Pith. sign in

REVIEW 3 major objections 6 minor 3 cited by

GFlowGR: Fine-tuning Generative Recommendation Frameworks with Generative Flow Networks

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read GFlowGR replaces the supervised fine-tuning of generative recommender LLMs with a GFlowNet flow-matching loss, making generation probability proportional to item utility and improving accuracy and diversity.

desk verdict A solid and useful GFlowNet fine-tuning recipe for generative recommenders whose empirical gains look real, but the reward-proportionality guarantee is overstated for the actual objective. read the letter →

arxiv 2506.16114 v3 pith:S6N4QQHG submitted 2025-06-19 cs.IR cs.AI

Yejing Wang , Shengyu Zhou , Jinyu Lu , Qidong Liu , Xinhang Li , Wenlin Zhang , Feng Li , Pengjie Wang
show 4 more authors
Chuan Yu Jian Xu Bo Zheng Xiangyu Zhao
This is my paper
classification cs.IRcs.AI
keywords generativerecommendationGFlowNetsLLMfine-tuningtrajectorybalanceexposurebiasrewardmodelingset-wise
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

Generative recommender systems that use an LLM to emit item identifiers are usually fine-tuned with next-token prediction on a single ground-truth item, so they never learn from the full set of candidates a recommender must rank. This paper tries to close that gap by treating each item as a multi-token generation trajectory and fine-tuning the LLM with a GFlowNet objective, which balances incoming and outgoing flow so that the probability of generating an item becomes proportional to a learned reward for that item. The reward fuses interaction signals, collaborative-model scores, a token-overlap format reward, and optionally the LLM's own confidence. On Yelp, Beauty, and Instruments, and across two backbones, the trajectory-balance variant consistently beats SFT, DPO, GRPO, S-DPO, SPRec, and IPA on hit rate and NDCG, and the authors report a 0.4 percent relative annual revenue gain from Taobao deployment.

What carries the argument

The engine is the GFlowNet flow-matching identity $F(s_l)P_F(s_{l+1}|s_l) = F(s_{l+1})P_B(s_l|s_{l+1})$, instantiated as detailed balance or trajectory balance losses. Because each item identifier is generated by deterministic token concatenation, the backward probability is 1 and the forward probability is just the LLM's token probability, so the loss reduces to a squared log-ratio involving the flow estimator $F(s_l)$ and the terminal reward $R(s_L)$. Around this sits a trajectory sampler (interaction logs, random negatives, collaborative-model-scored curriculum negatives, or LLM beam search) and a reward model that sums interaction signal, collaborative score, and a format reward counting shared tokens with the positive item. The final objective keeps the SFT loss and adds a weighted sum of GFlowNet losses over the $N$ sampled trajectories.

What would settle it

Run GFlowGR with two deliberately biased samplers—one always drawing very easy negatives and one always drawing very hard negatives—then measure, on held-out items, whether the model's generation probability stays proportional to the reward (e.g., whether the ratio $P(\tau)/R(s_L)$ is roughly constant across trajectories). If the ratio varies systematically with the sampler, the reward-proportional claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a GFlowNet fine-tuning stage is a drop-in replacement for the standard supervised fine-tuning stage of a generative recommender: construct trajectories from observed user interactions plus augmented candidates, score each terminated item with a behavior-aware reward, and train with detailed balance or trajectory balance while retaining the SFT loss. The GFlowNet constraint $P(\tau)\propto R(s_L)$ then aligns the model's token-level generation probabilities with item utility, giving token-level supervision that item-level DPO and GRPO style rewards lack. The authors argue this simultaneously fixes two mismatches: point-wise training versus set-wise serving, and equal treatment of interactions versus value-differentiated signals.

Load-bearing premise

The central premise is that training on trajectories chosen by external samplers—interaction logs, random draws, collaborative-model scores, or beam search—without any off-policy correction converges to a policy whose generation probability is proportional to the reward.

Editorial extensions

If this is right

  • Token-level credit assignment: every token in an item identifier receives a flow-consistent gradient, not just a single item-level reward.
  • Set-wise learning: training on clicked, presented-but-not-clicked, and augmented trajectories lets the model reduce exposure bias and produce diverse candidate lists.
  • Value-aligned ranking: higher-utility items are generated with proportionally higher probability, which the experiments show improves R@5, R@10, N@5, and N@10 over all baselines.
  • Drop-in compatibility: GFlowGR changes only the fine-tuning stage, so existing generative recommender backbones such as TIGER and LETTER can adopt it without touching inference.
  • Production evidence: the deployment at Taobao search advertising reports a 0.43 percent revenue lift in an online A/B test and a 0.4 percent relative annual revenue increase.

