Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Generative recommenders can identify items with variable-length sets of ordinary words from an LLM's own vocabulary — and this paper claims that beats fixed-length quantized identifiers on accuracy while removing autoregressive decoding.

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

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

T0 review · deepseek-v4-flash

2026-08-01 03:05 UTC pith:RUALOHEW

load-bearing objection A credible, well-tested proposal for replacing quantized SIDs with LLM-native variable-length tokens; the main caveat is a self-referential quality metric, not the held-out results. the 4 major comments →

arxiv 2607.25209 v1 pith:RUALOHEW submitted 2026-07-28 cs.IR cs.AI

VaLiDRec: Variable-Length LLM-Aligned Semantic IDs for Generative Recommendation

Shutong Qiao , Wei Yuan , Tong Chen , Hao Wang , Quoc Viet Hung Nguyen , Hongzhi Yin This is my paper
classification cs.IR cs.AI
keywords generative recommendationsemantic identifiersLLM-aligned tokenstoken-set predictionvariable-length SIDsequential recommendationcold-start recommendationgraph-aware prompt
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's central claim is that the standard interface between language models and recommenders — fixed-length identifiers built by clustering or quantizing item representations into artificial codes — is the wrong abstraction. VaLiDRec instead constructs each item's identifier as a variable-length set of native LLM vocabulary tokens selected from the item's own metadata: tokens are scored for salience and corpus discrimination, greedily pruned while a cosine semantic-retention score stays above a threshold, and refined to break collisions, so identifier length adapts to each item's semantic complexity. User modeling is then cast as parallel token-set prediction: a graph-aware soft prompt injects behavioral transitions, the LLM (fine-tuned with LoRA) scores all identifier tokens in one forward pass, and item relevance is the average of its token scores. If the reported results hold, the paper shows that explicit lexical semantics in the model's own vocabulary outperforms opaque quantized codebooks across eight baselines and four datasets, handles never-seen items better, and removes the main decoding cost. The broader significance for a general reader: 'generative recommendation' treats 'what word should identifier space speak?' as a design decision, and this paper argues the answer is the language the model already speaks.

Core claim

VaLiDRec claims identifiers need not be learned codes: each item's identifier is built from its metadata text by scoring tokens by hidden-state norm and inverse document frequency, then greedily deleting the least essential tokens while the cosine similarity between retained and full token representations stays above 0.95. Collisions are resolved by adding back informative tokens plus a non-semantic suffix. Recommendation becomes parallel token-set prediction: a graph-aware soft prompt injects transition signals, LoRA fine-tunes the LLM, and item scores average predicted probabilities over identifier tokens. Results: gains over eight baselines, better zero-shot cold-start ranking, 87.49× fas

What carries the argument

The load-bearing object is the variable-length semantic identifier itself: a set of native LLM vocabulary tokens selected from item metadata. It is constructed in three steps — importance scoring (contextual hidden-state norm weighted by inverse document frequency), semantic-quality-aware greedy pruning driven by the cosine quality score Q(R) = cos(mean(retained tokens), mean(all metadata tokens)) with threshold 0.95, and collision-aware refinement plus a non-semantic suffix for tie-breaking. The second stage replaces autoregressive generation with parallel token-set prediction: an item-transition graph summarized by recency-weighted pooling becomes a soft prompt, and item scores are the mea

Load-bearing premise

The whole construction leans on the assumption that the cosine similarity between the retained identifier tokens' mean representation and the full metadata's mean representation, at threshold 0.95, is enough to guarantee the identifier still contains the specific words needed to tell similar items apart — the paper validates this internal score, not whether pruned tokens would have changed the ranking.

What would settle it

Build or find two items whose metadata differ mainly in tokens that greedy pruning drops while Q stays above 0.95, and check whether a history pointing at one item separates their ranks; if such pairs are routinely tied or misranked, the threshold is not preserving discriminative information. A cleaner experiment: sweep τ from 0.90 to 1.00 and watch Recall@20 — if it is flat, the pruning premise is not load-bearing; if it collapses below 0.95, the premise is doing real work.

