acceptodds
Under review as a conference paper at ICLR 2027

AMINO: Generative Auto-bidding with Multimodal Trajectory Modeling and Initial Noise Optimization

Abstract

Auto-bidding is a critical technique in online advertising for achieving advertisers' commercial goals under constraints. Recent generative auto-bidding methods have shown strong potential by learning bidding policies directly from large-scale offline data. In this setting, auto-bidding trajectories are inherently multimodal, making conditional diffusion particularly suitable for trajectory modeling. However, our theoretical analysis reveals that conditional diffusion can suffer from mode concentration under the high target returns commonly used to steer generation toward high-value trajectories, potentially resulting in suboptimal bidding trajectories. Moreover, existing diffusion-based methods lack an explicit mechanism to exploit task feedback and improve beyond static offline data. We therefore propose AMINO (generative Auto-bidding with Multimodal trajectory modeling and Initial Noise Optimization), a novel framework that decouples performance optimization from multimodal trajectory modeling. Rather than modifying the diffusion planner, AMINO keeps it fixed and learns an initial noise policy via reinforcement learning for performance optimization. To address the challenges of delayed feedback and offline optimization in auto-bidding, we develop a tailored noise optimization pipeline, enabling reliable trajectory-level feedback and stable exploration in the continuous noise space. Extensive experiments on a large-scale real-world dataset demonstrate the state-of-the-art performance of AMINO.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Auto-bidding, Generative Models, Conditional Diffusion, Offline Reinforcement Learning

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_kvcjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces a generative auto‑bidding framework for online advertising built upon conditional diffusion trajectory modelling. The authors first identify and theoretically prove a mode‑concentration failure. Instead of fine‑tuning the full diffusion model, AMINO decouples multimodal trajectory generation and performance optimization. Evaluations are conducted on AuctionNet real‑world advertising benchmark. AMINO outperforms prior offline RL, Decision‑Transformer‑based, diffusion‑based and LLM‑based auto‑bidding baselines.

Critical strengths
  1. Novel theoretical proof of mode‑concentration under high return conditioning for diffusion‑based auto‑bidding. The paper derives Proposition 1 showing extreme target‑return conditioning collapses trajectory mode entropy. This explains a key practical failure mode widely ignored in existing diffusion auto‑bidding literature.

  2. Effective decoupled optimisation pipeline tailored to auto‑bidding constraints and delayed trajectory‑level feedback. By freezing the pretrained diffusion planner and only optimising initial noise via GNPO, AMINO preserves the pre‑learned multimodal trajectory distribution and avoids the risk of destroying multimodality that comes from end‑to‑end RL fine‑tuning of diffusion networks. Empirical improvements over state‑of‑art baselines validate the practical value.

Critical weaknesses
  1. Lack of online‑interaction / offline‑to‑online evaluation. All experiments run purely within offline‑simulated benchmark environments. In real auto‑bidding systems, market distributions usually shift after deployment. It remains untested whether the optimised noise policy degrades severely when facing real online distribution drift.

  2. Limited comparison against alternative latent‑space RL for diffusion. While the paper compares against CEM and GMM‑based baselines, it provides only limited discussion and empirical contrast against existing general‑purpose initial‑noise RL methods from broader diffusion literature. It is unclear how much performance gain comes from the auto‑bidding‑specific GNPO group‑relative formulation versus vanilla latent‑space RL adapted to this setting.

Questions for authors
  1. For Proposition 1’s mode‑concentration proof, the analysis assumes mode‑conditional returns follow Gaussian distributions. How well does this Gaussian assumption hold for real‑world auto‑bidding trajectory returns in AuctionNet?

  2. In GNPO, group‑relative advantage relies on sampling G noise candidates per history. How sensitive is final performance to group size G?

  3. Have you attempted adapting existing general latent‑noise RL algorithms to this auto‑bidding setup for direct comparison?

  4. What obstacles prevent adding online‑shift simulation experiments on AuctionNet to evaluate how AMINO’s noise policy degrades under distribution shift? Could simple fine‑tuning of only the noise policy address the online adaptation?

