Pith. sign in

REVIEW 4 major objections 4 minor 40 references

DEGR: Dual Exploration-Driven Generative Re-Ranking for Adaptive Cross-Request Context Bridging

T0 review · 4 major / 4 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 DEGR turns re-ranking into an adaptive cross-request bridge by balancing immediate and exploratory value, with live gains of 1.22% click-through rate and 0.20% page views.

desk verdict Competent industrial re-ranking paper with a sensible exploration objective and a real but small A/B win; the novelty is a new combination of known blocks, and the main weaknesses are missing error bars and a hand-wavy variance claim. read the letter →

arxiv 2608.04809 v1 pith:P5KXBD5X submitted 2026-08-05 cs.IR

Binglei Zhao , Xuanhua Yang , Xiwei Zhao , Sulong Xu This is my paper
classification cs.IR
keywords re-rankinggenerativeexploratoryrewardpreferenceoptimizationonlinerecommendatione-commercesequence-leveladaptiveexploration
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

Under a fixed, low-quality upstream supply, re-ranking has limited room to improve if it only optimizes immediate clicks. This paper proposes DEGR, a generator-only re-ranker that explicitly rewards exploratory exposure—sequences that keep the user scrolling and clicking later—alongside immediate engagement. The paper claims that this dual objective converts re-ranking from a passive corrective step into an adaptive bridge across user requests, and reports offline gains on public and production datasets plus online A/B gains of 1.22% UCTR and 0.20% PV. If correct, it means recommendation systems can extract value from weak candidate pools by preserving browsing potential rather than just pushing the least-bad items.

What carries the argument

The exploratory reward model is the central object. It takes a candidate sequence, produces item embeddings through DIN and position-aware self-attention, and emits two scores: item-wise immediate rewards $R_j$ for click/purchase and a sequence-level exploratory reward $R_e$ trained as a binary classifier over five historical labels (Terminate, Invalid Explore, Potential Interest, Instant Match, Deep Explore). These are fused as $R_\theta(\tau,q)=\alpha R_e+\sum_j(1-\alpha)\delta_j R_j$ with $\delta_j=1-\log_{M+1}(j)$, and the 'Potential Interest' training weight is boosted by $w_C=(4-3\log_2(1+\max_i pCTR_i))w'_C$ when upstream supply is weak. The generator's 'dual exploration' is the companion mechanism: supervised cross-entropy keeps it near the online distribution, an intra-cohort diversity constraint prevents parallel decoding heads from collapsing to similar items, and AR-ORPO uses reward-scaled odds-ratio preference terms over trajectories sampled by group beam search and heuristic sampling. Together these components let the generator output one sequence that balances immediate and exploratory value.

What would settle it

Run a controlled A/B test that fixes each request's candidate supply quality while randomizing the presentation order of DEGR's top-scoring exploratory sequences; if next-request click and next-exposure ratios do not increase relative to a position-randomized baseline, the exploratory reward is encoding position bias rather than real long-term interest.

Watch

Extended reading notes

Core claim

The paper's central claim is that re-ranking is better formulated as maximizing a mixture of immediate utility and exploratory value, with the mixture controlled by upstream supply quality: when the maximum predicted CTR in a request is low, the re-ranker should favor sequences that keep the user browsing instead of assembling the least-bad clickable order. The authors define an exploratory reward model that labels historical exposure sequences by whether the user terminated, scrolled, clicked, or clicked after scrolling, and dynamically up-weights the 'Potential Interest' class when Max-pCTR is low. They then train an encoder-decoder generator with a hybrid objective that combines supervised sequence distribution matching, an intra-cohort diversity penalty, and Adaptive Reward-weighted ORPO, in which rival sampled sequences are ordered by exploratory reward and the preference loss is scaled by softmax reward weights. The claimed result is that the trained generator behaves as a cross-request contextual bridge: offline it outperforms one-stage, two-stage, and generator-only baselines on public and production data, and online it gains 1.22% UCTR and 0.20% PV while adding about 3.2 ms to 99th-percentile serving latency.

