The reviewed record of science sign in
Pith

arxiv: 2605.27811 · v1 · pith:RGPPKI7W · submitted 2026-05-27 · cs.AI

Constrained Auto-Bidding via Generative Response Modeling

Eunseok Yang , Xingdong Zuo , Kyung-Min Kim This is my paper

Reviewed by Pith2026-06-29 13:10 UTCgrok-4.3pith:RGPPKI7Wopen to challenge →

classification cs.AI
keywords auto-biddinggenerative response modelingconstraint enforcementbid multiplieroptimality boundsreceding-horizon controlauction prediction
0
0 comments X

The pith

A generative response model predicts auction reactions to one bid multiplier, enabling an analytic controller with bounded optimality gap.

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

Auto-bidding must maximize advertiser value over long horizons while respecting budget and ratio constraints such as cost-per-acquisition amid uncertain traffic. Existing control methods react late and RL methods hide violations inside rewards. The paper shifts the target to learning responses instead of actions by training a history-conditioned sequence model, the Generative Response Model, that outputs future volume and aggregate cost-value curves as functions of a single multiplier. Under mild monotonicity the optimality gap to ideal per-tick control is bounded by the spread of per-tick marginal value-per-cost, the controller meets active constraints by a one-dimensional root find, and replanning keeps violations proportional to prediction error.

Core claim

The Generative Response Model jointly forecasts traffic volume and horizon-aggregate cost and value curves conditioned on a chosen bid multiplier. An analytic controller then solves each active constraint by a single 1D root-finding step on the predicted curves. The paper proves that this controller is exact whenever only one multiplier is used and that the optimality gap relative to full per-tick control is bounded by the dispersion of per-tick marginal value-per-cost under monotonicity; constraint violations under receding-horizon replanning are likewise bounded by the size of the prediction error.

What carries the argument

The Generative Response Model, a history-conditioned sequence model that outputs predicted traffic volume and aggregate cost-value curves as functions of a single bid multiplier, which the analytic controller then uses for root-finding constraint enforcement.

If this is right

  • The optimality gap to full per-tick control is bounded by per-tick marginal value-per-cost dispersion under monotonicity.
  • The analytic controller is exact for any single-multiplier problem.
  • Constraint violations under receding-horizon replanning scale directly with prediction error.
  • GRM yields measurably higher constraint stability and overall score than pacing or RL baselines on AuctionNet.

Where Pith is reading between the lines

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

  • Response modeling that separates prediction from control may improve robustness in other long-horizon constrained allocation tasks such as ad inventory or energy dispatch.
  • Empirical audits of marginal dispersion across real auction datasets would reveal how frequently the derived bound remains practically tight.
  • The same prediction-plus-root-find pattern could be tested on problems where multiple discrete actions must be collapsed to a single continuous parameter.

Load-bearing premise

A single bid multiplier remains sufficient to steer the whole horizon when the dispersion of per-tick marginal value-per-cost stays small enough for the derived bound to be useful.

What would settle it

Measure the realized optimality gap on live auction traffic when the GRM controller is applied and test whether the gap exceeds the per-tick marginal dispersion bound computed from the same data.

Figures

Figures reproduced from arXiv: 2605.27811 by Eunseok Yang, Kyung-Min Kim, Xingdong Zuo.

Figure 1
Figure 1. Figure 1: Overview of the Generative Response Model (GRM) framework. (Left) GRM encodes state-action history via a causal [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Robustness under distribution shift. (a) Competi [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Validation loss vs performance across training [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
read the original abstract

