Pith. sign in

REVIEW 3 major objections 6 minor 48 references

Reward Guided Decoding for Generative Recommendation

T0 review · 3 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Generative recommendation can be steered toward business value at decoding time by ranking each candidate code by log-probability plus reward, with no retraining.

desk verdict An industrially credible test-time reward controller for SID generative recommenders, with a standard Boltzmann-style derivation and a real Kuaishou deployment; the paper's theoretical packaging overreaches because the reward signal is a terminal-label classifier logit, not a proven value function. read the letter →

arxiv 2607.25344 v1 pith:DC2Q7SJT submitted 2026-07-28 cs.IR

Ruochen Yang , Yusheng Huang , Youfeng Zheng , Shuang Wen , Liangliang Chen , Pengbo Xu , Xiaoyu Zhang , Shijun Wang
show 6 more authors
Shuang Yang Zhaojie Liu Lantao Hu Wenwu Ou Jiawei Sheng Tingwen Liu
This is my paper
classification cs.IR
keywords generativerecommendationreward-guideddecodingbeamsearchKL-regularizedoptimizationsemanticIDtest-timecontrolalignment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the mismatch between generation likelihood and business value in generative recommendation can be fixed at decoding time rather than by retraining. It formulates reward-guided generation as a KL-regularized reward maximization problem and derives a closed-form Boltzmann distribution; ranking candidates by log P(j) + R(j)/β is the resulting decoding criterion. The base generator stays untouched, while a lightweight reward head attached to a frozen decoder supplies per-token value estimates. Offline results on Amazon categories and industrial logs show consistent gains, and an online A/B test on a live-streaming recommendation service reports higher CTR, watch time, and watch counts. If correct, this gives a controllable, switchable knob for aligning recommenders with changing business objectives.

What carries the argument

The carrying mechanism is the closed-form Boltzmann reward-shaped distribution Q*(j) ∝ P(j) exp(R(j)/β), which yields the ranking score log P(j) + R(j)/β. It is implemented with a lightweight chain-structured reward head attached to the frozen generator decoder: the head consumes candidate-code embeddings and cached prefix representations, produces per-level binary-cross-entropy logits, and converts them to unbounded log-odds values R(j) that are added to generator logits during beam search. This lets reward shape the search trajectory before high-reward, low-likelihood branches are pruned.

What would settle it

Take a setting where terminal rewards for every completed SID are known, train the reward head on per-step labels, and compare beam-search returns against an oracle that scores partial sequences by true expected terminal reward; if the logit-based score is not monotonically related to true expected reward, the optimality claim is unsupported.

Watch

Extended reading notes

Core claim

The paper's central claim is that value-guided decoding in generative recommendation has a closed-form optimal solution: solve Q* = argmax_Q E_Q[R(j)] − β·KL(Q||P), and the maximizing distribution is Q*(j) ∝ P(j)·exp(R(j)/β), so practical beam-search ranking reduces to comparing log P(j) + R(j)/β. This turns reward guidance into a decoding-time scoring rule rather than a retraining objective. The reward model is a chain-structured head on the frozen generator decoder whose per-level BCE logits are log-odds transformed into R(j), and the algorithm injects this signal into beam search before pruning. The paper validates the rule across three public Amazon datasets and an industrial live-stream

Load-bearing premise

The derivation assumes that the per-code classifier logit used as R(j) is a valid estimate of the expected downstream reward of a partially generated SID; without that, ranking by log P + R/β is a heuristic nudge rather than the optimal KL-regularized distribution.

Editorial extensions

If this is right

  • Decoding ranks candidates by log P(j) + R(j)/β, so reward enters before beam pruning and high-value low-likelihood items are no longer systematically discarded.
  • Business-objective switches become serving-time changes: replace or reweight the reward head and adjust β, with no generator retraining.
  • The reward-shaped distribution is optimal under a KL budget; larger β recovers likelihood-only decoding, so the method contains the base generator as a limit.
  • Offline Amazon benchmarks show RGD improving Recall@K and NDCG@K over likelihood-only generative baselines, and industrial data shows gains in HitRate and downstream reward metrics.
  • An online A/B test on live traffic reports page CTR +0.392%, watch time +0.689%, and watch counts +0.349% over the deployed baseline.

