Pith. sign in

REVIEW 3 major objections 5 minor 50 references

Closing the Performance Gap in Generative Recommenders with Collaborative Tokenization and Efficient Modeling

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper argues that generative recommenders can match or beat ID-based sequential recommenders once item tokenization carries collaborative signal and generation is split into a temporal encoder and a cheap pointwise decoder.

desk verdict Solid, well-engineered generative recommender paper that closes much of the gap on small benchmarks; the large-scale 'surpassing' claim relies on unreplicated single runs. read the letter →

arxiv 2508.14910 v1 pith:7LAPTPOA submitted 2025-08-12 cs.IR cs.LG

Simon Lepage , Jeremie Mary , David Picard This is my paper · ORCID
classification cs.IRcs.LG
keywords generativerecommendationsequentialsemanticIDtokenizationcontrastivelearningresidualvectorquantizationRQ-TransformerSASRecbaselineitemretrieval
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Generative recommender systems replace the usual rank-and-retrieve step with autoregressively generating a discrete code that names the next item. They have promised to remove large embedding tables and approximate-nearest-neighbor search, but in practice they have lagged behind well-tuned ID-based models such as SASRec. This paper claims the gap has two fixable causes: item tokenization ignores collaborative signals, and the standard encoder-decoder setup is needlessly expensive. It introduces COSETTE, a contrastive tokenization method that folds co-occurrence information into quantized item codes, and MARIUS, a lightweight architecture that separates timeline modeling from code decoding. Together they bring generative recommenders to parity on small benchmarks and ahead of a strong SASRec++ baseline on large Amazon 2023 datasets, at lower inference cost.

What carries the argument

Two mechanisms carry the argument. COSETTE is a residual-quantization tokenizer (RQ-VAE-style codebooks) with a three-part loss: reconstruction of the semantic item embedding, quantization, and a sigmoid contrastive loss over pairs of items in sampled timelines; co-occurring items are pulled together in the quantized latent space, and a post-training reallocation step resolves remaining code collisions. MARIUS is a two-stage generative model: a Temporal Transformer computes a context vector per position from one fused token per item, then a Depth Transformer predicts the next item's code tuple level by level, pointwise, with no cross-attention to the full history. This decoupling makes per-i

What would settle it

Re-run SASRec++ and MARIUS+COSETTE on the identical Amazon 2023 5-core leave-last-out splits across five seeds; if the R@10 margin falls inside one standard deviation, the 'surpassing' claim does not survive. Conversely, re-running the dagger-marked baselines under the paper's protocol would show whether the apparent gap reduction comes from tokenization and architecture or from looser baseline numbers.

Watch

Extended reading notes

Core claim

The paper's central claim is that the generative-recommender gap is not intrinsic but is produced by two correctable design choices, and it demonstrates the correction. On the tokenization side, COSETTE trains a residual quantizer with a reconstruction loss for semantic embeddings plus a contrastive loss that treats items co-occurring in sampled user timelines as positives; the learned code tuples therefore carry collaborative signal without needing a pretrained recommender. On the modeling side, MARIUS replaces the T5-style encoder-decoder with a causal Temporal Transformer that processes one token per item and a Depth Transformer that autoregressively predicts the code levels of the next i

Load-bearing premise

The gap-closing claim hinges on the comparability of the baseline numbers: several SASRec and TIGER figures are taken from prior papers that may have used different filtering, splits, or metric implementations, and the large-scale comparison is a single test run per model.

