Pith. sign in

REVIEW 4 major objections 5 minor 11 cited by

EGA-V2: An End-to-end Generative Framework for Industrial Advertising

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 A single generative model can replace the multi-stage cascade of industrial advertising, deciding which ads and creatives to show and what to charge them, all in one pass.

desk verdict A plausible industrial system paper whose architecture is worth discussing, but the central empirical claim is not established because the offline metrics are computed from the same reward model that selects the ads. read the letter →

arxiv 2505.17549 v3 pith:6ROQFBNY submitted 2025-05-23 cs.IR

Zuowu Zheng , Ze Wang , Fan Yang , Jiangke Fan , Teng Zhang , Yongkang Wang , Xingxing Wang This is my paper · ORCID
classification cs.IR
keywords generativeadvertisingend-to-endadauctiontoken-levelbiddingincentivecompatibilityex-postregretRQ-VAEtokenizationmulti-tokenpredictionrecommendation
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 claims that the many disjoint stages of an industrial ad system—recall, ranking, creative selection, auction, and ad placement—can be collapsed into one generative model that directly outputs the final ad list, the creatives, and the payments. The reason this would matter is that cascaded systems permanently discard candidates in early stages and cannot jointly optimize user experience and advertiser constraints. EGA-V2 couples a generative backbone trained on user behavior with a permutation-aware reward model, token-level bidding for allocation, and a learned payment network trained to minimize ex-post regret so that truthful bidding is approximately optimal. The paper reports that on a large location-based-services dataset, this unified model improves revenue (RPM), click-through rates, and advertiser truthfulness over a strong cascaded baseline and a generative-recommendation baseline. The gain, if it holds online, would make fully generative advertising architectures practical.

What carries the argument

The load-bearing machinery is the decoupled allocation-and-payment pipeline. Allocation runs on token-level bids: each token's weight is $w(a_j^i)=[\max(b_1,\dots,b_{N_i})]^\alpha+\beta$, normalized by a softmax over the codebook, and beam search generates candidate sequences that a permutation-aware reward model scores. Payment is computed separately at the POI level by a network $p=\hat{p}\odot b$ with $\hat{p}=\sigma(\mathrm{MLP}(S^*;B_-;Z\cdot\Theta))$, trained under the Lagrangian loss $-\sum_{y_i\in S^*} p_i \hat{r}_i^{\mathrm{pctr}} - \sum_i \lambda_i \mathrm{crgt}_i - \frac{\rho}{2}\sum_i(\mathrm{crgt}_i)^2$, alternating payment-network updates with multiplier updates on empirical ex-post regret. The RQ-VAE tokenizer, multi-token prediction heads, and two-phase training (interest-based pretraining then frozen-backbone auction post-training) supply the generative backbone; the ex-post regret Lagrangian is what carries the incentive-compatibility claim.

What would settle it

Recompute RPM and CTR on the held-out test days using real logged clicks and conversions instead of the reward model's predicted pCTR and pCVR, and compare EGA-V2 to MCA; if the reported advantage shrinks to noise, the paper's central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that allocation and payment can be decoupled inside a single generative advertising model, and that this decoupling preserves approximate incentive compatibility while improving revenue. Items are tokenized with RQ-VAE into hierarchical semantic tokens; a pretrained encoder-decoder with multi-token prediction generates both POI recommendations and creative tokens; a permutation-aware reward model scores full sequences; token-level bidding aggregates advertiser bids over the tokens; and a separate POI-level payment network charges advertisers using a sigmoid payment rate. The payment network is trained with a Lagrangian objective that maximizes expected payment while driving empirical ex-post regret toward zero. The paper states that this yields 'approximate incentive compatibility at the POI level.' In offline experiments, EGA-V2 reports RPM of 230.41, CTR-poi of 0.0612, CTR-img of 0.0583, and IC regret of 2.7%, beating both the multi-stage cascade (MCA) and generative recommendation (GR) baselines on all four metrics.

Load-bearing premise

The weakest load-bearing premise is that the reward model's predicted click and conversion rates faithfully reflect real user behavior, because these predicted values are used both to pick the generated sequences and to compute every offline metric reported in the paper.