5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

We sincerely thank Reviewer kvcj for the insightful comments. The remainders focus on concerns to address.


1. Real-World Evaluation and Adaptation under Distribution Shift (W1 & Q4)

Thank you for raising this important concern. As acknowledged in our Limitations section, safe online adaptation remains an important challenge due to distribution shifts and potentially unstable policy updates. Direct online deployment also faces operational safety risks and deployment costs. To further assess AMINO's practical applicability and adaptability to changing market conditions, we conducted two additional experiments.

(a) Evaluation on Real-World Industrial Data

To evaluate AMINO beyond the AuctionNet benchmark, we conducted additional experiments using recent production data collected in October 2026 from a large-scale e-commerce advertising platform. Unlike AuctionNet's benchmark simulator, this evaluation uses actual impression-level traffic, market prices, and predicted values from a production advertising system.

Under joint budget and ROAS constraints, we train the model offline using seven consecutive days of real bidding trajectories and evaluate it on a separate day. The evaluation follows the platform's bidding and budget-control rules, adjusting bidding parameters every 15 minutes and replaying auctions against recorded market prices. Performance is measured using normalized GMV with penalties for ROAS constraint violations.

Compared with DT, AMINO achieves the following improvements:

MetricGMVCostTarget CostROI
Ours vs. DT+6.17%+2.86%+3.31%+3.22%

These results provide further evidence of AMINO's effectiveness in practical advertising settings. While this remains an offline replay evaluation rather than an online A/B test, it incorporates real production market conditions. We have included these results in Section 6.7 of the revised manuscript.

(b) Adaptation under Distribution Shifts

To investigate whether fine-tuning only the noise policy can adapt to changing market conditions, we consider three distribution shifts in AuctionNet: (1) market prices increasing by 25%, (2) traffic volume decreasing by 20%, and (3) conversion probabilities decreasing by 20%.

Starting from the pretrained AMINO checkpoint, we freeze the diffusion planner and inverse dynamics model, updating only the noise policy through episode-level group-relative policy optimization using simulator rewards, without the trajectory evaluator. Adaptation is performed on a separate delivery period (period 8), while evaluation uses the original test period (period 7) under the same shift.

Distribution ShiftDTAMINOAMINO + Noise Policy AdaptationImprovement
Market Price +25%238275288+4.9%
Traffic Volume −20%302333351+5.5%
Conversion Probability −20%224253277+9.5%

The results show that AMINO maintains a substantial advantage over DT across all three distribution shifts, while fine-tuning only the noise policy yields further improvements of 4.9%–9.5%. These findings demonstrate the effectiveness of lightweight noise-policy adaptation under changing market conditions without retraining the diffusion planner, while safe adaptation in real online systems remains an important direction for future work. We have included these results in Appendix E.4 of the revised manuscript.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

2. Gaussian Assumption in Proposition 1 (Q1)

Thank you for the helpful question. Proposition 1 aims to provide a theoretical explanation for the potential limitations of conditioning diffusion planners on excessively high return-to-go targets. For analytical tractability, we assume that returns within each trajectory mode follow a Gaussian distribution. However, mode concentration can also occur under non-Gaussian distributions, provided that a unique right-tail-dominant mode exists.

Following the reviewer's suggestion, we further conducted empirical validation of the Gaussian assumption on AuctionNet. We selected three representative bidding behavior modes—Aggressive, Uniform, and Conservative—and grouped trajectories with similar observed histories at decision steps t∈12,24,36t\in{12,24,36}. Gaussian distributions were fitted to the future returns within each history–mode group.

We evaluate the goodness of fit using excess kurtosis (deviation from Gaussian kurtosis), the Kolmogorov–Smirnov (KS) statistic (maximum difference between empirical and fitted cumulative distributions), and Q95 Error (relative error in the 95th-percentile return). The results are summarized below.

ModeExcess KurtosisKSQ95 Error
Aggressive0.2480.1046.47%
Uniform-0.0110.1127.67%
Conservative0.0540.1136.32%

The results show that all three modes exhibit excess kurtosis close to zero, with 95th-percentile fitting errors of 6.32%–7.67%. Although the KS statistics indicate some deviations from exact Gaussianity, these findings support Gaussian distributions as a reasonable approximation of conditional returns in AuctionNet.

