acceptodds
Under review as a conference paper at ICLR 2027

Learning Revenue-Maximizing Auctions with Neural Affine Maximizer

Abstract

Learning truthful, revenue-maximizing auctions is a central challenge in automated mechanism design and differentiable economics. Existing learning approaches that guarantee truthfulness typically discretize the outcome space into a finite menu, which suffers from the “curse of dimensionality” in large-scale auctions. In this work, we propose *Neural Affine Maximizer* (NAM), a discretization-free mechanism for learning truthful auctions, as well as a novel algorithm to learn NAMs. NAM guarantees truthfulness by building on affine maximizer auctions (AMAs) while replacing the conventional finite menu with a boosting function over the outcome space. We then parameterize the boosting function with neural networks and derive unbiased gradient estimators to enable first-order optimization. Experiments on instances with up to buyers or goods show that NAM is either competitive with or outperforms existing approaches. Specifically, with buyers, goods, and additive valuations, NAM discovers a truthful auction achieving % higher revenue and lower training time compared with state-of-the-art baselines.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Automated Mechanism Design, Differentiable Economics, Deep Learning, Auctions

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_tor6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies revenue-maximizing auctions for multiple buyers and goods with additive valuations. It proposes Neural Affine Maximizer (NAM), which replaces a finite menu of allocation outcomes with a continuous boosting function over feasible probabilistic allocations. The boost is parameterized as the negative of an input convex neural network, making allocation and payment computation reducible to convex optimization. The ideal mechanism is dominant-strategy incentive compatible and individually rational when the required maximizations are solved exactly. The paper also establishes that concave boosts can reproduce finite affine maximizer mechanisms with equal bidder weights.

To learn the mechanism, the authors replace the allocation maximizations during training with Gibbs distributions and derive a covariance-based gradient estimator. Reflected Langevin dynamics supplies approximate samples for first-order optimization. Experiments cover several valuation distributions and settings with up to 15 buyers or 30 goods, including comparisons with learned truthful auctions and classical mechanisms. NAM is generally competitive and achieves substantial improvements over learned baselines in larger settings, including a reported 10.6% revenue gain over the strongest compared baseline for three buyers and 30 uniformly valued goods. Supplementary results include correlated and Bernoulli valuations, compute comparisons, and implementation details.

Critical strengths
  1. The continuous concave representation is a useful alternative to explicit outcome menus. Its ability to reproduce equal-weight affine maximizers while supporting convex allocation optimization gives the architectural choice a clear theoretical basis. This combination of expressiveness and tractable inference is the main methodological reason to favor acceptance.

  2. The empirical evidence supports the practical value of the representation. Revenue comparisons across multiple distributions, five independent runs, and training-time and memory measurements show benefits beyond a single example. The larger-instance gains over learned baselines make a cautiously positive assessment reasonable.

Critical weaknesses

The practical truthfulness guarantee needs clearer quantification. Inference uses a fixed number of projected-gradient iterations, and the reported objective precision is not clearly identified as a certified optimality gap rather than an empirical convergence measure. The authors acknowledge that solver error affects regret, but do not quantify the resulting incentive or IR error. Because guaranteed truthfulness is a central motivation, this limits confidence in the implemented mechanism. Clarifying the numerical guarantee and its implications would address this concern.

Questions for authors

What does the reported inference precision of approximately 10⁻⁶ measure, and how does it bound profitable misreporting and IR violations under the implemented payment rule? Please clarify using the existing solver analysis or validation; no new experiments are requested.

1 comment
AuthorsOpenReview ↗

Rebuttal to Reviewer tor6

Thank you for the positive assessment and the precise question.

The "≈10−6\approx10^{-6}" in the original appendix was an empirical convergence measure: we tuned the solver's learning rate schedule until the objective stopped changing in its first six significant digits. The revised manuscript replaces it with Appendix C.2, which (i) bounds profitable misreporting and IR violation by the solver's optimality gap and (ii) measures this gap against exact solutions.

(i) Bound on misreporting and IR violation (Proposition C.1). Under the implemented payment rule, where payments are computed from the solver's outputs, if the optimality gap at the type profile t\mathbf{t} is at most ε(t)\varepsilon(\mathbf{t}), then each buyer's gain from misreporting and each buyer's IR violation are at most ε(t)\varepsilon(\mathbf{t}).