Load-bearing premise

The load-bearing premise is that the five-tier exploratory labels derived from clicks, scrolls, and follow-up clicks in historical exposure logs capture true latent user value, so boosting 'Potential Interest' exposures will improve later engagement rather than merely optimizing artifacts of the logged ordering.

Editorial extensions

If this is right

  • A single trained generator can replace the generate-then-evaluate pipeline at serving time, with only a small latency increase, because the exploratory reward is distilled into the generator during training.
  • When upstream supply is poor (low Max-pCTR), boosting exploratory exposure is expected to raise the next-exposure ratio and the next-click ratio, turning a weak request into a bridge to future requests.
  • Adaptive reward-weighted ORPO should separate high-reward from low-reward sequences more cleanly than plain ORPO, because the softmax reward weights suppress gradient noise from irrelevant trajectories.
  • Removing the diversity constraint should hurt online performance more as the upstream candidate pool grows, since parallel decoding heads have more room to produce redundant items.

Reading between the lines

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

  • As an editorial extension, the five-tier label scheme could be reused as a feedback signal for upstream ranking or candidate generation, not just re-ranking, since it measures whether a request's supply preserved browsing potential.
  • Because the labels are derived from historical exposure logs, the method's gains may partly reflect position bias; a position-bias-corrected label variant is a natural next experiment the paper does not report.
  • The reported gain asymmetry—the reward model alone improves little while the generator improves more—suggests that much of the value comes from the generator's wider sequence-space exploration; a simpler deterministic explorer with the same reward model could isolate that contribution.
  • The same balancing principle should transfer to short-video feeds where scrolls are cheap and clicks are sparse, but the label weights and the Max-pCTR threshold would need re-estimation on that traffic.
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

4 major / 4 minor

Summary. The paper proposes DEGR, a generative re-ranking method for e-commerce recommendation. It trains an encoder-decoder generator under a hybrid supervised-reinforcement objective, guided by a learned exploratory reward model. The reward model scores sequences by combining item-wise immediate reward with a sequence-level exploratory reward; the exploratory labels are derived from historical click, scroll, and follow-up click behaviors, and an adaptive weight based on upstream Max-pCTR is used to emphasize exploration when upstream supply quality is low. The generator is trained with a cross-entropy term, an intra-cohort diversity regularization, and an adaptive reward-weighted ORPO (AR-ORPO) preference loss. Evaluations are reported on the Taobao dataset, a JD production dataset, and a 7-day JD online A/B test, with claimed improvements of up to 1.22% UCTR and 0.20% PV.

Significance. If the results hold, DEGR provides a practical demonstration that re-ranking can improve downstream browse and click behavior under constrained upstream supply by explicitly modeling exploratory value. The paper has several strengths: a clear architectural description, component-level ablations, complexity and latency analysis, and, most importantly, an online A/B test in a large deployed system, which is stronger evidence than offline metrics alone. The exploratory reward design and the adaptive weighting mechanism are also clearly tied to the paper's motivating problem. The significance is moderated, however, by the absence of uncertainty quantification, by the reliance on a proxy reward trained from logged behavior that may encode position bias, and by an unproven variance-reduction claim in the appendix.