Reading between the lines

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

  • Editorial extension: if per-step classifier logits were replaced by learned value estimates of the completed SID, the same Boltzmann rule would become a general test-time policy-improvement operator for any autoregressive recommender.
  • Editorial extension: the pre-merge mode first truncates each beam to top-M candidates by log P, so the method cannot revive candidates outside that likelihood window; a testable variant would relax or adapt M based on reward uncertainty.
  • Editorial extension: because the closed form is agnostic to the reward source, calibrated CTR predictions from a separate ranking model could be plugged in and compared against the paper's frozen-decoder reward head.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes RGD, a decoding-time reward-guided beam search for generative recommendation. It formulates the per-step decoding distribution as the solution of a KL-regularized reward maximization problem, derives the Boltzmann form Q*(j) ∝ P(j) exp(R(j)/β), and then implements this as a per-SID-level scoring rule (Eq. 10) in which a reward model head, trained with BCE against behavior labels, supplies R(j). The framework is evaluated on three public Amazon datasets, an industrial Kuaishou setting, and an online A/B test, with reported gains in Recall/NDCG, downstream reward metrics, and +0.392% CTR / +0.689% watch time / +0.349% watch counts in production. The core derivation in §3.2 is mathematically sound, but the paper's implementation and public experiments raise load-bearing questions about whether the trained classifier logit is a valid reward for the sequential decoding objective, whether the implemented search actually recovers low-logP/high-reward candidates, and whether the public offline results validate business-value alignment.

Significance. If the theoretical claim were fully established, RGD would be a useful test-time control method for generative recommendation, allowing flexible objective switching without retraining. The paper's strengths include a correct one-step Lagrange derivation, a clear and detailed algorithm, extensive ablations, and a real deployment with a two-week A/B test, which is rare and valuable. However, the paper substantially overreaches when it calls Eq. 10 'principled' and 'derived from the optimal KL-regularized reward-shaped distribution': the reward actually used is a terminal-label classifier logit, not an immediate per-token reward, and no value-function/Bellman argument closes that gap. In addition, the algorithm's pre-truncation to top-M by logP conflicts with the central motivation, and the public offline rewards are constructed from the ground-truth next item, making the 'business value' interpretation ambiguous. The empirical finding may still hold, but the paper as written needs either additional theory or scaled-back claims.

