Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Generative recommenders trained with token-level loss lose correct items to beam-search pruning; a prefix-aware training objective fixes the mismatch and measurably improves ranking.

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 ↗

Prefix-level pointwise and pairwise losses with adaptive worst-prefix weighting improve beam-search ranking in generative recommendation across multiple backbones.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Empirically promising and well-tested prefix-aware loss for beam-search generative recommendation, but the advertised lower-bound theorem is unproven and should be removed or fixed before the paper is accepted. the 3 major comments →

arxiv 2603.02730 v4 pith:MONJVEU3 submitted 2026-03-03 cs.IR

APAO: Bridging the Training-Inference Gap in Generative Recommendation via Adaptive Prefix-Aware Optimization

Yuanqing Yu , Yifan Wang , Weizhi Ma , Zhiqiang Guo , Min Zhang This is my paper
classification cs.IR
keywords generative recommendationbeam searchtraining-inference inconsistencyprefix-aware lossadaptive weightingsequential recommendationranking optimizationpreference alignment
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 reading

This paper takes aim at a contradiction at the heart of generative recommendation: during training the model always sees the ground-truth tokens, but during inference beam search repeatedly prunes away low-scoring prefixes, so the correct item can be eliminated before it is ever completed. The authors argue that this is not an incidental implementation detail but a structural training-inference inconsistency, and they propose Adaptive Prefix-Aware Optimization (APAO), which supervises every prefix length of the target item, not just the final token sequence. Two loss forms are offered, a pointwise likelihood over prefixes and a pairwise ranking loss against sampled negative prefixes, combined with an adaptive weighting scheme that concentrates training on whichever prefix is currently weakest. The paper claims that this prefix-aware objective optimizes a lower bound on beam-search recall, and reports consistent Recall and NDCG gains across backbones and datasets, plus a deployed pCTR improvement. A sympathetic reader would take the central insight to be: if beam search prunes prefixes, then training should make prefixes individually strong enough to survive pruning.

Core claim

The central claim is Theorem 1: optimizing the prefix-aware loss — the weighted sum of prefix-level pointwise or pairwise losses added to the standard cross-entropy loss — essentially optimizes a lower bound of the ranking metric, such as recall, under beam search. Beam-search success is a conjunction of per-step survival events: the target item's prefix must rank in the top K at every decoding step, whereas cross-entropy only optimizes average token likelihood. APAO replaces that single averaged objective with dense supervision over all prefix lengths, teaching the model to keep the correct prefix inside the beam at every step. The authors also show empirically that gains are largest for ea

What carries the argument

The load-bearing object is the prefix-aware loss: a weighted sum, over prefix lengths, of either a pointwise cross-entropy term or a pairwise ranking term that compares the target prefix against sampled negative prefixes at the same length. Around it, the paper builds an adaptive worst-prefix weighting rule that raises the weight of whichever prefix currently has the highest loss, focusing training on the link most likely to be pruned. Together these turn the beam-search success condition — a conjunction of per-step top-K constraints — into a differentiable surrogate. The proof wraps this in a lower-bound argument using a union bound over per-step failures and an exponential surrogate that u

Load-bearing premise

The method assumes that a small sample of uniformly random negative prefixes at each length faithfully represents the actual beam-search competition, so that pushing the positive prefix above those sampled negatives is enough to keep it inside the beam; if the true competitor is not in the sample, the claimed lower-bound guarantee is not supported by the proof.

What would settle it

Run APAO-Pairwise and the cross-entropy baseline on a dataset where, at each prefix length, a known adversarial prefix — one that scores just below the positive prefix but still inside the top-K beam — is withheld from the negative sample. If APAO's gain over cross-entropy disappears or reverses when that true competitor is absent from training, the sampling-proxy assumption is the weak link. Concretely, compare Recall@K under beam search with random negatives versus negatives that are explicitly drawn from the beam's top-K at each step.

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

If this is right

  • Training generative recommenders with prefix-level supervision should lift Recall and NDCG under beam search without changing the inference pipeline or adding serving latency.
  • The adaptive worst-prefix weighting puts training pressure on early-prefix failures, which is exactly where beam search is most fragile; removing early-prefix supervision is the most damaging ablation in the paper.
  • Because the method is training-side, it can be dropped onto any autoregressive generative recommender; the paper demonstrates gains on two backbone architectures and multiple tokenizers.
  • Smaller beam sizes become feasible: the paper reports that APAO at beam size 20 matches or beats baselines at beam size 100, making deployment cheaper.
  • The online A/B result — a 0.9% pCTR gain on a live platform — indicates the effect survives real traffic, not just offline benchmarks.