Editorial extensions

If this is right

  • If correct, the generative paradigm can be adopted without the usual performance penalty: on large Amazon 2023 benchmarks, MARIUS+COSETTE beats a carefully modernized SASRec++ baseline, taking top R@10 on nine of ten datasets.
  • Tokenization should be optimized for the recommendation task, not just for content reconstruction: contrastive co-occurrence alignment outperforms both reconstruction-only RQ-VAE and alignment to pretrained SASRec embeddings in the paper's comparisons.
  • The decoupled temporal/depth design makes inference substantially cheaper: at sequence length 50, MARIUS trains on roughly 33x more items per second and decodes about 3x faster than TIGER, with fewer hallucinated item IDs.
  • Generative recommenders shift predictions from the most popular items to mid-popular items, raising recall where the ID-based baseline's popularity bias hurts most.
  • The gap between generative and ID-based recommenders is not a fixed property of the paradigm; it can be closed by task-aligned tokenization and architectural efficiency.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A broader implication, left implicit: many published generative-versus-baseline comparisons may be against under-trained baselines; the paper's +50% gain from modernizing SASRec's loss and data augmentation suggests part of the historical gap is a baseline artifact.
  • The co-occurrence contrastive objective assumes collaborative signal lives in pairwise timelines; a natural next test is whether COSETTE still helps cold-start items with no co-occurrence history, where the reconstruction branch may dominate.
  • MARIUS's separation of sequence encoding from output decoding is a general recipe: any retrieval task with a large hierarchical output vocabulary, such as product search, code completion, or entity linking, could reuse the same temporal/depth split.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper addresses the performance gap between generative recommender systems and well-tuned ID-based baselines. It proposes COSETTE, a contrastive tokenization method that augments RQ-VAE with a co-occurrence-based contrastive loss, and MARIUS, an RQ-Transformer-style architecture with a Temporal Transformer for sequence context and a Depth Transformer for autoregressive code prediction. Experiments on Amazon 2014 and 2023 datasets compare against a modernized SASRec++ baseline and previous generative methods, reporting that the combination narrows or closes the gap on small benchmarks and outperforms SASRec++ on most large-scale datasets. The paper also includes ablations, efficiency measurements, and analyses of hallucination and popularity bias.

Significance. If the central claim holds, the paper makes a useful contribution: it demonstrates that generative recommenders can be competitive with a strong ID-based transformer baseline while offering inference advantages. The study has notable strengths: small-benchmark experiments report means and standard deviations over 5 runs, the SASRec++ baseline is carefully modernized and yields a large relative improvement over the official variant, COSETTE is ablated across loss components and hyperparameters, and the efficiency analysis of MARIUS versus TIGER is concrete. However, the headline large-scale superiority result rests on single unreplicated runs with small margins, and the tokenizer training split is underspecified. These issues must be resolved before the full strength of the conclusion can be accepted.

major comments (3)
  1. [§4.4.3, Table 11, Figure 3, Conclusion] The claim that the method 'even surpass[es] them significantly on large scale datasets' rests entirely on unreplicated comparisons in Table 11/Figure 3. No standard deviations, confidence intervals, or significance tests are reported for the 2023 datasets. The margins are small: e.g., R@10 Arts Crafts & Sewing 5.30 vs 5.09, Health & Household 3.42 vs 3.19, Office Products 4.98 vs 4.49, and Automotive is a counterexample (3.45 vs 3.53 in favor of SASRec++). Given that the small-benchmark tables report 5-run means and standard deviations, the absence of replication for the headline result prevents the reader from distinguishing a true superiority from run-to-run noise. Please add multiple runs with error bars and significance tests, or temper the conclusion to 'competitive on large scale.'
  2. [§3.1.2, §4.1] COSETTE's contrastive objective uses co-occurrence of items in batches of user timelines (Eq. 5-6), but the paper does not state whether these timelines come from the training split only. If validation/test interactions are included in tokenizer training, item codes can encode information about the held-out target items, leaking supervision into MARIUS. This is load-bearing because COSETTE's gains over RQ-VAE are attributed to collaborative signal. Please specify the split and, if necessary, retrain the tokenizer using only the training portion of the data.
  3. [Table 5 caption and dagger notation] The caption of Table 5 says dagger values are 'taken from [24,34],' yet the TIGER† row reports means and standard deviations over (apparently) 5 runs. It is unclear whether these numbers were re-run under the authors' protocol or copied from prior papers. If copied, differences in filtering, split, metric implementation, or item sets can change the measured gap; if re-run, the caption should say so. This matters for the 'closing the gap' claim relative to published generative recommenders, although the comparison with the paper's own SASRec++ baseline is less affected.
minor comments (5)
  1. [§3.1.2, Eq. (5)] The loss does not state whether i=j pairs are excluded. Since every item trivially co-occurs with itself, y_ii is always +1; if self-pairs are included in the positive set, the normalization and the contrastive objective are affected. Please specify the diagonal handling and the exact definition of Y_i.
  2. [§4.4.3] Cross-reference error: 'full tables provided in Appendix 11' should refer to a table or appendix label, e.g., 'Table 11 in Appendix B.'
  3. [§4.4.2] NV-Embed-v2 is mentioned as a stronger text encoder without a reference. Please add a citation or state that it is an external model.
  4. [§3.2, Eq. (9)] The logits are written as R^{L·K}; clarify how the loss is computed over the L code levels, e.g., whether a separate softmax per level is used and how partial decoding is supervised during training.
  5. [§4.4.4] The interpretive statements in Figure 4 (e.g., 'MARIUS is exploring alternative items' and 'overcome the frequency bias') are speculative without statistical tests across datasets. Please report the variance of these differences or phrase them as qualitative observations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical systems paper whose claims are supported by held-out evaluation, not by construction or self-citation.