major comments (3)
  1. [§3.2.2 (Eq. 10), §4.2.2 (Eqs. 17-18), Appendix A.2] The paper claims Eq. 10 is a 'principled ranking objective derived from the optimal KL-regularized reward-shaped distribution.' However, the R(j) inserted into Eq. 10 in the implementation is the logit of a per-level BCE classifier for the terminal behavior label y_m (Eqs. 17-18). This is not an immediate reward for choosing code j; it is a state-action value estimate under the behavior policy. Appendix A.2 proves only that the Boltzmann form uniquely maximizes the one-step objective Eq. 23 for a fixed R(j); it says nothing about composing per-step logits into a global policy that maximizes expected terminal reward under the KL constraint. A correct token-level decomposition would need a Bellman-style recursion (e.g., Q_l(s,j) = β log E_{j'~P} exp(Q_{l+1}(s,j,j')/β)) or an explicit FUDGE/Bayesian argument. Until this gap is closed, Eq. 10 is a heuristic reweighting of likelihood by a ter
  2. [Algorithm 1 (lines 5-8), §4.3.1] In every inference mode, each beam is first truncated to the top-M candidates by log P before any reward is computed (Alg. 1 lines 5-8). Thus the high-reward/low-logP candidates that motivate the paper ('pruned early during beam search') are still pruned before the reward can influence the search. In pre-merge mode, Eq. 10 can only reorder candidates within each beam's likelihood top-M; in post-merge mode, the expanded pool is similarly selected by logP. This contradicts the paper's narrative and means the optimality of the full-codebook Boltzmann solution (Eq. 8) does not carry over to the implemented search. The authors should either modify the algorithm so reward participates in the expansion (e.g., by scoring a reward-amplified candidate set) or explicitly acknowledge and analyze this truncation as a practical heuristic.
  3. [Appendix C.1 and §5.1] The offline public experiments define the reward-model positive as the ground-truth next item and use batch negatives (Appendix C.1). Consequently, the 'reward' in Table 1 is a second estimator of the same next-item distribution, not a business-value signal. The observed Recall/NDCG improvements may reflect the complementarity of discriminative and generative estimators of the same target, rather than 'aligning personalization and business value.' This confound is not acknowledged and weakens the external validity of the public benchmark results for the paper's central claim. The authors should either use a genuine multi-behavior reward label (as in the industrial setting) or re-frame the public experiments as a sanity check of reward-guided decoding rather than evidence for value alignment.
minor comments (6)
  1. [Table 1 footnote] The footnote states RPG* uses OpenAI text-embedding-3-large while other baselines use sentence-t5-base. This makes the RPG* comparison not apples-to-apples; the authors should use a consistent semantic encoder for all reproduced baselines or exclude the inconsistent row.
  2. [Algorithm 1 and §4.3.2] The notation for the expansion size is inconsistent: §4.3.2 says 'M_expand·K' while Algorithm 1 uses 'M' throughout. Please unify the notation and define the relationship between M and M_expand.
  3. [Figure 3] The caption states '∞ is equivalent to no reward guidance,' but the x-axis tick labels do not include ∞. Add a tick or clarify how β=∞ is represented.
  4. [Table 2] The 'Reward@604' metric is not defined in the main text; the caption should explain how the downstream ranking model computes this reward before presenting the results.
  5. [§3.2.1 and §4.1] The term 'value function R(j)' is used in §3.2.1 while §4.1 calls it 'reward model.' Since R(j) in Eq. 10 is a scalar reward, not a trained value function, the terminology should be made consistent to avoid implying a value-function interpretation that is not proven.
  6. [Throughout] Minor grammatical issues: e.g., 'Considering reward guided generation have shown' in the introduction. A careful proofread would improve clarity.

Circularity Check

1 steps flagged · score 4.0 of 10

Offline evaluation is partially circular: the reward model is trained on the ground-truth next item and then 'predicts' Recall/NDCG on that same ground-truth next item; the formal Boltzmann derivation itself is self-contained.

  1. fitted input called prediction [Appendix C.1 (Offline Public Experiment Settings) and §5.1.2]
    "Since offline public datasets only expose user-item interaction sequences without multi-behavior logs, we treat each ground-truth next item in the user sequence as a positive sample with implicit clicks, while negative samples are drawn from other items appearing within the same batch."

    The same ground-truth next item is both (a) the positive training label for the reward head via the per-level BCE losses in Eqs. 17-18 and (b) the target of the offline metrics Recall@K and NDCG@K. At inference, Eq. 10 ranks by log P(j) + R(j)/beta, where R(j) is the logit of this next-item classifier. The reported offline gains therefore largely show that an auxiliary classifier trained on the evaluation label can improve retrieval/ranking of that same label when added to the generator's log-likelihood. This is a supervised-ensemble effect rather than an independent validation of 'preference alignment and value guidance' as claimed in §5.1.2, so the offline 'prediction' is partly forced by construction.