Editorial extensions

If this is right

  • If EGA-V2 is right, ad systems no longer need separate recall, ranking, creative-selection, and auction modules; a single model can emit the final list, creatives, and payments.
  • A learned payment network trained with ex-post regret can keep advertisers roughly truthful at the POI level, with IC regret of 2.7% in offline tests versus 8.4% for a GSP-based generative baseline.
  • Token-level bidding gives operators two dials: α controls how strongly bids influence allocation, and β controls the ad-to-organic ratio, allowing a real-time trade-off between revenue and user experience.
  • The multi-phase training recipe—interest-based pretraining followed by frozen-backbone auction post-training—outperforms single-stage end-to-end training in the ablation, with RPM 230.41 versus 218.21.
  • Offline results position the architecture for online A/B testing, which the paper identifies as future work.

Reading between the lines

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

  • Editorial inference: the same allocation-payment decoupling could be tested in generative retrieval systems outside advertising; any generator that emits candidates under a budget could use a separate learned pricing head with an ex-post regret penalty.
  • Editorial inference: the IC guarantee is explicitly at the POI level and approximate, so an adversarial advertiser could still attempt to game the creative token or image branch, since creative-level payments are not directly constrained—this is an untested boundary.
  • Editorial inference: if the reward model is miscalibrated, the paper's own reported gains may be objective-fitting artifacts; a calibration check on out-of-sample real clicks would settle whether the architecture genuinely improves user experience or only the proxy.
  • Editorial inference: the ablation suggests max-aggregation of bids at token level is doing real work for IC; a monotone, differentiable alternative such as log-sum-exp with temperature could give similar revenue with smoother gradients.
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. EGA-V2 is an end-to-end generative framework for industrial advertising that unifies POI recommendation, creative generation, ad allocation, and payment computation in a single encoder-decoder model. The paper uses RQ-VAE tokenization, multi-token prediction, a permutation-aware reward model, token-level bidding, and a decoupled payment network trained with a Lagrangian objective on empirical ex-post regret. The claimed contributions are a multi-phase training strategy (interest-based pre-training followed by auction-based post-training) and approximately incentive-compatible payment at the POI level. Offline experiments on a Meituan industrial dataset report that EGA-V2 outperforms a multi-stage cascading architecture (MCA) and a generative recommendation baseline (GR) in RPM, CTR, and IC regret, with ablations attributing gains to each component.

Significance. If the empirical results hold, the paper would make a meaningful engineering contribution: it demonstrates a concrete way to fold retrieval, ranking, creative selection, allocation, and payment into a single generative model while respecting auction constraints. The strengths are the large-scale industrial dataset (200 million requests), the clear system description, and the ablation study that isolates the MTP module, multi-phase training, max-pooled token-level bidding, and the learned payment network. The main weakness is that all offline metrics are computed with predictions from the same reward model that selects the output sequence, making the evaluation partially self-referential. No online A/B test, calibration analysis, or debiased offline evaluation is provided, and the paper itself notes that online A/B testing is future work. The approximate-incentive-compatibility claim is also presented as a guarantee, although the optimization is non-convex and the regret is measured against the model's own predicted pCTR. The contribution is therefore promising but not yet established as reported.