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

If this is right

  • Fixed-length quantization codebooks are not necessary: building identifiers from the LLM's own vocabulary improves recall and NDCG across four datasets, with the largest gains on the sparsest data.
  • Identifier length can be matched to item complexity: pruning shrinks average SID length from about 15 tokens to 7.6–8.9 while raising the share of SIDs above the semantic-quality threshold from under 27% to over 83%.
  • Cold-start items that never appeared in training can still be ranked from their metadata alone, outperforming quantized-code baselines on zero-shot evaluation.
  • Inference becomes one LLM forward pass instead of beam search: 87.49× faster than the autoregressive baseline under the same backbone and hardware, with cost independent of SID length and beam size.
  • Suffix-based disambiguation makes every item's complete identifier unique (final collision rate 0) without contaminating semantic scores.

Where Pith is reading between the lines

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

  • Editorial: the strongest implicit claim is that the bottleneck in generative recommendation is vocabulary alignment rather than identifier capacity — if so, other fixed-code methods should benefit from mapping their codes into the model's native token space without retraining the whole pipeline.
  • Testable extension: sweep the retention threshold τ from 0.90 to 1.00 and measure ranking quality (not just the internal cosine score Q); if recall is flat across that range, the pruning premise is not actually load-bearing, and if it degrades sharply just below 0.95, the threshold is doing real work.
  • The 87.49× figure compares parallel token-set scoring against autoregressive generation with beam size 100; an equally informative benchmark would compare against a parallel-code baseline to separate the gains of parallelism from the gains of vocabulary alignment.
  • One unresolved variable: the SA2CRQ baseline was reproduced by the authors from the paper's description because no official code is released, so its numbers carry extra uncertainty and a different reproduction could shrink the reported margins.

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. VaLiDRec proposes a two-stage generative recommendation framework in which each item is represented by a variable-length set of native LLM vocabulary tokens chosen from its metadata. Stage 1 builds these semantic identifiers via token-importance scoring, semantic-quality-aware greedy pruning, and collision-aware refinement; Stage 2 fine-tunes an LLM with LoRA plus graph-aware soft prompts and trains it to predict target SID token sets in parallel. Item relevance is computed by averaging predicted token probabilities over each item's SID, avoiding autoregressive decoding and beam search. Experiments on four Amazon datasets claim consistent gains over eight sequential and generative baselines, better zero-shot item cold-start performance, and an 87.49x inference speedup over LC-Rec.

Significance. If the empirical claims hold, the framework is a meaningful step for generative recommendation: it replaces quantized codebooks with metadata-grounded, vocabulary-native variable-length identifiers, and the parallel token-scoring design is a practical efficiency improvement. The held-out next-item evaluation protects the recommendation claim from circularity, and the cold-start and efficiency studies address relevant deployment concerns. However, the paper's semantic-retention evidence is self-referential, several experimental reporting gaps remain, and the strongest efficiency comparison needs clarification. These issues are addressable but currently limit confidence in the headline 'more expressive' claim.