full rationale

The formal derivation is not circular. Section 3.2 defines a one-step KL-regularized reward maximization (Eq. 2) and solves it in closed form to obtain Q*(j) ∝ P(j)exp(R(j)/beta) (Eq. 8), with the ranking rule Eq. 10 following directly. Appendix A.2 gives a self-contained strict-concavity/uniqueness argument for that fixed-R problem. No load-bearing result is imported from a self-citation, and no uniqueness theorem from the authors' prior work is used to force the Boltzmann form; the same closed-form solution is a standard RLHF/controlled-decoding result. The main circularity-adjacent element is empirical: for the public datasets the reward head is trained with BCE against the ground-truth next item, which is exactly what Recall/NDCG measure, so the offline improvement is partly a next-item classifier ensemble rather than independent value-alignment evidence. The industrial experiments and the Kuaishou A/B test use real business labels and live metrics, providing independent support for the core claim. I also note two non-circular correctness caveats: the Appendix A.2 optimality proof covers a single-step distribution but not the multi-step beam-search composition, and Algorithm 1's pre-merge mode still truncates each beam to top-M by log P before reward scoring, so very low-likelihood high-reward candidates are not recovered. These affect the strength of the 'principled optimal' claim but do not make the derivation circular.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The framework adds a reward head and a β knob; no new physical or ontological entities. Its theoretical load is carried by a modeling assumption (per-step reward as value function) and by several unreported hyperparameters. The Boltzmann reweighting itself is standard and not counted as a free parameter.

free parameters (6)
  • beta (KL temperature) = not reported; tuned per dataset, see Fig. 3
    Controls reward strength in Q*(j)∝P(j)exp(R/β); central knob in Eq. 8/10, tuned on validation.
  • lambda_r (reward loss weight) = not reported
    Weights L_gen and L_r in Eq. 16; chosen by hand, no value or grid given.
  • alpha_m (multi-objective weights) = not reported
    Business weights for LTR reward in Eq. 19; set by business, no values given.
  • M (beam expansion size) = not reported
    Number of logP-top candidates per beam scored by reward in Alg. 1; affects whether high-reward low-logP items survive.
  • K (beam size) = 20 public; 302/604 industrial
    Beam width; public set to 20, industrial 302/604 (Table 5).
  • l' (hybrid first-layer count) = 1
    Hybrid mode uses pre-merge for first l' layers; production sets l'=1 (§4.3.3).
assumptions (5)
  • standard math P(j)>0 and R(j) bounded
    Required for strict concavity and unique optimum; stated in Appendix A.2.
  • domain assumption KL-regularized expected-reward objective Eq. 2 is the right criterion for decoding
    Modeling choice borrowed from RLHF/PPO; not derived from recommendation-specific utility.
  • domain assumption Per-level BCE logit R(j) estimates the value of appending code j
    Eq. 17-18 train a classifier of final behavior label y_m at each SID level; the paper does not prove this equals expected terminal reward.
  • domain assumption Log-odds transform Eq. 15 makes reward dimensionally consistent with logP
    Justified by a Bayesian posterior-update interpretation; assumes logit scale is compatible with generator logits.
  • domain assumption Beam-search cumulative score is the sum of per-step logP + R/β
    Alg. 1 scores partial candidates; the optimality proof covers a single token step, not multi-step beam accumulation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reward Guided Decoding for Generative Recommendation." pith.science (2026). https://pith.science/paper/DC2Q7SJT

@misc{pith2026260725344,
  author       = {Pith},
  title        = {Pith review of: Reward Guided Decoding for Generative Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DC2Q7SJT}},
  note         = {Machine review of arXiv:2607.25344}
}
read the original abstract

Generative recommendation formulates recommendation task into an SID sequence autoregressive generation paradigm, but the decoding process is often dominated by generation likelihood. This may conflict with real-world business objectives, where high-value candidates can receive low generation probability and be pruned early during beam search. Existing reranking or training-time alignment methods either intervene too late or require costly model retraining when business preferences change. To this end, we propose \textbf{R}eward \textbf{G}uided \textbf{D}ecoding, named \textbf{RGD}, a controllable decoding framework for industrial value-oriented generative recommendation. We formulate value-guided decoding as a KL-regularized reward maximization problem, deriving a closed-form reward guided decoding distribution that principledly combines generation probability with reward signals. RGD treats the base generator as a reference policy and introduces a reward model as a test-time controller, injecting reward into each decoding step to reshape the search trajectory without retraining the generator. Extensive offline and online experiments demonstrate the effectiveness of our approach for aligning personalization and business value. RGD has been deployed on the Kuaishou platform, bringing consistent improvements in real-world recommendation scenarios.

