acceptodds
Under review as a conference paper at ICLR 2027

Myerson-guided Machine Learning for Near-Optimal Auction Design

10 views

Abstract

Myerson characterizes the revenue-optimal auction for single-item settings, but the optimal mechanism remains unknown in general multi-bidder, multi-item environments. Neural mechanisms provide an expressive approach to this problem, yet some existing methods achieve only limited revenue gains over Item-wise Myerson in larger auction settings. Our profile-level analysis of ALGNet reveals complementary strengths: its gains on low-competition profiles are consistent with cross-item interactions and probabilistic allocations, but it increasingly underperforms Item-wise Myerson as item-level competition intensifies. Motivated by this observation, we propose Myerson-guided Machine Learning (M2L), a three-stage framework that uses Item-wise Myerson as a structural teacher. M2L first imitates Myerson's allocation and payment rules, then autonomously explores cross-item interactions and probabilistic allocations, and finally targets profiles on which the learned mechanism still underperforms Myerson while preserving its advantages elsewhere. Experiments across multi-item benchmarks show that M2L is competitive with, and often outperforms, existing neural approaches in expected revenue, substantially reduces below-Myerson profiles, and maintains low empirically estimated ex-post regret.

Primary area:
optimization
Keywords:
Optimal Auction Design, Deep Learning, Mechanism Design

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KkAzOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

There are by now a number of papers on using deep-learning techniques to design revenue-maximizing auction mechanisms. This is one of them. In this one, the authors note that item-by-item Myerson (Myerson is optimal for selling a single item) often outperforms an existing DL approach (ALGNet) in larger settings, and build on this insight. First they try to explain this in terms of "competitive" items. The authors then use item-level Myerson as a teacher, and eventually try to move beyond the teacher. With this, they (apparently) achieve state of the art results.

Critical strengths

See summary. Also: This is the best paper in my batch, though it must be said that the rest of my batch was just disappointingly bad with lots of (poor) LLM writing etc. This one is well written. So thank you to the authors for at least putting something decent into my batch and I hope in the future we can narrow down the set of submissions with desk rejections so that reviewers only get papers of at least this paper's quality to read.

Critical weaknesses

I mostly have questions so I'm putting this all in questions for authors.

Questions for authors

Regret vs revenue: clearly the more regret one allows the more revenue one can get. While the regret of the mechanisms in this paper doesn't seem all that bad, it still feels like somehow we should address the revenue-regret tradeoff. How much more revenue should I expect for an extra unit of regret?

On that note, some alternative approaches attain no regret (exact strategy-proofness), which is a nice property. I agree with the authors that we don't need to only consider such approaches, but there it feels especially uncomfortable to directly compare revenue. (There was at least one place where the authors said something like "zero regret is impossible" by which I think they meant zero regret is impossible in their framework, but it would be good to be clearer.)

... though actually it seems there are now methods to get from approximate strategy-proofness to exact strategy-proofness. Do they apply to your technique? That would be nice. See Guo et al. "Extending Myerson's optimal auctions to correlated bidders via neural network interpolation" (NeurIPS 2026)

Maybe I didn't fully understand but it wasn't clear to me to what extent the number of competitive items really explains item-wise Myerson doing better. It seems you show a correlation between these things, but there are probably lots of other things that correlate too. Should I take this as being the mechanism (in the other sense of the word mechanism) behind why it does better? Why?

One intuition is the following: when there are few bidders for lots of items, that's precisely when selling item by item does not do well -- that's where you want to bundle. Are you looking at the other extreme?

It is not clear to me whether you consider OD-VVCA to fall under the same framework or not. OD-VVCA is motivated more in automated mechanism design terms. But then it seems your paper also is an example of automated mechanism design. I don't have serious concerns about this either way, but perhaps you should give some more background on all this in the paper.

3 comments

1.Revenue–regret trade-off. How much additional revenue can a seller gain by allowing more regret?

Quantifying the exact revenue gain from a given increase in regret remains an open problem for general multi-bidder, multi-item auctions. For a single additive bidder, converting an approximately strategy-proof mechanism into an exactly strategy-proof one yields a known upper bound (Rahme et al., Auction Learning as a Two-Player Game, 2021; Balcan et al., Mechanism Design via Machine Learning, 2005). Specifically, letting P0P_0 denote the optimal expected revenue under exact strategy-proofness, allowing expected ex-post regret up to RR can increase optimal expected revenue by at most 2P0R+R2\sqrt{P_0 R} + R.

