Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Item identifiers that evolve with behavioral feedback beat fixed codes in generative recommenders.

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 02:44 UTC pith:DCECIQDN

load-bearing objection Novel idea—evolving item identifiers post-hoc with the model's own posteriors—but the evidence for the core mechanism is not yet airtight; send to review with expectation of major revision. the 3 major comments →

arxiv 2607.25329 v1 pith:DCECIQDN submitted 2026-07-28 cs.IR

Grevo: A Unified Generative Recommendation Framework with Evolutionary Item Indexing

Huanjie Wang , Liwei Guan , Zekai Sun , Hongwei Zhang , Honghui Bao This is my paper
classification cs.IR
keywords Generative recommendationSemantic identifiersItem indexingEvolutionary searchMultitask learningAutoregressive retrievalRecommender systemsPosterior evaluation
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.

Grevo claims that in generative recommendation, the item identifier—the code a model emits to retrieve an item—should be treated as an evolvable variable rather than as something fixed by a content-based tokenizer or jointly trained with the recommender. The paper builds a single sequence-to-sequence recommender with two tasks: generate the next item's code from a user's history (behavioral) and regenerate the same code from the item's own content embedding (semantic grounding). After training, that same model scores how easy each candidate code is to generate and to ground, and Grevo reassigns a small budget of 'high-risk' identifiers using tokens borrowed from behaviorally confused neighbors and underused codebook entries. Because edits respect the original vocabulary and code length, the evolved index is a drop-in replacement that transfers to other generative recommenders. On three Amazon benchmarks, Grevo reports consistent gains over strong generative baselines, with the evolution step adding gains on top of the multitask training.

Core claim

The paper's central claim is that the tokenizer that maps items to semantic identifiers can be reduced to a disposable initializer: once it has bootstrapped a codebook, the recommender itself can take over the role of refining the index. Grevo demonstrates this with a unified multitask model that predicts an item's identifier both from user behavior (BSG) and from its content representation (SSG); the two tasks share embeddings, so the semantic task enriches the token space without a separate tokenizer. Using this trained model as a posterior evaluator, Grevo then evolves a budgeted fraction of identifiers, scoring candidate edits by how much they improve behavioral generation likelihood, se

What carries the argument

The central mechanism is evolutionary item indexing, a round-based loop: pre-train the unified BSG/SSG recommender on the current index, collect posterior likelihood signals, enumerate candidate identifiers constrained to the fixed vocabulary and length, score each candidate by alignment gain (Δr_BSG + λ1 Δr_SSG + λ2 Δr_cons), and commit the highest-gain reassignments up to a budget. The candidate pool mixes the item's current token, tokens from items the model confuses with it, and lightly loaded codebook tokens. This turns identifier construction into a coordinate-descent-like discrete search over the index, with the model itself as the scorer and no second learnable model.

Load-bearing premise

The load-bearing premise is that the recommender's likelihood scores computed on the training contexts where an item appears are a faithful surrogate for held-out recommendation quality; if that posterior alignment does not transfer to unseen user histories, evolution could make identifiers easier to generate without making recommendations better.

What would settle it

Train Grevo on a dataset and, before retraining on the evolved index, take the round-0 candidate edits and score them on held-out validation sequences that were never used for scoring. If the average teacher-forced log-likelihood gain on held-out contexts does not correlate with the gains on training contexts, or if randomly selected budgeted edits achieve comparable held-out ranking gain, then the posterior surrogate is not the cause of the improvement.

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

If this is right

  • A generative recommender can absorb the tokenizer's role: after initial bootstrapping, index refinement needs no second learnable model, alignment losses, or alternating optimization.
  • Evolved identifiers are drop-in replacements: because vocabulary and code length are unchanged, the improved index can be fed to a different generative recommender, and the paper's transfer test shows it helps without touching that recommender's own training.
  • Behavior-driven index repair and multitask training are complementary; each contributes gains separately, so both matter for final quality.
  • The method makes progress round over round but converges after a few rounds, suggesting that a small, targeted budget of edits is better than large-scale re-indexing.
  • The 'semantic–behavioral gap' in generative recommendation can be narrowed post-hoc, offering a path to improving already-deployed generative recommenders by swapping their index.