major comments (4)
  1. [§4.1.2, §5.8 (Eqs. 7/10; Table 5)] Q is both the pruning criterion and the only semantic-retention measure. The stopping rule in Eq. (10) removes tokens only while Q(R_j\{i*}) ≥ τ = 0.95, and Table 5 then reports the fraction of refined SIDs with Q ≥ 0.95 as evidence of semantic preservation. This is a consistency check, not an external semantic validation. Mean-pooled final-layer hidden states can remain close after removing rare but item-discriminating tokens, so the SID may lose exactly the lexical cues needed to separate similar items; Eq. (29) assumes the retained tokens are sufficient. The Fig. 3 ablation cannot isolate this because the 'initial' SIDs also differ in length and noise. Please add an independent semantic/discriminative evaluation (e.g., item classification or retrieval from SID tokens, human/LLM-rated token informativeness, or a downstream comparison against an alternative pruning rule such as pure IDF
  2. [§5.2/§5.6, Table 2] The paper states that results are averaged over five runs with a paired t-test at p<0.05, but Table 2 reports no standard deviations, confidence intervals, or significance markers. Section 5.6 shows that the optimal ranking-loss weight α varies by dataset (0.8 for three datasets, 0.2 for Arts), yet the main table does not state which α was used for each dataset. This makes it hard to determine whether the reported 'consistent gains' are statistically robust or reflect per-dataset hyperparameter selection. Please report variance or confidence intervals, significance indicators, and the exact hyperparameter configuration for each main result, including the positive-class weight ω in Eq. (22), which is otherwise unspecified.
  3. [§5.7, Table 4] The 87.49x speedup claim may not be an apples-to-apples comparison. LC-Rec is measured with beam size 100 and O(BL) sequential steps, while VaLiDRec uses one forward pass plus O(Σ_j |S_j|) token-to-item aggregations. The table caption says 'online inference time' but does not state whether the 0.078s figure includes full-catalog item scoring and suffix tie-breaking, nor how the candidate item set is defined for each method. Please specify the exact measurement scope and, ideally, report end-to-end latency for both systems under the same candidate space and hardware conditions.
  4. [§5.3, Table 2] The SA2CRQ baseline is a self-reproduction: the footnote says 'As the original implementation has not been publicly released, we reproduce SA2CRQ based on the method described in the paper.' Without access to the original code, the reproduction may not match the reference implementation, so the comparison carries an unquantified risk. Please release the reproduction code or validate it against public results, and explicitly discuss this limitation when interpreting Table 2.
minor comments (5)
  1. [§5.8, Table 5] Please define 'Length Compression' and clarify what 'Initial' means. The average initial length (~15) is far below the candidate pool size (64), so the comparison in Table 5 is not between the full pre-pruning pool and the refined SID.
  2. [§5.2] Several training hyperparameters are not reported: the number of epochs, the learning rate for the graph-prompt/GraphSAGE modules, and the value of ω in Eq. (22). Providing these details would materially improve reproducibility.
  3. [General] No code or data splits are provided. Given the number of tunable components (τ, λ, α, β, Lmin, Lmax, candidate pool size, collision rounds), a code-release plan or detailed configuration files are important for verification.
  4. [§5.6, Figure 4] The sensitivity curves in Figure 4 are hard to read from the small panels; please also provide a table of the exact α values and corresponding metrics.
  5. [Eq. (18)] Equation (18) defines recency weights over T_u interactions, while Section 5.2 says only the most recent 20 historical items are used. Please reconcile these definitions.

Circularity Check

1 steps flagged

Minor self-referential semantic-quality evaluation; central ranking results are external and not circular.

specific steps
  1. self definitional [§4.1.2 (Eqs. 7, 10) and §5.8 (Table 5)]
    "A larger Q(R_j) indicates better preservation of the metadata semantics. ... removal continues only if |R_j|>L_min and Q(R_j\{i*\})≥τ. (10) ... As shown in Table 5, SID refinement consistently improves semantic quality ... indicating that semantic-quality-aware pruning effectively removes redundant tokens while preserving item semantics."

    The semantic-retention evidence in Table 5 uses the same function Q that the greedy pruning directly optimizes: pruning stops whenever the next deletion would lower Q below τ=0.95 (Eq. 10), and §5.8 then reports the resulting high Q values as evidence of 'preserving item semantics'. The claim of semantic preservation is therefore substantially guaranteed by the stop rule rather than verified by an independent semantic metric. This does not affect the main recommendation-accuracy results (Table 2), which are evaluated on held-out next-item interactions external to the SID-construction objective.