major comments (4)
  1. [§5.2, §5.3, Tables 2–3] The offline comparisons and ablations are reported without error bars, confidence intervals, or significance tests, and the absolute differences over the strongest baselines are small (e.g., Table 2: Taobao GAUC 0.6107 vs. 0.6037 for GReF; JD GAUC 0.6486 vs. 0.6403 for GReF). The online A/B test is the most compelling evidence, but the description in §5.3 does not state the traffic split, the number of users, the test statistic, or whether the reported gains are statistically significant. Please add repeated-seed results for offline tables and standard A/B reporting details for the online experiment; without these, the claim that DEGR 'outperforms SOTA' is not fully supported by the numbers as presented.
  2. [§4.1.2, Table 1, Eq. (4)] The exploratory reward model is trained on labels derived from historical exposure logs and follow-up clicks, so it may be confounded by position bias and by the behavior of the legacy re-ranking policy. In particular, Category C ('Potential Interest': no click, scroll, follow-up click) is upweighted when Max-pCTR is low, and the follow-up click is not causally attributed to the current request's exposure sequence. This creates a risk that the reward model rewards 'scroll-bait' sequences that reproduce legacy placement patterns rather than generating genuine cross-request exploration value. The online A/B test provides partial evidence that the learned policy transfers, but it does not isolate whether the gains come from the exploratory mechanism or from the proxy's confounds. Please add an analysis that addresses this, for example a position-bias-corrected reward label, a counterfactual or interleaving evaluation, or a comparison of the learned reward against long-term user value.
  3. [§4.2.2, Eq. (10), §A.2] Equation (10) as written is inconsistent with the gradient derivation in Appendix A.2. Equation (10) shows the argument of the sigmoid as log odds divided by a sum of raw odds: log σ(log odds(τ_i) / Σ_{j>i} odds(τ_j)). The appendix, however, derives gradients for g(τ_i,τ_j) = odds(τ_i) / Σ_{j>i} odds(τ_j), which is an odds ratio. These two expressions are not the same quantity. Please rewrite Equation (10) with explicit and consistent notation so that the exact AR-ORPO objective is unambiguous.
  4. [§A.2] The statement that 'assuming i.i.d. logit noise, Var(∇L_A) ≈ 1/(S−1) Var(∇L_ORPO)' is asserted without derivation. The gradient expressions are derived, but the variance approximation is not; it also relies on an independence assumption that is not established for the sampled trajectories, which share prefixes within group beam search. If this claim is retained, it needs a proper derivation and a statement of the assumptions under which it holds. If it is not essential, the sentence should be removed or weakened, because as written it makes a quantitative theoretical claim that the paper does not support.
minor comments (4)
  1. [§5.3, Figure 5] The online analysis in Figure 5 reports relative improvements over 'base PRM', but the production base system is not precisely defined in §5.3; please state clearly which baseline served the control traffic and how the Max-pCTR groups were constructed.
  2. [§5.1.2] The reward-model evaluation protocol generates only 16 sequences via heuristic sampling, which the paper itself notes limits coverage and diversity. This is acknowledged, but the limitation should be stated before the reward-model ablation results in §5.4, since it explains why ER and ER(w/o E) differ so little.
  3. [Table 3] The ablation labels 'DEGR(w/o E)', 'DEGR(w/ GBS)', 'DEGR(w/ ORPO)', and 'DEGR(w/o EDC)' are not all defined at first use in the table; please add a sentence in §5.4 or a table note explaining each variant's exact configuration.
  4. [§5.3] The acronyms UCTR and PV are defined, but 'PV' as 'exposure PV' is unusual; please clarify whether PV is page view count or exposure count per user, and whether the reported +0.20% is relative or absolute.

Circularity Check

2 steps flagged · score 4.0 of 10

Partial circularity: NCR/NER 'bridge' evidence restates the reward labels and Eq. 4 design, though the online UCTR/PV A/B test is external.

  1. fitted input called prediction [Section 4.1.2 (Table 1, Eq. 3, Eq. 5) vs. Section 5.3 (Figure 5)]
    "We define Max-pCTR as the maximum predicted CTR (predicted by upstream ranking) within the sequence, sCTR as the sequence-level click-through rate (the probability of at least one item receiving a click in sequence), Next Click Ratio (NCR) as the probability of producing a click after the current request, and Next Expo Ratio (NER) as the probability of continuing to browse after the current request. The subsequent improvement in NCR and NER demonstrates that our re-ranking can serve as an adaptive cross-request bridge."

    This 'demonstration' is not independent of the training signal: Table 1 defines the sequence-level label y_e from 'current clicks and scrolls, and subsequent click actions', with Category C = no click + scroll + follow-up click labeled 1. Hence R_e in Eq. 3 is a predictor of NCR-like next-click behavior (and B/D/E cover scroll and click continuation). The generator is trained by AR-ORPO (Eq. 10) to maximize R_theta = α R_e + Σ (1-α) δ_j R_j (Eq. 5). The NCR/NER gains in Figure 5 are therefore the same future-action quantities encoded in the reward labels, restated as post-hoc evidence of a 'cross-request bridge' rather than an emergent discovery. The online UCTR/PV A/B test is genuinely external and keeps this from being fully circular.

