arxiv: 2605.04559 · v1 · submitted 2026-05-06 · 💻 cs.IR

Recognition: unknown

Beyond Static Best-of-N: Bayesian List-wise Alignment for LLM-based Recommendation

Ruijun Chen , Chongming Gao , Jiawei Chen , Weiqin Yang , Xiangnan He

Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:03 UTC · model grok-4.3

classification 💻 cs.IR

keywords LLM recommendationBest-of-N alignmentBayesian dynamic estimationlist-wise metricsadaptive target distributiongradient decayindiscriminate supervision

0 comments

The pith

BLADE breaks the static Best-of-N upper bound in LLM-based recommendation by using Bayesian updates to create an adaptive supervision target.

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

The paper aims to show that static Best-of-N alignment in large language model recommenders loses effectiveness because fixed references cannot guide improvements beyond their range and the signal fades as the model gets better. BLADE addresses this by maintaining a Bayesian target distribution that blends past priors with fresh evidence from the current model's outputs, allowing the supervision to evolve alongside the policy. A sympathetic reader would care because this promises better optimization of hard-to-differentiate list metrics like NDCG, fairness, and diversity without the high cost of running Best-of-N at every inference step, leading to more accurate and balanced recommendations.

Core claim

BLADE (Bayesian List-wise Alignment via Dynamic Estimation) overcomes the limitations of indiscriminate supervision and gradient decay in BoN alignment by introducing a Bayesian framework that continuously updates the target distribution through the fusion of historical priors and dynamic evidence from the model's current rollouts, thereby constructing a self-evolving target that adapts to the model's growing capabilities and keeps the training signal informative.

What carries the argument

The Bayesian dynamic estimation mechanism in BLADE, which fuses historical priors with evidence from current model rollouts to update the target distribution for list-wise alignment.

If this is right

BLADE achieves sustained improvements in ranking metrics such as Recall and NDCG beyond what static methods can reach.
It delivers gains in complex list-wise metrics including fairness and diversity on real-world datasets.
The approach outperforms existing state-of-the-art baselines in LLM-based recommendation.
The self-evolving target prevents the loss of ranking guidance that occurs when candidates exceed the static reference's quality.

Where Pith is reading between the lines

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

If the Bayesian update continues to distinguish relative qualities effectively, it could allow training to continue productively even after the model surpasses initial references.
This dynamic alignment might apply to other areas where generative models need to optimize non-differentiable metrics without static bounds.
Reducing reliance on expensive inference-time search like Best-of-N could make high-quality list recommendations more practical for deployment.

Load-bearing premise

That fusing historical priors with dynamic evidence from the model's current rollouts will reliably produce an informative, non-degenerate training signal that continues to distinguish relative quality even as the policy improves.

What would settle it

An experiment where BLADE's performance plateaus at the same level as static Best-of-N alignment, or where the updated targets no longer provide distinguishable supervision signals after initial training, would indicate the central claim is incorrect.

Figures

Figures reproduced from arXiv: 2605.04559 by Chongming Gao, Jiawei Chen, Ruijun Chen, Weiqin Yang, Xiangnan He.

**Figure 1.** Figure 1: The efficiency bottleneck of Best-of-N in LLM4Rec. view at source ↗

**Figure 2.** Figure 2: Static BoN Alignment vs. BLADE. (Left) Unlike static view at source ↗

**Figure 3.** Figure 3: The overall training framework of BLADE. (Left) The pipeline illustrates how BLADE fuses the Static Prior with view at source ↗

