Pith. sign in

REVIEW 4 major objections 4 minor 38 references

A graph-structured decoder that expands the decoding space of non-autoregressive generative rerankers can escape the likelihood trap, improving both accuracy and diversity.

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 ↗

CONGRATS uses a DAG-structured positional decoder and evaluator-in-the-loop training to generate more diverse and accurate recommendation lists, showing offline and Kuaishou A/B gains.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A plausible, practically significant reranking method with an under-specified training objective; revise and send to review. the 4 major comments →

arxiv 2510.10127 v3 pith:HO25TY6M submitted 2025-10-11 cs.IR

Breaking the Likelihood Trap: Consistent Generative Recommendation with Graph-structured Model

Qiya Yang , Xiaoxi Liang , Zeping Xiao , Ying Cao , Yingjie Deng , Yuxin Ren , Yalong Wang , Yongqi Liu This is my paper
classification cs.IR
keywords generative rerankinglikelihood trapgraph-structured decodernon-autoregressive generationrecommendation diversitydifferentiable evaluatorGumbel-Softmaxdirected acyclic graph
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 targets the final reranking stage of recommender systems, where the goal is to reorder a set of candidates into a list that maximizes user engagement. It argues that non-autoregressive generative rerankers trained with maximum likelihood fall into a likelihood trap: they repeatedly recommend high-frequency, repetitive lists that look likely but feel low-quality. To break the trap, the authors replace the flat position-embedding decoder with a directed acyclic graph decoder, in which position slots become graph vertices and a sequence is a path through the graph. They also add a differentiable evaluator trained on real user feedback, using Gumbel-Softmax so the generator can be optimized against user preferences, not just likelihood. If the central claim is right, real-time generative reranking can simultaneously gain diversity and accuracy, as the paper demonstrates offline and in a five-day A/B test on a large-scale video platform.

Core claim

The paper's central claim is that the likelihood trap in NAR generative reranking is not inevitable: it stems from the narrow, flat decoding space of a fixed-length position decoder. By treating the m output positions as g > m graph vertices connected by a learned, lower-triangular transition matrix E, the model defines a joint distribution over sequences as a marginalization over all valid vertex paths. This expands the number of expressible sequences roughly from a single m-length order to a rich set of path-constrained orders, and the transition structure introduces implicit inter-item dependencies. The accompanying consistent training objective passes Gumbel-Softmax sampled lists through

What carries the argument

The central object is the graph-structured decoder, where position embeddings are redefined as g graph vertices and a lower-triangular softmax transition matrix E makes the vertex graph a directed acyclic graph. A path A = (a_1,...,a_m) with 1 = a_1 < ... < a_m = g selects an ordered subset of vertices; the sequence probability factorizes as P(Y,A|X) = P(A|X)P(Y|A,X). The output probability matrix P_g has shape n×g instead of n×m, multiplying the number of expressible sequences. The second mechanism is consistent differentiable training: Gumbel-Softmax reparameterization makes the sampled lists differentiable, allowing evaluator scores (cross-entropy losses from a PLE-based multi-task evalua

Load-bearing premise

The load-bearing premise is that the training loss in Eq. 11, a marginalization over all valid vertex paths, is what the model actually optimizes; the paper never specifies how this sum is computed or approximated, so if the implementation samples or approximates paths, the reported diversity and accuracy gains may come from a different objective than the derived one.

What would settle it

Construct a small setting (e.g., m=3, g=6) where all valid paths can be enumerated exactly, and compare the exact marginal loss training against a training that samples a single path or uses the Gumbel-approximated objective. If the two models perform equivalently on the evaluation metrics, then the exact path marginalization is not the source of the gains, undercutting the paper's explanation for why the graph decoder helps. Alternatively, ablate by replacing the learned transition matrix E with a uniform transition while keeping all else fixed: if metrics remain unchanged, the transition dep