We have included this empirical analysis in Appendix E.6 of the revised manuscript.


3. Group Size Sensitivity (Q2)

We conducted additional experiments to evaluate the sensitivity of GNPO to the group size GG, while keeping all other settings unchanged.

GG8163264128
Score351363374371366

The results demonstrate that GNPO is relatively insensitive to the group size. Despite a 16-fold variation in GG (from 8 to 128), the maximum relative Score difference is only 6.1%. Moreover, all configurations consistently outperform Diffusion Only (318) by at least 10.4%, demonstrating the robustness of GNPO across different group sizes.

We have included these results in Appendix E.4 of the revised manuscript.


4. Comparison with General Latent-Noise RL Methods (Q3 & W2)

Thank you for the valuable comments. We would like to clarify that AMINO's contributions extend beyond GNPO to a tailored noise optimization pipeline for auto-bidding, integrating a trajectory evaluator, GNPO, and conservative optimization. These components jointly address key challenges in directly applying general latent-noise RL methods to auto-bidding, including delayed trajectory-level feedback and unreliable reward estimation in offline settings.

To further validate its effectiveness, we adapt two representative latent-noise RL approaches to auto-bidding: Noise PPO-style, based on PPO, and DSRL-style, based on SAC-style actor–critic optimization[1][2l]. Their configurations are kept as consistent with AMINO as possible to ensure a fair comparison. The results are summarized below.

MethodDSRL-styleNoise PPO-styleNoise PPO-style + Reward ShapingAMINO
Score ↑339343351374

AMINO outperforms both approaches, achieving a 6.6% improvement over Noise PPO-style with reward shaping and a 10.3% improvement over DSRL-style. Notably, reward shaping improves Noise PPO-style from 343 to 351, further supporting the effectiveness of conservative optimization.

We have added these comparisons to Appendix E.3 and include Noise PPO in the Related Work section.


References:

[1] Ren, A. Z., et al. (2025). Diffusion Policy Policy Optimization. The Thirteenth International Conference on Learning Representations (ICLR).

[2] Wagenmaker, A., et al. (2025). Steering Your Diffusion Policy with Latent Space Reinforcement Learning. Proceedings of the 9th Conference on Robot Learning (CoRL), PMLR, 305, 258–282.

AuthorsOpenReview ↗

Summary Response

We sincerely thank you for the valuable comments and for recognizing the novelty of our theoretical analysis of mode concentration, the effectiveness of our decoupled optimization framework, and the performance improvements over existing methods.

We are also encouraged that the other two reviewers independently highlighted the novelty and significance of our main contributions. Reviewer Bage particularly appreciated the novel perspective of treating initial noise as a learnable policy and the comprehensive experimental evaluation. Reviewer y5SE emphasized the theoretically and empirically supported analysis of mode concentration, as well as the simplicity and effectiveness of separating multimodal trajectory generation from performance optimization.

Motivated by your suggestions, we have substantially strengthened our experimental evaluation, including additional experiments on real-world industrial advertising data, simulated distribution shifts and noise-policy adaptation, empirical validation of the Gaussian assumption, group-size sensitivity analysis, and direct comparisons with general latent-noise RL methods. These new results further support both the practical effectiveness and robustness of AMINO.

We hope these additional experiments and clarifications have adequately addressed your concerns and strengthened the evidence for our contributions. If you find our responses satisfactory, we would be sincerely grateful if you could reconsider your overall rating. We welcome any further discussion or suggestions.

Reviewer_y5SEOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces AMINO, a method for generative automated bidding that optimizes the initial noise of a frozen conditional diffusion planner. The authors observe that setting an excessively high target-return condition leads to mode collapse, which reduces the diversity of generated trajectories. To address this, AMINO employs a group-relative noise strategy, a trajectory-level evaluator, and conservative reward shaping. Experiments on AuctionNet demonstrate that AMINO consistently outperforms existing baselines for automated bidding.