full rationale

VaLiDRec's central claim—that LLM-native variable-length SIDs improve generative recommendation—is tested by held-out next-item recommendation (Recall/NDCG on four Amazon datasets), zero-shot cold-start ranking of unseen items, and wall-clock inference time against LC-Rec. None of these outcomes is a fitted parameter of the SID construction objective; the construction is unsupervised over item metadata, and recommendation quality is measured externally. The speedup comparison is mechanism-based (one parallel forward pass vs. O(BL) autoregressive decoding) and is not circular. No load-bearing self-citation or imported uniqueness theorem is present: related-work citations for quantized SID methods and baselines are external. The only self-referential element is the internal semantic-quality score Q: it is both the pruning objective (Eq. 7/10) and the success metric for the 'semantic retention' sub-claim (Table 5). Table 5 therefore demonstrates that the pruning algorithm achieves its own objective, not that the retained tokens preserve externally meaningful or discriminative semantics. This mildly inflates the semantic-retention contribution but does not undermine the externally evaluated ranking, cold-start, or efficiency results. Overall, the paper is self-contained against external benchmarks; the circularity is minor and non-load-bearing.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central empirical claim is measured on held-out interactions, so the free parameters above are model-selection knobs rather than fit-to-test constants. The most consequential assumptions are the internal semantic-quality proxy Q and the hidden-state-norm salience proxy; both are untested against external semantic benchmarks.

free parameters (7)
  • Semantic-quality threshold τ = 0.95
    Stopping threshold for greedy pruning; hand-selected, directly determines SID length and reported 'semantic retention'; no sensitivity analysis reported.
  • Distinctiveness coefficient λ = 0.05
    Trade-off between semantic preservation and IDF in Eq. (8); chosen by hand, no sensitivity analysis reported.
  • Ranking-loss weight α = 0.8 (three datasets), 0.2 (Arts)
    Tuned per dataset by inspecting validation curves in Figure 4; a free parameter fitted to data.
  • Contrastive-loss weight β = 0.1
    Fixed without sensitivity analysis; affects the L_item objective.
  • SID length bounds (Lmin, Lmax) = 2 and 16
    Hard constraints on identifier size; chosen by hand, affect representational capacity.
  • Candidate pool size = 64
    Number of token positions considered in Stage 1 before pruning; hand-set.
  • Collision expansion rounds = 8
    Maximum refinement rounds for collision resolution; hand-set.
axioms (5)
  • domain assumption The final-layer LLM hidden-state norm ||h_{j,i}||_2 is a valid proxy for token salience in item metadata (Eq. 4).
    Used to score and select candidate tokens in Stage 1; no external validation that high-norm tokens are the semantically informative ones.
  • domain assumption Cosine similarity between mean hidden states of a token subset and the full metadata (Q(R) in Eqs. 5-7) measures preservation of item semantics.
    This is the optimization target and the success metric in Table 5; there is no check against human judgment or downstream task utility.
  • domain assumption Consecutive interaction edges in user histories define a meaningful item-transition graph for collaborative preference modeling (Eq. 17, §4.2.1).
    Standard in sequential recommenders; plausible but unproved in this specific LLM-prompt setting.
  • domain assumption Recency-weighted pooling of graph-enhanced item representations summarizes user intent (Eq. 18).
    Used to build the soft prompt; the triangular weighting is one of many possible pooling schemes and is not derived.
  • standard math IDF computed over the item-metadata corpus is a reliable discriminativeness signal (Eq. 3).
    Standard TF-IDF idea; not ad hoc, but its suitability for short metadata texts is assumed.

pith-pipeline@v1.3.0-alltime-deepseek · 15255 in / 16245 out tokens · 157720 ms · 2026-08-01T03:05:44.233030+00:00 · methodology

0 comments
read the original abstract

Generative recommendation commonly represents items using fixed-length semantic identifiers (SIDs) constructed through clustering and quantization. However, these artificial codes may overcompress item semantics, remain misaligned with pretrained LLM vocabularies, and require costly autoregressive decoding. In light of this, we propose VaLiDRec, a generative recommendation framework based on variable-length, LLM-aligned semantic identifiers. VaLiDRec constructs SIDs directly from informative native LLM vocabulary tokens via token importance estimation, semantic-quality-aware pruning, and collision-aware refinement, allowing identifier lengths to adapt to item semantic complexity. To model user preferences, VaLiDRec incorporates graph-aware soft prompts and reformulates recommendation as token-set prediction with token-level item scoring, eliminating autoregressive SID generation and beam search. Experiments on four real-world datasets show that VaLiDRec consistently outperforms strong sequential and generative recommendation baselines across all evaluation metrics. It further achieves superior zero-shot item cold-start performance and 87.49$\times$ faster inference than LC-Rec. These results demonstrate that LLM-native variable-length semantic identifiers provide a more expressive and efficient paradigm for generative recommendation.