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

If this is right

  • If the central claim holds, non-autoregressive generative rerankers need not sacrifice diversity to gain speed; the DAG decoder is a principled way to enlarge the decoding space while keeping one-shot parallel inference.
  • The graph-transition mechanism may generalize beyond this specific architecture: any NAR sequence generator that uses positional embeddings could replace them with a larger vertex set and path-selection decoding, turning decoding into a differentiable graph search.
  • Training against an evaluator makes the reranker directly optimizable for business-relevant metrics such as watch time and likes, not merely for sequence likelihood; this could reduce the gap between offline perplexity-type objectives and online engagement.
  • The reported offline and online numbers suggest that diversity (lower repetition rate, higher item coverage, higher Distinct-2) and accuracy can improve together, not as trade-offs, when the decoding space is expanded.
  • Because all operations remain parallel, the approach is deployed in a real-time industrial setting (300M+ DAU app), implying that the added graph structure does not impose prohibitive latency in production.

Where Pith is reading between the lines

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

  • A key unverified assumption is whether the marginal loss in Eq. 11 is computed exactly or approximated; the paper does not give an algorithm for summing over all valid paths, so it is possible the actual trained objective differs from the derived one, and the reported gains may stem from the auxiliary loss or the Gumbel sampling rather than from the exact path marginalization.
  • The cross-attention in Eq. 4 uses keys K_v derived only from vertex embeddings V; as written, the attention weights do not depend on the candidate hidden states H_c, which contradicts the text's description of bidirectional cross-attention. This could mean the encoder-decoder interaction is weaker than claimed.
  • The same graph-structured decoding idea could be tested in other NAR text-generation tasks that suffer from repetition, providing a testable extension: if the mechanism works there, it is a general solution to the likelihood trap, not a recommender-specific trick.
  • A direct experiment measuring the likelihood of generated lists versus user engagement (e.g., likes or watch time) would quantify how large the trap actually is in practice; the paper's thesis implies these two quantities can diverge substantially.
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

4 major / 4 minor

Summary. The paper proposes CONGRATS, a generative reranking framework for recommender systems. It replaces the flat non-autoregressive decoder of NAR4Rec with a graph-structured decoder whose position embeddings form a directed acyclic graph (DAG): vertices are expanded by a factor λ and a transition matrix E defines a distribution over vertex paths. The generative loss L_gen in Eq. (11) marginalizes over all valid paths, and a second consistency loss L_cons in Eq. (14) is introduced by feeding generator samples into a differentiable evaluator trained on real user feedback. The authors report offline gains on Kuaishou and Avito (e.g., Recall@6 72.84% vs. 65.05% for NAR4Rec) and online A/B gains on Kuaishou for views, long views, and likes, together with diversity metrics such as Jaccard diversity, repetition rate, item coverage, and Distinct-2.

Significance. If the central claims hold, the paper would make a useful contribution by addressing the homogeneity of non-autoregressive generative reranking through an explicit graph-structured decoding space, while also coupling generation to an evaluator trained on real feedback. The online A/B results on a large-scale platform and the reported diversity improvements are concrete and practically relevant. The paper also ships a transparent architecture description and ablation studies. However, the as-written derivation leaves open whether the training objectives are actually computed as stated and whether gradients reach the transition parameters; until these points are resolved, the empirical gains cannot be confidently attributed to the proposed mechanism.