Reading between the lines

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

  • Editorial: if the external sampler is strongly biased and no off-policy correction is applied, the equal-weight loss in Eq. (7) could make the model track the sampler's distribution instead of $P(\tau)\propto R(s_L)$; an importance-weighting variant would be the direct test.
  • Editorial: the CM-free ablation suggests collaborative signals are helpful but not necessary, so a version using only LLM confidence for both sampling and reward is a natural next experiment the paper does not run.
  • Editorial: the same trajectory formulation transfers to other multi-token generative IR tasks, such as query suggestion or ad creative generation, where set-wise value-aware fine-tuning is desirable.
  • Editorial: the diversity result (smaller KL divergence to the normal distribution, larger collaborative-score range) implies exposure-bias metrics like long-tail coverage could be reported as a primary outcome in future work.
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 GFlowGR, a fine-tuning method for generative recommendation (GR) that augments the standard supervised fine-tuning (SFT) objective with a GFlowNet detailed-balance or trajectory-balance loss over sampled item-identifier trajectories. The method has three components: an adaptive trajectory sampler (interaction logs, random negatives, collaborative-model-based negatives, or LLM-based beam-search negatives), a reward model combining interaction type, collaborative-model score, and token overlap with the positive item, and the GFlowNet loss. Experiments on Yelp, Beauty, and Instruments with TIGER and LETTER backbones report consistent gains over SFT, GRPO, DPO, S-DPO, SPRec, and IPA, and the paper further reports a Taobao deployment with a claimed +0.43% revenue lift. The central theoretical claim is that training makes the generation probability of a trajectory proportional to its terminal reward, P(τ) ∝ R(s_L), thereby aligning generation probabilities with item utility.

Significance. If the empirical gains hold, GFlowGR would be a practical, drop-in fine-tuning upgrade for GR backbones and an early industrial validation of GFlowNet-style training for LLM-based recommenders. The method is modular, the public experiments cover multiple datasets and backbones, and the paper includes ablations and a deployment report. However, the theoretical motivation is not established by the actual training objective, and the experimental reporting omits standard deviations and a clear hyperparameter-selection protocol. The contribution is therefore currently an empirical recipe with strong results, rather than a demonstrated flow-consistency result.