Where Pith is reading between the lines

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

  • A natural extension is to replace uniform random negative sampling with hard-negative mining against the model's own beam, which would directly repair the gap between the sampled set and the true pruning competitors that the proof assumes.
  • The lower-bound argument is really about how faithfully the sampled negatives cover the beam's actual competition; if the same logic carries over, APAO-style prefix losses could apply to any constrained decoding scheme, such as top-p sampling or grammar-constrained generation, not just beam search.
  • The adaptive weighting rule is generic: it could be reused as a drop-in curriculum for other multi-task or multi-length objectives where the bottleneck shifts during training.
  • One testable prediction of the paper's mechanism is that the relative gain of APAO over cross-entropy should grow as the tokenizer becomes less informative, because weak tokens create more prefix-level failures; the paper's tokenizer robustness experiments are consistent with that.
Share X Bluesky LinkedIn Reddit HN

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 APAO, a training-side method for generative recommendation that adds prefix-level pointwise or pairwise losses to the standard cross-entropy objective, together with an adaptive weighting scheme that up-weights the currently worst-performing prefix. The motivation is to close the gap between teacher-forced token-level training and beam-search inference, where the ground-truth item can be pruned if any of its prefixes is not ranked in the top-K. Experiments on four datasets with TIGER and Llama backbones, ablations, and an industrial A/B test are reported, along with a theoretical claim (Theorem 1) that optimizing the prefix-aware loss optimizes a lower bound on beam-search recall.

Significance. If the empirical results are taken at face value, the paper makes a useful practical contribution: it identifies a real and underexplored source of train/inference mismatch in generative retrieval, proposes a simple training-side remedy with no added inference cost, and validates it across multiple architectures and datasets. The release of source code, the breadth of the empirical evaluation (four datasets, two backbones, ablations, sensitivity analyses, and an online A/B test), and the consistent gains over strong preference-optimization baselines are genuine strengths. However, the theoretical analysis in Section 4.4 and Appendix B is not merely a presentation issue: it is the advertised justification for the method's 'effectiveness.' The stated theorem is unproven as written and, as stated, vacuous. The publication value of the paper therefore hinges on the empirical evidence, which is solid but does not by itself establish the 'optimization consistency' claim.

major comments (3)
  1. [Appendix B, Eq. (21)] The inclusion A_m^c ⊆ {φ_m > 0} is asserted with N=100 uniformly sampled negative prefixes, while beam pruning at step m is decided against all K·|V| competing extensions. A positive prefix can fall outside the top-K because of a competitor that is not in the sampled set; in that case the failure event A_m^c occurs while φ_m ≤ 0, so the inclusion is false. The proof itself hedges with 'If the ranking failure occurs within the sampled set' and then drops this qualifier in the subsequent inequality. Thus Theorem 1 is unproven as written.
  2. [Appendix B, Eqs. (22)–(23)] Even if the inclusion were valid, the derived lower bound is vacuous. Since L_pair(m) = log(1 + e^{φ_m}) ≥ 0, we have exp(L_pair(m)) ≥ 1 for every m. Substituting into Eq. (22) gives I_Beam(y) ≥ 1 − Σ_m exp(L_pair(m)) ≤ 1 − T, which for the 4-token tokenizer is ≤ −3. A lower bound that is always negative cannot establish that minimizing the loss improves Recall@K. The final sentence of the proof ('thereby explicitly maximizing the lower bound') is therefore not supported by the displayed inequalities.
  3. [Appendix B, pointwise variant] The proof for the pointwise loss is dismissed as 'following an analogous logic.' This is not immediate: viewing the pointwise cross-entropy as a pairwise loss against all vocabulary tokens still compares only individual tokens, not complete competing prefixes at the same decoding step. The same sampling/competitor issue that invalidates the pairwise proof applies to the pointwise variant, so the pointwise claim in Theorem 1 is not established either.
minor comments (5)
  1. [Section 5.3.1, Figure 2] The text refers to 'Prefix 0' while the paper's notation in Sections 3 and 4 indexes prefixes as 1, ..., T. Please align the indexing or define the offset explicitly.
  2. [Table 1] The superscript '*' flags are used inconsistently: some cells are bold without '*'; please clarify which asterisks denote significance relative to the best baseline versus the CE baseline, and state the number of seeds for the t-test.
  3. [Appendix E, Algorithm 1] The pairwise loss in line 31 computes −log(pos_exp / (pos_exp + neg_sum_exp)), which is algebraically equivalent to Eq. (14) but differs in appearance. A one-line comment explaining the equivalence would avoid confusion.
  4. [Appendix D, Eq. (25)] The subscript N_m in the inner sum is not defined; it appears to denote the negative set for prefix m, but this should be stated explicitly, since in the main text N is defined once for all prefixes.
  5. [Section 5.4, Figure 7] The claim that 'the relative improvement grows steadily as the prefix length increases' should be verified across both datasets; on Office the improvement at prefix 0 appears comparable to or larger than at prefix 1 in the figure.