major comments (4)
  1. [§3.3, Eq. (14)] The generative loss L_gen marginalizes over all valid paths A∈Γ, but no algorithm is given for computing this sum. With g=24 and m=6 the number of increasing paths from vertex 1 to vertex 24 of length 6 is C(22,4)=7,315, and the text's statement that 'we compute the probability matrix P_g for all vertices in parallel' does not explain how the path marginal is evaluated. A standard forward dynamic program O(m·g²) would make Eq. (11) tractable, but it is not mentioned; if the implementation instead samples paths, the reported loss is a stochastic estimator whose variance and sample count are unstated. This matters because the diversity/accuracy benefits are attributed to learning transition probabilities under this marginal likelihood.
  2. [§3.3, Eqs. (13)–(14)] The consistency objective L_cons is claimed to 'enable the generator to be directly optimized based on the evaluator's assessed scores.' The only Gumbel-Softmax relaxation given is Eq. (13), which is applied to the item–vertex distribution H_c H_v^T. Path sampling from the transition matrix E (Eq. 6) is not relaxed, yet the evaluator scores 'periodically sampled lists' that include a selected path A. As written, gradients from L_cons do not reach E, so the central claim that evaluator feedback steers the graph transitions is unsupported. The authors should either provide a differentiable path-sampling procedure for E or explicitly state which parameters L_cons updates.
  3. [§3.2, Eq. (4)] The cross-attention formula as written uses Q_v and K_v both derived from V, so the attention weights do not depend on the candidate hidden states H_c. The text calls this 'bidirectional cross-attention,' which implies that H_c should be involved in the attention. This is likely a typographical slip (e.g., K_v should be a projection of H_c), but as written the decoder's cross-attention cannot encode candidate information, undermining the architecture description. Please correct the formula and verify that the reported experiments use the corrected version.
  4. [§4.2, Tables 2 and 3] All offline results are reported as point estimates with no variance, confidence intervals, or significance tests. For example, the claim of a 'significant increase of 7%' in Recall@6 (Table 2) cannot be assessed without knowing run-to-run variability or the number of seeds. Given that the primary baseline NAR4Rec is the authors' own prior work and the margins vary by a few percent, the authors should report standard deviations, multiple seeds, or paired significance tests to support the claimed improvements.
minor comments (4)
  1. [§3.2] The notation is inconsistent: P_g is defined as a matrix in Eq. (9), but later P_g^{a_i,y_i} is used as a probability entry. The subscript/superscript conventions should be harmonized.
  2. [Algorithm 1] The pseudocode is incomplete and contains undefined variables ('edges', 'tokens', 'i') and appears to conflate vertex and item indices. Since the lookahead search is central to inference, please provide a complete, self-contained algorithm.
  3. [§4.1] The description of the Kuaishou dataset says '60 candidate videos and 6 exposed items,' while the Avito description says 'first 21 days for training and last 7 days for testing.' More detail on how the reranking task is constructed from these logs (e.g., negative sampling, exposure definition) would improve reproducibility.
  4. [§4.4, Figure 3] The text says Recall@6 improves until λ=6, but the figure legend says 'Best λ=4.' While the final choice of λ=4 is consistent with the stated efficiency trade-off, the figure caption is confusing and should be clarified.

Circularity Check

0 steps flagged

No load-bearing circularity: the generator is optimized against an independently trained evaluator and real feedback data, and self-citations to NAR4Rec are backed by empirical comparisons and ablations. Remaining issues are correctness gaps, not definitional reductions.

full rationale

The paper's central derivation chain is not circular. The generator (graph-structured decoder, §3.2) is optimized by L_gen (Eq 11) and L_cons (Eq 14), where the evaluator is trained independently on real feedback labels (Eq 12) and the generator is then trained to maximize the evaluator's scores. That is an external, measurable objective, not a fitted constant or a definitional identity. The self-citations to NAR4Rec [22] and to the multi-generator setup [32] are present, but they are not load-bearing: NAR4Rec is used as an empirical baseline (Tables 2 and 3) and is ablated in Table 7, so the benefit of the graph decoder is established by measurement rather than by citation. The 'likelihood trap' premise is supported by external work [31,33]. The notable gaps—Eq 11's path marginal over Γ is not specified algorithmically (with g=24, m=6 it is C(22,4)=7315 paths), L_cons as written (Eqs 13-14) does not show differentiable sampling of the transition matrix E, and Eq 4's cross-attention weights depend only on V, not H_c—are correctness/completeness risks, not reductions of the claimed outputs to the model's inputs. Hence no circular step can be exhibited; the central claims stand or fall on the offline and online experiments.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central method rests on several assumptions carried from prior work: the likelihood-trap analogy from NLG, the reliability of an internal evaluator as a proxy for user preference, the tractability of path marginalization, and the adequacy of the offline metrics. The free parameters λ, τ, α are hand-set and λ is tuned on the reporting metric. No invented entities are introduced.

