Pith. sign in

REVIEW 4 major objections 5 minor 32 references

NGA: Non-autoregressive Generative Auction with Global Externalities for Advertising Systems

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read NGA, a non-autoregressive generative auction, models global externalities from ads and adjacent organic content to outperform CGA on revenue, click-through, conversion, and inference latency.

desk verdict A real engineering extension with an online A/B win, but the offline metrics are self-referential and the IC claim is asserted rather than established. read the letter →

arxiv 2506.05685 v1 pith:UOUZBZHB submitted 2025-06-06 cs.IR

Zuowu Zheng , Ze Wang , Fan Yang , Wenqing Ye , Weihua Huang , Wenqiang He , Teng Zhang , Xingxing Wang This is my paper · ORCID
classification cs.IR
keywords generativeauctionnon-autoregressivedecodingadexternalitiesorganiccontentincentivecompatibilityonlineadvertisingmechanismdesign
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

This paper argues that online ad auctions should treat the whole page—ads plus neighboring organic items—as the unit of allocation, and claims a new framework called NGA can do that without the latency cost of sequential generation. Existing generative auctions such as CGA model dependencies among ads but ignore organic content and generate slots one at a time. NGA instead scores all candidate items against all slots in one non-autoregressive pass, then uses a parallel multi-tower evaluator to score whole lists for reward and payment. The paper reports that, on a commercial location-based services platform, NGA beats CGA on offline revenue per mille (226.47 vs 221.49), CTR, CVR, and incentive-compatibility regret, and in a 2% traffic online A/B test it raises RPM by 1.9% while cutting response time by 6.5%. If these results hold, generative auctions can incorporate global page context and still serve in real time.

What carries the argument

The mechanism has three parts. First, a dual-encoder Generator: an item encoder turns each candidate ad or organic item into an embedding, a position encoder attends over those item embeddings to produce slot representations, and the two are combined into an allocation probability matrix over candidates and slots. Second, per-position constrained decoding: for each slot, the generator picks the highest-probability candidate that satisfies business constraints such as minimum ad position, ad density, and brand de-duplication. Third, a multi-tower parallel Evaluator: a list encoder embeds each candidate list, and three parallel towers predict list-wise pCTR, pCVR, and payment ratios (with payment equal to bid times the payment ratio), combined into a reward $R = \sum_i (\Theta_i \, \mathrm{bid}_i \, p_i + \alpha \, \Theta_i \, \Gamma_i)$. The generator is trained by policy gradient with reward defined as each item's marginal contribution to the evaluator's list reward.

What would settle it

Re-train the evaluator on allocations produced by NGA (or otherwise calibrate it on NGA-generated lists) and recompute the offline RPM, CTR, CVR, and IC regret; if the reported gains over CGA shrink or disappear, the improvements are an artifact of evaluator miscalibration rather than of the auction mechanism. A complementary test is an online A/B at higher traffic share (e.g., 10–20% instead of 2%) with a held-out revenue measurement.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a generative auction for ads can be made both more context-aware and faster by replacing autoregressive, position-by-position generation with a non-autoregressive allocation over a joint item-position probability matrix, and by extending the modeled context from the ad list to the surrounding organic content. The generator produces a probability matrix $z_{ij} = \mathrm{softmax}(\mathrm{ctr}_i \times \mathrm{bid}_i + \alpha \times \mathrm{ctr}_i \times \mathrm{cvr}_i + \mathbf{x}_i^\top \mathbf{t}_j)$, decodes each position by constrained argmax, and an evaluator with three parallel towers predicts list-wise pCTR, pCVR, and payment ratios for whole candidate lists. A policy-gradient step, using each item's marginal contribution to the evaluator's reward as the advantage, trains the generator. The claims are supported by offline experiments on 225 million page-views and a 2% traffic online A/B test, where NGA outperforms uGSP, DNA, and CGA on RPM, CTR, CVR, and IC regret, and runs faster than CGA.