Figures

Figures reproduced from arXiv: 2607.25344 by the authors.

Figure 1
Figure 1. Effect of 𝛽 on the Boltzmann reweighted distribu￾tion. Smaller 𝛽 sharpens the distribution toward high-reward candidates, while larger 𝛽 recovers the prior 𝑃. 3.2.2 Closed-Form Solution. We then solve this penalized ob￾jective in closed form. Introducing a Lagrange multiplier 𝜆 for the normalization constraint Í 𝑗 𝑄(𝑗) = 1, the Lagrangian is: L(𝑄, 𝜆) = ∑︁ 𝑗 𝑄(𝑗)𝑅(𝑗) − 𝛽 ∑︁ 𝑗 𝑄(𝑗) log 𝑄(𝑗) 𝑃 (𝑗) − 𝜆 ∑︁ 𝑗 𝑄(𝑗) − 1 ! .… view at source ↗
Figure 2
Figure 2. Overall architecture of RGD. (a) The reward model shares the frozen decoder with the base generator, simultaneously [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Effect of temperature 𝛽 on Sports and Beauty datasets. ∞ is equivalent to no reward guidance. computation, we design three inference strategies with different effectiveness-efficiency trade-offs. 4.3.1 Pre-merge Guidance. In the pre-merge mode, each beam evaluates the reward on its own top-𝑀 candidates before merging them into the global candidate pool. Reward therefore influences not only which candidate survives b… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Density distribution of generated candidates in the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The impact of different inference strategies on (a) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Effect of 𝛽 on the distribution entropy. In addition to the reweighted distribution shown in [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 8 linked inside Pith

  1. [1]

    Yimeng Bai, Chang Liu, Yang Zhang, Dingxian Wang, Frank Yang, Andrew Rabinovich, Wenge Rong, and Fuli Feng. 2025. Bi-Level Optimization for Gener- ative Recommendation: Bridging Tokenization and Generation.arXiv preprint arXiv:2510.21242(2025)

    arXiv 2025

  2. [2]

    Jiangxia Cao, Shen Wang, Yue Li, Shenghui Wang, Jian Tang, Shiyao Wang, Shuang Yang, Zhaojie Liu, and Guorui Zhou. 2024. Moment&Cross: Next- Generation Real-Time Cross-Domain CTR Prediction for Live-Streaming Recom- mendation at Kuaishou.arXiv preprint arXiv:2408.05709(2024)

    arXiv 2024

  3. [3]

    Jiangxia Cao, Pengbo Xu, Yin Cheng, Kaiwei Guo, Jian Tang, Shijun Wang, Dewei Leng, Shuang Yang, Zhaojie Liu, Yanan Niu, et al. 2025. Pantheon: Personalized multi-objective ensemble sort via iterative pareto policy optimization. InPro- ceedings of the 34th ACM International Conference on Information and Knowledge Management. 5575–5582

    2025

  4. [4]

    Sunhao Dai, Jiakai Tang, Jiahua Wu, Kun Wang, Yuxuan Zhu, Bingjun Chen, Bangyang Hong, Yu Zhao, Cong Fu, Kangle Wu, et al. 2025. Onepiece: Bringing context engineering and reasoning to industrial cascade ranking system.arXiv preprint arXiv:2509.18091(2025)

    arXiv 2025

  5. [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.arXiv preprint arXiv:2502.18965 (2025)

    arXiv 2025

  6. [6]