However, introducing multiple bidders fundamentally changes the strategic problem, because each bidder’s outcome depends on the others’ reports and all allocations must satisfy joint feasibility constraints. Even without a general theoretical formula, the seller can choose a mechanism based on application needs: GemNet and OD-VVCA provide exact DSIC when strict truthfulness is required, while M2L offers an effective alternative when a very small estimated regret is tolerable in pursuit of higher revenue.

Q2. What did we mean by “zero regret is impossible” ?

We acknowledge that this phrasing was imprecise. By this statement, we meant that models such as RegretNet, ALGNet, and M2L directly parameterize allocation and payment rules and penalize estimated regret during training. Consequently, this procedure cannot guarantee strictly zero regret in practice. We will revise the sentence accordingly in the revised paper.

Q3.Can the exact strategy-proofness procedure of Guo et al. be applied to M2L?

The procedure by Guo et al. cannot be directly applied to M2L. But adapting it to achieve exact strategy-proofness in our model is a promising direction for future work. We note that the preprint by Guo et al. was posted on arXiv after our submission, and we will discuss it in the final version.

Technically, their method verifies or repairs single-dimensional allocation monotonicity and then derives critical payments in a multi-unit setting where each bidder demands at most one unit and submits a single bid. In M2L, each bidder reports a value for each item and can change several bids jointly. Checking or repairing monotonicity one bid at a time cannot rule out profitable joint deviations across items. Therefore, enforcing cyclic monotonicity and recovering compatible payments in multi-item settings while preserving revenue remains a worthwhile research question.

Q4.Why does item-wise Myerson improve with more bidders, and what does the competitive-item count explain?

Feldman et al. (99% Revenue via Enhanced Competition, 2018) show that for additive bidders with independent items, running separate second-price auctions (and thus item-wise Myerson) achieves nearly all of the optimal expected revenue once the number of bidders is sufficiently large relative to the number of items.This makes item-wise Myerson a natural teacher in highly competitive markets. Palfrey (Bundling Decisions by a Multiproduct Monopolist with Incomplete Information, 1983) and Chakraborty (Bundling Decisions for Selling Multiple Objects, 1999) also show that bundling can earn more revenue with few bidders, while separate sales can earn more as the number of bidders grows.

Our competitive-item count quantifies the degree of competition in each valuation profile. A higher count means that more items attract at least two bids above the reserve and there are more competition. As more bidders enter, the expected count rises, and item-wise Myerson performs better.

Q5. Do OD-VVCA and M2L both fall under automated mechanism design?

Yes, both OD-VVCA and M2L fall under automated mechanism design, and the main distinction lies in what each method learns. OD-VVCA optimizes the parameters of a deterministic affine maximizer, whose structure determines allocations and payments and guarantees exact DSIC. In contrast, M2L learns both rules directly with neural networks, using item-wise Myerson as a guide and estimated regret as a training penalty. As a result, M2L offers greater expressive capacity than affine maximizers. We will clarify this distinction in the related work section.

Thank you, all your comments are in line with what I expected and so you have reinforced my positive review -- all my issues are addressed (well, some things are left open for future research but that is fine) as long as you ensure these points are clear in the final paper.

Thank you for your deeply insightful review. We truly appreciate the rigor of your comments and the constructive perspective you offered on the open directions for future work. We are grateful for your support and will make sure every point discussed is clearly reflected in the final version.

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

Recent work has used deep neural networks to learn high-revenue multi-bidder multi-item auctions. This is a surprisingly challenging problem in mechanism design, for which there has been almost no theoretical progress. In the single-item case, the Myerson auction is well-understood and gives optimal revenue (generally, it takes the form of a second-price auction with reserve price tuned depending on bidder valuation distributions). When selling multiple items, one can use an itemwise Myerson auction (important fact for this paper), but this is known to be suboptimal. Optimal auctions may involve very complex pricing and bundling decisions, but there has been little theoretical progress on characterizing them, hence the need to train neural networks to find approximate solutions.

The general idea of these approaches is to parameterize a direct revelation auction as a neural network outputting allocations and payments as a function of bids. Given sampled bidder valuations, train the network to maximize expected revenue, with an auxiliary loss term penalizing the regret, the gains from untruthful bidding (typically estimated using an inner loop where bids are optimized using PGD). Given the trained network, one can empirically estimate regret (this is somewhat tricky and easy to do in a flawed way), as well as revenue or other auction performance (usually straightforward to estimate from samples).