Load-bearing premise

The offline evaluation trusts the evaluator's predicted click, conversion, and payment ratios, trained from logs of the current auction, to stay correct for the new allocation lists NGA generates, even though the generator is optimized against that same evaluator.

Editorial extensions

If this is right

  • Ad-auction systems can incorporate organic feed context directly into the allocation objective without adding latency, since whole-list scoring and constraint decoding replace sequential generation.
  • Non-autoregressive decoding removes a scaling bottleneck: candidate counts and slot counts can grow without multiplying per-position sequential steps.
  • Parallel multi-tower evaluation makes list-wise payment computation cheap enough for real-time use, shortening response time (reported −6.5% in the online test).
  • Lower IC regret (Ψ = 2.9% vs 3.8% for CGA) suggests the mechanism preserves truth-telling incentives better than the autoregressive baseline.

Reading between the lines

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

  • The offline results are computed from the same evaluator that the generator is trained to maximize; fully convincing evidence would require counterfactual or calibrated evaluation of NGA's novel allocations.
  • Since the paper assumes organic ranking is fixed upstream, a natural extension is joint optimization of ad and organic placement; treating organic order as controllable might yield larger lifts than those reported.
  • The non-autoregressive decoder opens the door to per-request candidate sets much larger than 30 ads, which could change how auction frameworks are deployed in feeds.
  • The IC guarantee is heuristic: it comes from a Lagrangian penalty on empirical ex-post regret, not from a closed-form truthful mechanism, so incentives may degrade under distribution shift or strategic advertiser behavior.
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 / 5 minor

Summary. This paper introduces NGA, a non-autoregressive generative auction framework for online advertising. The framework combines a transformer-based generator that produces k-slot allocations from n ads and m organic items via per-position constrained decoding, and a multi-tower evaluator that predicts list-wise pCTR, pCVR, and payment ratios for candidate lists. The evaluator is trained with click, conversion, and payment losses, while the generator is trained by policy gradient on marginal contributions computed from the evaluator's reward. Offline experiments on 225M page views report RPM 226.47, CTR 5.83%, CVR 5.32%, and IC regret 2.9%, outperforming uGSP, DNA, and CGA; an online A/B test on 2% traffic reports +1.9% RPM, +1.6% CTR, +0.8% CVR, and -6.5% response time relative to a multi-objective CGA. All offline metrics are stated to be computed from reward-model predictions.

Significance. At face value, NGA addresses a real industrial deployment problem and demonstrates an online A/B test, which is uncommon in auction-mechanism papers. The non-autoregressive parallel design and the explicit inclusion of adjacent organic content are substantive extensions over CGA. If the reported results are trustworthy, the work would be useful to practitioners. However, the offline evidence is partly self-referential because the same evaluator supplies both the training reward and the evaluation labels; the online test lacks statistical detail; and the incentive-compatibility claim rests on an undefined ex-post regret term. These issues currently prevent the paper from providing a convincing validation of its central claim.

