Pith. sign in

REVIEW 3 major objections 4 minor 36 references

A shared generative retriever can serve multiple objectives by adding small per-objective LoRA decoders and coordinating beam search, lifting Recall@512 by up to 5.62% and production usage time by 0.37%.

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 ↗

Multi-Decoder OneRec attaches small LoRA experts for each objective to a shared generative recommender and uses quota-aware constrained beam search, beating single-decoder OneRec on offline recall and in production.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Real architectural novelty and a positive production A/B test, but the offline gains are in-sample selections because there is no validation split; read the offline numbers with caution. the 3 major comments →

arxiv 2607.26500 v1 pith:AARYYN4T submitted 2026-07-29 cs.IR

Multi-Decoder OneRec: Controllable Generative Retrieval for Multi-Objective Industrial Recommendation

You Wang , Zhao Liu , Guoping Tang , Yiqing Yang , Shuo Su , Jing Liu , Naifu Zhou , Xiaoyou Zhou
show 6 more authors
Wei Jiang Jian Liang Xiao Lv Ruiming Tang Liyin Hong Wenwu Ou
This is my paper
classification cs.IR
keywords generative recommendationmulti-objective retrievalindustrial recommender systemsSemantic IDLoRAconstrained beam searchcandidate quotaKwai26
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

Industrial recommenders must divide a fixed candidate budget across objectives such as watch time, likes, and cold-start coverage. This paper claims that a single generative retriever, which generates items as sequences of discrete Semantic IDs, can serve all these objectives if each objective gets its own lightweight Low-Rank Adaptation (LoRA) decoder on top of a shared base. Training each decoder with its own supervision while stopping gradients at the shared base, and then coordinating beam search across decoders with explicit quotas, keeps the objectives isolated yet complementary. If correct, this gives operators quota control without fragmenting models, and the paper reports Recall@512 gains of 1.69%-5.62% over the single-decoder baseline plus production gains in usage time and cold-start. It also releases a billion-scale benchmark, Kwai26, so the claims can be checked.

Core claim

Multi-Decoder OneRec's central claim is that a single Semantic-ID generative retriever can be made truly multi-objective: instead of training one decoder that mixes all objectives, or maintaining separate full models, the paper gives each objective a lightweight decoder built from LoRA adapters, a task-specific beginning-of-sequence token, and a residual to the shared SID embedding table. All objectives share the user-context encoder and the General Decoder, which provides a common SID prior. Training is isolated by gradient routing: exposure next-token prediction updates only the shared base, filtered next-token prediction updates event-based experts, and KL-regularized policy optimization

What carries the argument

The central mechanism is the combination of a shared Semantic-ID prior with isolated, low-rank adapters and quota-aware constrained decoding. The shared components are a user-context encoder and a General Decoder that learns the common SID distribution from exposure data. Each objective adds a LoRA expert (low-rank matrices on the attention projections), a task-specific BOS embedding, and an additive SID-embedding residual; these form an objective-specific decoder. During training, gradient routing (stop-gradient through the shared base) ensures each objective's loss touches only its own expert parameters. During inference, Multi-Decoder Constrained Beam Search assigns each decoder an explic

Load-bearing premise

The reported offline gains would be invalid if the hyperparameter choices (LoRA rank, KL weight, reward-history size, quotas, beam sizes, and constraint level) were overfit to the test split, since the evaluation used a single fixed test split with no validation set.

What would settle it

Reproduce the comparison on a fresh, temporally later split of Kwai26 (or another dataset) with all hyperparameters fixed in advance; if the 1.69%-5.62% Recall@512 gains over the single-decoder baseline do not persist, the framework-level claim is falsified. Also verify the budget-recovery mechanism: if MD-CBS fails to restore the full 512 unique candidates on a different catalog, the complementarity benefit would not hold.

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

If this is right

  • Adding a new objective is reduced to training a small LoRA adapter and setting a quota; the shared model and existing experts are untouched.
  • Quota allocation becomes an explicit product control: shifting quota to an objective measurably shifts Recall toward it (the paper's 'More WT' intervention raises watch-time recall by +0.31%).
  • Coordinated decoding recovers the full candidate budget: without constraint, only 207.67 unique candidates survive from independent routes, whereas MD-CBS restores the full 512.
  • Industrial systems can consolidate fragmented retrieval routes into one shared retriever, cutting parameter and maintenance overhead while keeping per-objective control (three experts add only 20% parameters).
  • The Kwai26 benchmark (1.31B records, 25.03M valid-SID items, predefined splits) gives the community a common testbed for multi-objective generative retrieval under a fixed budget.