Where Pith is reading between the lines

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

  • A concrete extension, not tested in the paper, is streaming settings where new items arrive: the same round-based evolution could reassign codes as behavioral feedback accumulates, using the model's own posteriors rather than retraining a tokenizer.
  • The posterior-evaluator idea could generalize beyond recommendation: any sequence target with a learnable codebook, such as hash codes in dense retrieval or label codes for classification, might be evolved by the model's own likelihoods after training.
  • A sharper test of the surrogate signal would score candidate edits on truly held-out histories before retraining, isolating whether training-context gains predict held-out ranking gains rather than reflecting the retrained model's familiarity with the new labels.
  • If the transfer result holds broadly, index evolution becomes a reusable artifact: one curated index could be built once and consumed by many recommenders, decoupling indexing research from architecture research.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Grevo, a generative recommendation framework in which item semantic identifiers (SIDs) are treated as a discrete variable that can be evolved after training, rather than produced by a frozen content tokenizer or a jointly trained second model. A single T5-style multitask model is trained on two tasks, Behavioral SID Generation (BSG) and Semantic SID Grounding (SSG); after training, the model's own posterior scores (Eqs. 4–6) are used to rank candidate SID reassignments, generated by mixing the item's current tokens with confusion-neighbor tokens and low-load codebook tokens (Eqs. 7–8). Edits are committed greedily under a budget ratio (Eq. 10). The lifecycle is pre-train, collect, enumerate, evolve, retrain. Experiments on three Amazon datasets compare against sequential and generative baselines, including TIGER, LETTER, and LC-Rec, with a transfer test in which LC-Rec is trained on TIGER+Grevo's evolved index. The paper reports consistent improvements from both the unified multitask model and the evolutionary indexing component.

Significance. If the central claim is upheld, Grevo offers a meaningful and practical contribution: a model-agnostic way to refine SID assignments after training, without a second trainable tokenizer, alignment losses, or alternating optimization. The framework is clearly described, the lifecycle is easy to follow, and the experimental design includes useful elements: evaluation on multiple backbones and datasets, an index-transfer test in which LC-Rec uses only the evolved SID, an ablation of the scoring signals and candidate sources, and a dedicated decodability test. These are strengths that go beyond a simple end-to-end comparison. However, the paper's central evidence for the selection mechanism is weakened by a circular validation in Section 4.3.2, by single-seed results with no variance or significance reporting, and by the absence of a random-relabeling control. The observed gains are modest in several comparisons, so these methodological issues are load-bearing for the paper's claim that posterior-guided evolution, rather than retraining or the candidate pool itself, drives improvement. The idea is promising and likely salvageable, but the current manuscript does not yet establish the c

