature tau = not reported
    Controls router hardness in Eq. (12); no value or schedule is provided.
  • LoRA rank and alpha = T5: r=8, alpha=32; Qwen3-0.6B: r=8, alpha=16
    Chosen in Section 4.1.3 for domain-specific fine-tuning; no sensitivity analysis.
  • RQ-VAE beta = not reported
    Balance between codebook learning and encoder update in Eq. (9); value absent from implementation details.
axioms (4)
  • domain assumption RQ-VAE residual quantization maps item embeddings into SIDs that preserve enough information for downstream next-item prediction.
    Section 2.2 presents RQ-VAE as the tokenizer; the entire generative recommendation pipeline depends on SID fidelity.
  • domain assumption Mean-pooled last-layer hidden states of LLaMA-7B or Qwen2.5-7B capture item semantics sufficient for cross-domain recommendation.
    Section 3.2.1 and Section 4.1.3 use these text embeddings as the only item input; no other item signal is used for tokenization.
  • domain assumption Merging each user's chronological interactions across domains into one stream preserves all information needed to predict the next item in each domain.
    Section 2.1 defines the task on the merged stream S-bar_u; if domain-specific ordering matters beyond the merged order, the formulation may lose signal.
  • domain assumption Causal transformer hidden states computed with placeholder tokens can be reused, together with a carried state vector, to predict future SID tokens in parallel without accuracy loss.
    Eqs. (19)-(21) assume h^1_l and h^2_l from placeholder-conditioned forward passes are valid substitutes for true autoregressive states of decoded tokens.

pith-pipeline@v1.3.0-alltime-deepseek · 16573 in / 10361 out tokens · 93498 ms · 2026-08-03T00:44:35.899012+00:00 · methodology

0 comments
read the original abstract

Cross-domain sequential recommendation (CDSR) aims to model users' dynamic interest transitions and sequential patterns across multiple domains. Recently, generative recommendation (GR) has emerged. It first learns semantic identifiers (SIDs) from item semantics and formulates recommendation as autoregressive generation. However, existing methods face two critical issues: (1) they ignore collaborative correlations across domains during tokenization, and (2) they adopt inefficient decoding strategies, such as beam search, during generation, which hinders real-time deployment. To address these limitations, we propose GenCDSR, an effective and efficient generative framework for CDSR. Specifically, we design a cross-domain hybrid tokenization mechanism with a multi-tower architecture to jointly capture cross-domain commonalities and domain-specific distinctions through hierarchical shared-specific and fine-grained codebooks. Furthermore, we develop a cross-domain serial-parallel decoding strategy that leverages the hierarchical SID structure to partially parallelize generation, significantly reducing inference latency while preserving generation consistency. Experiments on three public datasets show that GenCDSR achieves an average accuracy improvement of 1.5 percent and an average inference latency reduction of 85.1 percent compared with state-of-the-art baselines. The implementation code and datasets are available online: https://github.com/Applied-Machine-Learning-Lab/RecSys2026_GenCDSR.

Figures

Figures reproduced from arXiv: 2607.28659 by Chao Zhang, Lihua Zhang, Tianbo Huang, Xiangyu Zhao, Yuhao Wang, Yuxuan Hu, Ziwei Liu.

Figure 1
Figure 1. Figure 1: Overview of the proposed GenCDSR framework. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ablation study and feature fidelity comparison (higher is better for all subfigures). (a-d) Ablation results on the [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Impact of the codebook allocation between shared [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

55 extracted references · 14 linked inside Pith

  1. [1]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report.arXiv preprint arXiv:2309.16609(2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Jiangxia Cao, Xin Cong, Jiawei Sheng, Tingwen Liu, and Bin Wang. 2022. Con- trastive cross-domain sequential recommendation. InProceedings of the 31st ACM international conference on information & knowledge management. 138–147

    2022

  3. [3]

