acceptodds
Under review as a conference paper at ICLR 2027

Auto-Bidding with Disentangled Advertiser Profiles and Train-Free Adaptation

Abstract

Auto-bidding is a key component of modern advertising systems that provides a personalized bidding strategy for each advertiser. By characterizing each individual, profile-based methods achieve personalization and have proven effective in domains such as recommendation; however, despite the diverse bidding behavior of advertisers, their application to auto-bidding remains limited. A primary reason is that constructing and leveraging advertiser profiles face several challenges: extracting pure profiles is non-trivial, modeling shared and private information simultaneously is difficult, and profile updating and cold-start adaptation remain challenging. To tackle these issues, we propose **ADAPT**, an **A**uto-bidding framework with **D**isentangled **A**dvertiser **P**rofiles and **T**raining-free adaptation. ADAPT introduces a two-stage training paradigm and supports training-free adaptation. Specifically, (i) the first stage extracts pure stable and dynamic profiles via contrastive learning over the advertiser memory table; (ii) the second stage disentangles the dynamic profile into a shared strategy and a private strategy, and combines them with the static profile to jointly condition the bidding strategy; (iii) once trained, ADAPT constructs profiles for new advertisers and updates profiles of existing advertisers without retraining. Our experiments on a large-scale auto-bidding benchmark demonstrate that ADAPT consistently achieves superior performance, and ablation studies further validate the effectiveness of each module.

Primary area:
reinforcement learning
Keywords:
Auto-bidding, Generative Model, Disentangled Representation, Training-free Adaptation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YgytOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of personalized bidding strategies in online advertising auto-bidding and proposes the ADAPT framework. The core ideas include: (1) a two-stage training paradigm — in Stage 1, static and dynamic advertiser profiles are extracted from an advertiser memory bank via contrastive learning; in Stage 2, the dynamic profile is disentangled into a “common profile” and a “private profile” through a correlation loss, with all three profiles jointly conditioning the bidding decision; (2) Training-free Adaptation — for new advertisers, zero-shot cold start is achieved using the common profile, while for existing advertisers, profiles are updated online via exponential moving average (EMA) without retraining. On the real-world Alibaba AuctionNet dataset, ADAPT outperforms baselines including USCB, BCQ, CQL, IQL, DiffBid, DT, CDT, DT-score, and GAS under various budget settings.

Critical strengths

1.Practical significance of the problem setting. Auto-bidding is a core component of advertising systems. Introducing advertiser profiles into auto-bidding and addressing the challenges of their construction and utilization represents a research direction with genuine practical value. 2.Reasonable engineering consideration in training-adaptation decoupling. Decoupling model training from online profile updates and employing a training-free approach (EMA + common profile distillation) for online adaptation aligns well with industry needs for low maintenance costs. 3.Comprehensive experiments. The paper compares nine baseline methods on a real-world dataset and conducts multiple ablation studies, including profile effectiveness, disentanglement effectiveness, zero-shot cold start, and training-free updates. The appendix also includes hyperparameter sensitivity analysis and profile visualization.

Critical weaknesses

1.The fusion mechanism of the three profiles in Stage 2 is not clearly described. How do the static profile, common profile, and private profile interact with the Decision Transformer? Is it through concatenation, cross-attention, or another mechanism? Figure 1 shows a Cross Attention Layer, but the main text lacks corresponding formulas and dimension specifications.

2.Missing comparison with recent generative auto-bidding methods. The references mention LBM (Li et al., 2026, WWW), but this work is not included in the baseline comparisons. Given that LBM is a recent method combining large models with auto-bidding, the lack of comparison with it weakens the experimental persuasiveness.

3.CPA constraint satisfaction is not reported. The core constraint in auto-bidding is CPA (Cost Per Acquisition), yet the experimental metrics only report Score and Conversion, without reporting the CPA constraint satisfaction rate or Budget Consumption Rate. This makes the assessment of practical effectiveness incomplete.

4.The average cosine similarity of the common profile is as high as 0.9027 (Table 9), which may actually indicate insufficient disentanglement. If the common profiles of all advertisers are nearly identical, the role of the “common profile” may degenerate to a shared bias term, contributing little to personalized decision-making. The authors should further discuss this phenomenon and analyze the mutual information between common and private profiles to quantitatively demonstrate the disentanglement effect.

Questions for authors

1.Please provide the precise mathematical definition of the correlation loss L_cor, and explain how its gradient is backpropagated.

2.How do the methods perform in terms of CPA constraint satisfaction rate and budget consumption rate?

3.Does the average common profile similarity of 0.9027 indicate insufficient disentanglement? Please further analyze this from the mutual information perspective.