Critical strengths
  1. A well-grounded, novel diagnosis of the core failure mode. The paper identifies and names mode concentration: extrapolated target returns progressively collapse the conditional trajectory distribution onto a single behavioral mode. This is supported both theoretically (Proposition 1, showing the conditional mode entropy → 0 as y* → +∞) and empirically (Fig. 1: Best-of-N gains saturate as the target return grows because diversity is lost).
  2. The second contribution is a clean decoupling of performance optimization from multimodal modeling: by freezing the pretrained diffusion planner and optimizing only the initial noise via GNPO, the approach sidesteps the credit-assignment cost of RL-fine-tuning the entire denoiser (Black et al., 2024; Clark et al., 2024) while preserving the multimodality acquired during pretraining. The supporting machinery — a type-embedding / future-only trajectory evaluator, group-relative advantages, and conservative optimization — is purpose-built for auto-bidding's delayed, trajectory-level, constraint-bound feedback signals.
Critical weaknesses
  1. While the paper describes how the evaluator is trained, it fails to sufficiently demonstrate that the evaluator remains reliable on the trajectories generated during the GNPO (Generative Policy Optimization) process. The evaluation metrics reported are derived from a held-out subset of the offline dataset, which may significantly underestimate the distribution shift introduced by GNPO. Since the policy itself explicitly maximizes the evaluator’s score, the exploitation of the evaluator is a fundamental issue rather than a minor implementation detail. Although parameter clipping and reward shaping serve as useful heuristics, the paper provides no analysis of out-of-distribution (OOD) calibration, nor does it offer formal or empirical evidence that the optimized trajectories consistently remain within the region where the evaluator is reliable.
  2. AMINO samples multiple initial noise vectors, runs the diffusion planner for each candidate, and evaluates all generated trajectories before updating or selecting a candidate solution. Compared to baseline methods that rely on a single diffusion sample, this approach grants AMINO greater resources for both generation and evaluation. Although the paper includes a "Best-of-N" ablation study, this comparison fails to fully disentangle the benefits derived from a learned noise policy from those gained simply by increasing the number of candidate samples. The paper should report performance metrics normalized by computational cost, ensuring that all methods are compared under conditions where both the number of diffusion forward passes and the number of evaluator calls are held constant.
Questions for authors
  1. Is a single diagonal Gaussian sufficient to represent multiple, well-separated regions of high-quality (or low-noise) data? If "good" trajectories originate from distinct modes while "bad" trajectories lie between these modes, how does GNPO avoid averaging them into a single distribution that falls within a low-quality region?
  2. Given that only the first predicted state is ultimately converted into an executable action, what is the specific purpose of the full future sequence? Is it intended for long-term budget/CPA constraints, supervision by an evaluator, or both? Furthermore, how does it compare to short-horizon generation or single-step prediction—could similar performance be achieved at a lower inference cost?
3 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank Reviewer y5SE for the insightful comments. The remainders focus on concerns to address.


1. Evaluator Reliability on OOD Trajectories (W1)

Thank you for highlighting this concern. To assess evaluator reliability under OOD conditions, we collected actual bidding trajectories from 48 advertisers in the AuctionNet simulator by executing AMINO. We then applied the frozen evaluator to these trajectories and compared its predictions with the ground-truth Scores.

True ScorePred. ScoreMAEBiasSpearmanTop-25% Bias
37438218.6+7.595.7%+13.2

These results demonstrate strong ranking consistency and low prediction errors, with moderate overestimation for high-scoring trajectories, supporting the evaluator's reliability on trajectories induced by the optimized policy.

Nevertheless, we acknowledge that an offline-trained evaluator cannot guarantee accurate predictions for arbitrary OOD trajectories. Gaussian parameter clipping and conservative reward shaping help mitigate evaluator exploitation but cannot eliminate this risk entirely, which remains a limitation of our approach.

We have added these results to Section 6.4 of the revised manuscript.


2. Computational Budget and Fairness (W2)

We clarify that multiple noise samples and evaluator calls are required only during GNPO training. At inference, AMINO samples only one noise vector and uses the same 20 DDIM denoising steps as DiffBid, requiring only one additional noise-policy forward pass.