Where Pith is reading between the lines

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

  • A natural extension of this design is online adaptation: because each objective's expert is isolated, new objectives could be added, removed, or re-weighted at serving time without retraining the base, enabling rapid experimentation on new engagement signals.
  • The same pattern — a shared prior, isolated adapters, and constrained decoding with explicit output quotas — could transfer to other generative tasks that must fill a fixed-size output set while balancing competing criteria, such as diverse document bundling or multi-stakeholder content generation.
  • The +2.09% cold-start gain suggests that deliberately giving a small decoder its own SID-space exploration budget is a more direct lever for new-content coverage than any single-policy compromise that must balance cold-start against mature-content retrieval.
  • Because serving-time FLOPs grow with the number of decoders (2.23x for four routes), the practical limit of this approach may be decoding cost rather than parameter count; a future variant might share some decoding computation across routes while preserving the isolation that makes the experts complementary.
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 / 4 minor

Summary. The paper proposes Multi-Decoder OneRec, a generative retrieval framework for multi-objective industrial recommendation. It augments a shared user-context encoder and General Decoder with objective-specific LoRA experts, separate BOS embeddings, and SID embedding residuals, trained under gradient isolation with exposure NTP, target-filtered SFT, and a KL-regularized relative-reward policy optimization (L-GBPO) for the watch-time objective. At inference, explicit route quotas and Multi-Decoder Constrained Beam Search (MD-CBS) coordinate the route-specific decoders. The authors release Kwai26, a 1.31B-record benchmark with predefined train/test splits, and report offline Recall@512 gains of 1.69%–5.62% over single-decoder OneRec, together with a seven-day production A/B test showing positive gains on usage time, retention, interactions, and cold-start. The central claim is that this architecture achieves shared representation, isolated objective adaptation, and complementary candidate generation under a fixed retrieval budget.

Significance. If the empirical claims hold, the contribution is valuable: it offers a practical middle ground between fragmented multi-route retrieval and coupled single-decoder generative retrieval, and the public Kwai26 benchmark is a useful resource for the community. The paper is transparent about its split design, reports computational overhead, and includes a real-traffic A/B test, which is a strength. However, the offline evaluation currently does not support the headline quantitative gains because hyperparameter selection was performed on the same test set used for evaluation; the A/B test, while positive, does not validate the individual design choices. The evidence is suggestive but not yet conclusive.

major comments (3)
  1. [Appendix B / §5.1 / §5.5 / Appendix E.1] Appendix B states that after the leakage-safe split, 'no validation split or date/percentage alternative is used.' Yet the paper selects LoRA rank (Table 12), KL weight (Figure 3), reward-history size K (Table 13), route quotas/beam sizes (Figure 4), and MD-CBS level (Table 3) by comparing Recall@512 on the same 50,000 test sessions. The reported gains (Table 1: 1.69%–5.62%) and p<0.05 significance claims are therefore in-sample selection results, not out-of-sample estimates, and multiple comparisons inflate the chance of optimistic selection. The production A/B test tests only one fixed configuration and cannot validate these design comparisons. This is load-bearing for the paper's central quantitative claim. Please introduce a validation split for model selection and report test results for the chosen configuration, or otherwise demonstrate generalization to a fresh sample.
  2. [§4.3, Eqs. (13)–(17)] The L-GBPO objective in Eq. (14) uses an unclipped likelihood ratio and a nonstandard denominator in Eq. (13): for positive rewards the denominator becomes sg(p_t) whenever sg(p_t) > p_ref, zeroing the gradient on those tokens; for negative rewards the ratio is unbounded above. The paper does not provide a derivation or stability/off-policy justification for this objective. Since the Watch-time expert is one of the four offline metrics and a production route, this is central. Please add a theoretical justification, a stability analysis, or an ablation showing that the reported WT Recall gain is not an artifact of this particular objective.
  3. [Table 1 / §5.1] Table 1 reports that all gains over the strongest baseline are 'statistically significant (p<0.05)', but no significance-test procedure, confidence intervals, or correction for multiple comparisons is provided. Moreover, the p-values are conditional on hyperparameters selected on the same test set. Please specify the test used and, more importantly, separate model selection from evaluation so that the significance claims can be assessed.
minor comments (4)
  1. [§1, Introduction] The phrase 'objective-specific objective-specific decoders' is a duplicated-word error.
  2. [Figure 3] The x-axis label reads 'KL Loss Weight' while the text says 'KL weight'; please unify the terminology.
  3. [Table 5] The SFT baseline uses a '12s threshold' that does not appear in the hyperparameter list; please state how this threshold was chosen and whether it was also selected on the test set.
  4. [Appendix B] Please clarify whether the three-level SIDs are platform-provided fields or fitted on the 60-day window, and if the latter, whether the RQ-KMeans model is fit on training-period data only; this is important for leakage safety of the evaluation protocol.