    Jian Chen, Yuxuan Hu, Haifeng Lu, Wei Wang, Min Yang, Chengming Li, and Xiping Hu. 2025. MGHFT: Multi-Granularity Hierarchical Fusion Transformer for Cross-Modal Sticker Emotion Recognition. InProceedings of the 33rd ACM International Conference on Multimedia. 5794–5803

    2025

  4. [4]

    Jian Chen, Wei Wang, Yuzhu Hu, Junxin Chen, Han Liu, and Xiping Hu. 2024. Tgca-pvt: Topic-guided context-aware pyramid vision transformer for sticker emotion recognition. InProceedings of the 32nd ACM International Conference on Multimedia. 9709–9718

    2024

  5. [5]

    Shu Chen, Zitao Xu, Weike Pan, Qiang Yang, and Zhong Ming. 2024. A survey on cross-domain sequential recommendation.arXiv preprint arXiv:2401.04971 (2024)

    Pith/arXiv arXiv 2024

  6. [6]

    Rene De La Briandais. 1959. File searching using variable length keys. InPapers presented at the the March 3-5, 1959, western joint computer conference. 295–298

    1959

  7. [7]

    Markus Freitag and Yaser Al-Onaizan. 2017. Beam search strategies for neural machine translation. InProceedings of the First Workshop on Neural Machine Translation. 56–60

    2017

  8. [8]

    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

  9. [9]

    Jingtong Gao, Xiangyu Zhao, Bo Chen, Fan Yan, Huifeng Guo, and Ruiming Tang. 2023. AutoTransfer: Instance transfer for cross-domain recommendations. InProceedings of the 46th international ACM SIGIR conference on research and development in information retrieval. 1478–1487

    2023

  10. [10]

    Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Rozière, David Lopez-Paz, and Gabriel Synnaeve. 2024. Better & faster large language models via multi-token prediction.arXiv preprint arXiv:2404.19737(2024)

    Pith/arXiv arXiv 2024

  11. [11]

    Xiao Han, Zimo Zhao, Wanyu Wang, Maolin Wang, Zitao Liu, Yi Chang, and Xiangyu Zhao. 2026. Data efficient adaptation in large language models via continuous low-rank fine-tuning.Advances in Neural Information Processing Systems38 (2026), 165157–165182

    2026

  12. [12]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  13. [13]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models.Iclr1, 2 (2022), 3

    2022

  14. [14]

    Peiyu Hu, Wayne Lu, and Jia Wang. 2026. From ids to semantics: A generative framework for cross-domain recommendation with adaptive semantic tokeniza- tion. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 14874–14882

    2026

  15. [15]

    Yuxuan Hu, Jian Chen, Yuhao Wang, Zixuan Li, Jing Xiong, Pengyue Jia, Wei Wang, Chengming Li, and Xiangyu Zhao. 2026. Emotion and intention guided multi-modal learning for sticker response selection. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 14883–14891

    2026

  16. [16]

    Yuxuan Hu, Minghuan Tan, Chenwei Zhang, Zixuan Li, Xiaodan Liang, Min Yang, Chengming Li, and Xiping Hu. 2024. APTNESS: Incorporating Appraisal Theory and Emotion Support Strategies for Empathetic Response Generation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management(Boise, ID, USA)(CIKM ’24). Association for C...

    arXiv 2024

  17. [17]

    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

  18. [18]

    Pengyue Jia, Yichao Wang, Shanru Lin, Xiaopeng Li, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. 2024. D3: A methodological exploration of domain division, modeling, and balance in multi-domain recommendations. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 8553–8561

    2024

  19. [19]

    Jinqiu Jin, Yang Zhang, Fuli Feng, and Xiangnan He. 2025. Generative Multi- Target Cross-Domain Recommendation.arXiv preprint arXiv:2507.12871(2025)

    Pith/arXiv arXiv 2025

  20. [20]

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

    2018

  21. [21]

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregressive image generation using residual quantization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11523–11532

    2022

  22. [22]

    Jinming Li, Wentao Zhang, Tian Wang, Guanglei Xiong, Alan Lu, and Gerard Medioni. 2023. GPT4Rec: A generative framework for personalized recommen- dation and user interests interpretation.arXiv preprint arXiv:2304.03879(2023)