(ii) Optimality gap against exact solutions. With leaky-ReLU activations, the inner maximization of ASW(x;t,θ)\mathrm{ASW}(\mathbf{x};\mathbf{t},\theta) can be formulated exactly as an LP. We solve it with HiGHS on 100 random type profiles per setting and report the optimality gap of our inference procedure (Table 5). The gap is measured as the absolute value.

SettingMean gapMax gap
U2,2U _{2,2}1.72×10−61.72\times10^{-6}2.23×10−52.23\times10^{-5}
U3,10U _{3,10}1.81×10−61.81\times10^{-6}1.45×10−51.45\times10^{-5}
U3,30U _{3,30}3.50×10−63.50\times10^{-6}3.40×10−53.40\times10^{-5}
U15,10U _{15,10}5.16×10−55.16\times10^{-5}7.82×10−47.82\times10^{-4}

By (i), these gaps bound the gain from misreporting and the IR violation on the sampled profiles.

Reviewer_YqwyOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes the Neural Affine Maximizer (NAM): an affine maximizer auction whose finite menu is replaced by a continuous boosting function λ(x) over the whole feasible outcome space, parameterized as the negative of an ICNN so that the inner welfare maximization is concave. Truthfulness and IR follow from the standard AMA argument; Theorem 3.2 shows every unit-weight AMA is induced by some concave NAM. Training relaxes the argmax to a Gibbs distribution and uses a covariance-trick gradient estimator with reflected Langevin sampling. Experiments with up to 15 buyers or 30 goods report revenue ≥99% of the best baseline everywhere and +10.6% over AMenuNet on U_{3,30}.

Critical strengths
  • Clean idea: a concave boost over the continuous outcome space removes the (n+1)^m menu while keeping exact DSIC/IR and exact unit-supply feasibility.
  • The theory is correct; I checked Thm 3.2 (a nice LP-duality argument) and Prop 4.5.
  • Strong gains in the many-goods regime (U/Beta/I at m=30, ~+10% over the best learned baseline), supported by compute-matched curves (Fig. 5).
  • Inference cost is reported honestly; code for NAM and all learned baselines is released.
Critical weaknesses
  1. I was unable to reproduce several table entries from the configuration stated in the paper. Running the released code with the Table 3 settings (joint λ, d_h=10nm, β=512, T=32, η=0.04, B=16384, 10k updates, seeds 0–4, RTX 5080) gave: U_{3,3} 1.656±0.001 (paper 1.660 — reproduces); U_{3,10} 5.764±0.009 (paper 5.822 — all five seeds ~1% low, which places NAM below OD-VVCA's 5.821); U_{2,2} 0.777±0.211, U_{2,5} 1.926±0.459, C_{2,2} 0.704±0.265 (paper 0.873 / 2.312 / 0.893, sd ≤0.002). In the last three, 1–2 seeds collapse after ~200 updates, and re-running the same seed reproduces the collapse exactly, so this is not GPU non-determinism. Notably, U_{2,2} and U_{2,5} are exactly the settings for which the released code lowers the learning rate to 1e-4 (and β to 128 for U_{2,2}), which suggests Table 3 does not fully describe the configuration used. The code also defaults to a buyer-separable boost λ(x)=(1/n)Σ_i f(x_i) with hidden width 20·d_in, using the joint boost only for U_{2,2}. I therefore do not conclude the reported numbers are wrong, but the paper currently gives no per-table configuration or per-seed results, and the "consistent across seeds" claim (Sec. 5.2) is not supported under the stated configuration. I did not rerun the large settings.
  2. DSIC at inference is asserted rather than measured. Outcomes come from 500 steps of projected Nesterov ascent (App. C.1); no empirical regret, IR violation, or randomization ratio is reported. Since the ICNN uses leaky-ReLU, the inner maximization is exactly an LP, which would give exact outcomes and a way to quantify the solver gap.
  3. The paper credits Ma et al. [33] for the covariance trick and the overall algorithm design; the new elements are the continuous boost over an AMA's outcome space, the ICNN parameterization and Thm 3.2. Thm 3.2 is a modest step once an AMA is viewed as a boost on a finite set, and the claim in lines 223–224 that NAM "can approximate an optimal (unknown) AMA to arbitrary precision" does not follow directly from Thms 3.2 and 4.4: sup-norm closeness of λ does not by itself imply convergence of the induced allocations or revenue, since argmax outcomes jump at ties. A revenue-convergence statement (e.g. under a density assumption on F) would close this gap. Fixing w_i≡1 also narrows the class relative to general AMAs and to AMenuNet.
  4. The gains are concentrated in the few-buyers / many-goods regime. At n=15 NAM is tied with Itemwise Myerson (8.780 vs 8.750; 6.878 vs 6.875; 9.479 vs 9.482); on Bernoulli distributions it reaches only 93–97% of the optimum with high variance (B_{2,2}: 0.669±0.087 vs 0.720), plausibly because a concave boost plus finite-step sampling favours interior allocations. The abstract's "large-scale" framing should make clear that the gains are in m rather than n.
Questions for authors
  1. For each table entry: which λ parameterization and hyperparameters were used, and what are the per-seed revenues? Can you explain the diverging seeds and the ~1% gap on U_{3,10}?
  2. What are the empirical regret, IR violation and randomization ratio of the learned NAMs? Have you tried the exact LP formulation of the inner problem?
  3. Can you state (and prove) a revenue-convergence version of the expressiveness claim, and spell out precisely what is new relative to [33]?
  4. Which reflection rule was used in the experiments?