Circularity Check

0 steps flagged

No circularity found: APAO's objective is not defined in terms of its evaluation metric; the Theorem 1 proof gap is a correctness issue, not circularity.

full rationale

Walk of the derivation chain: the method trains Lunified = LCE + β Σ w_m L_m (Eq. 11) and evaluates on held-out Recall/NDCG under beam search; the evaluation metric is never reconstructed from fitted constants, so the empirical claims are not circular. The adaptive weights in Eq. (17) are a softmax reweighting by current prefix losses derived from a KL-regularized worst-prefix objective; this is part of the training dynamics, not a fitted parameter renamed as a prediction. β and η are tuned hyperparameters, not fit to the test metric. The theoretical claim (Theorem 1) is a surrogate-bound argument: it tries to show beam-failure indicators are dominated by exp(Lpair(m)), which is not an equality by construction between the loss and the metric. There is no load-bearing self-citation: the reference list contains no paper by the present authors; the closest prior work (S-DPO) is external and the paper explicitly differentiates its prefix-level granularity. The manuscript itself flags a limitation in Appendix B—'If the ranking failure occurs within the sampled set'—before using the unconditional inclusion A_m^c ⊆ {φ_m > 0} in Eq. (21); and the derived bound IBeam ≥ 1 - Σ_m exp(Lpair(m)) is vacuous because exp(Lpair(m)) ≥ 1 for the four-token setting. These are correctness gaps in the proof, but they do not make the derivation circular: the loss is not defined in terms of the target metric, and the empirical evaluation is self-contained against external benchmarks. Hence no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or formal entities. Its free parameters are standard training hyperparameters (beta, eta) plus the fixed negative-sample count. The main additional assumption is the sampling proxy connecting the pairwise loss to beam-search failure, which is not independently validated and is the source of the theoretical gap.

free parameters (3)
  • beta (prefix loss weight) = tuned per dataset/backbone over {0.05, 0.1, 0.2, 0.3, 0.4}; see Table 5
    Controls the trade-off between CE and the prefix-aware loss in Eq. (11); optimal value varies by dataset and backbone.
  • eta (adaptive weighting update rate) = tuned per dataset/backbone over {5e-6, 1e-5, 3e-5, 5e-5, 1e-4, 5e-4}; see Table 5
    Controls how quickly Eq. (17) shifts weight toward high-loss prefixes; tuned on each dataset.
  • number of negative samples N = 100 (fixed)
    All methods use 100 uniform random negatives; the pairwise loss and the lower-bound proof depend on this sampling budget.
axioms (5)
  • domain assumption Items are factorized autoregressively into T discrete tokens with P(item|history) = prod_t P(token_t | prefix, history).
    Foundation of the generative recommendation setup (Section 3.1, Eq. 1).
  • domain assumption Beam search scores a prefix as the sum of token log-probabilities and prunes to the top K at each step.
    Used to define the success condition in Eq. (9) and to motivate prefix-level losses.
  • ad hoc to paper A beam-search failure at step m can be detected by checking whether the positive prefix is beaten by one of the sampled negative prefixes.
    This is the load-bearing sampling assumption in Appendix B (Eq. 21). It is not justified: the true competitor in beam search may be outside the sampled set.
  • domain assumption Teacher-forcing cross-entropy is the appropriate base objective.
    Adopted from prior generative recommendation and LM training; the paper's method adds to it rather than replacing it.
  • domain assumption Four-level RQ-VAE tokenization provides adequate item codes.
    The experiments fix the tokenizer to a 4-level RQ-K-means quantizer; prefix lengths 1-4 are defined by this choice.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of APAO: Bridging the Training-Inference Gap in Generative Recommendation via Adaptive Prefix-Aware Optimization." pith.science (2026). https://pith.science/paper/MONJVEU3