    Kairui Fu, Changfa Wu, Kun Yuan, Binbin Cao, Dunxian Huang, Yuliang Yan, Junjun Zheng, Jianning Zhang, Silu Zhou, Jian Wu, et al. 2026. RankGR: Rank- Enhanced Generative Retrieval with Listwise Direct Preference Optimization in Recommendation.arXiv preprint arXiv:2602.08575(2026)

    arXiv 2026

  7. [7]

    Chengcheng Guo, Kuo Cai, Yu Zhou, Qiang Luo, Ruiming Tang, Han Li, Kun Gai, and Guorui Zhou. 2026. PROMISE: Process Reward Models Unlock Test-Time Scaling Laws in Generative Recommendations.arXiv preprint arXiv:2601.04674 (2026)

    arXiv 2026

  8. [8]

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

    2023

Show all 48 references
  1. [9]

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

    2025

  2. [10]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. InInternational conference on machine learning. PMLR, 2790–2799

    2019

  3. [11]

    Dietmar Jannach and Malte Ludewig. 2017. When recurrent neural networks meet the neighborhood for session-based recommendation. InProceedings of the eleventh ACM conference on recommender systems. 306–310

    2017

  4. [12]

    Jie Jiang, Yangru Huang, Zeyu Wang, Changping Wang, Yuling Xiong, Jun Zhang, and Huan Yu. 2026. Spend Search Where It Pays: Value-Guided Structured Sampling and Optimization for Generative Recommendation.arXiv preprint arXiv:2602.10699(2026)

    2026

  5. [13]

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

    2018

  6. [14]

    Ben Krause, Akhilesh Deepak Gotmare, Bryan McCann, Nitish Shirish Keskar, Shafiq Joty, Richard Socher, and Nazneen Fatema Rajani. 2021. Gedi: Generative discriminator guided sequence generation. InFindings of the Association for Computational Linguistics: EMNLP 2021. 4929–4952

    2021

  7. [15]

    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

  8. [16]

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s verify step by step. InInternational Conference on Learning Representations, Vol. 2024. 39578–39601

    2024

  9. [17]

    Xinkui Lin, Yongxiu Xu, Minghao Tang, Shilong Zhang, Hongbo Xu, Hao Xu, and Yubin Wang. 2025. REMOTE: A Unified Multimodal Relation Extraction Frame- work with Multilevel Optimal Transport and Mixture-of-Experts. InProceedings of the 33rd ACM International Conference on Multim...

    2025

  10. [18]

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

  11. [19]

    Xinchen Luo, Jiangxia Cao, Tianyu Sun, Jinkai Yu, Rui Huang, Wei Yuan, Hezheng Lin, Yichen Zheng, Shiyao Wang, Qigen Hu, et al . 2025. Qarm: Quantitative alignment multi-modal recommendation at kuaishou. InProceedings of the 34th ACM International Conference on Information and...

    2025

  12. [20]

    Sidharth Mudgal, Jong Lee, Harish Ganapathy, YaGuang Li, Tao Wang, Yanping Huang, Zhifeng Chen, Heng-Tze Cheng, Michael Collins, Trevor Strohman, et al

  13. [21]

    Jianmo Ni, Gustavo Hernandez Abrego, Noah Constant, Ji Ma, Keith Hall, Daniel Cer, and Yinfei Yang. 2022. Sentence-t5: Scalable sentence encoders from pre- trained text-to-text models. InFindings of the association for computational lin- guistics: ACL 2022. 1864–1874

    2022

  14. [22]

    Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural...