major comments (3)
  1. [Section 2, Section 3.1, Section 3.3, Algorithm 1] The paper's central claim that training yields P(τ) ∝ R(s_L) is not supported by the actual objective in Eq. (7). The GFlowNet losses in Eqs. (5)-(6) are evaluated on trajectories chosen by an external sampler (interaction logs, random draws, CM-based selection, or beam search), and each trajectory enters Eq. (7) with equal weight. Standard GFlowNet convergence results require minimizing the DB/TB loss over the full trajectory space or with on-policy sampling; minimizing the expected squared log-ratio under an arbitrary proposal q can be driven to zero on the support of q without matching R on unsampled items, and no importance weighting or coverage assumption is stated. In addition, the retained SFT term LGR(U,v) in Eq. (7) maximizes the probability of the positive item regardless of the rewards of other items, so even a perfectly converged GFlowNet term would not deliver a distribution proportional to R for the summed objective. The authors should either prove the proportionality claim under the stated sampling scheme, add an off-policy correction or coverage condition, or substantially temper the claims and provide empirical diagnostics, such as comparing P and R on held-out items.
  2. [Section 4.1, Table 2] Table 2 reports point estimates from three seeds but no standard deviations, and the claimed one-sided t-test significance (p < 0.05) cannot be verified from the reported numbers. The implementation details state that 'we find the optimal settings for both baselines and GFlowGR' without describing a held-out development protocol or the number of hyperparameter trials, which leaves open the possibility that the comparison favors the proposed method. Please report means and standard deviations (or confidence intervals), exact p-values, and a clear protocol for hyperparameter selection, such as tuning on the validation set with fixed seeds.
  3. [Section 3.2, Section 3.3] The GFlowNet component is underspecified for reproducibility. The flow estimator F(s_l) and the scalar Z in Eqs. (5)-(6) are described as learnable components, but their architecture, initialization, and training procedure are not given. Moreover, the reward model in Section 3.2 can assign R(s_L) = 0 to an augmented trajectory (e.g., r_a = 0, r_c = 0, r_sim = 0), which makes the logarithm in Eq. (6) undefined. The authors should specify a strictly positive lower bound on rewards or use a log-safe loss variant, and they should report the reward values and combination weights actually used in the experiments.
minor comments (6)
  1. [Abstract] The abstract states that experiments were conducted on 'two real-world datasets,' while the full abstract and Section 4.1 describe experiments on three datasets (Yelp, Beauty, and Instruments); please reconcile this inconsistency.
  2. [Introduction, Section 5] The introduction says the deployment delivers 'billion-level daily monetary gains' and a 0.4% improvement in annual revenue, while Section 5 reports a +0.43% lift in a 15-day A/B test on 10% of traffic; these figures should be harmonized and the time scale of the 'billion-level' claim clarified.
  3. [Section 4.7] The diversity analysis compares the generation distribution to 'the normal distribution' without defining what is meant by 'normal' here; reporting the standard deviation of per-item probabilities is not by itself a diversity metric, and the interpretation should be made precise.
  4. [Section 6] The novelty statement 'we are the first to explore GFlowNets for generative recommendations' appears to conflict with cited reference [10], which is described as applying flow-guided tuning to LLM-based recommenders; the distinction should be made explicit or the novelty claim should be softened.
  5. [Section 3.1] The interaction-log and random-sampling trajectory samplers do not specify how augmented samples are guaranteed to exclude positive items or how duplicates across sampled sets are handled; please clarify.
  6. [Figures 3 and 4] The red dashed line for SFT is said to fall outside the plotted range for some panels; including it within the axis range or annotating its value would make the parameter analysis more informative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GFlowGR applies standard GFlowNet objectives to an externally validated recommendation task; no prediction is defined from its own output.

full rationale

The derivation chain is self-contained and not circular. The paper takes the standard GR token-generation formulation (Eq. 1), maps the LLM's token probabilities to GFlowNet forward probabilities, and combines the SFT loss with DB/TB GFlowNet losses (Eqs. 5-7). The claimed property P(τ) ∝ R(s_L) is not assumed as an input; it is the intended convergence property of the GFlowNet objective, cited to external foundational work [1,2,38]. The trajectory sampler and reward model are designed from external signals (interaction logs, random draws, collaborative-model scores, LLM beam search, interaction levels, and token overlap with the training positive item). None of these is defined in terms of the evaluation metrics or the paper's reported gains. The token-overlap reward r_sim is computed against the ground-truth positive item during training, which is a normal supervised-label design rather than a fitted parameter being renamed as a prediction; at evaluation the target is held out under the standard last-item split. The only self-citation ([53]) appears in a background list about inference efficiency and is not load-bearing. The skeptical concern that off-policy sampling and the retained SFT term may prevent exact reward proportionality is a correctness or robustness issue, not circularity: the paper does not use its conclusion as an input to derive that conclusion. The main results are measured against external baselines on public datasets and production data, so no step reduces by construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard GFlowNet theory, the deterministic trajectory assumption, and two ad hoc design choices: the additive reward model and the CM-based curriculum. The free parameters (λ, K, r_a values, reward weights) are tuned or hand-set; the main text does not fix K for the headline results, which complicates reproduction. No new entities are postulated.