@misc{pith2026260302730,
  author       = {Pith},
  title        = {Pith review of: APAO: Bridging the Training-Inference Gap in Generative Recommendation via Adaptive Prefix-Aware Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MONJVEU3}},
  note         = {Machine review of arXiv:2603.02730}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generative recommendation has recently emerged as a promising paradigm for sequential recommendation. It formulates the task as an autoregressive generation process, predicting tokens of the next item conditioned on user interaction histories. Existing generative recommendation models are typically trained with token-level likelihood objectives such as cross-entropy loss, while employing beam search during inference to generate ranked candidates. However, this leads to a fundamental training-inference inconsistency: standard training assumes ground-truth tokens are always available, while beam search prunes low-probability branches during inference, causing the correct item to be prematurely discarded when its prefixes receive low scores. To address this issue, we propose the Adaptive Prefix-Aware Optimization (APAO) framework, which introduces prefix-level optimization losses to better align the training objective with the inference setting. Furthermore, we design an adaptive worst-prefix optimization strategy that dynamically focuses on the most vulnerable prefixes during training, thereby enhancing the model's ability to retain correct candidates under beam search constraints. We provide theoretical analyses to demonstrate the effectiveness and efficiency of our framework. Extensive experiments show that APAO consistently alleviates the training-inference inconsistency and improves performance across generative recommendation backbones. The source code is publicly available at https://github.com/yuyq18/APAO.

Figures

Figures reproduced from arXiv: 2603.02730 by Min Zhang, Weizhi Ma, Yifan Wang, Yuanqing Yu, Zhiqiang Guo.

Figure 1
Figure 1. Figure 1: Analysis of Beam Search. (a) Illustration of the beam [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ablation analysis on the Beauty dataset using the [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Weighting strategy analysis on the Beauty dataset [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Robustness analysis on the Beauty dataset using the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Parameter sensitivity analysis of APAO-pointwise [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Scalability analysis of beam size 𝐾 on the Beauty and Yelp datasets. 5.4 RQ3: Prefix-level Performance To evaluate whether APAO effectively bridges the training–inference gap at the granularity of beam expansion, we analyze prefix-level Recall@20 over the four semantic code levels produced by the 4-level tokenizer. Specifically, prefix indices 0–3 correspond to pro￾gressively longer semantic prefixes, with… view at source ↗
Figure 7
Figure 7. Figure 7: Prefix-level Recall@20 and relative improvement [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation

    cs.AI 2026-07 conditional novelty 6.0

    SIDs preserve coarse item organization but lose fine structure and prune good targets during generation; Item-Supported Decoding recovers up to 31.2% NDCG@10 without retraining.

  2. Bridging the Structural Gap: Adapting Autoregressive Generation for Recommendation

    cs.IR 2026-07 conditional novelty 6.0

    BARGE improves generative sequential recommendation by restoring item boundaries in the encoder and suppressing hierarchical semantic drift in decoding, outperforming prior generative baselines on public and industria...

  3. Closing the Indexing-Decoding Gap in Multimodal Generative Retrieval via Prefix Retention Optimization

    cs.IR 2026-06 unverdicted novelty 6.0

    PRO closes the indexing-decoding gap in multimodal generative retrieval via prefix ranking distillation, vocabulary scheduling, and geometric score fusion, improving target prefix retention on nine tasks.

Reference graph

Works this paper leans on

40 extracted references · 9 linked inside Pith · cited by 3 Pith papers

  1. [1]

    Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. 2024. A general theoreti- cal paradigm to understand learning from human preferences. InInternational Conference on Artificial Intelligence and Statistics. PMLR, 4447–4455

    2024

  2. [2]

    Zhuoxi Bai, Ning Wu, Fengyu Cai, Xinyi Zhu, and Yun Xiong. 2024. Aligning large language model with direct multi-preference optimization for recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 76–86

    2024

  3. [3]

    Yuxin Chen, Junfei Tan, An Zhang, Zhengyi Yang, Leheng Sheng, Enzhi Zhang, Xiang Wang, and Tat-Seng Chua. 2024. On softmax direct preference optimization for recommendation.Advances in Neural Information Processing Systems37 (2024), 27463–27489

    2024

  4. [4]

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences.Advances in neural information processing systems30 (2017)

    2017

  5. [5]

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization.arXiv preprint arXiv:2402.01306(2024)

    Pith/arXiv arXiv 2024

  6. [6]

    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

  7. [7]

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

  8. [8]

    Chi, Julian Mcauley, and Derek Zhiyuan Cheng

    Yupeng Hou, Jianmo Ni, Zhankui He, Noveen Sachdeva, Wang-Cheng Kang, Ed H. Chi, Julian Mcauley, and Derek Zhiyuan Cheng. 2025. ActionPiece: Contextually Tokenizing Action Sequences for Generative Recommendation. InProceedings of the 42nd International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 267), Aarti Singh, Maryam ...