**Figure 5.** Figure 5: Effect of the number of generated items (G) in view at source ↗

**Figure 7.** Figure 7: Training diagnostics under standalone fairness view at source ↗

**Figure 8.** Figure 8: Performance comparison on list-wise metrics. (a) view at source ↗

read the original abstract

Large Language Models have revolutionized recommender systems (LLM4Rec) by leveraging their generative capabilities to model complex user preferences. However, existing LLM4Rec methods primarily rely on token-level objectives, making it difficult to optimize list-level and non-differentiable metrics (e.g., NDCG, fairness) that define actual recommendation quality. While Best-of-N (BoN) directly optimizes these metrics during inference, its high computational cost hinders real-world deployment. To address this, BoN Alignment aims to distill the search capability into the model itself, yet current approaches suffer from two critical limitations: (1) Indiscriminate Supervision, where the static reference fails to distinguish the relative quality of candidates exceeding its empirical range, leading to a loss of ranking guidance; and (2) Gradient Decay, where the effective supervision signal rapidly diminishes as the evolving policy improves, resulting in inefficient optimization. To overcome these challenges, we propose BLADE (Bayesian List-wise Alignment via Dynamic Estimation). Unlike static approaches, BLADE introduces a Bayesian framework that continuously updates the target distribution by fusing historical priors with dynamic evidence from the model's current rollouts. This mechanism constructs a self-evolving target that adapts to the model's growing capabilities, ensuring the training signal remains informative throughout the learning process. Extensive experiments on three real-world datasets demonstrate that BLADE significantly outperforms state-of-the-art baselines. Crucially, it breaks the static performance upper bound, achieving sustained gains in both ranking accuracy (Recall, NDCG) and complex list-wise metrics (Fairness, Diversity). The code is available via https://github.com/RegionCh/BLADE.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

BLADE replaces the fixed reference in BoN alignment with a Bayesian-updated target from priors plus current rollouts, and the code release makes the update rule checkable.

read the letter

The main takeaway is that this paper gives a concrete way to make list-wise alignment dynamic instead of locking it to a static best-of-N reference. By fusing historical priors with evidence from the model's own rollouts, the target distribution moves as the policy improves, which is intended to keep the training signal from going flat on better candidates or losing gradient strength. That directly targets the two limitations they name in the abstract, and the experiments on three datasets show gains in Recall, NDCG, fairness, and diversity that exceed the static bound reported for prior methods. Releasing the code is useful here because it lets you inspect the exact fusion rule and any safeguards against degeneracy without having to guess from the high-level description. The central mechanism looks coherent on paper and the results are reported with enough baselines to make the comparison meaningful. The main soft spot is that the Bayesian update still appears to involve at least one weighting hyperparameter for prior versus evidence, and the abstract does not spell out how they chose or stabilized it across runs. If that weight turns out to be sensitive, the sustained gains could shrink on new data or different model scales. The paper does not claim parameter-free status, so this is a normal implementation detail rather than a hidden flaw. Readers working on LLM-based recommenders or on distilling search into policy will find the most direct value, especially anyone already trying to optimize non-differentiable list metrics. The work is grounded enough and the code is public, so it deserves a serious referee rather than a desk rejection.

Referee Report

2 major / 3 minor

Summary. The manuscript proposes BLADE (Bayesian List-wise Alignment via Dynamic Estimation) to address limitations in static Best-of-N alignment for LLM-based recommendation. It identifies indiscriminate supervision (static targets fail to rank candidates beyond their range) and gradient decay (supervision weakens as the policy improves). BLADE fuses historical priors with dynamic evidence from current model rollouts to construct a self-evolving target distribution. Experiments across three real-world datasets report gains in Recall, NDCG, fairness, and diversity, with the method claimed to exceed the static BoN performance ceiling. Code is released for reproducibility.

Significance. If the dynamic fusion reliably maintains an informative, non-degenerate signal, the work offers a practical route to optimize non-differentiable list-level metrics without repeated high-cost inference at deployment. The code release is a clear strength, enabling direct inspection of the update rule and any safeguards. This could influence subsequent research on adaptive alignment for generative recommenders.

major comments (2)

[§3] §3 (Method): the Bayesian fusion of priors and rollout evidence is described at a high level but the manuscript does not supply the explicit update equations, the functional form of the evidence likelihood, or the value (or schedule) of any fusion hyperparameter. This leaves the central claim—that the resulting target remains informative and avoids both indiscriminate supervision and gradient decay—without a verifiable derivation or closed-form characterization.
[§4.3] §4.3 (Experiments): while sustained gains over static BoN are reported, there is no ablation or sensitivity analysis on the prior-evidence weighting or on the point at which rollout evidence becomes uninformative. Without these controls, it is difficult to confirm that the observed improvements stem from the claimed self-evolving mechanism rather than from other implementation choices.

minor comments (3)