major comments (3)
  1. [Section 4.3.2, Figure 3] The 'held-out codeword decodability' test is contaminated by retraining. The round-1 model is trained with the evolved SIDs as target labels, so when the evolved SID is teacher-forced it is a label that appeared in training, while the original TIGER SID never appears as a target in round 1. The lower NLL and rank for the evolved SID are therefore expected from label substitution alone, independent of whether the new SID is intrinsically easier to predict from held-out user behavior. The sentence 'any improvement is attributable solely to the index edit and cannot reflect overfitting the round-0 posterior signal' is not justified: the retraining step already gives the evolved label an advantage. To support the surrogate claim, the authors need a control that separates the index edit from the change in training labels, for example scoring both SIDs with a single round-0 model without retra
  2. [Section 4.1 and Table 2] All experiments use a single seed (Section 4.1: 'seed 32'), and no error bars, variance, or significance tests are reported. The evolutionary gains are often modest (e.g., Beauty TIGER R@5 from 0.0428 to 0.0484; Toys TIGER R@5 from 0.0404 to 0.0455), so it is not possible to determine whether these differences are stable or noise. Additionally, the default hyperparameters are selected on Beauty (Section 4.4, Figure 4), the same dataset used for the ablation study, which further limits the strength of the cross-dataset generalization claim. Multiple seeds with standard deviations, or at least a paired significance test, are needed before the paper can claim consistent improvement.
  3. [Section 3.5 and Table 2] There is no control that isolates the selection mechanism in Eq. (10). The '+Grevo (no-evo)' control shows the effect of the unified BSG+SSG model with a fixed index, but the comparison between no-evo and full +Grevo conflates three factors: the posterior-guided selection, the extra retraining round on changed labels, and the candidate pool construction. Given that the evolution gains are modest, a random-relabeling control—where the same number of items and the same candidate sources are used but SIDs are reassigned at random, with the same retraining protocol—is needed to show that Eq. (10)'s scoring, rather than the extra round of optimization or the candidate pool itself, is responsible for the improvement. This is a load-bearing issue for the paper's central mechanism.
minor comments (5)
  1. [Section 3.3, Eq. (6)] The variables u_k and v_k are introduced in prose but not defined in the equation or immediately before Eq. (6). Please define them explicitly as per-layer BSG and SSG log-probabilities of the candidate token at layer k.
  2. [Section 3.2] The phrase 'training-after discrete refinement' is unclear; it appears to be a typo for 'training-free' or 'post-training' discrete refinement. Please rephrase.
  3. [Section 4.2, third bullet] Minor capitalization error: 'As a result, It provides' should be 'it provides'.
  4. [Figure 3 caption] The caption reports '357 changed items' while the default budget ratio 0.05 on Beauty would allow about 605 changes; the relationship between the budget and the actual number of changed items should be clarified, especially since Figure 3(a) says '357 changed items' and the text says 'for every edited item.'
  5. [Section 4.1, Implementation Details] The sentence 'where evolve round is 1' is awkwardly phrased; please clarify whether this means one evolution round or the first evolution round.

Circularity Check

1 steps flagged

Held-out decodability validation is circular: the round-1 model was retrained on the evolved SIDs used as labels, so the reported NLL gain is expected from label substitution and does not validate the Eq. (10) selection surrogate.

specific steps
  1. fitted input called prediction [Section 4.3.2 (Held-out Codeword Decodability), with scoring defined in Eqs. (4)-(10)]
    "With the round-1 model and user history fixed, we score the original TIGER SID and the evolved SID as two alternative labels for that item via teacher forcing, reporting the negative log-likelihood (NLL) and the rank of each correct token. Because the model, user history, and target item are all held fixed, the SID is the only variable in each paired comparison; any improvement is attributable solely to the index edit and cannot reflect overfitting the round-0 posterior signal."

    Candidate SIDs are chosen by Eq. (10), whose primary term is Δr_BSG = log pθ0(s'|C_i) − log pθ0(s_i|C_i) under the round-0 model on training contexts. The round-1 recommender is then retrained with the evolved SIDs as target labels, i.e., it is explicitly optimized to raise log pθ1(s'|·) for each edited item and it never sees s_i as a target label. The §4.3.2 comparison then reports log pθ1(s'|c) vs log pθ1(s_i|c) under that retrained model. The paper's claim that 'the model, user history, and target item are all held fixed' is misleading: pθ1 is not the round-0 model that produced the posterior scores; it is a model fitted to the new labels. Lower NLL/rank for s' is therefore expected from label substitution alone and does not test whether the Eq. (10) training-context posterior gain tran

