Pith. sign in

REVIEW 4 major objections 5 minor 8 cited by

Generative Recommendation with Semantic IDs: A Practitioner's Handbook

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A modular framework for generative recommendation with semantic IDs finds that several standard design choices—user tokens, RQ-VAE, constrained beam search—can be dropped without loss.

desk verdict A useful open-source framework and solid ablations, but the claimed reproduction of prior results is unverified—worth reviewing with a demand for a TIGER reproduction table and error bars. read the letter →

arxiv 2507.22224 v1 pith:AP43WEII submitted 2025-07-29 cs.IR

Clark Mingxuan Ju , Liam Collins , Leonardo Neves , Bhuvesh Kumar , Louis Yufeng Wang , Tong Zhao , Neil Shah This is my paper
classification cs.IR
keywords generativerecommendationsemanticIDstokenizationsequentialencoder-decodertransformerdataaugmentationablationstudyopen-sourceframework
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

The paper's goal is to make generative recommendation with semantic IDs (SIDs) easy to benchmark and extend by releasing GRID, a modular open-source framework that separates SID tokenization from next-item generation. Using GRID, the authors run systematic ablations and claim that several commonly used or costly components—user tokens, RQ-VAE tokenizers, constrained beam search—are not helping, and some hurt. They find that the components that matter most are the choice of encoder-decoder over decoder-only generation and sliding-window data augmentation. If these results hold, the field's default design choices shift toward simpler, cheaper pipelines, and GRID gives practitioners a shared base for comparing future proposals.

What carries the argument

The central object is GRID itself: a modular framework that splits GR with SIDs into a tokenization stage and a generation stage. Tokenizers implemented include RK-Means, R-VQ, and RQ-VAE over embeddings from pluggable semantic encoders; generation includes encoder-decoder and decoder-only transformers, user-token insertion, sliding-window augmentation, SID de-duplication, and constrained or free-form beam search. The framework's work is to make each design choice a swappable component so that ablations attribute performance differences to a single component.

What would settle it

Retrain the default TIGER-style pipeline in GRID on the same three datasets and compare against the numbers originally reported for TIGER; if the reproduced Recall@10 falls well short of the published values, the framework is not a neutral reference and the ablation findings cannot be attributed to the components.

Watch

Extended reading notes

Core claim

The central claim is that a faithful, modular implementation of the tokenization-then-generation pipeline, called GRID, changes what the field believes about which design choices drive performance in generative recommendation with semantic IDs. On the tokenization side, the paper reports that the simple residual mini-batch K-means (RK-Means) tokenizer matches or beats RQ-VAE even though RQ-VAE is trained five times longer, and that scaling the semantic encoder from 780M to 11B parameters yields only marginal gains. On the generation side, the paper reports that removing user tokens entirely gives the best or tied performance, that encoder-decoder transformers substantially outperform decoder-only transformers across all datasets, that sliding-window augmentation is essential (removing it cuts Recall@10 from 0.0597 to 0.0447 on Beauty), that SID de-duplication gives only a slight edge, and that unconstrained beam search matches constrained search at lower cost.

Load-bearing premise

The results stand or fall on whether GRID's reference implementation is faithful to the published TIGER-style pipeline and whether the fixed training settings are neutral across the components being compared; if the decoder-only model simply got less effective layers or a worse learning rate, the architecture gap would be an artifact.

Editorial extensions

If this is right

  • Default GR with SID pipelines can drop user tokens entirely; in the paper's Beauty experiments, zero user tokens gave the best Recall@10.
  • Encoder-decoder transformers, not decoder-only models, should be the default generation backbone; decoder-only models trailed on Beauty, Toys, and Sports.
  • Sliding-window augmentation is a necessary training component; without it, Recall@10 drops from 0.0597 to 0.0447 on Beauty.
  • The simpler RK-Means tokenizer is sufficient and preferable to RQ-VAE, which the paper trained five times longer without winning.
  • Constrained beam search is not worth its overhead; free-form beam search performed comparably with better efficiency.