free parameters (4)
  • λ (GFN loss weight) = 1.0 (best of {0.01, 0.1, 1, 10, 100})
    Balances SFT and GFlowNet losses in Eq. (7); tuned in Section 4.3.
  • K (number of augmented trajectories) = not explicitly stated for main results; studied over {0, 1, 3, 5}
    Controls trajectory set size; the default N is not fixed in the main experimental setup.
  • Interaction reward values r_a = 10 for liked, 1 for clicked, 0 for augmented
    Hand-chosen signal magnitudes in Section 3.2; no sensitivity analysis for these values.
  • Reward combination weights = 1, 1, 1 (sum), with learnable weighted-sum variant
    Default sum in Eq. (7); Table 4 shows the learnable weighted sum improves top-5 metrics, so the default is not optimal.
assumptions (4)
  • standard math GFlowNet flow-matching principle: DB and TB losses are valid training objectives for matching generation probability to reward.
    Invoked in Section 2 and Section 3.3; treated as background, not reproved.
  • domain assumption Item tokenization yields deterministic trajectories with unique predecessors, so P_B(s_l | s_{l+1}) = 1.
    Used in Section 2 when substituting P_B = 1 into Eqs. (5)-(6).
  • ad hoc to paper Interaction type, CM score, and token overlap can be combined into a scalar reward by addition.
    Section 3.2 defines r_a + r_c + r_sim; Table 4 shows a learnable weighted sum behaves differently, so the additive form is not neutral.
  • ad hoc to paper CM scores order negatives from easy to hard, making a curriculum for trajectory sampling.
    Section 3.1, CM-based sampling; no analysis of CM score reliability is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GFlowGR: Fine-tuning Generative Recommendation Frameworks with Generative Flow Networks." pith.science (2026). https://pith.science/paper/S6N4QQHG

@misc{pith2026250616114,
  author       = {Pith},
  title        = {Pith review of: GFlowGR: Fine-tuning Generative Recommendation Frameworks with Generative Flow Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S6N4QQHG}},
  note         = {Machine review of arXiv:2506.16114}
}
read the original abstract

Generative recommendations (GR), which usually include item tokenizers and generative Large Language Models (LLMs), have demonstrated remarkable success across a wide range of scenarios. The majority of existing research efforts primarily concentrate on developing powerful item tokenizers or advancing LLM decoding strategies to attain superior performance. However, the critical fine-tuning step in GR frameworks, which is essential for adapting LLMs to recommendation data, remains largely unexplored. Current approaches predominantly rely on either the next-token prediction loss of supervised fine-tuning (SFT) or recommendationspecific direct preference optimization (DPO) strategies. Both methods ignore the exploration of possible positive unobserved samples, which is commonly referred to as the exposure bias problem. To mitigate this problem, this paper treats the GR as a multi-step generation task and constructs a GFlowNets-based fine-tuning framework (GFlowGR). The proposed framework integrates collaborative knowledge from traditional recommender systems to create an adaptive trajectory sampler and a comprehensive reward model. Leveraging the diverse generation property of GFlowNets, along with sampling and heuristic weighting techniques, GFlowGR emerges as a promising approach to mitigate the exposure bias problem. Extensive empirical results on two real-world datasets and with two different GR backbones highlight the effectiveness and robustness of GFlowGR.

Figures

Figures reproduced from arXiv: 2506.16114 by the authors.

Figure 1
Figure 1. Overview of GR and GFlowNets, with the length of item identifiers [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Framework of GFlowGR. The visualized example shows a the learning on a [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Parameter study on 𝑁 and 𝜆 with TIGER on Beauty. The red dashed line denotes the performance of SFT, which falls outside the plotted range for the figures corresponding to 𝑁 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Parameter study on 𝑁 and 𝜆 with LETTER on Beauty. The red dashed line denotes the performance of SFT, which falls outside the plotted range for the figures corresponding to 𝑁 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Case study on user-2089 from Beauty with TIGER: We visualized three items from the top-10 list of each method for [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Deployed business scenarios in Taobao, illustrated [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Hierarchical Residual Policy Optimization for Generative Recommendations

    cs.IR 2026-08 conditional novelty 6.0 of 10

    HRPO decomposes item-level rewards into token-level 'residual credits' along semantic identifier hierarchies and optimizes the generator with a PPO-style objective, improving session utility in KuaiSim and production ...

  2. Interpretable Representation via LLM-Driven Generative Disentanglement for Local-Life Service Recommendation

    cs.IR 2026-07 conditional novelty 5.5 of 10

    Disentangling LLM hidden states into attribute-aligned geo and semantic slots before dual-stream residual quantization cuts SID collisions and improves local-life recommendation AUC.

  3. RecoWorld: Building Simulated Environments for Agentic Recommender Systems

    cs.IR 2025-09 conditional novelty 5.0 of 10

    A design proposal, not a tested system: a dual-view simulation loop in which an LLM-simulated user issues reflective instructions when about to disengage, and an instruction-following recommender adapts to maximize si...