    Pith/arXiv arXiv 2023

  23. [23]

    Pan Li and Alexander Tuzhilin. 2020. Ddtcdr: Deep dual transfer cross domain recommendation. InProceedings of the 13th international conference on web search and data mining. 331–339

    2020

  24. [24]

    Xinhang Li, Zhaopeng Qiu, Xiangyu Zhao, Zihao Wang, Yong Zhang, Chunxiao Xing, and Xian Wu. 2022. Gromov-wasserstein guided representation learning for cross-domain recommendation. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 1199–1208

    2022

  25. [25]

    Xiaopeng Li, Fan Yan, Xiangyu Zhao, Yichao Wang, Bo Chen, Huifeng Guo, and Ruiming Tang. 2023. Hamur: Hyper adapter for multi-domain recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 1268–1277

    2023

  26. [26]

    Zixuan Li, Binzong Geng, Jing Xiong, Yong He, Yuxuan Hu, Jian Chen, Ding- wei Chen, Xiyu Chang, Ngai Wong, Liang Zhang, et al. 2025. Ctr-sink: Atten- tion sink for language models in click-through rate prediction.arXiv preprint arXiv:2508.03668(2025)

    Pith/arXiv arXiv 2025

  27. [27]

    Zixuan Li, Jing Xiong, Fanghua Ye, Chuanyang Zheng, Xun Wu, Jianqiao Lu, Zhongwei Wan, Xiaodan Liang, Chengming Li, Zhenan Sun, et al. 2024. Uncer- taintyrag: Span-level uncertainty enhanced long-context modeling for retrieval- augmented generation.arXiv preprint arXiv:2410.02719(2024)

    Pith/arXiv arXiv 2024

  28. [28]

    Qidong Liu, Jiaxi Hu, Yutian Xiao, Xiangyu Zhao, Jingtong Gao, Wanyu Wang, Qing Li, and Jiliang Tang. 2024. Multimodal recommender systems: A survey. Comput. Surveys57, 2 (2024), 1–17

    2024

  29. [29]

    Qidong Liu, Xiangyu Zhao, Yejing Wang, Zijian Zhang, Howard Zhong, Chong Chen, Xiang Li, Wei Huang, and Feng Tian. 2025. Bridge the domains: Large language models enhanced cross-domain sequential recommendation. InProceed- ings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1582–1592

    2025

  30. [30]

    Ziwei Liu, Qidong Liu, Wanyu Wang, Yejing Wang, Tong Xu, Wei Huang, Chong Chen, Peng Chuan, and Xiangyu Zhao. 2025. LLM-EDT: Large Language Model Enhanced Cross-domain Sequential Recommendation with Dual-phase Training. arXiv preprint arXiv:2511.19931(2025)

    Pith/arXiv arXiv 2025

  31. [31]

    Ziru Liu, Jiejie Tian, Qingpeng Cai, Xiangyu Zhao, Jingtong Gao, Shuchang Liu, Dayou Chen, Tonghao He, Dong Zheng, Peng Jiang, et al. 2023. Multi-task recommendations with reinforcement learning. InProceedings of the ACM web conference 2023. 1273–1282

    2023

  32. [32]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. InInternational Conference on Learning Representations

    2019

  33. [33]

    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

  34. [34]

    Haokai Ma, Ruobing Xie, Lei Meng, Xin Chen, Xu Zhang, Leyu Lin, and Jie Zhou. 2024. Triple sequence learning for cross-domain recommendation.ACM Transactions on Information Systems42, 4 (2024), 1–29

    2024

  35. [35]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research21, 140 (2020), 1–67

    2020

  36. [36]

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

  37. [37]

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

  38. [38]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971(2023)

    Pith/arXiv arXiv 2023

  39. [39]

    Shoujin Wang, Liang Hu, Yan Wang, Longbing Cao, Quan Z Sheng, and Mehmet Orgun. 2019. Sequential recommender systems: challenges, progress and prospects.arXiv preprint arXiv:2001.04830(2019)