Reading between the lines

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

  • The user-token result suggests that in this paradigm personalization is already carried by the encoded interaction history and the semantics in the SIDs; an interesting extension would be testing whether a learned user embedding, not a hashed token, changes that conclusion.
  • The encoder-decoder advantage is consistent with the hypothesis that full-length dense attention over the user history matters; a decoder-only variant with explicit cross-attention to the history would test that mechanism directly.
  • The flat returns from scaling the semantic encoder imply the bottleneck sits in quantization or sequence generation; a natural next experiment is scaling the codebook size or sequence length together with encoder size to see whether larger encoders then pay off.
  • Because all results are on three 5-core-filtered categories, the natural stress test is cold-start and long-tail item regimes, where semantic IDs are supposed to help most; GRID makes that test straightforward.
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

4 major / 5 minor

Summary. The paper introduces GRID, an open-source modular framework for generative recommendation with semantic IDs (SIDs), and uses it to run a series of ablation studies on Amazon Beauty, Sports, and Toys datasets. The ablations cover SID tokenizer choice, semantic encoder size, tokenizer codebook dimensions, user tokens, encoder-decoder versus decoder-only architectures, sliding-window data augmentation, SID de-duplication, and constrained versus free-form beam search. The authors report several findings, including that RK-Means can outperform RQ-VAE, that user tokens do not help, that encoder-decoder models beat decoder-only models, and that sliding-window augmentation is important. The paper's central claims are that GRID provides a reliable reference implementation that can reproduce existing literature and that the experiments reveal overlooked architectural factors that substantially affect performance.

Significance. If the central claims are validated, GRID would be a valuable community resource: it is, to my knowledge, the first open-source framework designed specifically for prototyping generative recommendation with semantic IDs, and its modular design could accelerate research in this area. The empirical findings are also potentially actionable, especially the primacy of sliding-window augmentation and the suggestion that RQ-VAE, the default tokenizer in much prior work, may not be necessary. The paper's strengths include the release of open-source code, the use of multiple datasets and seeds, and a clear presentation of the pipeline components. However, the empirical contribution currently rests on an unverified assumption that the reference implementation is faithful to published methods, and several quantitative claims are made without measures of uncertainty.

major comments (4)
  1. [§1, §4 Setup] The paper asserts in Section 1 that GRID 'can reproduce results reported in existing literature,' but it never provides a direct comparison between its TIGER-style encoder-decoder instantiation and the published results of Rajput et al. (2023), even though Section 4 states that the same 5-core Amazon Beauty/Sports/Toys splits as reference [44] are used. Without such a reproduction table, the ablation findings in Tables 4-8 cannot be distinguished from implementation or hyperparameter artifacts. Please add a table comparing GRID's default TIGER instantiation against the published Recall/NDCG numbers, and clearly document any deviations in data processing, tokenizer training, or generation decoding.
  2. [§4 Setup, Table 5] The encoder-decoder versus decoder-only comparison is under-specified. The setup states that encoder-decoder models use '8 transformer layers in total (i.e., 4 in the encoder and 4 in the decoder),' but it does not state how many layers the decoder-only variant uses, nor whether the two architectures are matched in parameter count, FLOPs, or training budget. The large performance gap in Table 5 could therefore be an artifact of an unter-tuned or under-parameterized decoder-only model. Please specify the decoder-only configuration and, ideally, report a parameter-matched comparison.
  3. [§4 Setup, Tables 2-4] All experimental conclusions are based on point estimates averaged over five seeds, but no standard deviations, confidence intervals, or significance tests are reported. Many of the differences that drive conclusions are very small: for example, Table 2 shows L versus XL Recall@5 of 0.0429 versus 0.0422, and Table 4 shows 0 versus 8,000 user tokens with NDCG@10 of 0.0330 versus 0.0335. These differences are within the range of plausible seed-to-seed noise, yet the text interprets them as meaningful trends. Please report variance estimates or significance tests, and temper claims that depend on such small differences.
  4. [Table 7, Sports row] The Sports row of Table 7 reports 'No De-dup.' NDCG@5 as 0.0011 while 'With De-dup.' is 0.0124. This order-of-magnitude gap contradicts the text's claim that the two de-duplication strategies 'perform comparably.' If this is a typographical error (e.g., 0.0111), it must be corrected; if it is not, the conclusion about de-duplication needs to be revised, because the current table does not support the stated finding.
minor comments (5)
  1. [§4.2] There is a typo in the user-token subsection: 'i,.e., 0' should be 'i.e., 0'.
  2. [§3.1] The word 'tokenzier' should be 'tokenizer' in the SID tokenization paragraph.
  3. [§4] The notation 'Recall@K and NDCG@K for K∈5,10' should be written as 'K∈{5,10}'.
  4. [§4.2, Table 8] The text states that unconstrained beam search was 'significantly more efficient and computationally cheaper,' but no runtime measurements are reported in Table 8 or elsewhere. Please provide quantitative efficiency data or soften the claim.
  5. [References] Reference [41] is listed with '[n. d.]' and no venue; please provide a complete citation with year and public preprint identifier.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GRID's claims are empirical ablation results, not derivations; no step reduces by definition or self-citation to its own inputs.