full rationale

The paper's headline empirical claim — Grevo outperforms generative baselines — is supported by direct held-out Recall/NDCG in Table 2 and by ablations in Table 3, so the central benchmark result is not itself circular. The circularity is confined to the mechanism-level validation in §4.3.2, which the paper uses to argue that the Eq. (10) posterior-guided edit, rather than the extra retraining round or label substitution, makes identifiers genuinely more decodable. Because the round-1 model was trained on the evolved labels, the paired old-vs-new NLL comparison is not an independent check of the selection surrogate. This is a fitted-input-called-prediction pattern: the 'prediction' that evolved SIDs are easier to generate is partly forced by the retraining objective. The paper also lacks a random-relabeling control in Table 2, which would be needed to separate the effect of posterior-guided selection from any relabeling within the candidate pool; that is a confound rather than a definitional circularity. Self-citations (e.g., PIT, OneRec-V2, LETTER, OneReason) are used descriptively in related work and are not load-bearing; no uniqueness theorem or ansatz is imported from the authors' prior work. The LC-Rec index-transfer experiment is an independent, non-circular external test. Overall, one significant circular validation step warrants a partial-circularity score of 6, while the direct benchmark results keep the paper from being wholly circular.

Axiom & Free-Parameter Ledger

9 free parameters · 6 axioms · 0 invented entities

The paper does not present a formal derivation of why posterior-guided SID evolution improves recommendation; the framework is an algorithm whose success rests on empirical hyperparameters and domain assumptions. The central claim is carried by benchmark results, not by a first-principles argument.

free parameters (9)
  • SSG loss weight = 0.07
    Chosen for training; balances BSG and SSG losses in the unified model (Section 4.1).
  • Mutable layers = last two of K=4 layers
    Figure 4(a) shows two mutable layers give the strongest result; restricts evolution to lower-level codewords.
  • Budget ratio ρ = 0.05
    Figure 4(b) shows a plateau at small-to-moderate budgets; caps the number of evolved identifiers per round.
  • Alignment weight λ1 (SSG) = 0.25
    Figure 4(c) peaks at 0.25; controls how strongly semantic-grounding gain contributes to candidate score.
  • Alignment weight λ2 (consistency) = 0.15
    Figure 4(d) shows an intermediate optimum at 0.15; controls the cross-task consistency term.
  • Beam size for posterior diagnostics = 50
    Constrained beam search with beam size 50 used for collecting posterior signals (Section 4.1).
  • Retained user contexts per item = 32
    Up to 32 user contexts per target item are retained for exact alignment scoring (Section 4.1).
  • Candidate enumeration cap = 64
    Up to 64 candidates per item are enumerated from confusion-neighbor and low-load token sources (Section 4.1).
  • Evolution rounds = 1 (default in main results)
    Implementation details state 'evolve round is 1', though Figure 4(e) shows continued gains through round 3; the main table may under-use the method or be cost-limited.
axioms (6)
  • domain assumption Initial RQ-VAE/RQ-Kmeans semantic identifiers provide a useful starting manifold for evolution.
    Grevo discards the tokenizer after bootstrap and only edits a budgeted subset on a fixed vocabulary/length (Section 3.1, Eq. (1)); if the initial SIDs were poor, the constrained local search could not escape the manifold.
  • domain assumption The recommender's BSG and SSG posteriors are valid surrogate signals for recommendation quality.
    Candidate scoring in Section 3.5 (Eqs. (9)-(10)) uses log-probabilities from the trained model; the method assumes edits with higher posterior alignment on training contexts improve held-out recommendation.
  • ad hoc to paper Confusion-neighbor and low-load token pools cover the useful directions for SID repair.
    Candidate enumeration in Section 3.4 (Eqs. (7)-(8)) restricts search to the current token, tokens of posterior confusion neighbors, and low-load tokens; the ablation shows these heuristics matter.
  • domain assumption Multitask SSG training improves BSG recommendation through shared token representations.
    The unified model relies on SSG (Section 3.1, Eq. (3)) to let the recommender absorb the tokenizer's role; the benefit is empirically demonstrated but not derived.
  • domain assumption Leave-one-out evaluation on 5-core Amazon data is a meaningful proxy for recommendation performance.
    Section 4.1 states preprocessing follows [9, 28] and uses leave-one-out; all conclusions inherit the standard limitations of this benchmark protocol.
  • domain assumption Frozen LLaMA3-8B content embeddings provide a fair common basis for SID construction across baselines.
    Section 4.1 uses the same frozen LLaMA3-8B embeddings for TIGER, LETTER, LC-Rec, and OneRec-V2; fairness of comparisons depends on this choice.