major comments (4)
  1. [§5.1.2 and Eq. (24)] The evaluation loop closes: Section 5.1.2 states that "For offline experiments, evaluation metrics are computed using the predicted values from the reward model," and Eq. (24) uses the same reward model R to score and select the winning sequence S*. Therefore the RPM, CTR, and IC regret numbers in Table 2 are computed from the very model that the policy optimizes against. If R is miscalibrated or reward-hacked, the reported lifts (e.g., RPM 230.41 vs. 206.73 and IC regret 2.7% vs. 8.4%) can be artifacts of the optimization objective rather than genuine improvements in user behavior or advertiser utility. The paper should add a calibration analysis of R (e.g., reliability diagrams and expected calibration error), evaluate at least one metric from logged real outcomes or an unbiased offline estimator, and ideally report online A/B results; without this, the headline outperformance claim is not supported.
  2. [§4.4.2, Eqs. (31)–(32)] The post-training stage is underspecified with respect to which parameters are actually trained. The text says "Following the pre-training phase, we freeze its parameters and optimize the generative advertising model under auction constraints," but it does not state which parameters the policy-gradient loss LPG in Eq. (32) updates. If the pre-trained backbone is frozen, only the allocation weights in Eq. (22) are trainable, and alpha and beta are listed as hyperparameters; this would make the ``generative allocation training'' hard to interpret. The paper needs to state the exact parameterization of the policy, which parameters are frozen and which are updated in each step, and how gradients flow through Eq. (22).
  3. [§4.4.2, Eq. (33), and abstract] The abstract's claim of "guaranteeing approximate incentive compatibility" is stronger than what the method supports. The Lagrangian objective in Eq. (33) minimizes a penalty on empirical ex-post regret, but the paper acknowledges immediately after Eq. (35) that the objective is non-convex and convergence to the global optimum is not theoretically guaranteed. Moreover, the ex-post regret crgt_i and the utility u_i used to define the IC metric are computed with predicted pCTR from the reward model and with payments from the very payment network being optimized, so a low in-sample regret is partly a self-fulfilling outcome. The claim should be weakened to something like ``empirically reduces ex-post regret under model-estimated pCTR,'' or the paper should provide a formal IC guarantee or validation on held-out advertiser payoffs with real outcomes.
  4. [§4.4.2, Eq. (31)] The marginal contribution r_yi in Eq. (31) is not well defined because S*_{-i}, the ``best alternative ad sequence excluding y_i,'' is not specified. It is unclear how this alternative sequence is generated, whether it is recomputed after removing y_i from the beam search, and how the exclusion interacts with the token-level generative process. Since this quantity is the reward signal for the policy-gradient loss in Eq. (32), the training objective is currently not reproducible without additional specification.
minor comments (5)
  1. [Eq. (28)] The MTP loss uses Y_{1:t-1} in the conditioning term, but the summation index is i; this should presumably be Y_{1:i-1} to be consistent with the probabilistic decomposition in Eq. (13).
  2. [Footnote 3] The statement that reported results have been ``transformed in a way that preserves their statistical properties'' is too vague; the paper should clarify which statistical properties are preserved (e.g., ranks, relative differences, or margins) so readers can judge whether the reported absolute values and lift percentages are meaningful for comparison.
  3. [Section 1 and contributions] There are minor language issues, such as ``by an unified single generative model'' and ``Last but not least ,''; these should be corrected in a polish pass.
  4. [Figure 3] The figure label ``All Valid POIs'' is not explained in the text; the paper should define what makes a POI valid in the generative allocation step.
  5. [Footnote 2] The note that evaluation is limited to offline experiments because of computational complexity is an important limitation and deserves more discussion in the conclusion, especially because the paper describes the framework as ready for industrial deployment.

Circularity Check

2 steps flagged · score 6.0 of 10

Offline RPM/CTR/IC metrics are computed from the reward model and payment network being optimized, so the reported gains are partially self-fulfilling.

  1. fitted input called prediction [Section 5.1.2 (Evaluation Metrics); Eqs. (20), (24), (31)-(32)]
    "For offline experiments, evaluation metrics are computed using the predicted values from the reward model. ... ˆr_j = R(S^(j)). ... The final output is the sequence S∗ with the highest reward."

    The same reward model R provides the pCTR predictions in Eq. (20), selects the winning sequence S∗ in Eq. (24), and supplies the 'clicks' used to compute offline RPM and CTR in Section 5.1.2. The generative allocation policy is trained to maximize Σ_j b_j·ˆr^{pctr}_j (Eqs. 31-32), so the evaluation metric is essentially the objective being optimized. Reported lifts over MCA/GR therefore measure how well the model optimizes its own reward model, not an independent estimate of user behavior; no logged-outcome or debiased offline evaluation is provided.