major comments (4)
  1. [§4.1.2, §3.3.2] Section 4.1.2 states that all offline metrics are 'computed based on predictions from the reward model,' and Section 3.3.2 trains the generator to maximize exactly this reward model via Eqs. (17) and (18). Table 1 therefore compares allocations using a scorer that NGA has been optimized against, and NGA's allocation lists are out-of-distribution relative to the logged data that trained the evaluator. No calibration, inverse-propensity reweighting, or holdout validation on NGA-generated lists is provided, so the offline RPM/CTR/CVR gains could be artifacts of evaluator miscalibration rather than true improvements. Please add such validation or explicitly relegate offline numbers to diagnostic status.
  2. [§3.3.1, Eq. (15), §4.1.2] The payment loss in Eq. (15) uses crgt_i ('ex-post regret') without any definition, and the Lagrange penalty with lambda_i and rho is asserted to improve incentive compatibility without a derivation or finite-sample guarantee. The IC metric Psi in Section 4.1.2 is also defined in terms of the same undefined crgt_i, so the claimed IC regret of 2.9% in Table 1 is not verifiable. Please define crgt_i operationally and report truth-telling experiments (e.g., comparing advertiser utility under true versus manipulated bids) rather than relying on an unformalized penalty term.
  3. [§4.3, Table 2] Table 2 reports relative improvements of +1.9% RPM, +1.6% CTR, +0.8% CVR, and -6.5% RT from an online A/B test on 2% production traffic, but no confidence intervals, significance tests, experiment duration, or number of impressions are provided. Moreover, the baseline is described as a 'multi-objective version of CGA' that is not the CGA used in the offline comparison in Table 1, so the online and offline evaluations are not aligned. Please supply the missing statistical details and a precise description of the baseline; without these, the online result is suggestive but not conclusive.
  4. [§3.2.2, Eq. (6), Eq. (12)] The formulation of global externalities is incomplete for mixed ad/organic lists. Equations (6) and (12) use ctr_i, bid_i, and p_i for every item, but organic items do not have advertiser bids, and the manuscript never specifies how their presence enters the allocation probability, the reward, or the payment computation. As a result, the claimed joint modeling of ads and adjacent organic content is not formally defined. Please specify a separate treatment for organic items or clarify that they are used only as context features.
minor comments (5)
  1. [§2, Eq. (3)] The notation E_{a in A}[...] is confusing because A was previously introduced as an allocation rule; please use explicit random-variable notation for allocations.
  2. [§3.1.3, Eq. (6)] The softmax in Eq. (6) normalizes over candidate items for each position, but the paper does not state this explicitly; please also clarify whether an item can be selected at multiple positions before the constrained decoding is applied.
  3. [§4.2] There is a typo: 'golbal' should be 'global.' Also, the statement that the CVR lift is '6.4% higher' should be reconciled with the table's definition of lift percentage to avoid ambiguity.
  4. [§4.3] The claim that RT 'remains well within practical limits' is unsupported because no absolute latency is given; please report milliseconds or a similar concrete measure.
  5. [Header and metadata] The ACM reference format lists 'Woodstock ’18' and 2018, which appears to be a template artifact; please use the correct venue and year.

Circularity Check

1 steps flagged · score 6.0 of 10

Offline gains are measured by the same evaluator the generator is trained to maximize; online A/B provides only partial independent support.

  1. fitted input called prediction [Section 3.3.2 (Eqs. 17–18) and Section 4.1.2 (Evaluation Metrics)]
    "After convergence of the Evaluator, we adopt a policy gradient based method to train the Generator. ... r_yi = R_Y − R_Y−i ... LG = −1/|D| Σ Σ r_yi log Z_yi ... For offline experiments, all metrics are computed based on predictions from the reward model, and for commercial confidentiality reasons, the absolute value has been converted."

    The generator's policy-gradient objective (Eq. 18) maximizes marginal contributions defined by the evaluator reward R (Eq. 12), which aggregates the evaluator's predicted pCTR Θ, pCVR Γ, and payment ratio p. The paper then reports all offline metrics as 'computed based on predictions from the reward model.' Hence the offline RPM, CTR, CVR, and IC-regret improvements are not measured against independent user or advertiser behavior on NGA allocations; they are evaluations by the exact surrogate model NGA was trained to maximize. The evaluator was fitted to logged traffic from the existing auction, so NGA-generated lists are out-of-distribution, and no calibration, inverse-propensity, or counterfactual validation is provided.

full rationale

The paper's central weakness is not self-citation but the offline evaluation loop. The generator is optimized, via Eqs. 17–18, to increase the evaluator's reward R from Eq. 12, and then Section 4.1.2 states that all offline metrics are computed from that same reward model's predictions. This makes the headline offline improvements over CGA and DNA partially artifacts of the training objective rather than independent measurements. The online A/B test on 2% of traffic is an external benchmark and materially reduces the circularity burden, but it reports only relative lifts with no statistical detail and compares against a business-configured multi-objective CGA rather than the exact offline baseline. Self-citations in the references are present but not load-bearing for the main claims, so they do not raise the score. On balance, this is a partial circularity: the offline predictions reduce by construction to the fitted evaluator, while the online test provides independent but under-specified support.