pith-pipeline@v1.3.0-alltime-deepseek · 17080 in / 15243 out tokens · 147209 ms · 2026-08-01T02:44:05.816174+00:00 · methodology

0 comments
read the original abstract

Generative recommendation has recently emerged as a promising paradigm that reformulates retrieval as autoregressive generation over semantic identifiers (SIDs), achieving strong performance and drawing increasing attention as an alternative to matching. Despite this progress, SIDs are typically frozen by a content-based tokenizer before the recommender is trained, leaving a persistent gap between what best reconstructs an item's content and what a recommender can predict from user behavior. Recent end-to-end methods close this gap by jointly training the tokenizer and the recommender, but coupling the two destabilizes the identifier space and requires a second learnable model, alignment losses, and usually a delicate alternating-optimization schedule. To address this issue, we propose Grevo, a unified Generative recommendation framework with evolutionary item indexing, which treats the SID assignment itself as an evolvable discrete variable that adapts to behavioral feedback rather than as a tokenizer to be trained. Grevo builds on a single multitask recommender that unifies a behavioral SID generation task and a semantic SID grounding task, letting the recommender absorb the tokenizer's role. Through evolutionary item indexing, Grevo then uses the trained recommender itself as a posterior evaluator to reassign a budgeted set of high-risk identifiers under a fixed vocabulary and length. Together, these components turn index construction into a stable, feedback-driven search that requires no second learnable model, no alignment losses, and no alternating-optimization schedule. Extensive experiments on multiple real-world datasets demonstrate that Grevo consistently outperforms state-of-the-art generative recommendation methods.

Figures

Figures reproduced from arXiv: 2607.25329 by Honghui Bao, Hongwei Zhang, Huanjie Wang, Liwei Guan, Zekai Sun.

Figure 1
Figure 1. Figure 1: Comparison of GR indexing paradigms. (a) Static methods freeze the tokenizer before recommender training, leaving [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The Grevo per-round lifecycle: pre-train → collect → enumerate → evolve → retrain. The trained BSG/SSG model acts as a posterior evaluator that scores candidate SID reassignments, and the evolved index feeds the next round. Only the SID assignment on mutable layers is updated. and takes a coordinate-descent-like step in the discrete SID space, evolving a small fraction of identifiers per round toward the r… view at source ↗
Figure 3
Figure 3. Figure 3: Held-out SID decodability on Beauty. (a) Paired NLLs of original and evolved SIDs; points below [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hyper-parameter sensitivity and evolution-round analysis on Beauty. All five panels report absolute R@10 on the left [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 9 linked inside Pith

  1. [1]

    Yimeng Bai, Chang Liu, Yang Zhang, Dingxian Wang, Frank Yang, Andrew Rabi- novich, Wenge Rong, and Fuli Feng. 2026. Bi-Level Optimization for Generative Recommendation: Bridging Tokenization and Generation. InProceedings of the 49th International ACM SIGIR Conference on Research and Development in Infor- mation Retrieval. ACM, 53–64. doi:10.1145/3805712.3809632

    arXiv 2026

  2. [2]

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum Learning. InProceedings of the 26th Annual International Conference on Machine Learning, ICML 2009, Montreal, Quebec, Canada, June 14-18, 2009. ACM, 41–48. doi:10.1145/1553374.1553380

    arXiv 2009

  3. [3]

    Dietterich

    Thomas G. Dietterich. 2000. Ensemble Methods in Machine Learning. InMultiple Classifier Systems, First International Workshop, MCS 2000, Cagliari, Italy, June 21-23, 2000, Proceedings (Lecture Notes in Computer Science, Vol. 1857), Josef Kittler and Fabio Roli (Eds.). Springer, 1–15. doi:10.1007/3-540-45014-9_1

    doi:10.1007/3-540-45014-9_1 2000

  4. [4]