This paper provides an improved training pipeline for this general approach. Their key observation is that itemwise Myerson is a pretty good baseline. If we observe the learned auction underperforming itemwise Myerson, it's possible something may be amiss. In particular, the authors observe underperformance when there is a lot of competing demand for items (so Myerson can extract revenue, and clever bundling tricks are less important). The authors then use a three-stage pipeline: stage I consists of pretraining to initialize the auction closer to itemwise Myerson, stage II consists of ALGNet, a standard technique in the literature for auction training, and stage III consists of using itemwise Myerson again as a teacher model for the trained network. The authors show that their method results in better revenue and acceptable regret compared to a reasonable variety of baselines.

Critical strengths

The technique does actually work to improve performance of learned auctions on the standard benchmarks problems in this area. This is a useful contribution.

The experiments are carefully done. It is hard to prove regret has not been underestimated, but the method of You et al. is about as good as it gets for reliability.

The ablation studies in the appendices correctly answer several concerns one might have about the effectiveness of this technique.

Critical weaknesses

A complaint one could raise about this paper is that it is mostly just a small tweak to known methods. On the other hand, this small tweak works well.

Questions for authors

It might be fun to drop in AMenuNet or a similar approach for Stage II, if AMenuNet also shows a similar pattern of underperforming item-wise Myerson at small scales. Also, it's possible to initialize an affine maximizer by hand directly to be a separate Myerson auction, so perhaps stage I could be skipped.

Can you clarify more about where the neural network has an advantage on some profiles (because it can make clever bundling decisions) and where itemwise Myerson has an advantage (because if an item is competitive, a good Myerson auction gets lots of revenue)? This is sort of explained, but it's interesting and I would like more detail.

"However, these strictly constrained methods often limit the solution space, resulting in lower revenue than unconstrained approximations." As written and in its context, I think this sentence needs clarification that it applies to all the AMA-based methods but not to GEMNet.

4 comments

Q1. Can AMenuNet serve as the student, and Stage I be skipped?

Yes, AMenuNet can naturally serve as the student model, and Stage I can be skipped if the student model is an affine maximizer. We adopted our current neural architecture primarily because it is more expressive than AMenuNet and other affine maximizers. Fundamentally, our core contribution is architecture-agnostic: we identify profile-level revenue degradation that average revenue metrics obscure, and leverage this diagnosis to guide student optimization. Within this framework, the teacher can be flexibly substituted with other theoretical or empirical benchmarks (e.g., AMA or VVCA), while other deep-learning based models like AMenuNet can serve as the student, establishing a general paradigm for investigating optimal teacher-student pairings in auction design.

Q2. On which profiles does the neural mechanism have an advantage, and when does item-wise Myerson perform well?

A neural mechanism has more room to improve over item-wise Myerson when competition for individual items is weak. An item may go unsold at a low price on its own, yet still increase a bidder’s willingness to pay for a bundle. In such cases, the network can learn which items to group together and how to price each combination. This aligns with Palfrey (Bundling Decisions by a Multiproduct Monopolist with Incomplete Information, 1983) and Chakraborty (Bundling Decisions for Selling Multiple Objects, 1999), who show that bundling can earn more than separate sales when there are few bidders.

Conversely, item-wise Myerson performs well when many bidders compete for individual items. Feldman et al. (99% Revenue via Enhanced Competition, 2018) show that for additive bidders with independent items, running separate second-price auctions (and thus item-wise Myerson) achieves nearly all of the optimal expected revenue once the number of bidders is sufficiently large relative to the number of items. As the number of bidders grows,item-wise Myerson becomes an increasingly attractive choice in competitive markets due to its simplicity, training-free nature, and ease of deployment.

Q3. Does the restriction on the solution space also apply to GemNet?

Yes, we agree that GemNet and AMA-based methods impose different restrictions on the mechanism space. Specifically, GemNet offers each bidder a finite menu (300 entries by default, and 1,000 when there are more than five items). However, a finite menu can exclude the exact optimum: Daskalakis, Deckelbaum, and Tzamos (Strong Duality for a Multiple-Good Monopolist, 2017, §8.2.1) show that achieving optimal revenue can require infinitely many menu entries, even for a single additive bidder and two independent items. Thus, GemNet’s finite menu size also restricts the class of mechanisms it can represent. We will clarify this distinction in the revised paper.

Thank you for your responses!

I think including this intuition in the final paper and citing Palfrey and the other result would be helpful and interesting.

As for the issue of restricted solution spaces, this is a minor issue which is not fatal to the paper, but I still feel it should be corrected. For example, a neural network with smooth activations (as in this paper and many others) can represent a subset of infinite menu size auctions, while GemNet cannot. But with finite parameter size, the network may not be able to represent the optimal infinite menu size auction (e.g. the DDT17 optimal mechanism for the Beta distribution has infinite menu size and can't be represented by a Tanh network).