free parameters (4)
  • Graph size factor λ = 4
    Chosen based on Recall@6 trade-off in Figure 3; tuning on the reported benchmark inflates the headline offline gain.
  • Gumbel-Softmax temperature τ = 0.3
    Set in §4.1; no sensitivity analysis or justification.
  • Loss weight α = 0.5
    Balances L_con and L_gen in Eq 15; set by hand in §4.1.
  • Decoder depth L = 3
    Number of stacked blocks, fixed across datasets; no ablation reported.
axioms (5)
  • domain assumption The 'likelihood trap' observed in neural text generation (ref [33]) transfers to recommender reranking and explains list homogeneity.
    Introduced in §1 by analogy; no recommender-specific evidence cited.
  • domain assumption The trained evaluator's scores are a faithful proxy for user engagement, so optimizing the generator through the evaluator improves real-world metrics without reward hacking.
    Assumed in §3.3; no distribution-shift or reward-hacking analysis.
  • ad hoc to paper The loss in Eq 11 marginalizes over all valid DAG paths tractably.
    No algorithm or approximation is provided; without it the stated graph-decoder training objective is not what is implemented.
  • domain assumption The offline metrics (Recall@6/10, AUC, NDCG) and the Kuaishou 5%/5-day A/B setup adequately measure reranking quality.
    Standard in prior work [22,24], but no confidence intervals or significance tests accompany the offline numbers.
  • standard math Transformer attention, Gumbel-Softmax reparameterization, and lower-triangular masking behave as standard.
    Background from refs [27,13,11].

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Breaking the Likelihood Trap: Consistent Generative Recommendation with Graph-structured Model." pith.science (2026). https://pith.science/paper/HO25TY6M

@misc{pith2026251010127,
  author       = {Pith},
  title        = {Pith review of: Breaking the Likelihood Trap: Consistent Generative Recommendation with Graph-structured Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HO25TY6M}},
  note         = {Machine review of arXiv:2510.10127}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reranking, as the final stage of recommender systems, plays a crucial role in determining the final exposure, directly influencing user experience. Recently, generative reranking has gained increasing attention for formulating reranking as a holistic sequence generation task, implicitly modeling complex dependencies among items. However, most existing methods suffer from the likelihood trap, where high-likelihood sequences are often repetitive and perceived as low-quality by humans, thereby limiting user engagement. In this work, we propose Consistent Graph-structured Generative Recommendation (CONGRATS). We first introduce a novel Graph-structured Model, which enables the generation of more diverse sequences by exploring multiple paths. This design not only expands the decoding space to promote diversity, but also improves prediction accuracy by explicitly modeling item dependencies from graph transitions. Furthermore, we design a Consistent Differentiable Training method that incorporates an evaluator, allowing the model to learn directly from user preferences. Extensive offline experiments validate the superior performance of CONGRATS over state-of-the-art reranking methods. Moreover, CONGRATS has been evaluated on a large-scale video-sharing app, Kuaishou, with over 300 million daily active users, demonstrating that our approach significantly improves both recommendation quality and diversity, validating our effectiveness in practical industrial platforms.

Figures

Figures reproduced from arXiv: 2510.10127 by Qiya Yang, Xiaoxi Liang, Yalong Wang, Ying Cao, Yingjie Deng, Yongqi Liu, Yuxin Ren, Zeping Xiao.

Figure 1
Figure 1. Figure 1: An intuitive understanding of the likelihood trap. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our proposed Congrats, including the graph-structured model and the consistent differentiable training. In practice, we set the number of position vertices to 24. Note that both the vertex transition and sequence generation processes are in parallel, which can be deployed in a real-time industry platform. element in P to obtain the probability distribution over candidate items. This corresponds… view at source ↗
Figure 3
Figure 3. Figure 3: Effect of graph size factor (𝜆) on Recall@6. Increas￾ing 𝜆 improves recall up to 𝜆 = 6, after which the performance slightly drops. The best balance between accuracy and com￾putation cost is achieved at 𝜆 = 4 [PITH_FULL_IMAGE:figures/full_fig_p008_3.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

38 extracted references · 5 linked inside Pith

  1. [1]

    Sujoy Bag, Sri Krishna Kumar, and Manoj Kumar Tiwari. 2019. An efficient rec- ommendation generation using relevant Jaccard similarity.Information Sciences 483 (2019), 53–64

    2019

  2. [2]