4.Why was a 16-dimensional state and 1-dimensional action space chosen? Were larger dimensions attempted?

1 comment

We are pleased that our clarifications and additional experiments have helped address your concerns, and we sincerely appreciate your continued positive assessment. In particular, thank you for recognizing the genuine practical value of introducing advertiser profiles into personalized auto-bidding, the industry-aligned, low-maintenance design that decouples model training from training-free profile adaptation, and the comprehensive experimental evaluation, including extensive baseline comparisons, component ablations, zero-shot cold start, and training-free updates. Together, these strengths highlight ADAPT’s value in supporting personalized bidding and continued advertiser adaptation without advertiser-specific retraining. Your constructive feedback has helped us clarify the methodology and strengthen the empirical evidence in the revised manuscript.

Reviewer_HkJbOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how automated bidding can account for advertiser differences and adapt to new or changing advertisers without retraining. It proposes ADAPT, which uses advertiser profiles to guide a Decision Transformer. The method first learns category-level and advertiser-specific profiles from historical bidding trajectories through contrastive learning. It then derives common and private representations and uses them as prefixes for bid prediction. During deployment, existing profiles are updated through a weighted average of old profiles and new trajectory encodings. Experiments report the best bidding scores among the tested methods across five budget settings.

Critical strengths
  1. The paper addresses advertiser differences and adaptation without retraining.
  2. ADAPT reports the best scores across both AuctionNet variants and all five budget settings.
  3. The paper includes component ablations, profile perturbations, and tests on unseen advertisers.
Critical weaknesses
  1. For stage 1 learning, the contrastive pairs are not well justified. It is very important to design negative samples in contrastive learning. However, they just treat different advertisers as negatives, even when advertisers may have similar strategies. Whether this is reasonable is highly questionable. If we assume that advertisers serve as negative examples for one another and disentangle their representations, how can we ensure that there are some common features are learned during the second stage?

  2. The common/private feature split is also not sufficiently justified. The common profiles of different advertisers have a high similarity as shown in Figure 4 (90%). This raises a question: given that this feature is so similar across the board, why is it even necessary?

  3. The method mainly combines existing techniques. In particular, profile updating simply averages a new encoding with the old profile using a fixed weight. This is a basic smoothing rule, with little design specific to advertiser adaptation.

Questions for authors

Please see my weaknesses.

4 comments

Thank you again for your constructive feedback. As the rebuttal deadline approaches, please feel free to share any remaining concerns or questions about our responses. We would be happy to discuss them and provide further clarification.

Reviewer_HkJbOpenReview ↗

Thanks for the Reply.

Thank you for the authors' response. The additional explanations and experiments have helped address some of my concerns. Although I still have concerns about the rationale and novelty of the proposed method, given the promising experimental results, I am willing to raise my score. (A score of 2.5 might better reflect my evaluation.)

  • Negative sample selection in contrastive learning. I understand the authors' argument that such a simple negative sampling strategy can work. However, based on my experience, the selection of negative samples can significantly affect the final performance. The authors do not seem to have designed a negative sampling strategy specifically for this scenario. Instead, they directly apply contrastive learning and assume that different advertisers can serve as negative samples for each other. Using images as an analogy, this would be equivalent to treating any two different images as negative samples simply because they are not identical. This seems somewhat unreasonable to me, which is why I raised the question about negative sample selection in the first place. Given the promising experimental results, perhaps this strategy happens to work well due to the characteristics of the dataset. I would appreciate some case studies illustrating where the differences between advertisers actually come from. Are advertisers really so different from one another that they can all be treated as negative samples without any selection or filtering, without affecting performance? Alternatively, is the number of advertisers in the dataset sufficiently large that random sampling is unlikely to select similar advertisers?

  • Novelty of the adaptation. I understand that the authors consider applying the common + private framework to advertiser modeling as part of their novelty. However, this framework has already been applied in many scenarios, including product recommendation and user modeling, as the authors themselves mentioned. Meanwhile, I do not see any specific adaptations made for the advertiser scenario, such as the negative sample selection I mentioned above or the strategy of using averaging for adaptation. The same approach could be applied to advertisers, users, or products, and the underlying techniques are already well established. This makes me feel that the proposed method is essentially a straightforward application of existing techniques, without sufficiently considering what makes the advertiser scenario unique.

AuthorsOpenReview ↗

Further Clarifications on Negative Sampling and Method Novelty

We sincerely thank you for your positive reassessment and recognition of the promising experimental results. We are pleased that our clarifications and additional experiments have helped address several of your concerns. We further clarify the remaining questions below.