full rationale

I walked the derivation chain and found no step that reduces to its own input. COSETTE is trained with a reconstruction/quantization objective plus a contrastive loss over item co-occurrence in sampled timelines (Eqs. 3-7), and MARIUS is trained to autoregressively predict code tuples for the next item on training timelines; evaluation uses held-out penultimate/last-item splits ('We adopt the leave-last-out strategy to define validation and test splits'). The central comparison against SASRec++ is an independent baseline that the paper explicitly modernizes and tunes; quoting some prior results with daggers in Tables 5 and 6 affects comparability, but comparability concerns are correctness/robustness issues, not circularity. The 'surpassing significantly on large scale datasets' conclusion relies on Table 11/Figure 3, which reports single unreplicated runs with small margins; that is a statistical-support concern, explicitly outside the circularity definition. There are no load-bearing self-citations: none of the references are by Lepage/Mary/Picard, and the cited prior work (RQ-Transformer, TIGER, LETTER, CoST, LC-Rec, etc.) is external. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' own prior work, and no known result is merely relabeled. The paper is therefore self-contained as an empirical study; score 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on several tuned hyperparameters and domain assumptions about data and evaluation. No new physical entities are introduced; the methods are model architectures and training objectives.

free parameters (4)
  • lambda (contrastive loss weight) = 1e-3
    Weight for the collaborative loss in Eq. 7, tuned on validation to balance MSE and cross-entropy scales.
  • Codebook size K and number of levels L = K=256, L=4 (or 3+1 dedup)
    Chosen to keep vocabulary small while minimizing collisions; not derived from theory.
  • Temperature and bias initialization = t'=2, b=-8
    Initialized close to values reached at convergence, as described in Appendix A.1.
  • SASRec++ and MARIUS model sizes = varying (d=128-768, layers 2-10)
    Selected via grid search on validation data for each dataset, affecting the comparison baseline and the final method.
assumptions (5)
  • domain assumption The evaluation protocol (leave-last-out, 5-core filtering) is standard and fair.
    The paper uses this protocol to compare all methods; if it differs from protocols used in cited baselines, the comparison may be unfair.
  • domain assumption Numbers quoted from prior work were obtained under the same experimental protocol.
    Table 5 marks baselines with dagger as taken from [24,34]; assuming protocol equivalence is load-bearing for the gap-closing claim.
  • domain assumption Sentence-T5-XL embeddings faithfully represent item content.
    All tokenization methods start from these text embeddings; if they are weak, the semantic signal is weak.
  • domain assumption Co-occurrence of items in short user timelines captures collaborative signal relevant to next-item prediction.
    This is the core premise of the contrastive loss in Eq. 5-6.
  • standard math RQ-VAE with straight-through estimator is differentiable and trainable in the described way.
    The paper relies on standard neural quantization techniques from [16,38] without re-deriving them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Closing the Performance Gap in Generative Recommenders with Collaborative Tokenization and Efficient Modeling." pith.science (2026). https://pith.science/paper/7LAPTPOA