  2. fitted input called prediction [Section 4.4.2 Eq. (33); Section 5.1.2 IC Metric]
    "IC Metric: Ψ = 1/|D| Σ_{d∈D} Σ_{i∈k} crgt^d_i / u_i(v^d_i; b^d), where crgt^d_i denotes the empirical ex-post regret ... as defined in Equation (7). ... LPay = ... Σ_{y_i∈S*} λ_i crgt_i − ρ/2 Σ_{y_i∈S*} (crgt_i)^2."

    The reported IC metric Ψ is the same empirical ex-post regret crgt that the payment network is trained to minimize in Eq. (33). The payment network is fitted with a Lagrangian objective that drives crgt toward zero, and the paper then reports crgt/u as evidence of approximate incentive compatibility. Thus the low Ψ value (2.7%) is an in-sample training objective, not an independent counterfactual test of truthfulness; it is forced by construction rather than verified externally.

full rationale

EGA-V2's architecture and training losses are not themselves derived from the reported metrics, so this is not total circularity. The circularity is concentrated in the offline evaluation loop used to support the headline empirical claims. Eq. (20) defines the reward model's predicted pCTR; Eq. (24) uses that reward model to select the final sequence; Eqs. (31)-(32) train the generator to maximize revenue computed with those same predictions; and Section 5.1.2 explicitly states that offline RPM, CTR, and IC metrics are computed from the reward model's predicted values. Similarly, the IC metric Ψ is the averaged empirical ex-post regret crgt, which is exactly the quantity the payment network's loss in Eq. (33) is designed to minimize. The reported RPM increase (230.41 vs 206.73) and IC regret reduction (2.7% vs 8.4%) therefore partly reduce to the model optimizing its own evaluation function. The reward model is trained on real user feedback, which gives the loop some independent grounding, but no calibration or debiased offline evaluation is supplied. Overall, the central 'significantly outperforms' claim is partially circular, warranting a score of 6 rather than a higher score that would imply the derivation itself is an identity.

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

The free parameters are hyperparameters tuned by grid search or left unspecified; the axioms are domain assumptions that the user history predicts behavior, the reward model is a faithful proxy, and the economic mechanisms work as intended. No invented entities are introduced.

free parameters (5)
  • alpha (bid weight exponent) = 1.2
    Hyperparameter in Eq. (21) controlling the influence of bids on token allocation; tuned by grid search (Section 5.1.4) and central to the revenue/CTR trade-off.
  • beta (organic/ad balance) = 2
    Hyperparameter in Eq. (21) balancing ads versus organic content; tuned by grid search and directly affects the generated list composition.
  • beam width N_S = 64
    Number of candidate sequences generated by beam search (Section 4.3.2); a design choice that trades off allocation quality against latency, and the paper reports only the chosen value.
  • Lagrangian penalty rho = not reported
    Hyperparameter in the payment loss (Eq. 33) controlling the IC penalty strength; the paper does not state the value used.
  • number of sampled valuations N_v = not reported
    Number of valuations sampled from distribution F to estimate empirical ex-post regret in Eq. (7); value not specified.
assumptions (6)
  • domain assumption User historical behavior sequences are a reliable predictor of future click and conversion behavior.
    The pre-training objective predicts next exposed items from history (Section 4.1). The whole generative recommendation premise relies on this assumption.
  • domain assumption The learned reward model's predicted pCTR and pCVR are faithful proxies for real user engagement.
    All offline metrics (RPM, CTR, IC regret) are computed from reward model predictions (Section 5.1.2). If the reward model is miscalibrated, the experimental conclusions do not transfer to real behavior.
  • ad hoc to paper Aggregating item bids by taking the maximum over items sharing a token preserves advertiser incentives and supports an approximately incentive-compatible allocation.
    Eq. (21) defines the token bid as max over associated item bids. The paper asserts this alignment with token-level bidding theory (ref [6]) but provides no formal analysis of how this aggregation affects truthfulness.
  • ad hoc to paper Minimizing empirical ex-post regret with a Lagrangian penalty yields a payment rule that is approximately incentive compatible.
    Section 4.4.2 states the objective is non-convex and global convergence is not guaranteed, yet the abstract claims a guarantee. The approximation relies on the training procedure reaching a low-regret solution.
  • ad hoc to paper The marginal contribution of an item to revenue is correctly estimated by comparing the current sequence with the best alternative sequence without that item.
    Eq. (31) defines r_yi using the best alternative sequence S*_{-i}. The paper does not specify how S*_{-i} is computed or whether the approximation is unbiased.
  • domain assumption The organic content order is fixed and known a priori.
    Stated in footnote 1: the organic sequence is pre-ranked and its internal order remains fixed. This simplifies the allocation problem but may not hold in all production systems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EGA-V2: An End-to-end Generative Framework for Industrial Advertising." pith.science (2026). https://pith.science/paper/6ROQFBNY