Reference graph

Works this paper leans on

72 extracted references · 25 canonical work pages · cited by 3 Pith papers

  1. [1]

    Emmanuel Bengio, Moksh Jain, Maksym Korablyov, Doina Precup, and Yoshua Bengio. 2021. Flow network based generative models for non-iterative diverse candidate generation. Advances in Neural Information Processing Systems 34 (2021), 27381–27394

    work page 2021

  2. [2]

    Yoshua Bengio, Salem Lahlou, Tristan Deleu, Edward J Hu, Mo Tiwari, and Emmanuel Bengio. 2023. Gflownet foundations. Journal of Machine Learning Research 24, 210 (2023), 1–55

    work page 2023

  3. [3]

    Ben Chen, Xian Guo, Siyuan Wang, Zihan Liang, Yue Lv, Yufei Ma, Xinlong Xiao, Bowen Xue, Xuxin Zhang, Ying Yang, et al . 2025. OneSearch: A Preliminary Exploration of the Unified End-to-End Generative Framework for E-commerce Search. arXiv preprint arXiv:2509.03236 (2025)

    arXiv 2025

  4. [4]

    Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, and Xiangnan He. 2023. Bias and debias in recommender system: A survey and future directions. ACM Transactions on Information Systems 41, 3 (2023), 1–39

    2023

  5. [5]

    Yuxin Chen, Junfei Tan, An Zhang, Zhengyi Yang, Leheng Sheng, Enzhi Zhang, Xiang Wang, and Tat-Seng Chua. [n. d.]. On Softmax Direct Preference Opti- mization for Recommendation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    work page

  6. [6]

    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

  7. [7]

    Yijie Ding, Yupeng Hou, Jiacheng Li, and Julian McAuley. 2024. Inductive Generative Recommendation via Retrieval-based Speculation. arXiv preprint arXiv:2410.02939 (2024)

    arXiv 2024

  8. [8]

    Zichuan Fu, Xiangyang Li, Chuhan Wu, Yichao Wang, Kuicai Dong, Xiangyu Zhao, Mengchen Zhao, Huifeng Guo, and Ruiming Tang. 2025. A unified frame- work for multi-domain ctr prediction via large language models. ACM Transac- tions on Information Systems 43, 5 (2025), 1–33

    work page 2025

Show all 72 references
  1. [9]

    Chongming Gao, Ruijun Chen, Shuai Yuan, Kexin Huang, Yuanqing Yu, and Xiangnan He. 2025. SPRec: Self-Play to Debias LLM-based Recommendation. In Proceedings of the ACM on Web Conference 2025 . 5075–5084

    2025

  2. [10]

    Chongming Gao, Mengyao Gao, Chenxiao Fan, Shuai Yuan, Wentao Shi, and Xiangnan He. 2025. Process-supervised llm recommenders via flow-guided tuning. arXiv preprint arXiv:2503.07377 (2025)

    2025 arXiv

  3. [11]

    Chongming Gao, Shijun Li, Wenqiang Lei, Jiawei Chen, Biao Li, Peng Jiang, Xiangnan He, Jiaxin Mao, and Tat-Seng Chua. 2022. KuaiRec: A fully-observed dataset and insights for evaluating recommender systems. In Proceedings of the 31st ACM International Conference on Information...

    2022

  4. [12]

    Jingtong Gao, Bo Chen, Xiangyu Zhao, Weiwen Liu, Xiangyang Li, Yichao Wang, Wanyu Wang, Huifeng Guo, and Ruiming Tang. 2025. Llm4rerank: Llm-based auto-reranking framework for recommendations. In Proceedings of the ACM on Web Conference 2025. 228–239

    2025

  5. [13]

    Jingtong Gao, Yewen Li, Shuai Mao, Peng Jiang, Nan Jiang, Yejing Wang, Qing- peng Cai, Fei Pan, Peng Jiang, Kun Gai, et al. 2025. Generative auto-bidding with value-guided explorations. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development i...