1. Negative Sampling

Our negative pairs encode advertiser–trajectory correspondence, not mutually exclusive strategies. Different periods of the same advertiser are aligned with one individual profile, even when their observations and actions differ. Eq. (5) encourages correct correspondence without requiring profiles to be orthogonal or removing shared information. Related evidence exists in CLIP-based research: LoCoOp[1] identifies retained background information, while ID-like[2] Prompt Learning discusses cross-category semantic similarities despite contrastive pretraining with nonmatching image–text pairs. These findings support the compatibility of contrastive discrimination with retaining shared features, providing additional justification for our two-stage design.

Following your suggestion, we conducted a new advertiser-level case study and added it to the manuscript as Figure 6. Advertisers 2 and 13 have identical budgets (2250) and target CPAs (7), yet exhibit distinct stepwise bidding, spending, and conversion patterns: Advertiser 13 spends in two earlier bursts, whereas Advertiser 2 concentrates spending and conversions near the end. Thus, identical constraints do not imply interchangeable bidding histories. Figure 5 further shows that substituting profiles changes bidding behavior for the same target advertiser. Meanwhile, Table 9 shows highly similar common profiles, and shuffling common profiles causes a smaller Score decrease than shuffling private profiles (3.27% versus 5.45%). Together, these results support transferable shared context alongside advertiser-dependent conditioning, rather than assuming entirely different strategies across advertisers.

2. Novelty

Our novelty lies in extracting and maintaining actionable advertiser strategy priors, not in the decomposition or averaging operations themselves. In recommendation, LLM-based semantic profiling can exploit interpretable user attributes and interaction logs[3][4][5]. Auto-bidding, however, requires interpreting temporally coupled numerical return–state–action sequences without ready-made textual descriptions of strategy. Their strategic meaning lies in how decisions respond to opportunities under budget and CPA constraints. Extracting persistent, transferable decision context from these sequences is therefore a distinct challenge that cannot be addressed simply by transferring text-based profiling methods.

ADAPT addresses this challenge through cross-period contrastive learning that consolidates historical interactions into persistent advertiser memory. Advertiser-conditioned attention retrieves shared experience, while the private branch preserves advertiser-associated differences; both condition the DT alongside its current trajectory. Prototype-based initialization and subsequent EMA updates connect cold-start guidance with continued personalization while keeping encoder and bidding-model weights fixed. Our ablations and visualizations support the resulting profiles’ effectiveness and clearer shared/private functional roles. The contribution is this coordinated transformation of numerical bidding experience into reusable, personalized, and continuously maintained decision priors—not a simple combination of existing techniques.

[1] Miyai, A., Yu, Q., Irie, G., & Aizawa, K. (2023). LoCoOp: Few-shot out-of-distribution detection via prompt learning. In neural information processing systems (pp. 76298–76310).

[2] Bai, Y., Han, Z., Cao, B., Jiang, X., Hu, Q., & Zhang, C. (2024). ID-like prompt learning for few-shot out-of-distribution detection. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (pp. 17480–17489).

[3] Wang, L., Zhang, D., Yang, F., Zhao, P., Liu, J., Zhan, Y., ... & Zhang, Q. (2025, August). Lettingo: Explore user profile generation for recommendation system. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2 (pp. 2985-2995).

[4] Zhang, Y., Zhang, Y., Wang, Y., Chen, T., & Yin, H. (2026, August). ProEx: A Unified Framework Leveraging Large Language Model with Profile Extrapolation for Recommendation. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 (pp. 1940-1951).

[5] Zhang, J. (2024, November). Guided profile generation improves personalization with large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024 (pp. 4005-4016).

We hope these clarifications further address your concerns. If you have any additional questions, we would be happy to discuss them with you.

Reviewer_k7mwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ADAPT, a personalized auto-bidding framework that conditions a Decision Transformer on advertiser profiles. Stage 1 learns category embeddings and advertiser-specific profiles through trajectory-based contrastive learning. Stage 2 transforms individual profiles into common and private components, regularized by a decorrelation loss. At deployment, profiles are updated through an exponential moving average without gradient updates, while unseen advertisers use category information and a shared prototype. Experiments on the dense and sparse AuctionNet variants report improvements over nine baselines across five budget scales, supported by component ablations and adaptation evaluations.

Critical strengths
  1. A coherent approach to personalization and adaptation. ADAPT integrates historical advertiser information, cross-advertiser sharing, and gradient-free profile updates into a single bidding framework. Avoiding per-advertiser retraining is practically appealing, and the framework explicitly addresses both existing and unseen advertisers.