@misc{pith2026250517549,
  author       = {Pith},
  title        = {Pith review of: EGA-V2: An End-to-end Generative Framework for Industrial Advertising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ROQFBNY}},
  note         = {Machine review of arXiv:2505.17549}
}
read the original abstract

Traditional online industrial advertising systems suffer from the limitations of multi-stage cascaded architectures, which often discard high-potential candidates prematurely and distribute decision logic across disconnected modules. While recent generative recommendation approaches provide end-to-end solutions, they fail to address critical advertising requirements of key components for real-world deployment, such as explicit bidding, creative selection, ad allocation, and payment computation. To bridge this gap, we introduce End-to-End Generative Advertising (EGA-V2), the first unified framework that holistically models user interests, point-of-interest (POI) and creative generation, ad allocation, and payment optimization within a single generative model. Our approach employs hierarchical tokenization and multi-token prediction to jointly generate POI recommendations and ad creatives, while a permutation-aware reward model and token-level bidding strategy ensure alignment with both user experiences and advertiser objectives. Additionally, we decouple allocation from payment using a differentiable ex-post regret minimization mechanism, guaranteeing approximate incentive compatibility at the POI level. Through extensive offline evaluations we demonstrate that EGA-V2 significantly outperforms traditional cascaded systems in both performance and practicality. Our results highlight its potential as a pioneering fully generative advertising solution, paving the way for next-generation industrial ad systems.

Figures

Figures reproduced from arXiv: 2505.17549 by the authors.

Figure 1
Figure 1. (a) A typical cascade advertising system. (b) Our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the interest-based pre-training architecture. The pre-training model consists of encoder for modeling [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the proposed generative ad allocation and payment architecture. Token-level bidding aggregates item [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Overview of the proposed optimization and train [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Effect of hyperparameters of EGA-V2. We conduct a comprehensive study to evaluate the sensitivity of EGA-V2 with respect to key hyperparameters. The results are sum￾marized in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 11 Pith papers

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

  1. Gryphon-v2: One Model in Place of a Cascade - Generate-and-Rank Recommender with Rollout Distillation

    cs.IR 2026-08 conditional novelty 6.0 of 10

    A single generate-and-rank model with a distilled ranking module replaces a 15-plus stage production recommender cascade and lifts active users by 1.41% in an online A/B test.

  2. HGenPush: A Heterogeneous Generative Recommendation Architecture for Industrial Push Notification Systems

    cs.IR 2026-07 conditional novelty 6.0 of 10

    HGenPush jointly generates video and author recommendations via Chained-MTP and GSISPO preference alignment, raising Kuaishou push DAU by 0.181%.

  3. UniVA: Unified Value Alignment for Generative Recommendation in Online Advertising at Tencent

    cs.IR 2026-05 unverdicted novelty 6.0 of 10

    UniVA unifies value alignment in generative recommendation via a Commercial SID tokenizer, eCPM-aware RL decoder, and personalized beam search, reporting 37% offline Hit Rate gains and 1.5% online GMV lift on Tencent ...

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

    cs.IR 2025-06 conditional novelty 6.0 of 10

    GFlowGR fine-tunes generative recommender LLMs with GFlowNet losses and multi-signal rewards, beating SFT, DPO, and GRPO baselines on three datasets and in production.