Auto-bidding systems aim to maximize advertiser value over long horizons under budget constraints and ratio targets such as cost-per-acquisition, yet future traffic and auction dynamics are non-stationary and uncertain. Existing approaches face distinct limitations: control-based pacing reacts to deviations but cannot anticipate future conditions, while RL and generative methods fold constraints into reward signals, obscuring violations and degrading under distribution shift. We shift the learning target from actions to responses with the Generative Response Model (GRM), a history-conditioned sequence model that jointly predicts future traffic volume and horizon-aggregate cost/value curves as functions of a single bid multiplier. We show that under mild monotonicity conditions, the optimality gap relative to full per-tick control is bounded by the dispersion of per-tick marginal value-per-cost. Given predicted responses, a lightweight analytic controller enforces each active constraint via a 1D root-finding step. We prove this controller is exact for the single-multiplier problem and bound constraint violations under receding-horizon replanning in terms of prediction error. Experiments on AuctionNet show that GRM improves constraint stability and overall score compared to existing baselines.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. The paper proposes the Generative Response Model (GRM), a history-conditioned sequence model that predicts future traffic volume and horizon-aggregate cost/value curves as functions of a single bid multiplier. It claims that under mild monotonicity conditions the optimality gap to full per-tick control is bounded by the dispersion of per-tick marginal value-per-cost; that a lightweight analytic controller enforcing active constraints via 1D root-finding is exact for the single-multiplier problem; and that constraint violations under receding-horizon replanning are bounded by prediction error. Experiments on AuctionNet are reported to show improved constraint stability and overall score versus baselines.

Significance. If the stated bounds hold and the dispersion of per-tick marginal value-per-cost remains small under single-multiplier control, the GRM-plus-analytic-controller approach would offer a principled way to handle non-stationary traffic and ratio constraints without folding them into RL rewards. The shift from action prediction to response modeling is a clear conceptual contribution; explicit credit is due for the attempt to derive an optimality-gap bound and an exactness result for the controller, even though those derivations are not supplied in the abstract.

major comments (3)
  1. [Abstract] Abstract (optimality-gap claim): the bound on the optimality gap is expressed solely in terms of dispersion of per-tick marginal value-per-cost, yet the manuscript provides neither a derivation of the bound nor any reported measurement (or even summary statistic) of realized dispersion on AuctionNet under the GRM-predicted responses. Without this quantification the bound cannot be assessed for tightness or practical utility.
  2. [Abstract] Abstract (controller exactness): the claim that the analytic controller is exact for the single-multiplier problem rests on unspecified 'mild monotonicity conditions' and on the assumption that a single multiplier suffices to meet all active constraints; neither the precise monotonicity statement nor any verification that GRM outputs preserve it is supplied, leaving the exactness result unverifiable from the given text.
  3. [Abstract] Abstract (constraint-violation bound): the bound on constraint violations under receding-horizon replanning is stated to depend on prediction error, but no experimental protocol for measuring that error (e.g., how GRM predictions are compared to realized per-tick outcomes) or for implementing the baselines is described, so the reported improvements cannot be reproduced or stress-tested against the claimed guarantee.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for these constructive comments on the abstract. We will make revisions to address each point as detailed below.

read point-by-point responses

  1. Referee: [Abstract] Abstract (optimality-gap claim): the bound on the optimality gap is expressed solely in terms of dispersion of per-tick marginal value-per-cost, yet the manuscript provides neither a derivation of the bound nor any reported measurement (or even summary statistic) of realized dispersion on AuctionNet under the GRM-predicted responses. Without this quantification the bound cannot be assessed for tightness or practical utility.

    Authors: We agree with the observation. The abstract currently states the bound without including the derivation or the dispersion measurement. In the revision, we will incorporate a concise statement of the bound's derivation approach and report the realized dispersion on AuctionNet to allow assessment of its utility. revision: yes

  2. Referee: [Abstract] Abstract (controller exactness): the claim that the analytic controller is exact for the single-multiplier problem rests on unspecified 'mild monotonicity conditions' and on the assumption that a single multiplier suffices to meet all active constraints; neither the precise monotonicity statement nor any verification that GRM outputs preserve it is supplied, leaving the exactness result unverifiable from the given text.

    Authors: We agree that the monotonicity conditions and verification are not specified in the abstract. We will add the precise statement of the mild monotonicity conditions to the abstract and include verification that GRM outputs preserve monotonicity in the experiments. revision: yes

  3. Referee: [Abstract] Abstract (constraint-violation bound): the bound on constraint violations under receding-horizon replanning is stated to depend on prediction error, but no experimental protocol for measuring that error (e.g., how GRM predictions are compared to realized per-tick outcomes) or for implementing the baselines is described, so the reported improvements cannot be reproduced or stress-tested against the claimed guarantee.