    2025

  6. [14]

    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)

    2026

  7. [15]

    Xian Guo, Ben Chen, Siyuan Wang, Ying Yang, Chenyi Lei, Yuqing Ding, and Han Li. 2025. OneSug: The Unified End-to-End Generative Framework for E-commerce Query Suggestion. arXiv preprint arXiv:2506.06913 (2025)

    2025 arXiv

  8. [16]

    Wenyue Hua, Shuyuan Xu, Yingqiang Ge, and Yongfeng Zhang. 2023. How to index item ids for recommendation foundation models. In Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region . 195–204

    2023

  9. [17]

    Ting-Ji Huang, Jia-Qi Yang, Chunxu Shen, Kai-Qi Liu, De-Chuan Zhan, and Han- Jia Ye. 2024. Improving llms for recommendation with out-of-vocabulary tokens. arXiv preprint arXiv:2406.08477 (2024)

    2024 arXiv

  10. [18]

    Moksh Jain, Emmanuel Bengio, Alex Hernandez-Garcia, Jarrid Rector-Brooks, Bonaventure FP Dossou, Chanakya Ajit Ekbote, Jie Fu, Tianyu Zhang, Michael Kilgour, Dinghuai Zhang, et al. 2022. Biological sequence design with gflownets. In International Conference on Machine Learning...

    2022

  11. [19]

    Moksh Jain, Tristan Deleu, Jason Hartford, Cheng-Hao Liu, Alex Hernandez- Garcia, and Yoshua Bengio. 2023. Gflownets for ai-driven scientific discovery. Digital Discovery 2, 3 (2023), 557–577

    2023

  12. [20]

    Haoqiang Kang, Enna Sachdeva, Piyush Gupta, Sangjae Bae, and Kwonjoon Lee

  13. [21]

    Solomon Kullback. 1951. Kullback-leibler divergence

    1951

  14. [22]

    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

  15. [23]

    Xiaopeng Li, Bo Chen, Junda She, Shiteng Cao, You Wang, Qinlin Jia, Haiying He, Zheli Zhou, Zhao Liu, Ji Liu, et al. 2025. A survey of generative recommendation from a tri-decoupled perspective: Tokenization, architecture, and optimization. (2025)

    2025

  16. [24]

    Yongqi Li, Xinyu Lin, Wenjie Wang, Fuli Feng, Liang Pang, Wenjie Li, Liqiang Nie, Xiangnan He, and Tat-Seng Chua. 2024. A survey of generative search and recom- mendation in the era of large language models. arXiv preprint arXiv:2404.16924 (2024)

    2024 arXiv

  17. [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 Systems 43, 2 (2025), 1–47

    2025

  18. [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. arXiv preprint arXiv:2502.10833 (2025)

    2025 arXiv

  19. [27]

    Xinyu Lin, Chaoqun Yang, Wenjie Wang, Yongqi Li, Cunxiao Du, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2025. Efficient Inference for Large Language Model- based Generative Recommendation. In The Thirteenth International Conference on Learning Representations. https://openre...

    2025

  20. [28]

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

  21. [29]

    Qidong Liu, Xian Wu, Yejing Wang, Zijian Zhang, Feng Tian, Yefeng Zheng, and Xiangyu Zhao. 2024. Llm-esr: Large language models enhancement for long- tailed sequential recommendation. Advances in Neural Information Processing Systems 37 (2024), 26701–26727

    2024

  22. [30]

    Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Zijian Zhang, Feng Tian, and Yefeng Zheng. 2024. Large language model distilling medication recommendation model. arXiv preprint arXiv:2402.02803 (2024)

    2024 arXiv

  23. [31]

    Qidong Liu, Xiangyu Zhao, Yuhao Wang, Yejing Wang, Zijian Zhang, Yuqi Sun, Xiang Li, Maolin Wang, Pengyue Jia, Chong Chen, et al. 2025. Large Language Model Enhanced Recommender Systems: Methods, Applications and Trends. In Proceedings of the 31st ACM SIGKDD Conference on Know...

    2025

  24. [32]

    Qidong Liu, Xiangyu Zhao, Yejing Wang, Zijian Zhang, Howard Zhong, Chong Chen, Xiang Li, Wei Huang, and Feng Tian. 2025. Bridge the domains: Large language models enhanced cross-domain sequential recommendation. InProceed- ings of the 48th International ACM SIGIR Conference on...

    2025

  25. [33]

    Shuchang Liu, Qingpeng Cai, Zhankui He, Bowen Sun, Julian McAuley, Dong Zheng, Peng Jiang, and Kun Gai. 2023. Generative flow network for listwise rec- ommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 1524–1534

    2023

  26. [34]

    Zihan Liu, Yupeng Hou, and Julian McAuley. 2024. Multi-behavior generative recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management . 1575–1585

    2024

  27. [35]

    Ziru Liu, Shuchang Liu, Bin Yang, Zhenghai Xue, Qingpeng Cai, Xiangyu Zhao, Zijian Zhang, Lantao Hu, Han Li, and Peng Jiang. 2024. Modeling User Retention through Generative Flow Networks. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining ...