@misc{pith2026250814910,
  author       = {Pith},
  title        = {Pith review of: Closing the Performance Gap in Generative Recommenders with Collaborative Tokenization and Efficient Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7LAPTPOA}},
  note         = {Machine review of arXiv:2508.14910}
}
read the original abstract

Recent work has explored generative recommender systems as an alternative to traditional ID-based models, reframing item recommendation as a sequence generation task over discrete item tokens. While promising, such methods often underperform in practice compared to well-tuned ID-based baselines like SASRec. In this paper, we identify two key limitations holding back generative approaches: the lack of collaborative signal in item tokenization, and inefficiencies in the commonly used encoder-decoder architecture. To address these issues, we introduce COSETTE, a contrastive tokenization method that integrates collaborative information directly into the learned item representations, jointly optimizing for both content reconstruction and recommendation relevance. Additionally, we propose MARIUS, a lightweight, audio-inspired generative model that decouples timeline modeling from item decoding. MARIUS reduces inference cost while improving recommendation accuracy. Experiments on standard sequential recommendation benchmarks show that our approach narrows, or even eliminates, the performance gap between generative and modern ID-based models, while retaining the benefits of the generative paradigm.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 44 canonical work pages

  1. [1]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Cuturi, M. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems (2013)

    work page 2013

  2. [2]

    Moshi: a speech-text foundation model for real-time dialogue

    Défossez, A., Mazaré, L., Orsini, M., Royer, A., Pérez, P., Jégou, H., Grave, E., and Zeghidour, N. Moshi: a speech-text foundation model for real-time dialogue. Tech. rep., Kyutai, 2024

    work page 2024

  3. [3]

    A review of modern recommender systems using generative models (gen-recsys)

    Deldjoo, Y., He, Z., McAuley, J., Korikov, A., Sanner, S., Ramisa, A., Vidal, R., Sathiamoorthy, M., Kasirzadeh, A., and Milano, S. A review of modern recommender systems using generative models (gen-recsys). In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (2024)

    work page 2024

  4. [4]

    Recommender forest for efficient retrieval

    Feng, C., Li, W., Lian, D., Liu, Z., and Chen, E. Recommender forest for efficient retrieval. Advances in Neural Information Processing Systems (2022)

    work page 2022

  5. [5]

    Recommendation as language processing (RLP): A unified pretrain, personalized prompt & predict paradigm (P5)

    Geng, S., Liu, S., Fu, Z., Ge, Y., and Zhang, Y. Recommendation as language processing (RLP): A unified pretrain, personalized prompt & predict paradigm (P5). In Proceedings of the 16th ACM Conference on Recommender Systems (2022)

    work page 2022

  6. [6]

    InProceedings of the 39th International ACM SIGIR Conference on Research and Development in Information Retrieval (2016)

    He, X., Zhang, H., Kan, M.-Y., and Chua, T.-S.Fast matrix factorization for online recommendation with implicit feedback. InProceedings of the 39th International ACM SIGIR Conference on Research and Development in Information Retrieval (2016)

    work page 2016

  7. [7]

    A survey on user behavior modeling in recommender systems

    He, Z., Liu, W., Guo, W., Qin, J., Zhang, Y., Hu, Y., and Tang, R. A survey on user behavior modeling in recommender systems. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (2023)

    work page 2023

  8. [8]

    Session-based rec- ommendations with recurrent neural networks

    Hidasi, B., Karatzoglou, A., Baltrunas, L., and Tikk, D. Session-based rec- ommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939 (2015)

    arXiv 2015

Show all 50 references
  1. [9]