    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). InRecSys ’22: Sixteenth ACM Conference on Recommender Systems, Seattle, W A, USA, September 18 - 23, 2022, Jennifer Golbeck, F. Maxwell Harper, Vanessa Murdock, Michael ...

    arXiv 2022

  5. [5]

    Aaron Grattafiori, Abhimanyu Dubey, et al. 2024. The Llama 3 Herd of Models. arXiv:2407.21783 [cs] doi:10.48550/arXiv.2407.21783

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.21783 2024

  6. [6]

    Ruining He and Julian J. McAuley. 2016. Ups and Downs: Modeling the Visual Evolution of Fashion Trends with One-Class Collaborative Filtering. InProceed- ings of the 25th International Conference on World Wide Web, WWW 2016, Montreal, Canada, April 11 - 15, 2016, Jacqueline Bourdeau, Jim Hendler, Roger Nkambou, Ian Horrocks, and Ben Y. Zhao (Eds.). ACM, 5...

    arXiv 2016

  7. [7]

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

  8. [8]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2021. Billion-Scale Similarity Search with GPUs.IEEE Trans. Big Data7, 3 (2021), 535–547. doi:10.1109/TBDATA. 2019.2921572

    arXiv 2021

  9. [9]

    Wang-Cheng Kang and Julian J. McAuley. 2018. Self-Attentive Sequential Rec- ommendation. InIEEE International Conference on Data Mining, ICDM 2018, Singapore, November 17-20, 2018. IEEE Computer Society, 197–206. doi:10.1109/ ICDM.2018.00035

    arXiv 2018

  10. [10]

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregressive Image Generation using Residual Quantization. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. IEEE, 11513–11522. doi:10.1109/CVPR52688.2022.01123

    arXiv 2022

  11. [11]

    Lei Li, Yongfeng Zhang, Dugang Liu, and Li Chen. 2024. Large Language Mod- els for Generative Recommendation: A Survey and Visionary Discussions. In Proceedings of the 2024 Joint International Conference on Computational Linguis- tics, Language Resources and Evaluation, LREC/COLING 2024, 20-25 May, 2024, Torino, Italy, Nicoletta Calzolari, Min-Yen Kan, Vé...

    2024

  12. [12]

    Xiaopeng Li, Bo Chen, Junda She, Shiteng Cao, You Wang, Qinlin Jia, Haiying He, Zheli Zhou, Zhao Liu, Ji Liu, Zhiyang Zhang, Yu Zhou, Guoping Tang, Yiqing Yang, Chengcheng Guo, Si Dong, Kuo Cai, Pengyue Jia, Maolin Wang, Wanyu Wang, Shiyao Wang, Xinchen Luo, Qigen Hu, Qiang Luo, Xiao Lv, Chaoyi Ma, Ruiming Tang, Kun Gai, Guorui Zhou, and Xiangyu Zhao. 202...

    arXiv 2025

  13. [13]

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

  14. [14]

    Chen Ma, Peng Kang, and Xue Liu. 2019. Hierarchical Gating Networks for Se- quential Recommendation. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD 2019, Anchorage, AK, USA, August 4-8, 2019. ACM, 825–833. doi:10.1145/3292500.3330984

    arXiv 2019

  15. [15]

    Malkov and Dmitry A

    Yury A. Malkov and Dmitry A. Yashunin. 2020. Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs.IEEE Trans. Pattern Anal. Mach. Intell.42, 4 (2020), 824–836. doi:10.1109/TPAMI.2018. 2889473

    doi:10.1109/tpami.2018 2020

  16. [16]

    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. https://jmlr.org/papers/v21/20-074.html

    2020

  17. [17]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Kesha- van, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Mahesh Sathiamoorthy. 2023. Rec- ommender Systems with Generative Retrieval. InAdvances in Neural Infor- mation Processing Systems 36: Annual Conference on Neural Information Pro- ...