full rationale

The paper's central claims are that GRID is a modular, open-source framework and that controlled experiments reveal which architectural components matter for generative recommendation with semantic IDs. These are empirical findings generated by running the framework, not mathematical derivations, so the standard circularity patterns do not apply. No parameter is fitted to a subset of data and then reported as a prediction; no result is defined in terms of another result by construction; and no 'uniqueness theorem' is imported to force a choice. The self-citations in the reference list (e.g., several prior papers by the same authors) are contextual and are not load-bearing for any ablation conclusion; the key comparison baselines, such as TIGER [44], are external published work. The most substantive reviewer concern—that GRID's claim to 'reproduce results reported in existing literature' is not backed by a direct comparison table against published TIGER numbers—is a correctness or verification risk, not circularity: an unverified implementation can still be a genuinely independent empirical instrument. The paper is self-contained against external benchmarks in the sense that it defines its experimental setup, reports metrics, and compares design variants, so the honest finding is no significant circularity.

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

The central claims rest on a handful of experimental choices rather than fitted model parameters. The most consequential are the tokenizer training budgets, the unspecified decoder-only depth, and the early stopping rule. The paper's conclusions also assume the GRID implementations are faithful to prior work, which is not independently demonstrated. No new theoretical entities are introduced.

free parameters (4)
  • RQ-VAE training steps = 15k total vs 1k per layer for RK-Means/R-VQ
    Table 1's tokenizer comparison depends on this budget; the paper acknowledges RQ-VAE got 5x more steps but still lost, yet no convergence analysis is shown.
  • Decoder-only transformer depth = Unspecified; encoder-decoder uses 4 encoder + 4 decoder layers
    Table 5 compares architectures without stating whether the decoder-only model received 4, 8, or 12 layers, so the performance gap may reflect capacity allocation.
  • Early stopping patience = 10 validation intervals (100 steps each) without NDCG@10 improvement
    Section 4 Setup fixes this for all runs; a different patience could select different checkpoints and change the reported metrics.
  • Beam width = Not stated
    Section 3.1 mentions tunable beam width, but the paper does not report the value used for Table 8; conclusions about constrained vs unconstrained search depend on this setting.
assumptions (5)
  • domain assumption Amazon 5-core Beauty, Toys, and Sports with last-item test split are representative of sequential recommendation.
    All empirical conclusions in Section 4 are drawn from these three datasets; no other domains, scales, or interaction densities are tested.
  • domain assumption Mean-pooled Flan-T5 hidden states capture item semantics sufficiently for SID tokenization.
    Section 3.1 and Section 4 Setup fix this encoder; all tokenizers and recommenders operate on these embeddings, so encoder fidelity is assumed.
  • ad hoc to paper The GRID reference implementation of TIGER-style pipelines is faithful to published methods.
    Section 1 claims reproduction of existing literature but provides no side-by-side reproduction numbers, so baseline correctness is assumed.
  • domain assumption Next-token prediction with sliding-window augmentation is the appropriate training objective for comparing components.
    Section 3.1 and Section 4 Setup use this objective for every ablation; conclusions about user tokens and architecture may not transfer to other objectives.
  • ad hoc to paper Random-hash user tokens correctly instantiate TIGER's user token mechanism.
    Table 4's conclusion that removing user tokens is optimal depends on this implementation choice; TIGER's original assignment may differ in unspecified details.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Recommendation with Semantic IDs: A Practitioner's Handbook." pith.science (2026). https://pith.science/paper/AP43WEII