    2024

  28. [36]

    Ziwei Liu, Yejing Wang, Qidong Liu, Zijian Zhang, Chong Chen, Wei Huang, and Xiangyu Zhao. 2025. The Best of the Two Worlds: Harmonizing Semantic and Hash IDs for Sequential Recommendation. arXiv preprint arXiv:2512.10388 (2025)

    2025 arXiv

  29. [37]

    Haokai Ma, Ruobing Xie, Lei Meng, Fuli Feng, Xiaoyu Du, Xingwu Sun, Zhanhui Kang, and Xiangxu Meng. 2024. Negative Sampling in Recommendation: A Survey and Future Directions. arXiv preprint arXiv:2409.07237 (2024)

    2024 arXiv

  30. [38]

    Nikolay Malkin, Moksh Jain, Emmanuel Bengio, Chen Sun, and Yoshua Bengio

  31. [39]

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel

  32. [40]

    Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. 2023. A com- prehensive overview of large language models. arXiv preprint arXiv:2307.06435 (2023)

    2023 arXiv

  33. [41]

    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 Systems 36 (2023), 53728–53741

    2023

  34. [42]

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

  35. [43]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al . 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Aust...

    2024 arXiv

  36. [44]

    Ryoichi Takase, Masaya Tsunokake, Yuta Tsuchiya, and Shota Inuzuka. 2024. GFlowNet Fine-tuning for Diverse Correct Solutions in Mathematical Reasoning Tasks. arXiv preprint arXiv:2410.20147 (2024)

    2024 arXiv

  37. [45]

    Hanbing Wang, Xiaorui Liu, Wenqi Fan, Xiangyu Zhao, Venkataramana Kini, Devendra Pratap Yadav, Fei Wang, Zhen Wen, and Hui Liu. 2025. Rethinking large language model architectures for sequential recommendations. In Proceedings of the 14th International Joint Conference on Natu...

    2025

  38. [46]

    Jinpeng Wang, Jieming Zhu, and Xiuqiang He. 2021. Cross-batch negative sam- pling for training two-tower recommenders. In Proceedings of the 44th interna- tional ACM SIGIR conference on research and development in information retrieval . 1632–1636

    2021

  39. [47]

    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. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management . 2400–2409

    2024

  40. [48]

    Xin Wang, Yudong Chen, and Wenwu Zhu. 2021. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence 44, 9 (2021), 4555–4576

    2021

  41. [49]

    Yidan Wang, Zhaochun Ren, Weiwei Sun, Jiyuan Yang, Zhixiang Liang, Xin Chen, Ruobing Xie, Su Yan, Xu Zhang, Pengjie Ren, et al. 2024. Content-Based Collaborative Generation for Recommender Systems. In Proceedings of the 33rd ACM International Conference on Information and Know...

    2024

  42. [50]

    Yuhao Wang, Yichao Wang, Zichuan Fu, Xiangyang Li, Wanyu Wang, Yuyang Ye, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. 2024. Llm4msr: An llm- enhanced paradigm for multi-scenario recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowled...

    2024

  43. [51]

    Yuhao Wang, Xiangyu Zhao, Bo Chen, Qidong Liu, Huifeng Guo, Huanshuo Liu, Yichao Wang, Rui Zhang, and Ruiming Tang. 2023. PLATE: A prompt-enhanced paradigm for multi-scenario recommendations. In Proceedings of the 46th In- ternational ACM SIGIR Conference on Research and Devel...

    2023

  44. [52]

    Yejing Wang, Xiangyu Zhao, Tong Xu, and Xian Wu. 2022. Autofield: Automating feature selection in deep recommender systems. In Proceedings of the ACM Web Conference 2022. 1977–1986

    2022

  45. [53]

    Yejing Wang, Shengyu Zhou, Jinyu Lu, Ziwei Liu, Langming Liu, Maolin Wang, Wenlin Zhang, Feng Li, Wenbo Su, Pengjie Wang, et al. 2025. NEZHA: A Zero- sacrifice and Hyperspeed Decoding Architecture for Generative Recommenda- tions. arXiv preprint arXiv:2511.18793 (2025)

    2025

  46. [54]