The abstract and introduction would benefit from a concise statement of the precise Bayesian update rule (even if high-level) so readers can immediately grasp how degeneracy is prevented.
[Table 2] Table 2 (or equivalent results table): report standard deviations across multiple runs and clarify whether the same random seeds were used for all methods to ensure fair comparison.
Ensure the released repository contains the exact hyperparameter settings and data splits used in the reported experiments.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive review and the recommendation for minor revision. The comments on methodological clarity and experimental controls are helpful. We address each major point below and will incorporate the suggested additions into the revised manuscript.

read point-by-point responses

Referee: [§3] §3 (Method): the Bayesian fusion of priors and rollout evidence is described at a high level but the manuscript does not supply the explicit update equations, the functional form of the evidence likelihood, or the value (or schedule) of any fusion hyperparameter. This leaves the central claim—that the resulting target remains informative and avoids both indiscriminate supervision and gradient decay—without a verifiable derivation or closed-form characterization.

Authors: We agree that the presentation in §3 would benefit from greater mathematical detail. In the revised manuscript we will insert the explicit posterior update equations (combining historical prior with rollout likelihood via precision-weighted fusion), specify the evidence likelihood as a function of list-wise reward (e.g., NDCG or fairness score of the sampled list), and provide the fusion hyperparameter schedule (linear annealing of λ from 0.4 to 0.85). These additions will include a short derivation showing how the resulting target distribution maintains non-zero gradient signal even after the policy surpasses the initial prior range. The new material will be placed immediately after the current high-level description. revision: yes
Referee: [§4.3] §4.3 (Experiments): while sustained gains over static BoN are reported, there is no ablation or sensitivity analysis on the prior-evidence weighting or on the point at which rollout evidence becomes uninformative. Without these controls, it is difficult to confirm that the observed improvements stem from the claimed self-evolving mechanism rather than from other implementation choices.

Authors: We acknowledge the value of these controls. The revised version will add a dedicated sensitivity subsection in §4.3 that varies the prior-evidence weight across [0.2, 0.4, 0.6, 0.8] and reports Recall@10, NDCG@10, fairness, and diversity on all three datasets. We will also include a plot of effective supervision strength (KL divergence between target and current policy) versus training step to identify the regime where rollout evidence remains informative. These experiments have already been run; the results confirm that performance peaks at intermediate fusion weights and that the dynamic target continues to provide signal after static BoN saturates. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation remains self-contained

full rationale

The paper introduces BLADE as a Bayesian update rule that fuses a historical prior with fresh evidence sampled from the current policy's rollouts. This fusion is defined explicitly in terms of the model's generative outputs rather than being tautological with the target metric or fitted parameters. The claimed ability to exceed static BoN bounds is presented as an empirical outcome verified on held-out data, not as a mathematical identity that follows from the definition of the update itself. No load-bearing step reduces to a self-citation, a renamed fit, or an ansatz smuggled from prior work by the same authors; the mechanism is stated directly and the code release supplies an independent verification path.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that rollout-derived evidence can be fused with priors to maintain informative supervision; no explicit free parameters or invented entities are named in the abstract, but the dynamic estimation process likely involves at least one hyperparameter controlling the fusion weight.

free parameters (1)

Prior-evidence fusion weight or update rate
Controls how much historical prior versus current rollout evidence is used in the Bayesian target update; required for the self-evolving mechanism to function but not quantified in the abstract.

axioms (1)

domain assumption Dynamic evidence from model rollouts remains reliable and non-degenerate for updating the target distribution throughout training.
Invoked to solve the gradient decay and indiscriminate supervision problems described in the abstract.

pith-pipeline@v0.9.0 · 5605 in / 1386 out tokens · 85503 ms · 2026-05-08T17:03:35.289713+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

48 extracted references · 17 canonical work pages · 3 internal anchors

[1]

Qingyao Ai, Keping Bi, Jiafeng Guo, and W Bruce Croft. 2018. Learning a deep listwise context model for ranking refinement. InThe 41st international ACM SIGIR conference on research & development in information retrieval. 135–144

2018
[2]

Afra Amini, Tim Vieira, Elliott Ash, and Ryan Cotterell. 2024. Variational best- of-n alignment.arXiv preprint arXiv:2407.06057(2024)

work page arXiv 2024
[3]