@misc{pith2026250722224,
  author       = {Pith},
  title        = {Pith review of: Generative Recommendation with Semantic IDs: A Practitioner's Handbook},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AP43WEII}},
  note         = {Machine review of arXiv:2507.22224}
}
read the original abstract

Generative recommendation (GR) has gained increasing attention for its promising performance compared to traditional models. A key factor contributing to the success of GR is the semantic ID (SID), which converts continuous semantic representations (e.g., from large language models) into discrete ID sequences. This enables GR models with SIDs to both incorporate semantic information and learn collaborative filtering signals, while retaining the benefits of discrete decoding. However, varied modeling techniques, hyper-parameters, and experimental setups in existing literature make direct comparisons between GR proposals challenging. Furthermore, the absence of an open-source, unified framework hinders systematic benchmarking and extension, slowing model iteration. To address this challenge, our work introduces and open-sources a framework for Generative Recommendation with semantic ID, namely GRID, specifically designed for modularity to facilitate easy component swapping and accelerate idea iteration. Using GRID, we systematically experiment with and ablate different components of GR models with SIDs on public benchmarks. Our comprehensive experiments with GRID reveal that many overlooked architectural components in GR models with SIDs substantially impact performance. This offers both novel insights and validates the utility of an open-source platform for robust benchmarking and GR research advancement. GRID is open-sourced at https://github.com/snap-research/GRID.

Figures

Figures reproduced from arXiv: 2507.22224 by the authors.

Figure 1
Figure 1. (a) The overarching architecture of GRID. GRID modularizes all intermediate steps in the workflow of GR with SIDs to [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 8 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. OneShot: Index-in-Ranking with Neural Scoring for Large-Scale Retrieval

    cs.IR 2026-07 conditional novelty 6.0 of 10

    OneShot trains a hierarchical item index and a neural ranking score jointly, reporting +20% offline recall and 10x efficiency over k-means ANN at Instagram.

  2. OneShot: Index-in-Ranking with Neural Scoring for Large-Scale Retrieval

    cs.IR 2026-07 conditional novelty 6.0 of 10

    OneShot trains hierarchical item codebooks jointly with the ranking loss, enabling nonlinear neural scoring in billion-scale retrieval and reporting +20% recall, 10x fewer dense-ranked items, and live Instagram gains.

  3. Beyond Fixed Depths and Widths: Optimizing Textual Decoding Tries in LLM-based Generative Recommendation

    cs.IR 2026-07 conditional novelty 6.0 of 10

    BONSAI constructs variable-depth, low-branching decoding tries for LLM-based generative recommendation and reports 16–22% relative gains over state-of-the-art baselines.

  4. SIDInspector: A Mapping-First Diagnostic Resource for Semantic-ID Tokenizers

    cs.IR 2026-06 accept novelty 6.0 of 10

    SIDInspector provides a standardized adapter contract and mapping-level probes for Semantic-ID tokenizers, with empirical contrasts showing high aliasing in GRID-style exports and superior prefix alignment from determ...

  5. Gated Bidirectional Linear Attention for Generative Retrieval

    cs.IR 2026-06 unverdicted novelty 6.0 of 10

    GBLA extends kernelized linear attention with local causal mixing, key gating, and gated RMSNorm; a 1:2 hybrid with self-attention matches full bidirectional self-attention quality on Yandex Music data while deliverin...

  6. Understanding Generative Recommendation with Semantic IDs from a Model-scaling View

    cs.AI 2025-09 conditional novelty 6.0 of 10

    Semantic-ID-based generative recommenders saturate as model size grows, while directly using an LLM as the recommender keeps improving with scale and learns collaborative filtering signals better.

  7. Sequential Data Augmentation for Generative Recommendation

    cs.LG 2025-09 conditional novelty 6.0 of 10

    GenPAS unifies common data augmentation strategies for generative recommendation as special cases of a bias-controlled stochastic sampling process and demonstrates gains in accuracy, data efficiency, and parameter eff...

  8. FORGE: Forming Semantic Identifiers for Generative Retrieval in Industrial Datasets

    cs.IR 2025-09 conditional novelty 5.0 of 10

    FORGE shows that balancing codebook usage and adding multimodal side information improves semantic identifiers for generative retrieval, validated offline and on Taobao.