Circularity Check

0 steps flagged

No significant circularity: the central claim is an empirical benchmark against an external baseline plus a production A/B test; the main caveat is test-set hyperparameter selection, which is a generalization concern, not a circular derivation.

full rationale

Multi-Decoder OneRec's central claim is empirical: four Recall@512 metrics and an online A/B test compared with single-decoder OneRec, an external published baseline (Deng et al., arXiv:2502.18965) that shares no authors with this paper. The architecture — shared encoder and General Decoder, LoRA experts, gradient isolation, and MD-CBS — is not defined in terms of the metrics it reports; Eqs. (10)-(18) specify training losses and Eq. (19) specifies cross-route masking, none of which encode the test Recall values. No fitted parameter is relabeled as a prediction, and no 'uniqueness theorem' or self-authored ansatz is invoked to force the design. The paper does cite OneRec-V2 [36] and CAPTS [37], which overlap with the present author list, but these citations support related-work context such as reward normalization and trigger selection rather than the paper's main result; the baseline OneRec [5] itself is external. The most notable methodological limitation is stated in Appendix B: 'no validation split or date/percentage alternative is used.' Hyperparameters such as LoRA rank, KL weight, reward-history size K, and CBS level were therefore selected on the same 50,000 test sessions (Tables 12-13, Figures 3-4). This weakens the offline gains as out-of-sample evidence and is a legitimate fairness/generalization concern, but it is not circular reasoning: the reported numbers are not equivalent by construction to a fitted input. The production A/B test provides an independent, though single-configuration, check. Thus the derivation chain is self-contained and no circular step can be exhibited.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The framework rests on several domain assumptions about SID codebooks, exposure priors, and user-context sufficiency, plus ad hoc reward-ratio stabilizers. No new physical or conceptual entities are introduced; the LoRA experts, BOS embeddings, and SID residuals are architectural components rather than invented entities.

free parameters (7)
  • LoRA rank r_t = 32
    Selected by ablation on the test set; rank 32 is best on all four Recall@512 metrics (Table 12).
  • KL regularization weight λ_KL = 1.0
    Chosen from {0, 0.01, 0.5, 1.0, 2.0} on the basis of test-set Watch-time Recall and SID legal rate (Figure 3).
  • Reward-history size K = 500
    Ablation shows K=500 improves WT Recall over K=8; selected on the test set (Table 13).
  • Route quotas and beam sizes = q: 86/85/85/256; beam: 86/171/256/512
    Hand-set operating point; Figure 4 intervenes on quotas/beams and reports deltas, so the default is a tuned configuration rather than derived.
  • MD-CBS constraint level d = 2 (L3 CBS)
    L3 CBS beats L2 CBS on all metrics; selected as default after ablation (Table 3).
  • SID codebook size = 3 levels × 8,192 codes
    RQ-KMeans tokenization hyperparameter; determines retrieval granularity and the number of items with valid SIDs (25.03M).
  • SFT threshold for watch-time baseline = 12s
    Used in the Table 5 comparison; not a parameter of the proposed method but a free choice in the evaluation.
axioms (6)
  • standard math Autoregressive factorization of SID generation (Eq. 3)
    The probability of an SID sequence is decomposed level by level; standard for generative retrieval.
  • domain assumption Exposure-sample NTP creates a general SID prior that is a good reference for all objectives
    The General Decoder is trained on served/exposed items only, and KL regularization anchors each expert to this distribution; if the exposure distribution is biased, the reference prior is biased.
  • domain assumption RQ-KMeans SIDs preserve enough semantic structure that exact-SID de-duplication is safe
    L3 CBS removes only complete SID tuples; this assumes SID-to-item mapping is meaningful enough that avoiding exact duplicates does not over-prune useful candidates.
  • domain assumption User-context features (recent 20, Long-View 256, watch-time 500) are sufficient for all objectives
    One encoder serves all decoders; if the fixed history views omit signal needed by a particular objective, that objective's expert cannot recover it.
  • ad hoc to paper The stop-gradient denominator p_old in Eq. (13) is a valid stabilizer for the policy ratio
    The max(p_ref, sg(p_t)) / max(p_ref, 1-sg(p_t)) construction is introduced without derivation or sensitivity analysis; it shapes the Watch-time expert's gradient but is not independently justified.
  • ad hoc to paper L-GBPO with unclipped likelihood ratios is a valid off-policy objective for logged watch-time data
    Eq. (14) uses a raw importance-style ratio without clipping or off-policy correction; the paper relies on the logged data being sufficiently on-policy for this to behave.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Decoder OneRec: Controllable Generative Retrieval for Multi-Objective Industrial Recommendation." pith.science (2026). https://pith.science/paper/AARYYN4T