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

The central claim rests mainly on domain assumptions about the auction environment and on the ad hoc design of the reward model and payment rule. The fixed organic order limits the claimed 'global externalities', and the IC guarantee is only an approximation. No new physical or conceptual entities are introduced.

free parameters (5)
  • alpha = 5
    Hand-chosen balance coefficient between revenue and order volume in Eq (3) and Eq (6); affects the generator's allocation scores and is not tuned on a holdout set.
  • rho
    Hyperparameter for the IC penalty term in Eq (15); value is not reported.
  • w1, w2, w3
    Weights for the evaluator loss in Eq (16); values are not reported.
  • lambda_i
    Lagrange multipliers for the IC constraints in Eq (15); learned during training but values are not reported.
  • transformer layers L
    Number of layers in the item, position, and list encoders; not specified in the implementation details.
assumptions (6)
  • domain assumption The organic sequence is pre-ranked by an upstream module and its internal order remains fixed.
    Footnote 1 in Section 2; the paper models organic content as fixed context, so 'global externalities' only captures organic effects on ads, not joint optimization of organic placement.
  • domain assumption Advertisers have private click values and submit bids, with utility (v_i - p_i) * ctr_i.
    Standard auction model assumed in Eq (1); underlies the IC definition and the regret-based training objective.
  • ad hoc to paper The evaluator's predicted pCTR, pCVR, and payment ratios generalize to counterfactual allocation lists.
    Sections 3.2.2 and 4.1.2; offline metrics and generator training rely on this generalization, and the paper gives no calibration or out-of-distribution validation.
  • ad hoc to paper The allocation probability in Eq (6), a softmax over an additive score, is an appropriate generator model.
    Section 3.1.3; a design choice not derived from auction theory or from the stated objective.
  • ad hoc to paper The payment rule pay_i = bid_i * p_i with a learned ratio p_i approximates incentive compatibility when trained with a Lagrange penalty.
    Section 3.3.1 Eq (15); the paper provides no formal IC guarantee, only a heuristic penalty on an undefined ex-post regret.
  • standard math The policy gradient estimator in Eq (18) is unbiased under the standard REINFORCE assumptions.
    Section 3.3.2; the generator update uses a standard REINFORCE-style gradient with the reward model held fixed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NGA: Non-autoregressive Generative Auction with Global Externalities for Advertising Systems." pith.science (2026). https://pith.science/paper/UOUZBZHB

@misc{pith2026250605685,
  author       = {Pith},
  title        = {Pith review of: NGA: Non-autoregressive Generative Auction with Global Externalities for Advertising Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UOUZBZHB}},
  note         = {Machine review of arXiv:2506.05685}
}
read the original abstract

Online advertising auctions are fundamental to internet commerce, demanding solutions that not only maximize revenue but also ensure incentive compatibility, high-quality user experience, and real-time efficiency. While recent learning-based auction frameworks have improved context modeling by capturing intra-list dependencies among ads, they remain limited in addressing global externalities and often suffer from inefficiencies caused by sequential processing. In this work, we introduce the Non-autoregressive Generative Auction with global externalities (NGA), a novel end-to-end framework designed for industrial online advertising. NGA explicitly models global externalities by jointly capturing the relationships among ads as well as the effects of adjacent organic content. To further enhance efficiency, NGA utilizes a non-autoregressive, constraint-based decoding strategy and a parallel multi-tower evaluator for unified list-wise reward and payment computation. Extensive offline experiments and large-scale online A/B testing on commercial advertising platforms demonstrate that NGA consistently outperforms existing methods in both effectiveness and efficiency.