Reference graph

Works this paper leans on

82 extracted references · 21 canonical work pages · cited by 7 Pith papers

  1. [44]

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

  2. [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)

    arXiv 2023

  3. [2]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM Conference on Recommender Systems. 1007–1014

    2023

  4. [3]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners.Advances in neural information processing systems33 (2020), 1877–1901

    2020

  5. [4]

    Runjin Chen, Mingxuan Ju, Ngoc Bui, Dimosthenis Antypas, Stanley Cai, Xi- aopeng Wu, Leonardo Neves, Zhangyang Wang, Neil Shah, and Tong Zhao

  6. [5]

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al

  7. [6]

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2024. Scaling instruction-finetuned language models.Journal of Machine Learning Research25, 70 (2024), 1–53

    2024

  8. [7]

    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)

    arXiv 2025

Show all 82 references
  1. [8]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human...

    2019

  2. [9]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprin...

    2020 arXiv

  3. [10]

    Patrick Esser, Robin Rombach, and Bjorn Ommer. 2021. Taming transformers for high-resolution image synthesis. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 12873–12883

    2021

  4. [11]

    Christopher Fifty, Ronald G Junkins, Dennis Duan, Aniketh Iyengar, Jerry W Liu, Ehsan Amid, Sebastian Thrun, and Christopher Ré. 2024. Restructuring vector quantization with the rotation trick.arXiv preprint arXiv:2410.06424(2024)

    2024 arXiv

  5. [12]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). InProceedings of the 16th ACM conference on recommender systems. 299–315

    2022

  6. [13]

    Benjamin Ghaemmaghami, Mustafa Ozdal, Rakesh Komuravelli, Dmitriy Korchev, Dheevatsa Mudigere, Krishnakumar Nair, and Maxim Naumov. 2022. Learning to Collide: Recommendation System Model Compression with Learned Hash Func- tions.ArXivabs/2203.15837 (2022). https://api.semantic...

    2022 arXiv

  7. [14]

    Carlos A Gomez-Uribe and Neil Hunt. 2015. The netflix recommender system: Algorithms, business value, and innovation.ACM Transactions on Management Information Systems (TMIS)(2015)

    2015

  8. [15]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al . 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

    2025 arXiv

  9. [16]

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick

  10. [17]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  11. [18]

    Yupeng Hou, Jiacheng Li, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley

  12. [19]

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

    2023

  13. [20]

    Bowen Jin, Hansi Zeng, Guoyin Wang, Xiusi Chen, Tianxin Wei, Ruirui Li, Zhengyang Wang, Zheng Li, Yang Li, Hanqing Lu, et al. 2023. Language models as semantic indexers.arXiv preprint arXiv:2310.07815(2023)

    2023 arXiv

  14. [21]

    Clark Mingxuan Ju, Leonardo Neves, Bhuvesh Kumar, Liam Collins, Tong Zhao, Yuwei Qiu, Qing Dou, Sohail Nizam, Sen Yang, and Neil Shah. 2025. Revisiting Self-attention for Cross-domain Sequential Recommendation.arXiv preprint arXiv:2505.21811(2025)

    2025 arXiv

  15. [22]

    Bridging Language and Items for Retrieval and Recommendation.arXiv preprint arXiv:2403.03952(2024)

    2024 arXiv

  16. [23]

    Mingxuan Ju, William Shiao, Zhichun Guo, Yanfang Ye, Yozen Liu, Neil Shah, and Tong Zhao. 2024. How Does Message Passing Improve Collaborative Filtering? arXiv preprint arXiv:2404.08660(2024)

    2024 arXiv

  17. [24]

    Mingxuan Ju, Wenhao Yu, Tong Zhao, Chuxu Zhang, and Yanfang Ye. 2022. Grape: Knowledge graph enhanced passage reader for open-domain question answering. arXiv preprint arXiv:2210.02933(2022)

    2022 arXiv

  18. [25]

    Mingxuan Ju, Tong Zhao, Qianlong Wen, Wenhao Yu, Neil Shah, Yanfang Ye, and Chuxu Zhang. 2022. Multi-task self-supervised graph neural networks enable stronger task generalization.arXiv preprint arXiv:2210.02016(2022)

    2022 arXiv

  19. [26]