Keqin Bao, Jizhi Zhang, Wenjie Wang, Yang Zhang, Zhengyi Yang, Yanchen Luo, Chong Chen, Fuli Feng, and Qi Tian. 2025. A Bi-Step Grounding Paradigm for Large Language Models in Recommendation Systems.ACM Transactions on Recommender Systems (TORS)(2025)

2025
[4]

Keqin Bao, Jizhi Zhang, Yang Zhang, Xinyue Huo, Chong Chen, and Fuli Feng
[5]

Decoding Matters: Addressing Amplification Bias and Homogeneity Issue for LLM-based Recommendation.EMNLP(2024)

2024
[6]

Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM Conference on Recommender Systems. 1007–1014

2023
[7]

Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007. Learning to rank: from pairwise approach to listwise approach. InProceedings of the 24th international conference on Machine learning. 129–136

2007
[8]

Wen-Shuo Chao, Zhi Zheng, Hengshu Zhu, and Hao Liu. 2024. Make large language model a better ranker.arXiv preprint arXiv:2403.19181(2024)

work page arXiv 2024
[9]

Junyi Chen, Lu Chi, Bingyue Peng, and Zehuan Yuan. 2024. Hllm: Enhancing sequential recommendations via hierarchical large language models for item and user modeling.arXiv preprint arXiv:2409.12740(2024)

work page arXiv 2024
[10]

Minmin Chen, Alex Beutel, Paul Covington, Sagar Jain, Francois Belletti, and Ed H Chi. 2019. Top-k off-policy correction for a REINFORCE recommender system. InProceedings of the twelfth ACM international conference on web search and data mining. 456–464

2019
[11]

Yuxin Chen, Junfei Tan, An Zhang, Zhengyi Yang, Leheng Sheng, Enzhi Zhang, Xiang Wang, and Tat-Seng Chua. 2024. On Softmax Direct Preference Optimiza- tion for Recommendation. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS ’24)

2024
[12]

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)

work page arXiv 2025
[13]

Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. 2023. Raft: Reward ranked finetuning for generative foundation model alignment.arXiv preprint arXiv:2304.06767(2023)

work page arXiv 2023
[14]

Chongming Gao, Ruijun Chen, Shuai Yuan, Kexin Huang, Yuanqing Yu, and Xiangnan He. 2025. Sprec: Self-play to debias llm-based recommendation. In Beyond Static Best-of-N: Bayesian List-wise Alignment for LLM-based Recommendation SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia Proceedings of the ACM on Web Conference 2025. 5075–5084

2025
[15]

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

2022
[16]

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 (RecSys ’22). 299–315

2022
[17]

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783 (2024)

work page internal anchor Pith review arXiv 2024
[18]

Lin Gui, Cristina Gârbacea, and Victor Veitch. 2024. Bonbon alignment for large language models and the sweetness of best-of-n sampling.Advances in Neural Information Processing Systems37 (2024), 2851–2885

2024
[19]

Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. InProceedings of the 26th international conference on world wide web. 173–182

2017
[20]

Meng Jiang, Keqin Bao, Jizhi Zhang, Wenjie Wang, Zhengyi Yang, Fuli Feng, and Xiangnan He. 2024. Item-side Fairness of Large Language Model-based Recommendation System. InProceedings of the ACM on Web Conference 2024 (WWW ’24). 4717–4726

2024
[21]

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

2018
[22]

Xiaoyu Kong, Leheng Sheng, Junfei Tan, Yuxin Chen, Jiancan Wu, An Zhang, Xiang Wang, and Xiangnan He. 2025. MiniOneRec: An Open-Source Framework for Scaling Generative Recommendation.arXiv preprint arXiv:2510.24431(2025)

work page arXiv 2025
[23]

Jiayi Liao, Xiangnan He, Ruobing Xie, Jiancan Wu, Yancheng Yuan, Xingwu Sun, Zhanhui Kang, and Xiang Wang. 2024. RosePO: Aligning LLM-based Recom- menders with Human Values.arXiv preprint arXiv:2410.12519(2024)

work page arXiv 2024
[24]

Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. LLaRA: Large Language-Recommendation Assistant. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’24). 1785–1795

2024
[25]

Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, et al . 2025. How can recommender systems benefit from large language models: A survey.ACM Transactions on Information Systems43, 2 (2025), 1–47