  5. OneSug: The Unified End-to-End Generative Framework for E-commerce Query Suggestion

    cs.IR 2025-06 conditional novelty 6.0 of 10

    A unified encoder-decoder model with prefix representation enhancement and reward-weighted DPO outperforms Kuaishou's online multi-stage query suggestion system in offline and live A/B evaluations.

  6. UniGD: A Unified Generative-Discriminative Framework for Industrial Retrieval

    cs.AI 2026-08 conditional novelty 5.0 of 10

    UniGD couples generative retrieval with explicit relevance scoring in one model, reporting +5.78% ad revenue, 33.1% lower latency at Kuaishou, and improved Recall@10 on NQ320K and MS300K.

  7. TSGR: Taobao Search Generative Retrieval

    cs.IR 2026-07 conditional novelty 5.0 of 10

    TSGR makes Taobao search retrieval value-aware by ordering item IDs by clicks and query intent and re-ranking generated candidates with a value module, improving HR@1000 and GMV.

  8. Noise is not always detrimental: the capacity of quantum batteries is enhanced in black holes

    quant-ph 2026-04 unverdicted novelty 5.0 of 10

    Hawking radiation is claimed to enhance quantum battery capacity for bipartite mixed states, while environmental noise generally degrades it in type-dependent ways.

  9. FORGE: Forming Semantic Identifiers for Generative Retrieval in Industrial Datasets

    cs.IR 2025-09 conditional novelty 5.0 of 10

    FORGE shows that balancing codebook usage and adding multimodal side information improves semantic identifiers for generative retrieval, validated offline and on Taobao.

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

    cs.IR 2025-06 conditional novelty 5.0 of 10

    NGA is a non-autoregressive generative auction that models effects of adjacent organic content and computes rewards and payments in parallel, reporting gains in RPM, CTR, CVR, and latency over CGA.

  11. Coarse-to-Fine Long-term Interest Modeling for Generative Recommendation

    cs.IR 2026-02 conditional novelty 4.0 of 10

    GLASS extends generative retrieval with a tiered long-term interest vector and a first-SID-keyed search of long histories, reporting consistent gains over Tiger and DualGR on two public datasets.

Reference graph

Works this paper leans on

44 extracted references · 20 canonical work pages · cited by 11 Pith papers

  1. [1]

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

    work page 2014

  2. [2]

    Dagui Chen, Qi Yan, Chunjie Chen, Zhenzhe Zheng, Yangsu Liu, Zhenjia Ma, Chuan Yu, Jian Xu, and Bo Zheng. 2022. Hierarchically constrained adaptive ad exposure in feeds. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management . 3003–3012

    work page 2022

  3. [3]

    Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2020. Au- toregressive entity retrieval. arXiv preprint arXiv:2010.00904 (2020)

    arXiv 2020

  4. [4]

    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

  5. [5]

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

    work page 2020

  6. [6]

    Paul Duetting, Vahab Mirrokni, Renato Paes Leme, Haifeng Xu, and Song Zuo

    work page

  7. [7]

    Paul Dütting, Zhe Feng, Harikrishna Narasimhan, David Parkes, and Sai Srivatsa Ravindranath. 2019. Optimal auctions through deep learning. In International Conference on Machine Learning . PMLR, 1706–1715

    work page 2019

  8. [8]

    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 review 97, 1 (2007), 242–259

    work page 2007

Show all 44 references
  1. [9]

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

    2012

  2. [10]

    Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Rozière, David Lopez-Paz, and Gabriel Synnaeve. 2024. Better & faster large language models via multi-token prediction. arXiv preprint arXiv:2404.19737 (2024)

    2024 arXiv

  3. [11]

    Siyu Gu and Xiangrong Sheng. 2022. On Ranking Consistency of Pre-ranking Stage. arXiv preprint arXiv:2205.01289 (2022)

    2022 arXiv

  4. [12]

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

    2014

  5. [13]

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

  6. [14]

    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

  7. [15]

    Deep automated mechanism design for integrating ad auction and allocation Woodstock ’18, June 03–05, 2018, Woodstock, NY Zheng et al. in feed. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1211–1220

    2018

  8. [16]

    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. In Proceedings of the ACM Web Conference 2022 . 401– 409

    2022

  9. [17]

    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. In Proceedings of the 45th International ACM SIGIR Conference on Research and Devel...