Figures

Figures reproduced from arXiv: 2607.25209 by Hao Wang, Hongzhi Yin, Quoc Viet Hung Nguyen, Shutong Qiao, Tong Chen, Wei Yuan.

Figure 1
Figure 1. Figure 1: Comparison of traditional fixed-length SID se [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of VaLiDRec. history. Given H𝑢, the sequential recommendation task is to predict the next item 𝑗 + 𝑢 ∈ I. Each item 𝑗 ∈ I is associated with metadata 𝑥𝑗 . From 𝑥𝑗 , we construct a variable-length semantic token set 𝑆𝑗 ⊆ V, where V denotes the pretrained LLM vocabulary. We formulate next-item recommendation as predicting the semantic token set of the target item: 𝑝𝜃  𝑆𝑗 + 𝑢 | H𝑢  . (1) A… view at source ↗
Figure 3
Figure 3. Figure 3: Ablation results on four datasets. (a) Luxury (b) Scientific (c) Instruments (d) Arts [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity to the ranking-loss weight 𝛼. components provide complementary gains, with the graph prompt contributing the largest improvement. 5.6 Hyperparameter Study We investigate the sensitivity of the item-level ranking loss weight 𝛼, which balances token-level SID prediction and direct item rank￾ing. We vary 𝛼 from 0.1 to 1.0 while fixing all other hyperparame￾ters. As shown in [PITH_FULL_IMAGE:figur… 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

44 extracted references · 12 linked inside Pith

  1. [1]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wang Wenjie, Fuli Feng, and Xiangnan He. 2023. Large language models for recommendation: Progresses and future directions. InProceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region. 306–309

    2023

  2. [2]

    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

  3. [3]

    Wenzhuo Cheng, Menghang Gong, Qixin Guo, Hang Zheng, Zhaobin Yang, Jianguo Lou, and Zhengwei Zheng. 2026. CapsID: Soft-Routed Variable-Length Semantic IDs for Generative Recommendation.arXiv preprint arXiv:2605.05096 (2026)

    Pith/arXiv arXiv 2026

  4. [4]

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

    Pith/arXiv arXiv 2025

  5. [5]

    Xinyu Du, Huanhuan Yuan, Pengpeng Zhao, Jianfeng Qu, Fuzhen Zhuang, Guan- feng Liu, Yanchi Liu, and Victor S Sheng. 2023. Frequency enhanced hybrid attention network for sequential recommendation. InProceedings of the 46th international ACM SIGIR conference on research and development in information retrieval. 78–88

    2023

  6. [6]

    Junchen Fu, Xuri Ge, Alexandros Karatzoglou, Ioannis Arapakis, Suzan Ver- berne, Joemon M Jose, and Zhaochun Ren. 2026. Differentiable Semantic ID for Generative Recommendation.arXiv preprint arXiv:2601.19711(2026)

    Pith/arXiv arXiv 2026

  7. [7]

    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

  8. [8]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783 (2024)

    Pith/arXiv arXiv 2024

  9. [9]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs.Advances in neural information processing systems30 (2017)

    2017

  10. [10]

    B Hidasi. 2015. Session-based Recommendations with Recurrent Neural Networks. arXiv preprint arXiv:1511.06939(2015)

    Pith/arXiv arXiv 2015

  11. [11]

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

    2023

  12. [12]

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

    2025

  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]

    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

  15. [15]

    Jian Jia, Jingtong Gao, Ben Xue, Junhao Wang, Qingpeng Cai, Quan Chen, Xi- angyu Zhao, Peng Jiang, and Kun Gai. 2025. From principles to applications: A comprehensive survey of discrete tokenizers in generation, comprehension, recommendation, and information retrieval.arXiv preprint arXiv:2502.12448 (2025)