    2025

  9. [9]

    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

  10. [10]

    Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated gain-based evaluation of IR techniques.ACM Transactions on Information Systems (TOIS)20, 4 (2002), 422–446

    2002

  11. [11]

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

    2018

  12. [12]

    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

  13. [13]

    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

  14. [14]

    Xinyu Lin, Chaoqun Yang, Wenjie Wang, Yongqi Li, Cunxiao Du, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua. 2025. Efficient Inference for Large Lan- guage Model-based Generative Recommendation. InInternational Conference on Learning Representations, Y. Yue, A. Garg, N. Peng, F. Sha, and R. Yu (Eds.), Vol. 2025. 91672–91697. https://proceedings.iclr.cc/paper_...

    2025

  15. [15]

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

  16. [16]

    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

  17. [17]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

    Pith/arXiv arXiv 2017

  18. [18]

    OpenAI. 2023. GPT-4 Technical Report.arXiv preprint arXiv:2303.08774(2023)

    Pith/arXiv arXiv 2023

  19. [19]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback.Advances in neural information processing systems35 (2022), 27730–27744

    2022

  20. [20]

    Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. 2022. Focus on the Common Good: Group Distributional Robustness Follows.ICLR(2022)

    2022

  21. [21]

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al . 2018. Improving language understanding by generative pre-training. (2018)

    2018

  22. [22]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems36 (2023), 53728–53741

    2023

  23. [23]

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

  24. [24]

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

  25. [25]

    Weiwei Sun, Lingyong Yan, Zheng Chen, Shuaiqiang Wang, Haichao Zhu, Pengjie Ren, Zhumin Chen, Dawei Yin, Maarten Rijke, and Zhaochun Ren. 2023. Learning to tokenize for generative retrieval.Advances in Neural Information Processing Systems36 (2023), 46345–46361

    2023

  26. [26]

    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

  27. [27]

    Bohao Wang, Feng Liu, Jiawei Chen, Xingyu Lou, Changwang Zhang, Jun Wang, Yuegang Sun, Yan Feng, Chun Chen, and Can Wang. 2025. Msl: Not all tokens are what you need for tuning llm as a recommender. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1912–1922

    2025

  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. 2400–2409

    2024

  29. [29]

    Shiguang Wu, Zhaochun Ren, Xin Xin, Jiyuan Yang, Mengqi Zhang, Zhumin Chen, Maarten de Rijke, and Pengjie Ren. 2025. Constrained Auto-Regressive Decoding Constrains Generative Retrieval. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2429–2440

    2025

  30. [30]

    Chaoqun Yang, Xinyu Lin, Wenjie Wang, Yongqi Li, Teng Sun, Xianjing Han, and Tat-Seng Chua. 2025. EARN: Efficient Inference Acceleration for LLM-based Generative Recommendation by Register Tokens. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 3483–3494

    2025

  31. [31]

    Tianchi Yang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, and Qi Zhang. 2023. Auto search indexer for end-to-end document retrieval. arXiv preprint arXiv:2310.12455(2023)

    Pith/arXiv arXiv 2023

  32. [32]

    Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. 2022. Glm-130b: An open bilingual pre-trained model.arXiv preprint arXiv:2210.02414(2022)

    Pith/arXiv arXiv 2022

  33. [33]

    Jianyang Zhai, Zi-Feng Mai, Chang-Dong Wang, Feidiao Yang, Xiawu Zheng, Hui Li, and Yonghong Tian. 2025. Multimodal Quantitative Language for Generative Recommendation.arXiv preprint arXiv:2504.05314(2025)

    Pith/arXiv arXiv 2025

  34. [34]

    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

  35. [35]

    Guorui Zhou, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Qiang Luo, Qian- qian Wang, Qigen Hu, Rui Huang, Shiyao Wang, et al. 2025. OneRec Technical Report.arXiv preprint arXiv:2506.13695(2025)

    arXiv 2025

  36. [36]

    Guorui Zhou, Hengrui Hu, Hongtao Cheng, Huanjie Wang, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Lu Ren, Liao Yu, et al. 2025. OneRec-V2 Technical Report.arXiv preprint arXiv:2508.20900(2025). A Closed-form Update for Adaptive Worst-prefix Optimization Proposition 2 (Closed-form update via KKT).Consider the convex program max 𝑤∈Δ𝑇 𝐹(𝑤):= 𝑇∑︁ 𝑚=1 𝑤𝑚𝐿...

    Pith/arXiv arXiv 2025

  37. [2015]

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

    Pith/arXiv arXiv 2015

  38. [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

  39. [2023]

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

    2023

  40. [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

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.