Figures

Figures reproduced from arXiv: 2506.05685 by the authors.

Figure 1
Figure 1. Overview of the NGA architecture. In this section, we introduce the overall framework of the pro￾posed Non-autoregressive Generative Auction with global external￾ities (NGA), as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 21 canonical work pages

  1. [1]

    Yoram Bachrach, Sofia Ceppi, Ian A Kash, Peter Key, and David Kurokawa. 2014. Optimising trade-offs among stakeholders in ad auctions. InProceedings of the fifteenth ACM conference on Economics and computation. 75–92

    work page 2014

  2. [2]

    Yuan Deng, Sébastien Lahaie, Vahab Mirrokni, and Song Zuo. 2020. A data- driven metric of incentive compatibility. InProceedings of The Web Conference

    work page 2020

  3. [3]

    Benjamin Edelman, Michael Ostrovsky, and Michael Schwarz. 2007. Internet advertising and the generalized second-price auction: Selling billions of dollars worth of keywords.American economic review97, 1 (2007), 242–259

    2007

  4. [4]

    Nicola Gatti, Alessandro Lazaric, and Francesco Trovo. 2012. A truthful learning mechanism for contextual multi-slot sponsored search auctions with externalities. InProceedings of the 13th ACM Conference on Electronic Commerce. 605–622

    work page 2012

  5. [5]

    Arpita Ghosh and Mohammad Mahdian. 2008. Externalities in online advertising. InProceedings of the 17th international conference on World Wide Web. 161–168

    work page 2008

  6. [6]

    Patrick Hummel and R Preston McAfee. 2014. Position auctions with externalities. InWeb and Internet Economics: 10th International Conference, WINE 2014, Beijing, China, December 14-17, 2014. Proceedings 10. Springer, 417–422

    work page 2014

  7. [7]

    Bernard J Jansen and Tracy Mullen. 2008. Sponsored search: an overview of the concept, history, and technology.International Journal of Electronic Business6, 2 (2008), 114–131

    work page 2008

  8. [8]

    Diederik P Kingma. 2014. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980(2014)

    arXiv 2014

Show all 32 references
  1. [9]

    Ningyuan Li, Yunxuan Ma, Yang Zhao, Zhijian Duan, Yurong Chen, Zhilin Zhang, Jian Xu, Bo Zheng, and Xiaotie Deng. 2023. Learning-based ad auction design with externalities: the framework and a matching-based approach. InProceedings of the 29th ACM SIGKDD Conference on Knowledg...

    2023

  2. [10]

    Xuejian Li, Ze Wang, Bingqi Zhu, Fei He, Yongkang Wang, and Xingxing Wang

  3. [11]

    Guogang Liao, Xuejian Li, Ze Wang, Fan Yang, Muzhi Guan, Bingqi Zhu, Yongkang Wang, Xingxing Wang, and Dong Wang. 2022. NMA: neural multi-slot auctions with externalities for online advertising.arXiv preprint arXiv:2205.10018 (2022)

    2022 arXiv

  4. [12]

    Guogang Liao, Xiaowen Shi, Ze Wang, Xiaoxu Wu, Chuheng Zhang, Yongkang Wang, Xingxing Wang, and Dong Wang. 2022. Deep page-level interest net- work in reinforcement learning for ads allocation. InProceedings of the 45th International ACM SIGIR Conference on Research and Develo...