    Clark Mingxuan Ju, Leonardo Neves, Bhuvesh Kumar, Liam Collins, Tong Zhao, Yuwei Qiu, Qing Dou, Yang Zhou, Sohail Nizam, Rengim Aykan Ozturk, et al

  20. [27]

    Dongmoon Kim, Kun-su Kim, Kyo-Hyun Park, Jee-Hyong Lee, and Keon Myung Lee. 2007. A music recommendation system with a dynamic k-means clustering algorithm. InSixth international conference on machine learning and applications (ICMLA 2007). IEEE, 399–403

    2007

  21. [28]

    Matthew Kolodner, Mingxuan Ju, Zihao Fan, Tong Zhao, Elham Ghazizadeh, Yan Wu, Neil Shah, and Yozen Liu. 2024. Robust training objectives improve embedding-based retrieval in industrial recommendation systems.arXiv preprint arXiv:2409.14682(2024)

    2024 arXiv

  22. [29]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization tech- niques for recommender systems.Computer42, 8 (2009), 30–37

    2009

  23. [30]

    Zhirui Kuai, Zuxu Chen, Huimu Wang, Mingming Li, Dadong Miao, Binbin Wang, Xusong Chen, Li Kuang, Yuxing Han, Jiaxing Wang, et al. 2024. Breaking the Hourglass Phenomenon of Residual Quantization: Enhancing the Upper Bound of Generative Retrieval.arXiv preprint arXiv:2407.21488(2024)

    2024 arXiv

  24. [31]

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

    2018

  25. [32]

    Blerina Lika, Kostas Kolomvatsos, and Stathes Hadjiefthymiades. 2014. Facing the cold start problem in recommender systems.Expert systems with applications 41, 4 (2014), 2065–2073

    2014

  26. [33]

    Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, et al . 2025. How can recommender systems benefit from large language models: A survey.ACM Transactions on Information Systems43, 2 (2025), 1–47

    2025

  27. [34]

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

  28. [35]

    Zihan Liu, Yupeng Hou, and Julian McAuley. 2024. Multi-behavior generative recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 1575–1585

    2024

  29. [36]

    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

  30. [37]

    Donald Loveland, Xinyi Wu, Tong Zhao, Danai Koutra, Neil Shah, and Mingxuan Ju. 2025. Understanding and Scaling Collaborative Filtering Optimization from the Perspective of Matrix Rank. InProceedings of the ACM on Web Conference

    2025

  31. [38]

    Xinchen Luo, Jiangxia Cao, Tianyu Sun, Jinkai Yu, Rui Huang, Wei Yuan, Hezheng Lin, Yichen Zheng, Shiyao Wang, Qigen Hu, et al . 2024. QARM: Quantita- tive Alignment Multi-Modal Recommendation at Kuaishou.arXiv preprint arXiv:2411.11739(2024). Conference’17, July 2017, Washing...

    2024 arXiv

  32. [39]

    Zhongyu Ouyang, Mingxuan Ju, Soroush Vosoughi, and Yanfang Ye. 2025. Non- parametric Graph Convolution for Re-ranking in Recommendation Systems. arXiv preprint arXiv:2507.09969(2025)

    2025 arXiv

  33. [40]

    arXiv preprint arXiv:2409.05546(2024)

    End-to-End Learnable Item Tokenization for Generative Recommendation. arXiv preprint arXiv:2409.05546(2024)

    2024 arXiv

  34. [41]

    Fabian Paischer, Liu Yang, Linfeng Liu, Shuai Shao, Kaveh Hassani, Jiacheng Li, Ricky TQ Chen, Zhang Gabriel Li, Xiaoli Gao, Wei Shao, et al. [n. d.]. Preference Discerning in Generative Sequential Recommendation. ([n. d.])