@misc{pith2026260726500,
  author       = {Pith},
  title        = {Pith review of: Multi-Decoder OneRec: Controllable Generative Retrieval for Multi-Objective Industrial Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AARYYN4T}},
  note         = {Machine review of arXiv:2607.26500}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Industrial recommender systems build candidate pools by assigning explicit quotas to objective-specific retrieval routes. This design offers quota control but increasingly fragments modeling, training, and serving as the route set grows. Semantic-ID-based generative retrieval provides a unified alternative, yet a single decoder entangles objective policies and limits candidate complementarity. We propose Multi-Decoder OneRec, a controllable framework that combines shared representations, isolated objective adaptation, and coordinated decoding. All objectives share a user-context module and the General Decoder, while each objective adds an isolated, parameter-efficient LoRA expert. During training, exposure-sample next-token prediction (NTP) updates the shared base, target-filtered NTP updates the event-based experts, and Kullback-Leibler (KL)-regularized policy optimization updates the Watch-time expert; gradient routing isolates these updates, and the General Decoder supplies a stop-gradient reference. At inference, explicit route quotas allocate the fixed budget and Multi-Decoder Constrained Beam Search reduces cross-route overlap. We publicly release Kwai26, a large-scale multi-objective benchmark with 1.31 billion raw item-level records, 31.85 million Item-ID entries, and 25.03 million items with valid Semantic IDs, together with predefined splits and an evaluation protocol. Under the same 512-item retrieval budget, Multi-Decoder OneRec improves over the single-decoder OneRec baseline by 1.69%-5.62% across four Recall@512 metrics. In a production A/B test, it yields relative gains of 0.37% in app usage time per device, 0.19% in Day-7 retained users, 0.19% in devices with at least one share, and 2.09% in new-content Cold-Start. These results show that generative retrieval can combine shared modeling with objective-specific control and complementary candidate generation.

Figures

Figures reproduced from arXiv: 2607.26500 by GuoPing Tang, Jian Liang, Jing Liu, Liyin Hong, Naifu Zhou, Ruiming Tang, Shuo Su, Wei Jiang, Wenwu Ou, Xiao Lv, Xiaoyou Zhou, Yiqing Yang, You Wang, Zhao Liu.

Figure 1
Figure 1. Figure 1: Comparison of multi-objective retrieval paradigms. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Multi-Decoder OneRec. (a) Shared architecture: user history is encoded once by the user-context module [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: KL-weight sensitivity. Removing KL collapses recall [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Recall@512 deltas from the default configuration. [PITH_FULL_IMAGE:figures/full_fig_p009_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

36 extracted references · 2 canonical work pages

  1. [1]

    Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. 2017. Guided Open Vocabulary Image Captioning with Constrained Beam Search. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. 936–945. doi:10.18653/v1/D17-1098

    work page doi:10.18653/v1/d17-1098 2017

  2. [3]