    Hou, Y., He, Z., McAuley, J., and Zhao, W. X. Learning vector-quantized item representation for transferable sequential recommenders. InTheWebConf (2023)

    2023

  2. [10]

    Bridging language and items for retrieval and recommendation

    Hou, Y., Li, J., He, Z., Y an, A., Chen, X., and McAuley, J. Bridging language and items for retrieval and recommendation. arXiv preprint arXiv:2403.03952 (2024)

    2024 arXiv

  3. [11]

    In Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region (2023)

    Hua, W., Xu, S., Ge, Y., and Zhang, Y.How to index item ids for recommendation foundation models. In Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region (2023)

    2023

  4. [12]

    Billion-scale similarity search with GPUs

    Johnson, J., Douze, M., and Jégou, H. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data (2019)

    2019

  5. [13]

    Self-attentive sequential recommendation

    Kang, W.-C., and McAuley, J. Self-attentive sequential recommendation. In 2018 IEEE International Conference on Data Mining (ICDM) (2018)

    2018

  6. [14]

    B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

    2001 arXiv

  7. [15]

    Turning dross into gold loss: is bert4rec really better than sasrec? In Proceedings of the 17th ACM Conference on Recommender Systems (2023)

    Klenitskiy, A., and V asilev, A. Turning dross into gold loss: is bert4rec really better than sasrec? In Proceedings of the 17th ACM Conference on Recommender Systems (2023)

    2023

  8. [16]

    Autoregressive image gener- ation using residual quantization

    Lee, D., Kim, C., Kim, S., Cho, M., and Han, W.-S. Autoregressive image gener- ation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)

    2022

  9. [17]

    Text is all you need: Learning language representations for sequential recommendation

    Li, J., W ang, M., Li, J., Fu, J., Shen, X., Shang, J., and McAuley, J. Text is all you need: Learning language representations for sequential recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (2023)

    2023

  10. [18]

    Embedding optimization for training large-scale deep learning recommendation systems with embark

    Liu, S., Zheng, N., Kang, H., Simmons, X., Zhang, J., Langer, M., Zhu, W., Lee, M., and W ang, Z. Embedding optimization for training large-scale deep learning recommendation systems with embark. InProceedings of the 18th ACM Conference on Recommender Systems (2024)

    2024

  11. [19]

    Hierarchical gating networks for sequential recom- mendation

    Ma, C., Kang, P., and Liu, X. Hierarchical gating networks for sequential recom- mendation. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining (2019), pp. 825–833

    2019

  12. [20]

    Image-based recom- mendations on styles and substitutes

    McAuley, J., Targett, C., Shi, Q., and V an Den Hengel, A. Image-based recom- mendations on styles and substitutes. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval (2015)

    2015

  13. [21]

    In Findings of the Association for Computational Linguistics (2022)

    Ni, J., Hernandez Abrego, G., Constant, N., Ma, J., Hall, K., Cer, D., and Y ang, Y.Sentence-T5: Scalable sentence encoders from pre-trained text-to-text models. In Findings of the Association for Computational Linguistics (2022)

    2022

  14. [22]

    Oord, A. v. d., Li, Y., and Vinyals, O. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  15. [23]

    TIGER reviewing process

    OpenReview. TIGER reviewing process. https://openreview.net/forum?id= BJ0fQUU32w

  16. [24]

    H., Vu, T., Heldt, L., Hong, L., Tay, Y., Tran, V

    Rajput, S., Mehta, N., Singh, A., Keshavan, R. H., Vu, T., Heldt, L., Hong, L., Tay, Y., Tran, V. Q., Samost, J., Kula, M., Chi, E. H., and Sathiamoorthy, M. Recommender systems with generative retrieval. In Thirty-seventh Conference on Neural Information Processing Systems (2023)

    2023

  17. [25]