    Pith/arXiv arXiv 2019

  40. [40]

    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

  41. [41]

    Xinhua Wang, Houping Yue, Zizheng Wang, Liancheng Xu, and Jinyu Zhang. 2023. Unbiased and robust: External attention-enhanced graph contrastive learning for cross-domain sequential recommendation. In2023 IEEE International Conference on Data Mining Workshops (ICDMW). IEEE, 1526–1534

    2023

  42. [42]

    Yejing Wang, Zhaocheng Du, Xiangyu Zhao, Bo Chen, Huifeng Guo, Ruiming Tang, and Zhenhua Dong. 2023. Single-shot feature selection for multi-task recommendations. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 341–351. RecSys ’26, September 27-October 02, 2026, Minneapolis, MN, USA Yuxuan...

    2023

  43. [43]

    Yejing Wang, Shengyu Zhou, Jinyu Lu, Ziwei Liu, Langming Liu, Maolin Wang, Wenlin Zhang, Feng Li, Wenbo Su, Pengjie Wang, et al. 2025. NEZHA: A Zero- sacrifice and Hyperspeed Decoding Architecture for Generative Recommenda- tions.arXiv preprint arXiv:2511.18793(2025)

    arXiv 2025

  44. [44]

    Shiwen Wu, Fei Sun, Wentao Zhang, Xu Xie, and Bin Cui. 2022. Graph neural networks in recommender systems: a survey.ACM computing surveys55, 5 (2022), 1–37

    2022

  45. [45]

    Jing Xiong, Zixuan Li, Chuanyang Zheng, Zhijiang Guo, Yichun Yin, Enze Xie, Zhicheng Yang, Qingxing Cao, Haiming Wang, Xiongwei Han, et al. 2024. Dq-lore: Dual queries with low rank approximation re-ranking for in-context learning. In International Conference on Learning Representations, Vol. 2024. 41179–41203

    2024

  46. [46]

    Wujiang Xu, Qitian Wu, Runzhong Wang, Mingming Ha, Qiongxu Ma, Linxun Chen, Bing Han, and Junchi Yan. 2024. Rethinking cross-domain sequential recommendation under open-world assumptions. InProceedings of the ACM Web Conference 2024. 3173–3184

    2024

  47. [47]

    Zitao Xu, Shu Chen, Weike Pan, and Zhong Ming. 2025. A multi-view graph contrastive learning framework for cross-domain sequential recommendation. ACM Transactions on Recommender Systems3, 4 (2025), 1–28

    2025

  48. [48]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    Pith/arXiv arXiv 2025

  49. [49]

    Chao Zhang, Yuhao Wang, Derong Xu, Haoxin Zhang, Yuanjie Lyu, Yuhao Chen, Shuochen Liu, Tong Xu, Xiangyu Zhao, Yan Gao, et al. [n. d.]. Tearag: A token- efficient agentic retrieval-augmented generation framework.ACM Transactions on Information Systems([n. d.])

  50. [50]

    Kangning Zhang, Jiarui Jin, Yingjie Qin, Ruilong Su, Jianghao Lin, Yong Yu, and Weinan Zhang. 2024. Learning ID-free Item Representation with Token Crossing for Multimodal Recommendation.arXiv preprint arXiv:2410.19276(2024)

    Pith/arXiv arXiv 2024

  51. [51]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 1435–1448

    2024

  52. [52]

    Yongchun Zhu, Zhenwei Tang, Yudan Liu, Fuzhen Zhuang, Ruobing Xie, Xu Zhang, Leyu Lin, and Qing He. 2022. Personalized transfer of user preferences for cross-domain recommendation. InProceedings of the fifteenth ACM international conference on web search and data mining. 1507–1515

    2022

  53. [2015]

    Session-based recommendations with recurrent neural networks.arXiv preprint arXiv:1511.06939(2015)

    Pith/arXiv arXiv 2015

  54. [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

  55. [2023]

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

    2023