    2019

  15. [23]

    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

  16. [24]

    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

  17. [25]

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

  18. [26]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  19. [27]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

    2024 arXiv

  20. [28]

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

    2023

  21. [29]

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

  22. [30]

    Shen Wang, Yusheng Huang, Ruochen Yang, Shuang Wen, Pengbo Xu, Jiangxia Cao, Yueyang Liu, Kuo Cai, Chengcheng Guo, Shiyao Wang, et al. 2026. OneLive: Dynamically Unified Generative Framework for Live-Streaming Recommenda- tion.arXiv preprint arXiv:2602.08612(2026)

    2026

  23. [31]

    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

  24. [32]

    Dekai Sun, Yiming Liu, Jiafan Zhou, Xun Liu, Chenchen Yu, Yi Li, Jun Zhang, Huan Yu, and Jie Jiang. 2026. OneRanker: Unified Generation and Ranking with One Model in Industrial Advertising Recommendation.arXiv preprint arXiv:2603.02999(2026)

    2026

  25. [33]

    Yejing Wang, Shengyu Zhou, Jinyu Lu, Ziwei Liu, Langming Liu, Maolin Wang, Wenlin Zhang, Feng Li, Wenbo Su, Pengjie Wang, et al. 2026. Nezha: A zero- sacrifice and hyperspeed decoding architecture for generative recommendations. InProceedings of the ACM Web Conference 2026. 8073–8082

    2026

  26. [34]

    Zhipeng Wei, Kuo Cai, Junda She, Jie Chen, Minghao Chen, Yang Zeng, Qiang Luo, Wencong Zeng, Ruiming Tang, Kun Gai, et al. 2026. Oneloc: Geo-aware generative recommender systems for local life service. InProceedings of the Nineteenth ACM International Conference on Web Search ...

    2026

  27. [35]

    Yi Xu, Moyu Zhang, Chenxuan Li, Zhihao Liao, Haibo Xing, Hao Deng, Jinxin Hu, Yu Zhang, Xiaoyi Zeng, and Jing Zhang. 2026. Mmq: Multimodal mixture- of-quantization tokenization for semantic id generation and user behavioral adaptation. InProceedings of the Nineteenth ACM Inter...

    2026

  28. [36]

    Kevin Yang and Dan Klein. 2021. FUDGE: Controlled text generation with future discriminators. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 3511–3535

    2021

  29. [37]

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

    2024

  30. [38]

    Yuhao Yang, Zhi Ji, Zhaopeng Li, Yi Li, Zhonglin Mo, Yue Ding, Kai Chen, Zijian Zhang, Jie Li, LIU LIN, et al . 2026. Sparse meets dense: Unified generative recommendations with cascaded sparse-dense representations.Advances in Neural Information Processing Systems38 (2026), 9...

    2026

  31. [39]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations.arXiv Conference acronym ’XX, June 03–05, 20...

    2024 arXiv

  32. [40]

    Kun Zhang, Jingming Zhang, Wei Cheng, Yansong Cheng, Jiaqi Zhang, Hao Lu, Xu Zhang, Haixiang Gan, Jiangxia Cao, Tenglong Wang, et al. 2026. OneMall: One Model, More Scenarios–End-to-End Generative Recommender Family at Kuaishou E-Commerce.arXiv preprint arXiv:2601.21770(2026)

    2026

  33. [41]

    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)

    2025 arXiv

  34. [42]

    Ruochen Yang, Xiaodong Li, Jiawei Sheng, Jiangxia Cao, Xinkui Lin, Shen Wang, Shuang Yang, Zhaojie Liu, and Tingwen Liu. 2026. From Agnostic to Specific: Latent Preference Diffusion for Multi-Behavior Sequential Recommendation. In Proceedings of the 32nd ACM SIGKDD Conference ...

    2026

  35. [43]

    Yu Zhou, Chengcheng Guo, Kuo Cai, Ji Liu, Qiang Luo, Ruiming Tang, Han Li, Kun Gai, and Guorui Zhou. 2026. Gems: Breaking the long-sequence barrier in generative recommendation with a multi-stream decoder.arXiv preprint arXiv:2602.13631(2026). A Formula Proof A.1 Equivalent Pr...

    2026

  36. [47]

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

    2020

  37. [2017]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347 (2017)

    2017 arXiv

  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]

    Controlled decoding from language models.arXiv preprint arXiv:2310.17022 (2023)

    2023 arXiv

  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

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.