MethodDDIM StepsNoise SamplesInference Time (ms/decision)
Diffusion Only201147.8
AMINO201150.5

The additional noise-policy forward pass incurs only 2.7 ms per decision. Therefore, AMINO's performance gains stem from the learned noise policy rather than an increased number of candidate samples.


3. Expressiveness of the Gaussian Noise Policy (Q1)

The generation process of AMINO can be summarized as:

Observed History → Noise Policy → Gaussian Noise → Diffusion Planner → Future Trajectory

Although the noise policy is a single diagonal Gaussian, the nonlinear diffusion planner provides multimodal generation capability, allowing different noise samples to produce distinct trajectory modes.

Regarding the averaging concern, GNPO optimizes the noise policy through trajectory-level rewards rather than directly averaging high-quality noise samples. If the policy assigns probability to low-quality regions between high-quality modes, the evaluator provides lower scores, and GNPO encourages the policy to shift probability toward higher-reward regions rather than their low-quality average. This is supported by our CEM-based regression comparison in Appendix E.1, where CEM achieves a Score of 257, compared with 374 for GNPO.


4. Full-Horizon Trajectory Planning (Q2)

Full-horizon generation provides two key benefits: (1) It captures the temporal correlation between current bidding decisions and future budget allocation, enabling long-term planning. (2) It allows the trajectory evaluator to assess future cumulative returns and CPA constraint satisfaction, providing long-term feedback for GNPO optimization.

To examine the impact of planning horizon, we conducted additional experiments using only the diffusion planner, varying the horizon HH during both training and inference while keeping all other settings unchanged.

Horizon HH1122448
Score289305311318
Inference Time (ms/decision)144.2145.1145.9147.8

Full-horizon planning improves the Score by 10.0% over single-step prediction, with only 3.6 ms additional inference latency. Computation is dominated by 20 DDIM denoising steps, while increasing the horizon incurs relatively little additional parallel computation. This latency is well within the requirements of industrial auto-bidding systems, where bidding parameters are typically adjusted at coarse intervals (e.g., every 30 minutes) rather than for individual impressions.


We sincerely thank the reviewer again for the helpful feedback, which has contributed to improving our work. We hope our responses have adequately addressed your concerns and welcome any further questions or suggestions.

AuthorsOpenReview ↗

Summary Response

We sincerely thank you for the insightful feedback and for recognizing the novelty and theoretical motivation of mode concentration, as well as the effectiveness of our decoupled noise optimization framework and its task-specific design.

We are encouraged that the other reviewers also acknowledge these contributions. Reviewer Bage highlighted the novel perspective of learning an initial-noise policy rather than directly optimizing individual noise vectors, along with the breadth of our experimental evaluation. Reviewer kvcj similarly recognized the novel theoretical analysis of mode concentration, the effectiveness of our tailored optimization pipeline, and its superior empirical performance.

Following your suggestions, we have conducted additional experiments to evaluate the trajectory evaluator on AMINO-generated trajectories, clarified the computational fairness of inference, and investigated the expressiveness of the Gaussian noise policy and the benefits of full-horizon planning. These results provide further evidence supporting the reliability and effectiveness of our approach. We hope our responses have addressed your concerns and would sincerely appreciate your reconsideration of the overall assessment. We would be happy to discuss any remaining questions.

Dear Reviewer,

As the rebuttal period is approaching its deadline, we would greatly appreciate it if you could share any remaining concerns soon so we can address them in time. If our responses have resolved your concerns, we kindly invite you to reconsider your score.

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

The paper argues that conditioning diffusion bidders on target returns far above the offline maximum collapses the generated trajectory distribution onto a single mode, which it calls mode concentration and formalizes in Proposition 1. AMINO instead freezes the pretrained planner at a moderate target return and optimizes the DDIM initial noise: GNPO updates a history-conditioned Gaussian noise policy by group-relative advantage, using a trajectory evaluator (future return times a CPA penalty) as the reward, with parameter clipping, entropy regularization and reward shaping for stability. Evaluated on AuctionNet and AuctionNet-Sparse over five budget ratios against ten baselines, with component ablations.