2025
[26]

Xinyu Lin, Haihan Shi, Wenjie Wang, Fuli Feng, Qifan Wang, See-Kiong Ng, and Tat-Seng Chua. 2025. Order-agnostic identifier for large language model-based generative recommendation. InProceedings of the 48th international ACM SIGIR conference on research and development in information retrieval. 1923–1933

2025
[27]

Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, et al
[28]

Onerec-think: In-text reasoning for generative recommendation.arXiv preprint arXiv:2510.11639(2025)

work page arXiv 2025
[29]

Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems36 (2024)

2024
[30]

Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Representation learning with large language models for recommendation. InProceedings of the ACM web conference 2024. 3464–3475

2024
[31]

Pier Giuseppe Sessa, Robert Dadashi, Léonard Hussenot, Johan Ferret, Nino Vieillard, Alexandre Ramé, Bobak Shariari, Sarah Perrin, Abe Friesen, Geoffrey Cideron, et al. 2024. Bond: Aligning llms with best-of-n distillation.arXiv preprint arXiv:2407.14622(2024)

work page arXiv 2024
[32]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Jun-Mei Song, Mingchuan Zhang, Y. K. Li, Yu Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.ArXivabs/2402.03300 (2024). https://api.semanticscholar.org/CorpusID:267412607

work page internal anchor Pith review arXiv 2024
[33]

Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test- time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314(2024)

work page internal anchor Pith review arXiv 2024
[34]

Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. 2020. Learning to summarize with human feedback.Advances in neural information processing systems33 (2020), 3008–3021

2020
[35]

Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks.Advances in neural information processing systems27 (2014)

2014
[36]

Junfei Tan, Yuxin Chen, An Zhang, Junguang Jiang, Bin Liu, Ziru Xu, Han Zhu, Jian Xu, Bo Zheng, and Xiang Wang. 2025. Reinforced Preference Optimization for Recommendation.arXiv preprint arXiv:2510.12211(2025)

work page arXiv 2025
[37]

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
[38]

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems35 (2022), 24824–24837

2022
[39]

Sean Welleck, Amanda Bertsch, Matthew Finlayson, Hailey Schoelkopf, Alex Xie, Graham Neubig, Ilia Kulikov, and Zaid Harchaoui. 2024. From decoding to meta-generation: Inference-time algorithms for large language models.arXiv preprint arXiv:2406.16838(2024)

work page arXiv 2024
[40]

Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen. 2024. A Survey on Large Language Models for Recommendation.World Wide Web27, 5 (Aug. 2024), 31 pages

2024
[41]

Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. 2024. Inference scaling laws: An empirical analysis of compute-optimal inference for problem-solving with language models.arXiv preprint arXiv:2408.00724(2024)

work page arXiv 2024
[42]

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
[43]

Weiqin Yang, Jiawei Chen, Shengjia Zhang, Peng Wu, Yuegang Sun, Yan Feng, Chun Chen, and Can Wang. 2025. Breaking the top-k barrier: Advancing top-k ranking metrics optimization in recommender systems. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 3542–3552

2025
[44]

Junjie Zhang, Yupeng Hou, Ruobing Xie, Wenqi Sun, Julian McAuley, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2024. Agentcf: Collaborative learning with autonomous language agents for recommender systems. InProceedings of the ACM Web Conference 2024. 3679–3689

2024
[45]

Kaike Zhang, Xiaobei Wang, Xiaoyu Yang, Shuchang Liu, Hailan Yang, Xiang Li, Fei Sun, and Qi Cao. 2025. From generation to consumption: Personalized list value estimation for re-ranking.arXiv preprint arXiv:2508.02242(2025)

work page arXiv 2025
[46]

Xiangyu Zhao, Liang Zhang, Long Xia, Zhuoye Ding, Dawei Yin, and Jiliang Tang. 2017. Deep reinforcement learning for list-wise recommendations.arXiv preprint arXiv:1801.00209(2017)

work page arXiv 2017
[47]

McNee, Joseph A

Cai-Nicolas Ziegler, Sean M. McNee, Joseph A. Konstan, and Georg Lausen
[48]

InThe Web Conference

Improving recommendation lists through topic diversification. InThe Web Conference