    Edoardo Botta, Jaewon Yang, Yi-Ping Hsu, Laksh Bhasin, Yilin Chen, Prabhat Agarwal, Anirudhan Badrinath, Jiajing Xu, and Charles Rosenberg. 2026. PinRec: Unified Generative Retrieval for Pinterest Recommender Systems. InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM. doi:10.1145/3770855.3818462

    arXiv 2026

  3. [4]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. InProceedings of the 10th ACM Conference on Recommender Systems. 191–198. doi:10.1145/2959100.2959190

    arXiv 2016

  4. [5]

    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.CoRRabs/2502.18965 (2025). arXiv:2502.18965 [cs.IR] doi:10.48550/arXiv.2502.18965

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.18965 2025

  5. [6]

    Chenxiao Fan, Chongming Gao, Yaxin Gong, Haoyan Liu, Fuli Feng, and Xiangnan He. 2026. Uncertainty-aware Generative Recommendation. InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. 12 pages. doi:10.1145/3770855.3817975

    arXiv 2026

  6. [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. doi:10.1145/3523227.3546767

    arXiv 2022

  7. [9]

    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). PMLR, 24004–24024. h...

    2025

  8. [10]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations. https: //openreview.net/forum?id=nZeVKeeFYf9

    2022

  9. [11]

    Junjie Huang, Jiarui Qin, Jianghao Lin, Ziming Feng, Weinan Zhang, and Yong Yu. 2025. Unleashing the Potential of Multi-Channel Fusion in Retrieval for Personalized Recommendations. InProceedings of the ACM Web Conference 2025. 483–494. doi:10.1145/3696410.3714753

    arXiv 2025

  10. [12]

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning Deep Structured Semantic Models for Web Search Using Clickthrough Data. InProceedings of the 22nd ACM International Conference on Information and Knowledge Management. 2333–2338. doi:10.1145/2505515. 2505665

    doi:10.1145/2505515 2013

  11. [13]

    Bowen Jin, Hansi Zeng, Guoyin Wang, Xiusi Chen, Tianxin Wei, Ruirui Li, Zhengyang Wang, Zheng Li, Yang Li, Hanqing Lu, Suhang Wang, Jiawei Han, and Xianfeng Tang. 2024. Language Models as Semantic Indexers. InProceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PMLR, 22244–22259. https:...

    2024

  12. [14]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Rec- ommendation. InProceedings of the 2018 IEEE International Conference on Data Mining. 197–206. doi:10.1109/ICDM.2018.00035

    arXiv 2018

  13. [15]

    Yifan Liu, Yaokun Liu, Zelin Li, Zhenrui Yue, Gyuseok Lee, Ruichen Yao, Yang Zhang, and Dong Wang. 2026. Learning Decomposed Contextual Token Represen- tations from Pretrained and Collaborative Signals for Generative Recommenda- tion. InProceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1175–11...

    arXiv 2026

  14. [16]

    Zihan Liu, Yupeng Hou, and Julian McAuley. 2024. Multi-Behavior Generative Recommendation. InProceedings of the 33rd ACM International Conference on In- formation and Knowledge Management. 1575–1585. doi:10.1145/3627673.3679730

    arXiv 2024

  15. [17]

    Zhao Liu, Yichen Zhu, Yiqing Yang, Xiao Lv, Guoping Tang, Rui Huang, Qiang Luo, Ruiming Tang, and Guorui Zhou. 2026. DiffGRM: Diffusion-based Generative Recommendation Model. InProceedings of the ACM Web Conference 2026. ACM, 5853–5864. doi:10.1145/3774904.3792156

    arXiv 2026

  16. [18]

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H. Chi. 2018. Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture- of-Experts. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 1930–1939. doi:10.1145/3219819.3220007

    arXiv 2018

  17. [19]

    Ping Nie, Yujie Lu, Shengyu Zhang, Ming Zhao, Ruobing Xie, William Yang Wang, and Yi Ren. 2022. MIC: Model-agnostic Integrated Cross-channel Recom- mender. InProceedings of the 31st ACM International Conference on Information and Knowledge Management. 3400–3409. doi:10.1145/3511808.3557081

    arXiv 2022

  18. [20]

    Christiano, Jan Leike, and Ryan Lowe

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. 2022. Training Language Models to Follow Instructions with Hu- ma...