    Authors: We agree that the experimental protocol is not described in the abstract. We will revise the abstract to briefly outline the protocol for measuring prediction error and implementing baselines, ensuring the improvements can be reproduced and tested against the guarantee. revision: yes

Circularity Check

0 steps flagged

No significant circularity; theoretical bounds and exactness proof are independent of fitted inputs

full rationale

The paper states a bound on the optimality gap in terms of per-tick marginal value-per-cost dispersion under monotonicity, proves the 1D root-finding controller is exact for the single-multiplier problem, and bounds violations by prediction error. These are presented as mathematical derivations from the problem formulation and GRM outputs rather than reductions by construction, self-definition, or self-citation chains. The GRM is trained to predict responses used as controller inputs; the claims do not rename fitted quantities as predictions or rely on load-bearing self-citations. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on an unstated monotonicity condition whose precise form is not given in the abstract, plus the modeling assumption that a scalar multiplier suffices for the entire horizon. No free parameters or invented physical entities are described.

axioms (1)
  • domain assumption Mild monotonicity conditions on per-tick marginal value-per-cost
    Invoked to bound the optimality gap relative to per-tick control; exact statement and location not provided in abstract.

pith-pipeline@v0.9.1-grok · 5724 in / 1327 out tokens · 26408 ms · 2026-06-29T13:10:45.223644+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 1 canonical work pages

  1. [1]

    Suzan Ece Ada, Erhan Oztop, and Emre Ugur. 2024. Diffusion policies for out-of- distribution generalization in offline reinforcement learning.IEEE Robotics and Automation Letters(2024)

    2024

  2. [2]

    Gagan Aggarwal, Ashwinkumar Badanidiyuru, Santiago R Balseiro, Kshipra Bhawalkar, Yuan Deng, Zhe Feng, Gagan Goel, Christopher Liaw, Haihao Lu, Mohammad Mahdian, et al. 2024. Auto-bidding and auctions in online advertising: A survey.ACM SIGecom Exchanges22, 1 (2024), 159–183

    2024

  3. [3]

    Santiago R Balseiro, Omar Besbes, and Gabriel Y Weintraub. 2015. Repeated auctions with budgets in ad exchanges: Approximations and design.Management Science61, 4 (2015), 864–884

    2015

  4. [4]

    Santiago R Balseiro, Kshipra Bhawalkar, Zhe Feng, Haihao Lu, Vahab Mirrokni, Balasubramanian Sivan, and Di Wang. 2024. A Field Guide for Pacing Budget and ROS Constraints. InProceedings of the 41st International Conference on Machine Learning. PMLR, 2607–2638

    2024

  5. [5]

    Han Cai, Kan Ren, Weinan Zhang, Kleanthis Malialis, Jun Wang, Yong Yu, and Defeng Guo. 2017. Real-time bidding by reinforcement learning in display adver- tising. InProceedings of the tenth ACM international conference on web search and data mining. 661–670

    2017

  6. [6]

    Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. 2021. Decision transformer: Reinforcement learning via sequence modeling.Advances in neural information processing systems34 (2021), 15084–15097

    2021

  7. [7]

    Shuang Chen, Qisen Xu, Liang Zhang, Yongbo Jin, Wenhao Li, and Linjian Mo

  8. [8]

    InProceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems

    Model-Based Reinforcement Learning for Auto-bidding in Display Adver- tising. InProceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems. 1560–1568

    2023

  9. [9]

    Ye Chen, Pavel Berkhin, Bo Anderson, and Nikhil R Devanur. 2011. Real-time bidding algorithms for performance-based display ad allocation. InProceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining. 1307–1315

    2011

  10. [10]

    Vincent Conitzer, Christian Kroer, Debmalya Panigrahi, Okke Schrijvers, Nico- las E Stier-Moses, Eric Sodomka, and Christopher A Wilkens. 2022. Pacing equilibrium in first price auction markets.Management Science68, 12 (2022), 8515–8535

    2022

  11. [11]