    2023

  18. [18]

    Prajit Ramachandran, Barret Zoph, and Quoc V. Le. 2017. Searching for Activation Functions. arXiv:1710.05941 [cs.NE] https://arxiv.org/abs/1710.05941

    Pith/arXiv arXiv 2017

  19. [19]

    Zihua Si, Zhongxiang Sun, Jiale Chen, Guozhang Chen, Xiaoxue Zang, Kai Zheng, Yang Song, Xiao Zhang, Jun Xu, and Kun Gai. 2024. Generative Retrieval with Semantic Tree-Structured Identifiers and Contrastive Learning. InProceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific...

    2024

  20. [20]

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

  21. [21]

    Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to Sequence Learning with Neural Networks. InAdvances in Neural Information Pro- cessing Systems 27: Annual Conference on Neural Information Processing Sys- tems 2014, December 8-13 2014, Montreal, Quebec, Canada, Zoubin Ghahra- mani, Max Welling, Corinna Cortes, Neil D. Lawrence, and Kilian Q....

    2014

  22. [22]

    Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Jian Wu, and Yuning Jiang. 2025. Think Before Recommend: Unleashing the Latent Reasoning Power for Sequential Recommendation. arXiv:2503.22675 [cs] doi:10.48550/arXiv. 2503.22675

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.22675 2025

  23. [23]

    Jiaxi Tang and Ke Wang. 2018. Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding. InProceedings of the Eleventh ACM International Conference on Web Search and Data Mining, WSDM 2018, Marina Del Rey, CA, USA, February 5-9, 2018, Yi Chang, Chengxiang Zhai, Yan Liu, and Yoelle Maarek (Eds.). ACM, 565–573. doi:10.1145/3159652.3159656

    arXiv 2018

  24. [24]

    OneRec Team, Biao Yang, Boyang Ding, Chenglong Chu, Dunju Zang, Fei Pan, Han Li, Hao Jiang, Honghui Bao, Huanjie Wang, Jian Liang, Jiangxia Cao, Jiao Ou, Jiaxin Deng, Jinghao Zhang, Kun Gai, Lu Ren, Peiru Du, Pengfei Zheng, Rongzhou Zhang, Ruiming Tang, Shiyao Wang, Siyang Mao, Siyuan Lou, Teng Shi, Wei Yuan, Wenlong Xu, Xingchen Liu, Xingmei Wang, Xinqi ...

    Pith/arXiv arXiv 2018

  25. [25]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guil- laume Lample. 2023. LLaMA: Open and Efficient Foundation Language Models. arXiv:2302.13971 [cs] doi:10.48550/arXiv.2302.13971

    Pith/arXiv arXiv doi:10.48550/arxiv.2302.13971 2023

  26. [26]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. InAdvances in Neural Information Processing Systems 30: An- nual Conference on Neural Information Processing Systems 2017, December 4- 9, 2017, Long Beach, CA, USA, Isabelle Guyon, Ulrike von Luxbur...

    2017

  27. [27]

    Huanjie Wang, Xinchen Luo, Honghui Bao, Zixing Zhang, Lejian Ren, Yun- fan Wu, Hongwei Zhang, Liwei Guan, and Guang Chen. 2026. PIT: A Dy- namic Personalized Item Tokenizer for End-to-End Generative Recommendation. arXiv:2602.08530 [cs.IR] https://arxiv.org/abs/2602.08530

    arXiv 2026

  28. [28]

    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, CIKM 2024, Boise, ID, USA, October 21-25, 2024, Edoardo Serra and Francesca Spezzano (Eds....

    arXiv 2024

  29. [29]

    Wenjie Wang, Xinyu Lin, Fuli Feng, Xiangnan He, and Tat-Seng Chua. 2023. Generative Recommendation: Towards Next-generation Recommender Paradigm. arXiv:2304.03516 [cs.IR] doi:10.48550/arXiv.2304.03516