    Irwan Bello, Sayali Kulkarni, Sagar Jain, Craig Boutilier, Ed Chi, Elad Eban, Xiyang Luo, Alan Mackey, and Ofer Meshi. 2018. Seq2Slate: Re-ranking and slate optimization with RNNs.arXiv preprint arXiv:1810.02019(2018)

    Pith/arXiv arXiv 2018

  3. [3]

    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

    2016

  4. [4]

    Angela Fan, Edouard Grave, and Armand Joulin. [n. d.]. Reducing Transformer Depth on Demand with Structured Dropout. InInternational Conference on Learn- ing Representations

  5. [5]

    Yufei Feng, Binbin Hu, Yu Gong, Fei Sun, Qingwen Liu, and Wenwu Ou. 2021. GRN: Generative Rerank Network for Context-wise Recommendation.arXiv preprint arXiv:2104.00860(2021)

    Pith/arXiv arXiv 2021

  6. [6]

    Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. 2019. Mask-Predict: Parallel Decoding of Conditional Masked Language Models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP). 6112–6121

    2019

  7. [7]

    Xudong Gong, Qinlin Feng, Yuan Zhang, Jiangling Qin, Weijie Ding, Biao Li, Peng Jiang, and Kun Gai. 2022. Real-time Short Video Recommendation on Mobile Devices. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 3103–3112

    2022

  8. [8]

    Jiatao Gu, James Bradbury, Caiming Xiong, Victor OK Li, and Richard Socher. 2018. Non-Autoregressive Neural Machine Translation. InInternational Conference on Learning Representations

    2018

  9. [9]

    Jiatao Gu, Changhan Wang, and Junbo Zhao. 2019. Levenshtein transformer. Advances in Neural Information Processing Systems32 (2019)

    2019

  10. [10]

    Fei Huang, Tianhua Tao, Hao Zhou, Lei Li, and Minlie Huang. 2022. On the learn- ing of non-autoregressive transformers. InInternational conference on machine learning. PMLR, 9356–9376

    2022

  11. [11]

    Fei Huang, Hao Zhou, Yang Liu, Hang Li, and Minlie Huang. 2022. Directed acyclic transformer for non-autoregressive machine translation. InInternational Conference on Machine Learning. PMLR, 9410–9428

    2022

  12. [12]

    Iris AM Huijben, Wouter Kool, Max B Paulus, and Ruud JG Van Sloun. 2022. A review of the gumbel-max trick and its extensions for discrete stochasticity in machine learning.IEEE transactions on pattern analysis and machine intelligence 45, 2 (2022), 1353–1371

    2022

  13. [13]

    Eric Jang, Shixiang Gu, and Ben Poole. 2016. Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144(2016)

    Pith/arXiv arXiv 2016

  14. [14]

    Lukasz Kaiser, Samy Bengio, Aurko Roy, Ashish Vaswani, Niki Parmar, Jakob Uszkoreit, and Noam Shazeer. 2018. Fast Decoding in Sequence Models Using Dis- crete Latent Variables. InProceedings of the International Conference on Machine Learning (ICML). 2390–2399

    2018

  15. [15]

    Jindřich Libovický and Jindřich Helcl. 2018. End-to-End Non-Autoregressive Neural Machine Translation with Connectionist Temporal Classification. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP). 3016–3021

    2018

  16. [16]

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. 2024. Awq: Activation-aware weight quantization for on-device llm compression and accel- eration.Proceedings of machine learning and systems6 (2024), 87–100

    2024

  17. [17]

    Xiao Lin, Xiaokai Chen, Chenyang Wang, Hantao Shu, Linfeng Song, Biao Li, and Peng Jiang. 2024. Discrete conditional diffusion for reranking in recommendation. InCompanion Proceedings of the ACM Web Conference 2024. 161–169

    2024

  18. [18]

    Xiangyu Liu, Chuan Yu, Zhilin Zhang, Zhenzhe Zheng, Yu Rong, Hongtao Lv, Da Huo, Yiqing Wang, Dagui Chen, Jian Xu, et al. 2021. Neural auction: End-to- end learning of auction mechanisms for e-commerce advertising. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 3354–3364