  35. [42]

    Donald Loveland, Mingxuan Ju, Tong Zhao, Neil Shah, and Danai Koutra. 2025. On the Role of Weight Decay in Collaborative Filtering: A Popularity Perspective. arXiv preprint arXiv:2505.11318(2025)

    2025 arXiv

  36. [43]

    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

  37. [45]

    Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Representation learning with large language models for recommendation. InProceedings of the ACM Web Conference 2024. 3464–3475

    2024

  38. [46]

    Fabian Paischer, Liu Yang, Linfeng Liu, Shuai Shao, Kaveh Hassani, Jiacheng Li, Ricky Chen, Zhang Gabriel Li, Xialo Gao, Wei Shao, et al. 2024. Preference Dis- cerning with LLM-Enhanced Generative Retrieval.arXiv preprint arXiv:2412.08604 (2024)

    2024

  39. [47]

    Aravind Sankar, Yozen Liu, Jun Yu, and Neil Shah. 2021. Graph neural networks for friend ranking in large-scale social platforms. InProceedings of the Web Conference 2021. 2535–2546

    2021

  40. [48]

    Enrico Palumbo, Gustavo Penha, Andreas Damianou, José Luis Redondo García, Timothy Christopher Heath, Alice Wang, Hugues Bouchard, and Mounia Lal- mas. 2025. Text2Tracks: Prompt-based Music Recommendation via Generative Retrieval.arXiv preprint arXiv:2503.24193(2025)

    2025 arXiv

  41. [49]

    William Shiao, Mingxuan Ju, Zhichun Guo, Xin Chen, Evangelos E Papalexakis, Tong Zhao, Neil Shah, and Yozen Liu. 2025. Improving Out-of-Vocabulary Hashing in Recommendation Systems. InCompanion Proceedings of the ACM on Web Conference 2025. 2521–2530

    2025

  42. [50]

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

    2024

  43. [51]

    Juntao Tan, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Zelong Li, and Yongfeng Zhang. 2024. Idgenrec: Llm-recsys alignment with textual id learning. InProceed- ings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 355–364

    2024

  44. [52]

    Yi Tay, Vinh Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, et al. 2022. Transformer memory as a differentiable search index.Advances in Neural Information Processing Systems 35 (2022), 21831–21843

    2022

  45. [53]

    Steffen Rendle, Walid Krichene, Li Zhang, and John Anderson. 2020. Neural collaborative filtering vs. matrix factorization revisited. InProceedings of the 14th ACM Conference on Recommender Systems. 240–248

    2020

  46. [54]

    Aaron Van Den Oord, Oriol Vinyals, et al. 2017. Neural discrete representation learning.Advances in neural information processing systems30 (2017)

    2017

  47. [55]

    J Ben Schafer, Joseph Konstan, and John Riedl. 1999. Recommender systems in e-commerce. InProcs. of ACM conference on Electronic commerce

    1999

  48. [56]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  49. [57]

    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

  50. [58]

    Kilian Weinberger, Anirban Dasgupta, John Langford, Alex Smola, and Josh Attenberg. 2009. Feature hashing for large scale multitask learning. InProceedings of the 26th annual international conference on machine learning. 1113–1120

    2009

  51. [59]

    Xinyi Wu, Donald Loveland, Runjin Chen, Yozen Liu, Xin Chen, Leonardo Neves, Ali Jadbabaie, Mingxuan Ju, Neil Shah, and Tong Zhao. 2025. GraphHash: Graph Clustering Enables Parameter Efficiency in Recommender Systems. InProceedings of the ACM on Web Conference 2025. 357–369

    2025

  52. [60]

    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)

    2023 arXiv

  53. [61]

    Liu Yang, Fabian Paischer, Kaveh Hassani, Jiacheng Li, Shuai Shao, Zhang Gabriel Li, Yun He, Xue Feng, Nima Noorshams, Sem Park, et al. 2024. Unifying Gen- erative and Dense Retrieval for Sequential Recommendation.arXiv preprint arXiv:2411.18814(2024)

    2024 arXiv

  54. [62]

    Manasi Vartak, Arvind Thiagarajan, Conrado Miranda, Jeshua Bratman, and Hugo Larochelle. 2017. A meta-learning perspective on cold-start recommendations for items.Advances in neural information processing systems30 (2017)

    2017

  55. [63]

    Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. 2022. Generate rather than retrieve: Large language models are strong context generators.arXiv preprint arXiv:2209.10063(2022)

    2022 arXiv

  56. [64]

    Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. 2023. Where to go next for recommender systems? id- vs. modality-based recommender models revisited. InProceedings of the 46th International ACM SIGIR Conference on Research and Devel...