Critical strengths

[S1] Method design. Separating performance optimization from distribution modeling, and acting only on the sampling entry point while leaving the planner untouched, avoids credit assignment through the denoising chain and keeps the pretrained generative behavior intact.

[S2] Novelty of the framing. Treating the initial noise as a learnable policy, rather than searching for or regressing onto a single good noise vector, is a useful shift of perspective for conditional generation in sequential decision making.

[S3] Experimental completeness. The study spans dense and sparse feedback, five budget levels, and four families of baselines (online RL, offline RL, decision transformers, diffusion planners), with component ablations and a check on the reward model's own prediction quality.

Critical weaknesses

[W1] The main text never states the target return actually used, its relation to γmax\gamma_{max}, or the guidance strength. It is therefore impossible to tell whether AMINO's planner still operates in the regime Proposition 1 describes. If the chosen condition sits inside the data range, the method avoids mode concentration by configuration rather than resolving it, and the framing would need revising.

[W2] The noise policy is confined to a clipped region of the latent space, while the denoiser was only ever trained on samples from a standard Gaussian prior. In high dimensions a clipped Gaussian can sit almost entirely off that prior's typical set. No diagnostic of the learned noise distribution is reported, yet this is the most direct route by which the planner could be pushed out of distribution.

[W3] Removing parameter clipping collapses Score from 374 to 67 and degrades CPA Ratio from 0.91 to 1.92. The headline result therefore rests on constraint thresholds whose values and sensitivity are not reported in the main text.

[W4] Table 2 is internally inconsistent. A CPA Ratio of 0.91 is below one, which implies Score equals Conversions, i.e. 409, but 374 is reported; the Diffusion Only row shows the same discrepancy (318 against 359).

[W5] The evaluator target in Eq. 4 is functionally identical to the test metric Score, so AMINO optimizes a surrogate of the evaluation objective that no baseline has access to.

[W6] Proposition 1 is stated over the posterior of the data distribution, whereas the phenomenon of interest is the learned model's behavior outside its training support. Its tail-dominance criterion also follows from the Gaussian assumption rather than from any property of bidding returns.

[W7] DiffBid scores 152 at dense 100%, below IQL at 281 and below the paper's own Diffusion Only at 318, which is the same architecture. The claim that diffusion baselines are weak cannot be separated from a reproduction artifact.

Questions for authors

[Q1] What target return and guidance strength does AMINO use, and how do they compare with γmax\gamma_{max}?

[Q2] How does the learned noise distribution compare with the standard Gaussian prior, and does performance survive projecting samples back onto that prior?

[Q3] Please resolve the Score and CPA Ratio inconsistency, and report the distribution of CPA Ratio rather than its mean, since a mean below one is compatible with many violating episodes.

[Q4] How much of the advantage remains if the evaluator target is changed so that it no longer shares the functional form of Score?

[Q5] What target return and tuning protocol were used for DiffBid?

5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

We sincerely thank Reviewer Bage for the insightful comments. The remainders focus on concerns to address.


1.Target Return and Guidance Strength (W1 & Q1)

As reported in Appendix A.2 (Table 4), AMINO uses a target return of 1.0 and a CFG guidance strength of 0.5. The maximum return in the training dataset is ymax⁡=0.78y_{\max}=0.78 (all returns are normalized using the same scale). Our target return is only 1.28 times the maximum observed return, providing moderate high-return guidance while preserving trajectory diversity. In contrast, DiffBid uses a target return of 4.0 and a guidance strength of 1.2. Such aggressive return conditioning can lead to severe mode concentration, as demonstrated in Fig. 1.

Importantly, as discussed at the end of Section 4, mode concentration can be avoided in the first place rather than addressed after it occurs. AMINO therefore adopts moderate return conditioning to preserve trajectory multimodality, while optimizing the initial noise through GNPO to improve bidding performance. Our goal is not to restore diversity under extreme return conditioning, but to achieve performance improvements without relying on such aggressive guidance.