    2022

  5. [13]

    Guogang Liao, Ze Wang, Xiaoxu Wu, Xiaowen Shi, Chuheng Zhang, Yongkang Wang, Xingxing Wang, and Dong Wang. 2022. Cross DQN: Cross deep Q network for ads allocation in feed. InProceedings of the ACM Web Conference 2022. 401– 409

    2022

  6. [14]

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

    2021

  7. [15]

    Roger B Myerson. 1981. Optimal auction design.Mathematics of operations research6, 1 (1981), 58–73

    1981

  8. [16]

    Junyan Qiu, Ze Wang, Fan Zhang, Zuowu Zheng, Jile Zhu, Jiangke Fan, Teng Zhang, Haitao Wang, and Xingxing Wang. 2025. One Model to Rank Them All: Unifying Online Advertising with End-to-End Learning.arXiv preprint arXiv:2505.19755(2025)

    2025 arXiv

  9. [17]

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

  10. [18]

    Xiaowen Shi, Ze Wang, Yuanying Cai, Xiaoxu Wu, Fan Yang, Guogang Liao, Yongkang Wang, Xingxing Wang, and Dong Wang. 2023. MDDL: A Framework for Reinforcement Learning-based Position Allocation in Multi-Channel Feed. InProceedings of the 46th International ACM SIGIR Conference ...

    2023

  11. [19]

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

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

  12. [20]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  13. [21]

    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

  14. [22]

    Ze Wang, Guogang Liao, Xiaowen Shi, Xiaoxu Wu, Chuheng Zhang, Yongkang Wang, Xingxing Wang, and Dong Wang. 2022. Learning List-wise Representation in Reinforcement Learning for Ads Allocation with Multiple Auxiliary Tasks. In Proceedings of the 31st ACM International Conferenc...

    2022

  15. [23]

    Yiqing Wang, Xiangyu Liu, Zhenzhe Zheng, Zhilin Zhang, Miao Xu, Chuan Yu, and Fan Wu. 2022. On designing a two-stage auction for online advertising. In Proceedings of the ACM Web Conference 2022. 90–99

    2022

  16. [24]

    Ruobing Xie, Shaoliang Zhang, Rui Wang, Feng Xia, and Leyu Lin. 2021. Hierar- chical reinforcement learning for integrated recommendation. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 4521–4528

    2021

  17. [25]

    Ze Wang, Guogang Liao, Xiaowen Shi, Xiaoxu Wu, Chuheng Zhang, Bingqi Zhu, Yongkang Wang, Xingxing Wang, and Dong Wang. 2022. Hybrid Transfer in Deep Reinforcement Learning for Ads Allocation. InProceedings of the 31st ACM International Conference on Information & Knowledge Man...

    2022

  18. [26]

    Zhanhao Zhang. 2021. A survey of online auction mechanism design using deep learning approaches.arXiv preprint arXiv:2110.06880(2021)

    2021 arXiv

  19. [27]

    Jinyun Yan, Zhiyuan Xu, Birjodh Tiwana, and Shaunak Chatterjee. 2020. Ads allocation in feed via constrained optimization. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 3386– 3394

    2020

  20. [28]

    Xiangyu Zhao, Changsheng Gu, Haoshenglun Zhang, Xiwang Yang, Xiaobing Liu, Jiliang Tang, and Hui Liu. 2021. Dear: Deep reinforcement learning for online advertising impression in recommender systems. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 750–758

    2021

  21. [29]

    Zhilin Zhang, Xiangyu Liu, Zhenzhe Zheng, Chenrui Zhang, Miao Xu, Junwei Pan, Chuan Yu, Fan Wu, Jian Xu, and Kun Gai. 2021. Optimizing multiple performance metrics with deep GSP auctions for e-commerce advertising. InProceedings of the 14th ACM International Conference on Web ...

    2021

  22. [30]

    Ruitao Zhu, Yangsu Liu, Dagui Chen, Zhenjia Ma, Chufeng Shi, Zhenzhe Zheng, Jie Zhang, Jian Xu, Bo Zheng, and Fan Wu. 2024. Contextual Generative Auction with Permutation-level Externalities for Online Advertising.arXiv preprint arXiv:2412.11544(2024)

    2024 arXiv

  23. [31]

    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

  24. [2024]

    InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Deep automated mechanism design for integrating ad auction and allocation in feed. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1211–1220

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.