  2. self definitional [Section 4.1.2, Eq. 4 vs. Section 5.3, Figure 5]
    "It increases positive weights as Max-pCTR decreases, prioritizing exploratory exposure. Since category B samples have exploratory value under supply-constrained conditions, we mask sample B when Max-pCTR≤ 0.01 during training to eliminate negative penalties, thereby promoting scroll and follow-up action implicitly."

    The paper's Figure 5 conclusion, 'Under a low Max-pCTR caused by upstream constraints, DEGR enhances exploratory exposure (reflected in the improvement of NER)', is a direct consequence of Eq. 4: the loss explicitly up-weights Category C and masks Category B in exactly the low-Max-pCTR regime. The adaptive balancing behavior is inserted into the training loss by definition and then reported as evidence that DEGR 'adaptively balances' exploratory value under low-quality supply. This is reading the design choice back from the output, not an independent confirmation; it does not by itself establish that the exploratory reward is the right proxy, which is why the online A/B is the decisive evidence.

full rationale

The paper's online A/B test (1.22% UCTR, 0.20% PV) is an external, non-circular benchmark: live-traffic metrics were not used to fit the reward model or generator, and they test the whole system against the production baseline. Neither the reward-model labels nor the generator training objectives contain UCTR/PV directly, so the headline online result is independent. The partial circularity is confined to the mechanistic narrative around 'cross-request contextual bridging.' The reward model's sequence-level label is defined from follow-up clicks and scrolls (Table 1), i.e., from the same NCR/NER quantities that Section 5.3 later presents as evidence that the bridge works; and Eq. 4 explicitly boosts Category C when Max-pCTR is low, so the observed low-supply NER improvement is a read-back of the loss design. The paper's own caveat in Section 5.1.2 ('offline metrics do not guarantee online alignment') is honest and tempers the concern, but it does not remove the internal self-confirmation in the NCR/NER analysis. Self-citations (e.g., ref. [34]) are not load-bearing; no uniqueness theorem is imported. Overall, the central online claim has independent content, but the 'exploratory bridge' evidence is partially circular, so a score of 4 is appropriate.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method is built on a learned exploratory reward whose labels and weights are chosen by the authors, plus several tuned hyperparameters. None of these choices is validated against an external benchmark, and the only independent check is a proprietary online A/B test.

free parameters (6)
  • alpha (Eq.5) = 0.2
    Weight balancing sequence-level exploratory reward and item-wise immediate rewards in Eq.5; set by hand without sensitivity analysis.
  • beta and gamma (Eq.11) = beta=2.0, gamma=0.01
    Loss weights for intra-cohort regularization and AR-ORPO; chosen by experiments.
  • temperature t (Eq.10) = 0.2
    Softmax temperature for reward soft weights; decided by experiments per Section 5.1.4.
  • quantile count S = 3
    Number of quantiles in AR-ORPO; selected by experiments per Figure 7.
  • label weights and Eq.4 constants = w_E > w_D, w_C rescaling constants 4 and 3
    Hand-designed weights for label categories A-E and adaptive boosting formula; no principled derivation.
  • sample B masking threshold = 0.01 Max-pCTR
    Threshold below which invalid-exploration samples are masked; chosen heuristically.
assumptions (4)
  • domain assumption Logged future actions (click, scroll, follow-up click) are valid labels for exploratory value.
    Section 4.1.2, Table 1 defines the reward labels from future behavior; if these actions are driven by position bias or the old policy, the reward is biased.
  • domain assumption Upstream pCTR is a reliable indicator of supply quality and a suitable dynamic weight.
    Eq.4 and Eq.5 use Max-pCTR to rescale exploratory weights; the paper's Figure 4 evidence is correlational and aggregated.
  • ad hoc to paper i.i.d. logit noise assumption used for the variance reduction claim.
    Appendix A.2 derives Var(∇L_A)≈1/(S-1)Var(∇L_ORPO) assuming i.i.d. logit noise; no justification or empirical check is provided.
  • domain assumption Exposure logs are usable for reward training despite known exposure bias.
    Section 5.1.2 acknowledges exposure bias but the reward model is trained on the same biased logs; no debiasing is applied.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DEGR: Dual Exploration-Driven Generative Re-Ranking for Adaptive Cross-Request Context Bridging." pith.science (2026). https://pith.science/paper/P5KXBD5X