    Y., Guo, W., Liu, Y., Guo, H., Lian, D., Tang, R., and Chen, E

    Shen, T., W ang, H., Wu, C., Chin, J. Y., Guo, W., Liu, Y., Guo, H., Lian, D., Tang, R., and Chen, E. Optimizing sequential recommendation models with scaling laws and approximate entropy. arXiv preprint arXiv:2412.00430 (2024)

    2024 arXiv

  18. [26]

    Better generalization with semantic ids: A case study in ranking for recommendations

    Singh, A., Vu, T., Mehta, N., Keshavan, R., Sathiamoorthy, M., Zheng, Y., Hong, L., Heldt, L., Wei, L., Tandon, D., Chi, E., and Yi, X. Better generalization with semantic ids: A case study in ranking for recommendations. In Proceedings of the 18th ACM Conference on Recommende...

    2024

  19. [27]

    Improved Deep Metric Learning with Multi-class N-pair Loss Objective

    Sohn, K. Improved Deep Metric Learning with Multi-class N-pair Loss Objective. In Advances in Neural Information Processing Systems (2016)

    2016

  20. [28]

    In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (2019)

    Sun, F., Liu, J., Wu, J., Pei, C., Lin, X., Ou, W., and Jiang, P.BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (2019)

    2019

  21. [29]

    In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining (2018)

    Tang, J., and W ang, K.Personalized top-n sequential recommendation via convo- lutional sequence embedding. In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining (2018)

    2018

  22. [30]

    Q., Dehghani, M., Ni, J., Bahri, D., Mehta, H., Qin, Z., Hui, K., Zhao, Z., Gupta, J., Schuster, T., Cohen, W

    Tay, Y., Tran, V. Q., Dehghani, M., Ni, J., Bahri, D., Mehta, H., Qin, Z., Hui, K., Zhao, Z., Gupta, J., Schuster, T., Cohen, W. W., and Metzler, D. Transformer memory as a differentiable search index. InProceedings of the 36th International Conference on Neural Information Pr...

    2022

  23. [31]

    Neural discrete representation learning

    V an Den Oord, A., Vinyals, O., et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (2017)

    2017

  24. [32]

    N., Kaiser, Ł., and Polosukhin, I

    V aswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I. Attention is all you need. Advances in neural information processing systems (2017)

    2017

  25. [33]

    Learnable item tokenization for generative recommendation

    W ang, W., Bao, H., Lin, X., Zhang, J., Li, Y., Feng, F., Ng, S.-K., and Chua, T.-S. Learnable item tokenization for generative recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management (2024)

    2024

  26. [34]

    Eager: Two-stream generative recommender with behavior- semantic collaboration

    W ang, Y., Xun, J., Hong, M., Zhu, J., Jin, T., Lin, W., Li, H., Li, L., Xia, Y., Zhao, Z., and Dong, Z. Eager: Two-stream generative recommender with behavior- semantic collaboration. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (2024)

    2024

  27. [35]

    A survey on large language models for recommendation

    Wu, L., Zheng, Z., Qiu, Z., W ang, H., Gu, H., Shen, T., Qin, C., Zhu, C., Zhu, H., Liu, Q., Xiong, H., and Chen, E. A survey on large language models for recommendation. World Wide Web (2024)

    2024

  28. [36]

    Session-based recom- mendation with graph neural networks

    Wu, S., Tang, Y., Zhu, Y., W ang, L., Xie, X., and Tan, T. Session-based recom- mendation with graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence (2019)

    2019

  29. [37]

    Uniaudio: An audio foundation model toward universal audio generation

    Y ang, D., Tian, J., Tan, X., Huang, R., Liu, S., Chang, X., Shi, J., Zhao, S., Bian, J., Wu, X., et al. Uniaudio: An audio foundation model toward universal audio generation. In Proceedings of the 41st International Conference on Machine Learning (2024)

    2024

  30. [38]