2.Learned Noise Distribution and Prior Deviation (W2 & Q2)

Thank you for the valuable comments. We first clarify that parameter clipping constrains the predicted Gaussian parameters, rather than directly clipping the sampled noise vectors. Therefore, the learned noise distribution retains unbounded Gaussian support.

To quantify the distributional shift, we evaluate the learned noise policy against the standard Gaussian prior using mean squared shift (MSS), average variance (AvgVar), per-dimension KL divergence (KL/dim), and the proportion of samples falling within the standard Gaussian's 95% radial interval (RC95).

DistributionMSSAvgVarKL/dimRC95
Standard Gaussian01.000095.08%
Learned Noise Policy0.1350.8370.14989.82%

The results show that 89.82% of learned noise samples remain within the standard Gaussian's 95% radial interval, while the per-dimension KL divergence is 0.149. These findings indicate that the learned policy introduces a controlled distributional shift rather than an extreme deviation from the prior's typical radial region. We have included these results in Appendix E.5 of the revised manuscript.

Furthermore, the clipping sensitivity analysis in our response to W3 shows that even when the mean clipping threshold is reduced from 0.5 to 0.05, AMINO achieves a Score of 334, outperforming standard Gaussian sampling (318) by 5.0%. This suggests that substantial mean shifts are not necessary for effective noise optimization. Recovering the exact standard Gaussian distribution would remove the learned sampling policy and reduce AMINO to Diffusion Only.


3. Parameter Clipping Sensitivity (W3)

As reported in Appendix A.2 (Table 6), we clip the predicted noise mean to [−uψ,uψ][-u_\psi,u_\psi], with uψ=0.5u_\psi=0.5. To further investigate the sensitivity to this threshold, we evaluate AMINO with different values of uψu_\psi:

uψu_\psi0.050.10.30.50.81.05.0
Score334359369374357307177

The results show that AMINO maintains strong performance across a range of clipping thresholds around uψ=0.5u_\psi=0.5, indicating that its effectiveness does not depend on a precisely tuned threshold. However, overly restrictive clipping can limit exploration in the noise space, while overly relaxed clipping allows larger mean shifts that may deviate substantially from the pretrained Gaussian prior, leading to performance degradation.

We have added these results to Appendix E.4 of the revised manuscript.


4. Score and CPA Ratio Consistency (W4 & Q3)

We clarify that Score, Conversions, and CPA Ratio are computed independently for each evaluation instance before averaging, rather than deriving Score from the averaged metrics.

As described in Appendix D.1, we follow AuctionNet's round-robin evaluation protocol, where the evaluated policy sequentially replaces each of the 48 bidding agents.

Therefore, a mean CPA Ratio below 1 does not imply that all evaluation instances satisfy the CPA constraint. Instances exceeding the constraint incur penalties, resulting in a lower average Score than average Conversions. To further clarify this, we report the distribution of CPA Ratio, the results show that 34.93% of evaluation instances violate the CPA constraint:

StatisticMeanMedianP25P75P90Violation Rate (>1)
CPA Ratio0.910.910.721.041.2134.93%
AuthorsOpenReview ↗

Official Comment by Authors (2/2)

5. Evaluator Objective (W5 & Q4)

In CPA-constrained auto-bidding, Score jointly measures conversion performance and CPA constraint satisfaction, making it a natural optimization objective. Both the CPA penalty and Score can be directly computed from the offline dataset. Therefore, this objective is not exclusive to AMINO and can also be adopted by other methods, without requiring any information beyond the offline dataset.

For diffusion-based methods, the target return can be replaced with the CPA-penalized Score as the conditioning signal. We conducted this experiment with Diffusion Only, improving its Score from 318 to 322 and reducing its CPA Ratio from 0.99 to 0.97. Moreover, GAS, the strongest baseline, already employs a Score-based Q-function to guide policy optimization [1].