On the other hand, if one allows an arbitrary parameter count, a neural network can approximate the optimal solution arbitrarily closely. This property is true of general neural networks used in RegretNet-style approaches, and is also true of GemNet. However, it is not true of the AMA approaches in multi-bidder settings.

Therefore, when drawing this distinction, I think it makes sense to group GemNet with the other universal approximators.

Thank you for your professional review and for pointing out promising directions for our future work. You also raise a very valid point regarding the solution space, and we will update the final version of the paper accordingly by:

Grouping GemNet with the other universal approximators in the Introduction and Related Work section, while clarifying the distinction between finite-parameter expressiveness and asymptotic universal approximation.

Adding the intuition regarding which profiles the neural mechanism has an advantage on and when item-wise Myerson performs well, along with citations to Palfrey (1983) and Feldman et al. (2018).

We hope these updates fully address your concerns and support your final assessment.

Thanks, my concerns have been addressed and my assessment remains positive.

Reviewer_Xw9pOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Myerson-guided Machine Learning (M2L) for learning high-revenue multi-item auctions with low empirically estimated regret. Motivated by a profile-level diagnosis of ALGNet’s underperformance relative to Item-wise Myerson under stronger competition, the framework combines Myerson imitation, adversarial revenue–regret optimization, and selective refinement through teacher-based forward intervention and straight-through gradients.

Experiments on additive auctions with uniform and truncated-normal valuations show revenue improvements over the evaluated direct-parameterization baselines and fewer below-Myerson profiles. The final student is evaluated without teacher intervention. Additional experiments examine initialization, continued training, parameter counts, and computational costs.

Critical strengths
  1. Informative diagnosis and a well-motivated use of economic structure. The profile-level analysis identifies where an expressive neural mechanism loses revenue to a classical benchmark and relates these failures to competition intensity. This provides a concrete motivation for incorporating auction-theoretic structure into learning and offers insight beyond aggregate revenue comparisons.

  2. Meaningful empirical improvements with useful validation controls. The results are supported by four independent runs, student-only evaluation, and a common, comparatively strong misreport-search protocol. Appendix E also controls the additional training budget and learning rate, showing that ordinary continuation of Stage II does not reproduce the refinement gains. These checks strengthen the evidence that the proposed intervention provides a useful training signal.

Critical weaknesses
  1. The relationship between the Stage-III objective and the deployed student remains insufficiently explained. Stage III evaluates regret on a report-dependent teacher–student mechanism but propagates gradients through, and ultimately deploys, the student alone. Switching between mechanisms can introduce incentive effects absent from either component, so the intervened regret need not track student-only regret. The final evaluation supports the reported empirical outcomes, but limited analysis or diagnostic evidence connects the surrogate updates to improvements in the deployed mechanism.

  2. The distinctive contribution of selective intervention and straight-through optimization is not adequately established. The continued-training control is valuable, but the paper does not compare against simpler teacher-guidance alternatives, such as selective Myerson imitation or a revenue-gap penalty. It also does not separate intervention in the revenue term from intervention in the regret term. Consequently, it remains unclear which component produces the gains and whether simpler guidance could achieve comparable results. This limits the evidence for the central methodological contribution.

Questions for authors
  1. What connects the straight-through updates under the intervened mechanism to improvement of the student-only objective? In particular, how do intervened and student-only regret evolve during Stage III, and does their relationship differ between teacher-dominated and student-dominated profiles? Such diagnostics would help assess whether the training signal consistently addresses weaknesses of the deployed student.

  2. Under the same student architecture and training budget, how does Stage III compare with selective Myerson imitation or a revenue-gap penalty at comparable estimated regret? Additionally, what happens when the intervention is applied only to revenue while regret is always computed on the student? These comparisons would materially clarify the source of improvement and the value of the proposed refinement procedure.

3 comments

Weakness 1: Why Replace Student Regret with Mixed Mechanism Regret ?