@misc{pith2026260804809,
  author       = {Pith},
  title        = {Pith review of: DEGR: Dual Exploration-Driven Generative Re-Ranking for Adaptive Cross-Request Context Bridging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P5KXBD5X}},
  note         = {Machine review of arXiv:2608.04809}
}
read the original abstract

In industrial recommendation systems, the re-ranking stage balances business objectives and diversity for sequence-level optimization while modeling contextual information. However, constrained by fixed upstream supply, existing methods fail to deliver further effectiveness gains, especially under low-quality supply. To overcome this, re-ranking can actively balance immediate and exploratory value, for instance, by prioritizing exploratory exposure under low-quality supply to preserve browsing potential and facilitate serendipitous conversions. Therefore, we propose a Dual Exploration-Driven Generative Re-Ranking (DEGR) method. DEGR adopts a hybrid supervised-reinforcement exploration and optimization paradigm, guided by an exploratory reward model that adaptively balances immediate and exploratory value. The hybrid optimization paradigm integrates three key components: supervised learning, exploration diversity constraint, and adaptive reward-weighted ORPO for preference optimization. Through this dual exploration, the generator ultimately acts as an adaptive cross-request contextual bridge. Offline and online experiments indicate that DEGR outperforms SOTA methods, achieving improvements of up to 1.22% UCTR and 0.20% PV in the JD E-commerce recommendation system.

Figures

Figures reproduced from arXiv: 2608.04809 by the authors.

Figure 1
Figure 1. An example of different re-ranking sequences under [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Offline training and online serving processes of our [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Architecture of Dual Exploratory-Driven Generative Re-ranking. (A) The encoder-decoder generator is trained using a [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Relationships between Max-pCTR and sCTR, Next [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Figure5. A comprehensive evaluation shows that DEGR achieves [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Quantile-Metric Relationships in AR-ORPO. Larger [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 7 canonical work pages

  1. [1]

    Bruce Croft

    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(Ann Arbor, MI, USA)(SIGIR ’18). Association for Computing Machinery, New York, NY, USA, 135–144. https://doi.org/10.1145/3209978.3209985

    arXiv 2018

  2. [2]

    Irwan Bello, Sayali Kulkarni, Sagar Jain, Craig Boutilier, Ed Chi, Elad Eban, Xiyang Luo, Alan Mackey, and Ofer Meshi. 2019. Seq2Slate: Re-ranking and Slate Optimization with RNNs. arXiv:1810.02019 [cs.IR] https://arxiv.org/abs/1810. 02019

    arXiv 2019

  3. [3]

    Lee, Deming Chen, and Tri Dao

    Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D. Lee, Deming Chen, and Tri Dao. 2024. MEDUSA: Simple LLM inference acceleration framework with multiple decoding heads. InProceedings of the 41st International Conference on Machine Learning(Vienna, Austria)(ICML’24). JMLR.org, Article 203, 27 pages

    work page 2024

  4. [4]

    Chi Chen, Hui Chen, Kangzhi Zhao, Junsheng Zhou, Li He, Hongbo Deng, Jian Xu, Bo Zheng, Yong Zhang, and Chunxiao Xing. 2022. EXTR: Click-Through Rate Prediction with Externalities in E-Commerce Sponsored Search. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Washington DC, USA)(KDD ’22). Association for Computing M...

    arXiv 2022

  5. [5]

    Sirui Chen, Yuan Wang, Zijing Wen, Zhiyu Li, Changshuo Zhang, Xiao Zhang, Quan Lin, Cheng Zhu, and Jun Xu. 2023. Controllable Multi-Objective Re-ranking with Policy Hypernetworks. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining(Long Beach, CA, USA)(KDD ’23). Association for Computing Machinery, New York, NY, USA, 385...