    Soundstream: An end-to-end neural audio codec, 2021

    Zeghidour, N., Luebs, A., Omran, A., Skoglund, J., and Tagliasacchi, M. Soundstream: An end-to-end neural audio codec, 2021

    2021

  31. [39]

    In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (2023)

    Zhai, J., Gong, Z., W ang, Y., Sun, X., Y an, Z., Li, F., and Liu, X.Revisiting neural retrieval on accelerators. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (2023)

    2023

  32. [40]

    Actions speak louder than words: trillion-parameter sequential transducers for generative recommendations

    Zhai, J., Liao, L., Liu, X., W ang, Y., Li, R., Cao, X., Gao, L., Gong, Z., Gu, F., He, J., Lu, Y., and Shi, Y. Actions speak louder than words: trillion-parameter sequential transducers for generative recommendations. In Proceedings of the 41st International Conference on Mac...

    2024

  33. [41]

    Multi- modal quantitative language for generative recommendation

    Zhai, J., Mai, Z.-F., W ang, C.-D., Y ang, F., Zheng, X., Li, H., and Tian, Y. Multi- modal quantitative language for generative recommendation. In The Thirteenth International Conference on Learning Representations (2025)

    2025

  34. [42]

    Sigmoid loss for language image pre-training

    Zhai, X., Mustafa, B., Kolesnikov, A., and Beyer, L. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision (2023)

    2023

  35. [43]

    X., and Wen, J.-R

    Zhang, G., Hou, Y., Lu, H., Chen, Y., Zhao, W. X., and Wen, J.-R. Scaling law of large sequential recommendation models. In Proceedings of the 18th ACM Conference on Recommender Systems (2024)

    2024

  36. [44]

    S., Xu, J., W ang, D., Liu, G., and Zhou, X

    Zhang, T., Zhao, P., Liu, Y., Sheng, V. S., Xu, J., W ang, D., Liu, G., and Zhou, X. Feature-level deeper self-attention network for sequential recommendation. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (2019)

    2019

  37. [45]

    Recommender systems in the era of large language models (llms)

    Zhao, Z., Fan, W., Li, J., Liu, Y., Mei, X., W ang, Y., Wen, Z., W ang, F., Zhao, X., Tang, J., and Li, Q. Recommender systems in the era of large language models (llms). IEEE Transactions on Knowledge and Data Engineering (2024)

    2024

  38. [46]

    X., Chen, M., and Wen, J.-R

    Zheng, B., Hou, Y., Lu, H., Chen, Y., Zhao, W. X., Chen, M., and Wen, J.-R. Adapting large language models by integrating collaborative semantics for rec- ommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) (2024)

    2024

  39. [47]

    X., Zhu, Y., W ang, S., Zhang, F., W ang, Z., and Wen, J

    Zhou, K., W ang, H., Zhao, W. X., Zhu, Y., W ang, S., Zhang, F., W ang, Z., and Wen, J. S3-rec: Self-supervised learning for sequential recommendation with mutual information maximization. In CIKM ’20: The 29th ACM International Conference on Information and Knowledge Manageme...

    2020

  40. [48]

    Learning tree-based deep model for recommender systems

    Zhu, H., Li, X., Zhang, P., Li, G., He, J., Li, H., and Gai, K. Learning tree-based deep model for recommender systems. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2018). Conference’17, July 2017, Washington, DC, USA Sim...

    2018

  41. [49]

    CoST: Contrastive quantiza- tion based semantic tokenization for generative recommendation

    Zhu, J., Jin, M., Liu, Q., Qiu, Z., Dong, Z., and Li, X. CoST: Contrastive quantiza- tion based semantic tokenization for generative recommendation. InProceedings of the 18th ACM Conference on Recommender Systems (2024)

    2024

  42. [50]

    Generative pre-trained speech language model with efficient hierarchical transformer

    Zhu, Y., Su, D., He, L., Xu, L., and Yu, D. Generative pre-trained speech language model with efficient hierarchical transformer. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (2024). Closing the Performance Gap in Generative Recomme...

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.