    Pith/arXiv arXiv 2025

  16. [16]

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

    2018

  17. [17]

    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

  18. [18]

    Jing Li, Pengjie Ren, Zhumin Chen, Zhaochun Ren, Tao Lian, and Jun Ma. 2017. Neural attentive session-based recommendation. InProceedings of the 2017 ACM on Conference on Information and Knowledge Management. 1419–1428

    2017

  19. [19]

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

    2025

  20. [20]

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

  21. [21]

    Jingzhe Liu, Liam Collins, Jiliang Tang, Tong Zhao, Neil Shah, and Clark Mingx- uan Ju. 2025. Understanding generative recommendation with semantic ids from 9 Conference’17, July 2017, Washington, DC, USA Shutong Qiao et al. a model-scaling view.arXiv preprint arXiv:2509.25522(2025)

    Pith/arXiv arXiv 2025

  22. [22]

    Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. 2024. Finite scalar quantization: Vq-vae made simple. InInternational Conference on Learning Representations, Vol. 2024. 51772–51783

    2024

  23. [23]

    Aleksandr V Petrov and Craig Macdonald. 2023. Generative sequential recom- mendation with gptrec.arXiv preprint arXiv:2306.11114(2023)

    Pith/arXiv arXiv 2023

  24. [24]

    Haohao Qu, Wenqi Fan, Zihuai Zhao, and Qing Li. 2025. Tokenrec: Learning to tokenize id for llm-based generative recommendations.IEEE Transactions on Knowledge and Data Engineering(2025)

    2025

  25. [25]

    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

  26. [26]

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

  27. [27]

    Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factor- izing personalized markov chains for next-basket recommendation. InProceedings of the 19th international conference on World wide web. 811–820

    2010

  28. [28]

    Gerard Salton and Christopher Buckley. 1988. Term-weighting approaches in automatic text retrieval.Information processing & management24, 5 (1988), 513–523

    1988

  29. [29]

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

  30. [30]

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

    2018

  31. [31]

    Hao Wang, Wei Guo, Luankang Zhang, Jin Yao Chin, Yufei Ye, Huifeng Guo, Yong Liu, Defu Lian, Ruiming Tang, and Enhong Chen. 2025. Generative large recom- mendation models: emerging trends in llms for recommendation. InCompanion Proceedings of the ACM on Web Conference 2025. 49–52

    2025

  32. [32]

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

    Pith/arXiv arXiv 2026

  33. [33]

    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

  34. [34]

    Wenjie Wang, Xinyu Lin, Fuli Feng, Xiangnan He, and Tat-Seng Chua. 2023. Generative recommendation: Towards next-generation recommender paradigm. arXiv preprint arXiv:2304.03516(2023)

    Pith/arXiv arXiv 2023

  35. [35]

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

    2024

  36. [36]

    Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan. 2019. Session-based recommendation with graph neural networks. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 346–353

    2019

  37. [37]

    Xu Xie, Fei Sun, Zhaoyang Liu, Shiwen Wu, Jinyang Gao, Jiandong Zhang, Bolin Ding, and Bin Cui. 2022. Contrastive learning for sequential recommendation. In 2022 IEEE 38th international conference on data engineering (ICDE). IEEE, 1259– 1273

    2022

  38. [38]

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

    Pith/arXiv arXiv 2024

  39. [39]

    Zhen Zhang, Jujia Zhao, Xinyu Ma, Xin Xin, Maarten de Rijke, and Zhaochun Ren. 2026. Cold-Starts in Generative Recommendation: A Reproducibility Study. arXiv preprint arXiv:2603.29845(2026)

    Pith/arXiv arXiv 2026

  40. [40]

    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

  41. [41]

    Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S3-rec: Self-supervised learning for se- quential recommendation with mutual information maximization. InProceedings of the 29th ACM international conference on information & knowledge management. 1893–1902. 10

    2020

  42. [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

  43. [2023]

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

    2023

  44. [2025]

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

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