    work page arXiv 2023

  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:2502.18965 [cs.IR] https://arxiv.org/abs/2502.18965

    arXiv 2025

  7. [7]

    Rahul Dey and Fathi M. Salem. 2017. Gate-variants of Gated Recurrent Unit (GRU) neural networks. In2017 IEEE 60th International Midwest Symposium on Circuits and Systems (MWSCAS). 1597–1600. https://doi.org/10.1109/MWSCAS. 2017.8053243

    work page arXiv 2017

  8. [8]

    Yufei Feng, Yu Gong, Fei Sun, Junfeng Ge, and Wenwu Ou. 2021. Revisit Rec- ommender System in the Permutation Prospective. arXiv:2102.12057 [cs.IR] https://arxiv.org/abs/2102.12057

    arXiv 2021

Show all 40 references
  1. [9]

    Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Rozière, David Lopez-Paz, and Gabriel Synnaeve. 2024. Better & Faster Large Language Models via Multi-token Prediction. arXiv:2404.19737 [cs.CL] https://arxiv.org/abs/2404.19737

    2024 arXiv

  2. [10]

    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(Atlanta, GA, USA)(CIKM...

    2022

  3. [11]

    Jiwoo Hong, Noah Lee, and James Thorne. 2024. ORPO: Monolithic Preference Optimization without Reference Model. arXiv:2403.07691 [cs.CL] https://arxiv. org/abs/2403.07691

    2024 arXiv

  4. [12]

    Lillicrap, Jonathan J

    Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Manfred Otto Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. 2015. Continuous control with deep reinforcement learning.CoRRabs/1509.02971 (2015). https: //api.semanticscholar.org/CorpusID:16326763

    2015 arXiv

  5. [13]

    Zhijie Lin, Zhuofeng Li, Chenglei Dai, Wentian Bao, Shuai Lin, Enyun Yu, Haoxi- ang Zhang, and Liang Zhao. 2025. GReF: A Unified Generative Framework for Efficient Reranking via Ordered Multi-token Prediction. arXiv:2510.25220 [cs.IR] https://arxiv.org/abs/2510.25220

    2025

  6. [14]

    Shuchang Liu, Qingpeng Cai, Zhankui He, Bowen Sun, Julian McAuley, Dong Zheng, Peng Jiang, and Kun Gai. 2023. Generative Flow Network for List- wise Recommendation. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining(Long Beach, CA, USA)(KDD ...

    2023

  7. [15]

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H. Chi. 2018. Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture- of-Experts. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining(London, Unit...

    2018

  8. [16]

    Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. SimPO: Simple Preference Optimization with a Reference-Free Reward. InAdvances in Neural Information Processing Systems (NeurIPS)

    2024

  9. [17]

    Changhua Pei, Yi Zhang, Yongfeng Zhang, Fei Sun, Xiao Lin, Hanxiao Sun, Jian Wu, Peng Jiang, Junfeng Ge, Wenwu Ou, and Dan Pei. 2019. Personalized re-ranking for recommendation. InProceedings of the 13th ACM Conference on Recommender Systems(Copenhagen, Denmark)(RecSys ’19). A...

    2019

  10. [18]

    Manning, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. arXiv:2305.18290 [cs.LG] https://arxiv.org/ abs/2305.18290

    2024 arXiv

  11. [19]

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

  12. [20]

    John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel

  13. [21]

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

  14. [22]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeek- Math: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300 [cs.CL] https://arxiv.org/abs/2402.03300

    2024 arXiv

  15. [23]

    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 ...

    2023

  16. [24]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. InProceedings of the 31st International Conference on Neural Information Processing Systems(Long Beach, California, US...