    2022

  10. [18]

    Han Liu, Yinwei Wei, Xuemeng Song, Weili Guan, Yuan-Fang Li, and Liqiang Nie. 2024. Mmgrec: Multimodal generative recommendation with transformer model. arXiv preprint arXiv:2404.16555 (2024)

    2024 arXiv

  11. [19]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  12. [20]

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

  13. [21]

    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. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge ...

    2021

  14. [22]

    Juntao Tan, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Zelong Li, and Yongfeng Zhang. 2024. Idgenrec: Llm-recsys alignment with textual id learning. In Proceed- ings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 355–364

    2024

  15. [23]

    Yubao Tang, Ruqing Zhang, Jiafeng Guo, and Maarten de Rijke. 2023. Recent advances in generative information retrieval. In Proceedings of the Annual In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region . 294–297

    2023

  16. [24]

    Zihua Si, Zhongxiang Sun, Jiale Chen, Guozhang Chen, Xiaoxue Zang, Kai Zheng, Yang Song, Xiao Zhang, Jun Xu, and Kun Gai. 2024. Generative Retrieval with Semantic Tree-Structured Identifiers and Contrastive Learning. In Proceedings of the 2024 Annual International ACM SIGIR Co...

    2024

  17. [25]

    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

  18. [26]

    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

  19. [27]

    Yi Tay, Vinh Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, et al. 2022. Transformer memory as a differentiable search index. Advances in Neural Information Processing Systems 35 (2022), 21831–21843

    2022

  20. [28]

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

    2021

  21. [29]

    Yue Xu, Qijie Shen, Jianwen Yin, Zengde Deng, Dimin Wang, Hao Chen, Lixi- ang Lai, Tao Zhuang, and Junfeng Ge. 2023. Multi-channel Integrated Recom- mendation with Exposure Constraints. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 5338–5349

    2023

  22. [30]

    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

  23. [31]

    Yuhao Yang, Zhi Ji, Zhaopeng Li, Yi Li, Zhonglin Mo, Yue Ding, Kai Chen, Zijian Zhang, Jie Li, Shuanglong Li, et al. 2025. Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations. arXiv preprint arXiv:2503.02453 (2025)

    2025 arXiv

  24. [32]

    Zhiguang Yang, Liufang Sang, Haoran Wang, Wenlong Chen, Lu Wang, Jie He, Changping Peng, Zhangang Lin, Chun Gan, and Jingping Shao. 2024. Parallel ranking of ads and creatives in real-time advertising systems. In Proceedings of the AAAI Conference on Artificial Intelligence , ...

    2024

  25. [33]

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

    2020

  26. [34]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152 (2024)

    2024 arXiv

  27. [35]

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

    2021 arXiv

  28. [36]

    Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. 2021. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30 (2021), 495–507

    2021

  29. [37]

    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

  30. [38]

    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. In Proceedings of the AAAI conference on artificial intelligence , Vol. 35. 750–758

    2021

  31. [39]

    Zhixuan Zhang, Yuheng Huang, Dan Ou, Sen Li, Longbin Li, Qingwen Liu, and Xiaoyi Zeng. 2023. Rethinking the role of pre-ranking in large-scale e-commerce searching system. arXiv preprint arXiv:2305.13647 (2023)

    2023 arXiv

  32. [40]

    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

  33. [41]

    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

  34. [42]

    Zhishan Zhao, Jingyue Gao, Yu Zhang, Shuguang Han, Siyuan Lou, Xiang-Rong Sheng, Zhe Wang, Han Zhu, Yuning Jiang, Jian Xu, et al. 2023. COPR: Consistency- Oriented Pre-Ranking for Online Advertising. In Proceedings of the 32nd ACM International Conference on Information and Kn...

    2023

  35. [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

  36. [2024]

    In Proceedings of the ACM Web Conference 2024

    Mechanism design for large language models. In Proceedings of the ACM Web Conference 2024. 144–155

    2024

Pith tools

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