  2. Consistent empirical improvements with useful ablations. ADAPT achieves the highest reported scores across all ten benchmark settings. Profile removal/shuffling experiments, the correlation-loss ablation, and adaptation evaluations provide evidence that the proposed components contribute beyond an unconditioned Decision Transformer.

Critical weaknesses
  1. Advertiser-level profiles may conflate heterogeneous campaigns. In practical advertising systems, one advertiser often runs multiple campaigns with different objectives, budgets, creatives, and lifecycle stages. A newly launched campaign may have sparse conversion feedback, while an established campaign may experience declining conversion performance. These differences motivate my concern about using a single advertiser-level profile. The paper does not clarify whether an AuctionNet “advertiser” represents a multi-campaign account or a single campaign-level bidding agent, nor how the model accommodates within-advertiser heterogeneity. If profiles are account-level, their changes may reflect campaign composition rather than advertiser preferences; if the AuctionNet benchmark is effectively campaign-level, applicability to multi-campaign accounts remains unestablished.

  2. The additional information provided by profiles is not sufficiently identified. In deployed systems, conversion prediction models already estimate impression-level conversion probabilities conditioned on the ad and traffic context, while bidding controllers handle budget pacing and cost constraints. This does not make bidding-side profiles inherently redundant, but it raises the question of what information they add. The paper’s state already contains current and historical predicted conversion-probability means. Since campaigns with identical means can differ in their probability distributions and relationships with auction prices, profile gains could compensate for information lost through state aggregation rather than capture advertiser preferences. The current experiments do not distinguish these explanations, leaving the necessity and interpretation of the proposed profile modeling unclear.

  3. The transition from cold-start priors to actual personalization is unspecified. A practical cold-start mechanism should not only provide initial bidding guidance, but also become advertiser-specific as observations accumulate. ADAPT initializes unseen advertisers using a category embedding, a global common-profile prototype, and a zero private profile. This is prior-based guidance rather than an advertiser-specific profile at initialization. The paper does not explain how subsequent trajectories create a new memory entry or when the model switches to personalized profiles. Consequently, the reported zero-shot results validate initial guidance, but not the complete new-advertiser adaptation process.

  4. The evidence does not establish semantic disentanglement. Separating shared bidding knowledge from advertiser-specific information is potentially valuable because the former should transfer across advertisers while the latter should preserve individual differences. However, Eq. (8) only penalizes the centered cosine similarity of the two vectors within each sample. Decorrelated vectors can still encode the same underlying information through different transformations. The reported similarity patterns therefore do not establish that the components have the claimed shared/private semantics. Without stronger validation or a capacity-matched alternative, the improvement may arise from additional profile transformations and regularization rather than semantic disentanglement.

  5. The update mechanism may not adequately handle realistic nonstationarity. Campaign launches, creative fatigue, and traffic shifts can change bidding conditions abruptly, so adaptation must balance historical stability against responsiveness. ADAPT uses a fixed EMA coefficient of 0.99, which may lag behind such changes, but the paper does not report controlled evaluations of abrupt shifts or campaign deterioration. There is also a representation-consistency concern: Eq. (12) averages a raw memory embedding with a LayerNorm-normalized encoder output, although the training objective does not explicitly ensure compatible scales. These issues require clarification and validation before the update can be considered reliable adaptation.

  6. The benchmark provides limited evidence for deployment at realistic advertiser scale. The proposed method relies on cross-advertiser information sharing through a memory bank, making both population diversity and population size relevant to its effectiveness. Yet both evaluated benchmark variants contain only 48 advertisers; a large number of impressions does not compensate for limited advertiser diversity. The paper reports neither larger-population evaluations nor real-world online A/B experiments. Thus, the results support effectiveness within this benchmark, but provide limited evidence for scalability and robustness under live feedback and changing advertiser/campaign populations.

Questions for authors

See the Weakness.

2 comments

Thank you again for your constructive feedback. As the rebuttal deadline approaches, please feel free to share any remaining concerns or questions about our responses. We would be happy to discuss them and provide further clarification.

Dear Reviewer k7mw,

Thank you again for your constructive feedback. With one day remaining before the rebuttal deadline, we would appreciate the opportunity to address any remaining concerns.

Our clarifications and additional experiments have helped address overlapping concerns from the other reviewers: Reviewer HkJb raised their score to 3, while Reviewer Ygyt maintained their score of 3. We have also added Figure 6, showing distinct bidding and spending patterns between advertisers with identical budget and CPA constraints, providing further context for your question about the information represented by advertiser profiles.

Please let us know if further clarification would be helpful. We would be happy to address your questions and discuss them with you.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.