    2023

  57. [65]

    Myana, Deepak Dilipkumar, Suvadip Paul, Ikuhiro Ihara, Prasang Upadhyaya, Ferenc Huszár, and Wenzhe Shi

    Caojin Zhang, Yicun Liu, Yuanpu Xie, Sofia Ira Ktena, Alykhan Tejani, Ak- shay Gupta, Pranay K. Myana, Deepak Dilipkumar, Suvadip Paul, Ikuhiro Ihara, Prasang Upadhyaya, Ferenc Huszár, and Wenzhe Shi. 2020. Model Size Reduc- tion Using Frequency Based Double Hashing for Recomm...

    2020

  58. [66]

    Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, et al. 2024. Recommender systems in the era of large language models (llms).IEEE Transactions on Knowledge and Data Engineering(2024)

    2024

  59. [67]

    Longtao Xiao, Haozhao Wang, Cheng Wang, Linfei Ji, Yifan Wang, Jieming Zhu, Zhenhua Dong, Rui Zhang, and Ruixuan Li. 2025. Progressive Collaborative and Semantic Knowledge Fusion for Generative Recommendation.arXiv preprint arXiv:2502.06269(2025)

    2025

  60. [68]

    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

  61. [69]

    Yuhao Yang, Zhi Ji, Zhaopeng Li, Yi Li, Zhonglin Mo, Yue Ding, Kai Chen, Zijian Zhang, Jie Li, Shuanglong Li, et al. 2025. Sparse meets dense: Unified generative recommendations with cascaded sparse-dense representations.arXiv preprint arXiv:2503.02453(2025)

    2025 arXiv

  62. [70]

    Bowen Zheng, Hongyu Lu, Yu Chen, Wayne Xin Zhao, and Ji-Rong Wen. 2025. Universal Item Tokenization for Transferable Generative Recommendation.arXiv preprint arXiv:2504.04405(2025)

    2025 arXiv

  63. [71]

    Peilin Zhou, You-Liang Huang, Yueqi Xie, Jingqi Gao, Shoujin Wang, Jae Boum Kim, and Sunghun Kim. 2024. Is contrastive learning necessary? a study of data augmentation vs contrastive learning in sequential recommendation. In Proceedings of the ACM Web Conference 2024. 3854–3863

    2024

  64. [72]

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

  65. [73]

    Jing Zhu, Mingxuan Ju, Yozen Liu, Danai Koutra, Neil Shah, and Tong Zhao. 2025. Beyond Unimodal Boundaries: Generative Recommendation with Multimodal Semantics.arXiv preprint arXiv:2503.23333(2025)

    2025 arXiv

  66. [74]

    Zhe Zhao, Lichan Hong, Li Wei, Jilin Chen, Aniruddh Nath, Shawn Andrews, Aditee Kumthekar, Maheswaran Sathiamoorthy, Xinyang Yi, and Ed Chi. 2019. Recommending what video to watch next: a multitask ranking system. InPro- ceedings of the 13th ACM conference on recommender syste...

    2019

  67. [76]

    Bowen Zheng, Enze Liu, Zhongfu Chen, Zhongrui Ma, Yue Wang, Wayne Xin Zhao, and Ji-Rong Wen. 2025. Pre-training Generative Recommender with Multi- Identifier Item Tokenization.arXiv preprint arXiv:2504.04400(2025)

    2025 arXiv

  68. [80]

    InProceedings of the 18th ACM Conference on Recommender Systems

    CoST: Contrastive Quantization based Semantic Tokenization for Genera- tive Recommendation. InProceedings of the 18th ACM Conference on Recommender Systems. 969–974

  69. [82]

    Yongxin Zhu, Bocheng Li, Yifei Xin, and Linli Xu. 2024. Addressing representa- tion collapse in vector quantized models with one linear layer.arXiv preprint arXiv:2411.02038(2024)

    2024

  70. [2016]

    InProceedings of the 1st workshop on deep learning for recommender systems

    Wide & deep learning for recommender systems. InProceedings of the 1st workshop on deep learning for recommender systems. 7–10

  71. [2022]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 16000–16009

  72. [2023]

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

    2023

  73. [2024]

    Enhancing item tokenization for generative recommendation through self-improvement.arXiv preprint arXiv:2412.17171(2024)

    2024 arXiv

  74. [2025]

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

    Learning Universal User Representations Leveraging Cross-domain User Intent at Snapchat. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 4345–4349

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.