    2021

  19. [19]

    Changhua Pei, Yi Zhang, Yongfeng Zhang, Fei Sun, Xiao Lin, Hanxiao Sun, Jian Wu, Peng Jiang, Junfeng Ge, Wenwu Ou, et al. 2019. Personalized re-ranking for recommendation. InProceedings of the 13th ACM conference on recommender systems. 3–11

    2019

  20. [20]

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

  21. [21]

    Ali Razavi, Aaron van den Oord, and Oriol Vinyals. 2019. Generating Diverse High-Fidelity Images with VQ-VAE-2. InAdvances in Neural Information Process- ing Systems (NeurIPS)

    2019

  22. [22]

    Yuxin Ren, Qiya Yang, Yichun Wu, Wei Xu, Yalong Wang, and Zhiqiang Zhang

  23. [23]

    Yuxin Ren, Zihan Zhong, Xingjian Shi, Yi Zhu, Chun Yuan, and Mu Li. 2023. Tailoring Instructions to Student’s Learning Levels Boosts Knowledge Distillation. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1990–2006

    2023

  24. [24]

    Xiaowen Shi, Fan Yang, Ze Wang, Xiaoxu Wu, Muzhi Guan, Guogang Liao, Wang Yongkang, Xingxing Wang, and Dong Wang. 2023. PIER: Permutation-Level Interest-Based End-to-End Re-ranking Framework in E-commerce. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4823–4831

    2023

  25. [25]

    Siqi Sun, Yu Cheng, Zhe Gan, and Jingjing Liu. 2019. Patient Knowledge Dis- tillation for BERT Model Compression. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 4323–4332

    2019

  26. [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 Rec- ommender Systems(Virtual Event, Brazil)(RecSys ’20). Association for Computing Machinery, New York, NY, USA, 269–278. doi:10.1145/338...

    arXiv 2020

  27. [27]

    A Vaswani. 2017. Attention is all you need.Advances in Neural Information Processing Systems(2017)

    2017

  28. [28]

    Benyou Wang, Yuxin Ren, Lifeng Shang, Xin Jiang, and Qun Liu. [n. d.]. Exploring extreme parameter compression for pre-trained language models. InInternational Conference on Learning Representations

  29. [29]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & cross network for ad click predictions. InProceedings of the ADKDD’17. 1–7

    2017

  30. [30]

    Shuli Wang, Xue Wei, Senjie Kou, Chi Wang, Wenshuai Chen, Qi Tang, Yinhua Zhu, Xiong Xiao, and Xingxing Wang. 2025. NLGR: Utilizing Neighbor Lists for Generative Rerank in Personalized Recommendation Systems. InCompanion Proceedings of the ACM on Web Conference 2025. 530–537

    2025

  31. [31]

    Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2019. Neural text generation with unlikelihood training.arXiv preprint arXiv:1908.04319(2019)

    Pith/arXiv arXiv 2019

  32. [32]

    Hailan Yang, Zhenyu Qi, Shuchang Liu, Xiaoyu Yang, Xiaobei Wang, Xiang Li, Lantao Hu, Han Li, and Kun Gai. 2025. Comprehensive List Generation for Multi-Generator Reranking. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2298–2308

    2025

  33. [33]

    Hugh Zhang, Daniel Duckworth, Daphne Ippolito, and Arvind Neelakantan

  34. [34]

    Kai Zheng, Haijun Zhao, Rui Huang, Beichuan Zhang, Na Mou, Yanan Niu, Yang Song, Hongning Wang, and Kun Gai. 2024. Full stage learning to rank: A unified framework for multi-stage systems. InProceedings of the ACM Web Conference

    2024

  35. [38]

    Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009

    3621–3631. Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009

    2007

  36. [2020]

    arXiv:2004.10450 [cs.CL] https://arxiv.org/abs/2004.10450

    Trading Off Diversity and Quality in Natural Language Generation. arXiv:2004.10450 [cs.CL] https://arxiv.org/abs/2004.10450

    Pith/arXiv arXiv 2004

  37. [2023]

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

    2023

  38. [2024]

    In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Non-autoregressive generative models for reranking recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5625–5634

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.