    Ying Cui, Ruofei Zhang, Wei Li, and Jianchang Mao. 2011. Bid landscape forecast- ing in online ad exchange marketplace. InProceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining. 265–273

    2011

  12. [12]

    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

  13. [13]

    Scott Fujimoto, David Meger, and Doina Precup. 2019. Off-policy deep rein- forcement learning without exploration. InInternational conference on machine learning. PMLR, 2052–2062

    2019

  14. [14]

    Jingtong Gao, Yewen Li, Shuai Mao, Peng Jiang, Nan Jiang, Yejing Wang, Qing- peng Cai, Fei Pan, Peng Jiang, Kun Gai, et al. 2025. Generative auto-bidding with value-guided explorations. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 244–254

    2025

  15. [15]

    Aritra Ghosh, Saayan Mitra, Somdeb Sarkhel, Jason Xie, Gang Wu, and Viswanathan Swaminathan. 2019. Scalable bid landscape forecasting in real- time bidding. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 451–466

    2019

  16. [16]

    Jiayan Guo, Yusen Huo, Zhilin Zhang, Tianyu Wang, Chuan Yu, Jian Xu, Bo Zheng, and Yan Zhang. 2024. AIGB: Generative Auto-bidding via Conditional Diffusion Modeling. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5038–5049

    2024

  17. [17]

    Yue He, Xiujun Chen, Di Wu, Junwei Pan, Qing Tan, Chuan Yu, Jian Xu, and Xiaoqiang Zhu. 2021. A unified solution to constrained bidding in online display advertising. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2993–3001

    2021

  18. [18]

    Haruka Kiyohara, Kosuke Kawakami, and Yuta Saito. 2021. Accelerating offline reinforcement learning application in real-time bidding and recommendation: Potential use of simulation.arXiv preprint arXiv:2109.08331(2021)

    work page arXiv 2021

  19. [19]

    Ilya Kostrikov, Ashvin Nair, and Sergey Levine. 2021. Offline Reinforcement Learning with Implicit Q-Learning. InDeep RL Workshop NeurIPS 2021

    2021

  20. [20]

    Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. 2020. Conserva- tive q-learning for offline reinforcement learning.Advances in Neural Information Processing Systems33 (2020), 1179–1191

    2020

  21. [21]

    Haoming Li, Yusen Huo, Shuai Dou, Zhenzhe Zheng, Zhilin Zhang, Chuan Yu, Jian Xu, and Fan Wu. 2024. Trajectory-wise Iterative Reinforcement Learning Framework for Auto-bidding. InProceedings of the ACM on Web Conference 2024. 4193–4203

    2024

  22. [22]

    Kaiyuan Li, Pengyu Wang, Yunshan Peng, Pengjia Yuan, Yanxiang Zeng, Rui Xiang, Yanhua Cheng, Xialong Liu, and Peng Jiang. 2025. EBaReT: Expert-guided Bag Reward Transformer for Auto Bidding. InCompanion Proceedings of the ACM on Web Conference 2025. 1104–1108

    2025

  23. [23]

    Yewen Li, Shuai Mao, Jingtong Gao, Nan Jiang, Yunjian Xu, Qingpeng Cai, Fei Pan, Peng Jiang, and Bo An. 2025. GAS: Generative Auto-bidding with Post-training Search. InCompanion Proceedings of the ACM on Web Conference 2025. 315–324

    2025

  24. [24]

    Zuxin Liu, Zijian Guo, Yihang Yao, Zhepeng Cen, Wenhao Yu, Tingnan Zhang, and Ding Zhao. 2023. Constrained decision transformer for offline safe reinforcement learning. InInternational Conference on Machine Learning. PMLR, 21611–21630

    2023

  25. [25]

    Zhiyu Mou, Yusen Huo, Rongquan Bai, Mingzhou Xie, Chuan Yu, Jian Xu, and Bo Zheng. 2022. Sustainable online reinforcement learning for auto-bidding. Advances in Neural Information Processing Systems35 (2022), 2651–2663

    2022

  26. [26]