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.03516 2023

  30. [30]

    Le, Ed H

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V. Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-Consistency Improves Chain of Thought Reasoning in Language Models. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net. https://openreview.net/forum?id=1PL1NIMMrw

    2023

  31. [31]

    Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, and Zhenhua Dong. 2024. EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024, Ricardo B...

    arXiv 2024

  32. [32]

    Zhengyi Yang, Xiangnan He, Jizhi Zhang, Jiancan Wu, Xin Xin, Jiawei Chen, and Xiang Wang. 2023. A Generic Learning Framework for Sequential Recommenda- tion with Distribution Shifts. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023, Taipei, Taiwan, July 23-27, 2023. ACM, 331–340....

    arXiv 2023

  33. [33]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, Yinghai Lu, and Yu Shi. 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations. InForty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. Op...

    2024

  34. [34]

    Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, Yingqian Min, Zhichao Feng, Xinyan Fan, Xu Chen, Pengfei Wang, Wendi Ji, Yaliang Li, Xiaoling Wang, and Ji-Rong Wen. 2021. RecBole: Towards a Unified, Comprehensive and Efficient Framework for Recommendation Algorithms. InCIKM ’21: Th...

    arXiv 2021

  35. [35]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting Large Language Models by Integrating Col- laborative Semantics for Recommendation. In40th IEEE International Conference on Data Engineering, ICDE 2024, Utrecht, The Netherlands, May 13-16, 2024. IEEE, 1435–1448. doi:10.1109/ICDE60146.2024.00118

    arXiv 2024

  36. [36]

    Guorui Zhou, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Qiang Luo, Qianqian Wang, Qigen Hu, Rui Huang, Shiyao Wang, Weifeng Ding, Wuchao Li, Xinchen Luo, Xingmei Wang, Zexuan Cheng, Zixing Zhang, Bin Zhang, Boxuan Wang, Chaoyi Ma, Chengru Song, Chenhui Wang, Di Wang, Dongxue Meng, Fan Yang, Fangyu Zhang, Feng Jiang, Fuxing Zhang, Gang Wang, Guowang ...

    doi:10.48550/arxiv.2506.13695 2025

  37. [37]

    Guorui Zhou, Hengrui Hu, Hongtao Cheng, Huanjie Wang, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Lu Ren, Liao Yu, Pengfei Zheng, Qiang Luo, Qianqian Wang, Qigen Hu, Rui Huang, Ruiming Tang, Shiyao Wang, Shujie Yang, Tao Wu, Wuchao Li, Xinchen Luo, Xingmei Wang, Yi Su, Yunfan Wu, Zexuan Cheng, Zhanyu Liu, Zixing Zhang, Bin Zhang, Boxuan Wang, Chaoyi ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.20900 2025

  38. [38]

    Kuan Zou and Aixin Sun. 2025. A Survey of Real-World Recommender Systems: Challenges, Constraints, and Industrial Perspectives. arXiv:2509.06002 [cs] doi:10. 48550/arXiv.2509.06002

    Pith/arXiv arXiv doi:10.48550/arxiv.2509.06002 2025

  39. [2016]

    In International Conference on Learning Representations

    Session-based Recommendations with Recurrent Neural Networks. In International Conference on Learning Representations. arXiv:1511.06939 [cs.LG] https://arxiv.org/abs/1511.06939

    Pith/arXiv arXiv

  40. [2019]

    InProceedings of the 28th ACM International Con- ference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Repre- sentations from Transformer. InProceedings of the 28th ACM International Con- ference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019. ACM, 1441–1450. doi:10.1145/3357384.3357895

    arXiv 2019

  41. [2025]

    Generative Recommender with End-to-End Learnable Item Tokenization. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2025, Padua, Italy, July 13-18, 2025, Nicola Ferro, Maria Maistro, Gabriella Pasi, Omar Alonso, Andrew Trotman, and Suzan Verberne (Eds.). ACM, 729–739. doi:10.1145/372...

    arXiv 2025