Furthermore, we evaluate an alternative evaluator target in Fig. 4. By removing the CPA penalty and training the evaluator to predict Conversions alone, AMINO still achieves a Score of 362, maintaining state-of-the-art performance with improvements of 10.0% over DT and 13.8% over Diffusion Only. This demonstrates that the effectiveness of initial noise optimization does not depend on matching the exact functional form of the evaluation metric.


6. DiffBid Baseline and Experimental Settings (W7 & Q5)

We clarify that the DiffBid results in Table 1 are taken from previously published evaluations [1][2], rather than our own reproduction. Both studies report a Score of 152 for DiffBid on AuctionNet Dense at the 100% budget setting.

DiffBid uses a target return of 4.0 and a CFG guidance strength of 1.2. The main difference from our Diffusion Only planner lies in the training procedure. As described in Section 5.1, we randomly sample a decision step tt during training, treating s1:ts_{1:t} as observed history and st+1:Ts_{t+1:T} as the future to be generated. For both training and inference, the observed prefix is clamped to the ground-truth history at every diffusion step, i.e., τ1:tk←ht\tau_{1:t}^{k}\leftarrow h_t, such that diffusion is performed only over the unknown future. This ensures consistency between training and inference.

To verify this difference, we conducted an additional controlled experiment incorporating only history-aware trajectory completion into a DiffBid-style planner, while keeping all other settings unchanged. The resulting Score reaches 321, comparable to our Diffusion Only (318), while AMINO further improves it to 374 through initial noise optimization.


7. Scope of Proposition 1 and Gaussian Assumption (W6)

Proposition 1 is formulated on the posterior mode distribution of offline data, which the diffusion planner is trained to approximate. The proposition aims to provide a theoretical explanation for the limitations of excessively high return conditioning, thereby motivating an alternative approach to performance optimization. It is not intended as a theoretical guarantee for AMINO itself, but rather as an analysis of the limitations that motivate its design.

Regarding the Gaussian assumption, in many practical auto-bidding systems, including the one considered in this work, a fixed bidding policy is deployed for operational stability, and offline trajectories are collected from the same policy across multiple days. For a given advertiser under similar conditions, trajectory variations primarily arise from stochastic environmental factors, such as traffic fluctuations and auction outcomes. Since these variations aggregate numerous impression-level random effects, the resulting deviations in future returns can be reasonably approximated by a Gaussian distribution [3].

Importantly, mode concentration does not fundamentally depend on Gaussianity. The essential condition is the existence of a unique right-tail-dominant mode. Under this condition, Bayes' rule implies that the posterior probability of the dominant mode approaches 1 as the target return increases. The Gaussian assumption merely provides a tractable sufficient condition.


References:

[1] Li, Y., et al. (2025). GAS: Generative Auto-Bidding with Post-Training Search. Companion Proceedings of the ACM Web Conference 2025.

[2] Gao, J., et al. (2025). Generative Auto-Bidding with Value-Guided Explorations. Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 244–254).

[3] Mou, Z., et al. (2026). Enhancing Generative Auto-Bidding with Offline Reward Evaluation and Policy Search. The Fourteenth International Conference on Learning Representations (ICLR).


We sincerely thank the reviewer again for the helpful feedback, which has contributed to improving our work. We hope our responses have adequately addressed your concerns and welcome any further questions or suggestions.

AuthorsOpenReview ↗

Summary Response

We sincerely thank you for the constructive feedback and for recognizing the novelty of treating initial noise as a learnable policy, the decoupled optimization design, and the comprehensive experimental evaluation.

We are also encouraged that the other reviewers share a positive assessment of AMINO's core contributions. Reviewer y5SE particularly recognized the theoretically and empirically supported analysis of mode concentration and the effectiveness of separating multimodal generation from performance optimization. Reviewer kvcj likewise highlighted the novel theoretical analysis of mode concentration, the tailored optimization framework, and the performance improvements over existing methods.

In response to your concerns, we have provided additional analyses of the learned noise distribution and clipping sensitivity, clarified the evaluation metrics and theoretical assumptions, and conducted further controlled comparisons with existing methods. We hope these clarifications and new results have adequately addressed your concerns. If so, we would be grateful if you could reconsider your overall assessment of our work. We welcome any further questions or suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.