    Weitong Ou, Bo Chen, Xinyi Dai, Weinan Zhang, Weiwen Liu, Ruiming Tang, and Yong Yu. 2023. A Survey on Bid Optimization in Real-Time Bidding Display Advertising.ACM Transactions on Knowledge Discovery from Data18, 3 (2023), 1–31

    2023

  27. [27]

    Rafael Figueiredo Prudencio, Marcos ROA Maximo, and Esther Luna Colombini

  28. [28]

    A survey on offline reinforcement learning: Taxonomy, review, and open problems.IEEE Transactions on Neural Networks and Learning Systems(2023)

    2023

  29. [29]

    Kefan Su, Yusen Huo, Zhilin Zhang, Shuai Dou, Chuan Yu, Jian Xu, Zongqing Lu, and Bo Zheng. 2024. AuctionNet: A Novel Benchmark for Decision-Making in Large-Scale Games. InThe Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track. https://arxiv.org/abs/2412. 10798

    2024

  30. [30]

    Yumin Su, Min Xiang, Yifei Chen, Yanbiao Li, Tian Qin, Hongyi Zhang, Yasong Li, and Xiaobing Liu. 2024. Spending programmed bidding: Privacy-friendly bid optimization with roi constraint in online advertising. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5731–5740

    2024

  31. [31]

    Faraz Torabi, Garrett Warnell, and Peter Stone. 2018. Behavioral Cloning from Observation. InProceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence. International Joint Conferences on Artificial Intelligence Organization

    2018

  32. [32]

    Yuchen Wang, Kan Ren, Weinan Zhang, Jun Wang, and Yong Yu. 2016. Functional bid landscape forecasting for display advertising. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 115–131

    2016

  33. [33]

    Chao Wen, Miao Xu, Zhilin Zhang, Zhenzhe Zheng, Yuhui Wang, Xiangyu Liu, Yu Rong, Dong Xie, Xiaoyang Tan, Chuan Yu, et al. 2022. A cooperative-competitive multi-agent framework for auto-bidding in online advertising. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. 1129–1139

    2022

  34. [34]

    Di Wu, Xiujun Chen, Xun Yang, Hao Wang, Qing Tan, Xiaoxun Zhang, Jian Xu, and Kun Gai. 2018. Budget constrained bidding by model-free reinforcement learning in display advertising. InProceedings of the 27th ACM International Conference on Information and Knowledge Management. 1443–1451

    2018

  35. [35]

    Hao Yu, Michael Neely, and Xiaohan Wei. 2017. Online convex optimization with stochastic constraints.Advances in Neural Information Processing Systems 30 (2017)

    2017

  36. [36]

    Shuai Yuan, Jun Wang, and Xiaoxue Zhao. 2013. Real-time bidding for online advertising: measurement and analysis. InProceedings of the seventh international workshop on data mining for online advertising. 1–8

    2013

  37. [37]

    Wei Zhang, Yanjun Han, Zhengyuan Zhou, Aaron Flores, and Tsachy Weissman

  38. [38]

    Advances in Neural Information Processing Systems35 (2022), 21329–21341

    Leveraging the hints: Adaptive bidding in repeated first-price auctions. Advances in Neural Information Processing Systems35 (2022), 21329–21341

    2022

  39. [39]

    Weinan Zhang, Yifei Rong, Jun Wang, Tianchi Zhu, and Xiaofan Wang. 2016. Feedback control of real-time display advertising. InProceedings of the Ninth ACM International Conference on Web Search and Data Mining. 407–416

    2016

  40. [40]

    Jun Zhao, Guang Qiu, Ziyu Guan, Wei Zhao, and Xiaofei He. 2018. Deep rein- forcement learning for sponsored search real-time bidding. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1021–1030

    2018

  41. [41]

    Xiangyu Zhao, Long Xia, Jiliang Tang, and Dawei Yin. 2019. Deep reinforcement learning for search, recommendation, and online advertising: a survey.ACM SIGWEB Newsletter2019, Spring (2019), 1–15. KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Yang et al. A Full Proofs A.1 Proof of Theorem 5.2 (Structural Gap) We bound the budget-only gap (th...

    2019