    2022

  19. [21]

    Matt Post and David Vilar. 2018. Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies. Association for Computational Linguistics, 1314–1324. doi:10.18653/v1/N18-1119

    doi:10.18653/v1/n18-1119 2018

  20. [22]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheswaran Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. InAdvances in Neural Information Processing Systems, Vol. 36. https://proceedings.neurips.cc/paper_file...

    2023

  21. [23]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.CoRR abs/2402.03300 (2024). arXiv:2402.03300 [cs.CL] doi:10.48550/arXiv.2402.03300

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.03300 2024

  22. [24]

    Teng Shi, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Enyun Yu. 2025. GenSAR: Unifying Balanced Search and Recommendation with Gen- erative Retrieval. InProceedings of the 19th ACM Conference on Recommender Systems. 124–134. doi:10.1145/3705328.3748071

    arXiv 2025

  23. [25]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential Recommendation with Bidirectional En- coder Representations from Transformer. InProceedings of the 28th ACM In- ternational Conference on Information and Knowledge Management. 1441–1450. doi:10.1145/3357384.3357895

    arXiv 2019

  24. [26]

    Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progres- sive Layered Extraction (PLE): A Novel Multi-Task Learning (MTL) Model for Personalized Recommendations. InProceedings of the 14th ACM Conference on Recommender Systems. 269–278. doi:10.1145/3383313.3412236

    arXiv 2020

  25. [27]

    Vijayakumar, Michael Cogswell, Ramprasaath R

    Ashwin K. Vijayakumar, Michael Cogswell, Ramprasaath R. Selvaraju, Qing Sun, Stefan Lee, David Crandall, and Dhruv Batra. 2018. Diverse Beam Search for Improved Description of Complex Scenes. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 32. 7371–7379. doi:10.1609/aaai.v32i1.12340 9 KDD ’27, August 2027, San Jose, CA, USA You Wang et al

    doi:10.1609/aaai.v32i1.12340 2018

  26. [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 In- formation and Knowledge Management. 2400–2409. doi:10.1145/3627673.3679569

    arXiv 2024

  27. [29]

    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. ACM, 3245–3254. doi:10.1145/3637528.3671775

    arXiv 2024

  28. [30]

    Yejing Wang, Shengyu Zhou, Jinyu Lu, Qidong Liu, Xinhang Li, Wenlin Zhang, Feng Li, Pengjie Wang, Chuan Yu, Jian Xu, Bo Zheng, and Xiangyu Zhao. 2026. GFlowGR: Fine-tuning Generative Recommendation Frameworks with Gener- ative Flow Networks. InProceedings of the 49th International ACM SIGIR Con- ference on Research and Development in Information Retrieval...

    arXiv 2026

  29. [31]

    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. ACM, 2429–2440. doi:10.1145/3726302.3729934

    arXiv 2025

  30. [32]

    Ruobing Xie, Zhijie Qiu, Jun Rao, Yi Liu, Bo Zhang, and Leyu Lin. 2020. Internal and Contextual Attention Network for Cold-start Multi-channel Matching in Rec- ommendation. InProceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence. 2732–2738. doi:10.24963/ijcai.2020/379

    work page doi:10.24963/ijcai.2020/379 2020

  31. [33]

    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. doi:10.1145/3711896.3736919

    arXiv 2025

  32. [34]

    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. InProceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Rese...

    2024

  33. [35]

    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

  34. [36]

    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

  35. [37]

    Xiaoyou Zhou, Yuqi Liu, Zhao Liu, Xiao Lv, Bo Chen, Ruiming Tang, and Guorui Zhou. 2026. CAPTS: Channel-Aware, Preference-Aligned Trigger Selection for Multi-Channel Item-to-Item Retrieval. InProceedings of the 20th ACM Conference on Recommender Systems. arXiv:2602.12564 [cs.IR] https://arxiv.org/abs/2602. 12564 To appear

    Pith/arXiv arXiv 2026

  36. [38]

    Yanyan Zou, Junbo Qi, Lunsong Huang, Yu Li, Kewei Xu, Jiahao Gao, Binglei Zhao, Xuanhua Yang, Sulong Xu, and Shengjie Li. 2026. GenRec: A Preference- Oriented Generative Framework for Large-Scale Recommendation. InProceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 6 pages. doi:10.1145/3805...

    arXiv 2026

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.