    Zhipeng Wei, Kuo Cai, Junda She, Jie Chen, Minghao Chen, Yang Zeng, Qiang Luo, Wencong Zeng, Ruiming Tang, Kun Gai, et al. 2025. OneLoc: Geo-Aware Genera- tive Recommender Systems for Local Life Service.arXiv preprint arXiv:2508.14646 (2025)

    2025 arXiv

  47. [55]

    Ji Yang, Xinyang Yi, Derek Zhiyuan Cheng, Lichan Hong, Yang Li, Simon Xiaom- ing Wang, Taibai Xu, and Ed H Chi. 2020. Mixed negative sampling for learning two-tower neural networks in recommendations. In Companion proceedings of the web conference 2020 . 441–447

    2020

  48. [56]

    Fangxu Yu, Lai Jiang, Haoqiang Kang, Shibo Hao, and Lianhui Qin. 2024. Flow of reasoning: Efficient training of llm policy with divergent thinking. arXiv preprint arXiv:2406.05673 (2024)

    2024 arXiv

  49. [57]

    Chao Zhang, Haoxin Zhang, Shiwei Wu, Di Wu, Tong Xu, Xiangyu Zhao, Yan Gao, Yao Hu, and Enhong Chen. 2025. Notellm-2: Multimodal large representation models for recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 . 2815–2826

    2025

  50. [58]

    Zijian Zhang, Shuchang Liu, Ziru Liu, Rui Zhong, Qingpeng Cai, Xiangyu Zhao, Chunxu Zhang, Qidong Liu, and Peng Jiang. 2025. Llm-powered user simulator for recommender system. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 13339–13347

    2025

  51. [59]

    Xiangyu Zhao, Liang Zhang, Zhuoye Ding, Long Xia, Jiliang Tang, and Dawei Yin

  52. [60]

    Xiangyu Zhao, Liang Zhang, Zhuoye Ding, Dawei Yin, Yihong Zhao, and Jiliang Tang. 2017. Deep Reinforcement Learning for List-wise Recommendations. arXiv preprint arXiv:1801.00209 (2017)

    2017 arXiv

  53. [61]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) . IEEE, 1435–1448

    2024

  54. [62]

    Bowen Zheng, Hongyu Lu, Yu Chen, Wayne Xin Zhao, and Ji-Rong Wen. 2025. Universal Item Tokenization for Transferable Generative Recommendation.arXiv preprint arXiv:2504.04405 (2025)

    2025 arXiv

  55. [63]

    Zuowu Zheng, Ze Wang, Fan Yang, Jiangke Fan, Teng Zhang, and Xingxing Wang. 2025. EGA: A Unified End-to-End Generative Framework for Industrial Advertising Systems. arXiv preprint arXiv:2505.17549 (2025)

    2025 arXiv

  56. [64]

    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

  57. [65]

    Jieming Zhu, Mengqun Jin, Qijiong Liu, Zexuan Qiu, Zhenhua Dong, and Xiu Li

  58. [72]

    InProceedings of the 18th ACM Conference on Recommender Systems

    CoST: Contrastive Quantization based Semantic Tokenization for Genera- tive Recommendation. InProceedings of the 18th ACM Conference on Recommender Systems. 969–974

  59. [2015]

    In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval

    Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval. 43–52

  60. [2018]

    In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining

    Recommendations with Negative Feedback via Pairwise Deep Reinforce- ment Learning. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . ACM, 1040–1048

  61. [2022]

    Advances in Neural Information Processing Systems 35 (2022), 5955–5967

    Trajectory balance: Improved credit assignment in gflownets. Advances in Neural Information Processing Systems 35 (2022), 5955–5967

    2022

  62. [2023]

    Advances in Neural Information Processing Systems 36 (2023), 10299–10315

    Recommender systems with generative retrieval. Advances in Neural Information Processing Systems 36 (2023), 10299–10315

    2023

  63. [2024]

    arXiv preprint arXiv:2409.05546 (2024)

    End-to-End Learnable Item Tokenization for Generative Recommendation. arXiv preprint arXiv:2409.05546 (2024)

    2024 arXiv

  64. [2025]

    arXiv preprint arXiv:2503.06514 (2025)

    GFlowVLM: Enhancing Multi-step Reasoning in Vision-Language Models with Generative Flow Networks. arXiv preprint arXiv:2503.06514 (2025)

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.