    2017

  17. [25]

    Vijayakumar, Michael Cogswell, Ramprasaath R

    Ashwin K. Vijayakumar, Michael Cogswell, Ramprasaath R. Selvaraju, Qing Sun, Stefan Lee, David J. Crandall, and Dhruv Batra. 2016. Diverse Beam Search: Decoding Diverse Solutions from Neural Sequence Models.CoRRabs/1610.02424 (2016). arXiv:1610.02424 http://arxiv.org/abs/1610.02424

    2016 arXiv

  18. [26]

    Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. 2017. Pointer Networks. arXiv:1506.03134 [stat.ML] https://arxiv.org/abs/1506.03134

    2017 arXiv

  19. [27]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & Cross Network for Ad Click Predictions. InProceedings of the ADKDD’17(Halifax, NS, Canada) (ADKDD’17). Association for Computing Machinery, New York, NY, USA, Article 12, 7 pages. https://doi.org/10.1145/3124749.3124754

    2017

  20. [29]

    Zhe Wang, Liqin Zhao, Biye Jiang, Guorui Zhou, Xiaoqiang Zhu, and Kun Gai

  21. [30]

    Yunjia Xi, Weiwen Liu, Jieming Zhu, Xilong Zhao, Xinyi Dai, Ruiming Tang, Weinan Zhang, Rui Zhang, and Yong Yu. 2022. Multi-Level Interaction Reranking with User Behavior History. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Infor...

    2022

  22. [31]

    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...

    2025

  23. [32]

    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:2508.02242 [cs.IR] https://arxiv.org/abs/ 2508.02242

    2025 arXiv

  24. [33]

    Wenhao Zhang, Yuexiang Xie, Yuchang Sun, Yanxi Chen, Guoyin Wang, Yaliang Li, Bolin Ding, and Jingren Zhou. 2025. On-Policy RL Meets Off-Policy Experts: Harmonizing Supervised Fine-Tuning and Reinforcement Learning via Dynamic Weighting. arXiv:2508.11408 [cs.LG] https://arxiv....

    2025

  25. [34]

    Binglei Zhao, Houying Qi, Guang Xu, Mian Ma, Xiwei Zhao, Feng Mei, Sulong Xu, and Jinghe Hu. 2025. A Hybrid Cross-Stage Coordination Pre-ranking Model for Online Recommendation Systems. InCompanion Proceedings of the ACM on Web Conference 2025(Sydney NSW, Australia)(WWW ’25). ...

    2025

  26. [35]

    Guorui Zhou, Hengrui Hu, Hongtao Cheng, Huanjie Wang, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Lu Ren, Liao Yu, Pengfei Zheng, Qiang Luo, Qianqian Wang, Qigen Hu, Rui Huang, Ruiming Tang, Shiyao Wang, Shujie Yang, Tao Wu, Wuchao Li, Xinchen Luo, Xingmei Wang, Yi Su, Yu...

    2025 arXiv

  27. [36]

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click- Through Rate Prediction. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining(...

    2018

  28. [37]

    Tao Zhuang, Wenwu Ou, and Zhirong Wang. 2018. Globally optimized mutual influence aware ranking in e-commerce search. InProceedings of the 27th Interna- tional Joint Conference on Artificial Intelligence(Stockholm, Sweden)(IJCAI’18). AAAI Press, 3725–3731. A Appendix A.1 AR-OR...

    2018

  29. [2015]

    High-Dimensional Continuous Control Using Generalized Advantage Estimation.Computer ence(2015)

    2015

  30. [2017]

    https://api.semanticscholar.org/CorpusID:28695052 KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea

    Proximal Policy Optimization Algorithms.ArXivabs/1707.06347 (2017). https://api.semanticscholar.org/CorpusID:28695052 KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Binglei Zhao, Xuanhua Yang, Xiwei Zhao, and Sulong Xu

    2017 arXiv

  31. [2020]

    https://api.semanticscholar.org/CorpusID:220919785

    COLD: Towards the Next Generation of Pre-Ranking System.ArXiv abs/2007.16122 (2020). https://api.semanticscholar.org/CorpusID:220919785

    2020 arXiv

  32. [2024]

    InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining(Barcelona, Spain)(KDD ’24)

    Non-autoregressive Generative Models for Reranking Recommendation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining(Barcelona, Spain)(KDD ’24). Association for Computing Machinery, New York, NY, USA, 5625–5634. https://doi.org/10.1145/3637...

Pith tools

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