Using only revenue together with the student mechanism’s own regret provides merely a scalar signal of revenue change. This signal is too weak: the student knows it should increase revenue, but lacks directional guidance on how to adjust its rules to achieve it. Mixed regret introduces a directional driving force toward itemwise Myerson, since the closer the student mechanism aligns with Myerson, the lower the mixed regret becomes, while preserving the flexibility of the neural network. Rather than forcing the student to replicate Myerson, it allows the student to stay close to Myerson while freely optimizing both revenue and regret. Conversely, directly penalizing the student using the teacher’s allocation and payment outputs imposes an overly rigid constraint: although itemwise Myerson is a helpful teacher, it is not the optimal mechanism for auctions with multiple items, so strict imitation sacrifices flexibility and lowers revenue. Our 5×105 \times 10 experimental results directly confirm this explanation: compared to our method’s revenue of 6.972 (with 0.00516 regret), guiding training solely with the revenue difference achieves only 6.897 (with 0.00521 regret), and directly imitating Myerson’s rules achieves only 6.882 (with 0.00518 regret), demonstrating that our approach attains both higher revenue and lower regret.

Q1.1:What connects the straight-through updates under the intervened mechanism to improvement of the student-only objective?

Our cosine-similarity analysis showed that the straight-through gradient of the intervened loss was consistently aligned with the student-only gradient. In this 5×105\times10 run, we evaluated both gradients on the same 1000 held-out profiles every two epochs. Cosine similarity was positive at all 30 checkpoints, ranging from 0.533 to 0.885. Thus, at each checkpoint, the straight-through descent direction also pointed downhill for the student-only validation loss. Over Stage III, student-only validation loss fell from −2.5028 to −2.5179.

The selection rule explains this alignment. When the auction selects the student mechanism under both truthful reporting and a unilateral misreport, both the revenue and the bidder’s gain from misreporting match their student-only counterparts in value and gradient. Across all 30 checkpoints (150,000 unilateral deviation checks in total, covering all 1,000 profiles and 5 bidders per checkpoint regardless of which mechanism the truthful profile selects), this joint case, where both the truthful profile and the misreport are routed to the student, was the most frequent outcome at every checkpoint, accounting for 77.7% of all checks.

Q1.2:How do intervened and student-only regret evolve during Stage III, and does their relationship differ between teacher-dominated and student-dominated profiles?

In this 5×105\times10 run, both intervened regret and student-only regret rose early in training. Intervened regret increased from 0.0210.021 to 0.0240.024 at epoch 4, then fluctuated and ended at 0.0230.023. Meanwhile, student-only regret rose from 0.00270.0027 to a peak of 0.00400.0040 at epoch 20, before generally declining to 0.00310.0031 by epoch 60.

Breaking down the profiles reveals that there were some differences in this relationship across subsets. At epochs 0,2,…,600, 2, \ldots, 60, we grouped the same 1,000 profiles based on which mechanism was selected under truthful bidding:

Teacher-dominated profiles. Both regret measures rose early in training. Student-only regret rose from 0.00230.0023 to a peak of 0.00370.0037 at epoch 20, then generally declined to 0.00280.0028 by epoch 60. Intervened regret remained elevated after its initial rise, ending at 0.0100.010 compared with 0.00490.0049 at the start.

Student-dominated profiles. Student-only regret rose from 0.00310.0031 to a peak of 0.00410.0041 at epoch 8, then returned close to its initial level, ending at 0.00310.0031. Intervened regret fell early and then fluctuated at a lower level, ending at 0.0250.025 compared with 0.0340.034 at the start.

We note that a good model is to achieve a balance of regret and revenue at the same time, instead of only one of them.

Q2. How does Stage III compare with a revenue-gap penalty, selective Myerson imitation and with intervention on revenue alone?

We conducted experiments for both methods in the 5×105\times10 setting. The same Stage II checkpoint was used for all three Stage III variants. We evaluated each final student on 10,000 valuation profiles. Our experimental results show that the full Stage III achieved the largest revenue.

Training methodStudent revenueRegret
Stage II6.8180.00473
Full Stage III6.9720.00516
Revenue-gap penalty6.9220.00534
Revenue-only intervention6.8970.00521
Selective Myerson imitation6.8820.00518

Thank you again for your feedback and suggestions on Stage III. With the author-reviewer discussion deadline approaching, we wanted to kindly follow up to check whether our response including the gradient-alignment diagnostics, subset regret trajectories, and ablation comparisons against simpler teacher-guidance baselines has resolved your concerns.

Please let us know if there is anything else we can clarify before the discussion closes. If our new experiments address your questions, we would greatly appreciate it if you could reconsider your evaluation of our paper.

Thank you for the additional experiments and I have raised my score accordingly.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

Thank you very much for your time, constructive feedback, and thoughtful engagement throughout the review and discussion period. We also sincerely thank the Area Chair for overseeing the discussion process.

Now all three reviewers are satisfied with our responses and supportive of our work. We truly appreciate your encouraging assessments and insightful suggestions, and we will carefully incorporate all new experiments and clarifications from the rebuttal into the final version of the paper.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.