Minor: the "bold within one sd" rule is undefined for baselines without a reported sd.

Overall: An elegant idea with correct theory and convincing many-goods results. My concerns are that the stated configuration does not reproduce several table entries (most likely because Table 3 is incomplete), truthfulness at inference is not verified, and the expressiveness and "large-scale" claims are stated more strongly than supported. I lean toward rejection in the current form and would raise my score given per-table configurations that reproduce the results together with regret diagnostics.

2 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer Yqwy

Thank you for carefully running our released code and for the detailed feedback.

W1/Q1: Reproducibility. Table 3 now lists the default hyperparameters, and the new Table 4 lists the per-setting departures, following NAMArgs.from_setting in the released code. The key difference from your runs is the boost: all settings except U2,2U _{2,2} use the buyer-separable boost λNAM(x;θ)=−1n∑if(xi;θ)\lambda^{\mathrm{NAM}}(\mathbf{x};\theta)=-\frac{1}{n}\sum _i f(x _i;\theta) (Eq. 33) with hidden width dh=20md _h=20m, while U2,2U _{2,2} uses the joint boost λNAM(x;θ)=−f(x;θ)\lambda^{\mathrm{NAM}}(\mathbf{x};\theta)=-f(\mathbf{x};\theta) with dh=80d _h=80. The main motivation to use separable boost is to provide an inductive bias that accelerates training.

We re-ran U2,2U _{2,2}, U2,5U _{2,5} and C2,2C _{2,2} with these configurations. Per-seed revenues (seeds 0–4) are listed below:

  • U2,2U _{2,2}: 0.870, 0.869, 0.869, 0.874, 0.874 (0.871±0.0020.871\pm0.002)
  • U2,5U _{2,5}: 2.312, 2.316, 2.313, 2.312, 2.315 (2.314±0.0022.314\pm0.002)
  • C2,2C _{2,2}: 0.893, 0.892, 0.894, 0.892, 0.895 (0.893±0.0010.893\pm0.001)

W1/Q1: Seed collapse on U2,2,U2,5,C2,2U _{2,2}, U _{2,5}, C _{2,2} and the gap on U3,10U _{3,10}. Your runs follow the original Table 3, which differs from the configurations we used in these settings. We attribute the collapse to this mismatch. Our configurations are: U2,2U _{2,2} uses learning rate 10−410^{-4} and β=128\beta=128; U2,5U _{2,5} uses the separable boost with learning rate 10−410^{-4}; C2,2C _{2,2} uses the separable boost (Table 4). With them no seed collapses, as the per-seed values above show. Under the configurations of Tables 3–4, the standard deviation is at most 0.009 on all UU, Beta, II and CC settings, which supports the consistency claim in Section 5.2.

Your 5.7645.764 on U3,10U _{3,10} comes from the joint boost, which converges more slowly and has not converged within the 10k-update training budget.

W2/Q2: Inference precision and regret. Following your suggestion, we solve the inner maximization exactly: the leaky-ReLU ICNN admits an exact LP formulation, which we solve with HiGHS on 100 random type profiles per setting (Appendix C.2, Table 5). The mean solver gap is 1.72×10−61.72\times10^{-6}, 1.81×10−61.81\times10^{-6}, 3.50×10−63.50\times10^{-6} and 5.16×10−55.16\times10^{-5} on U2,2U _{2,2}, U3,10U _{3,10}, U3,30U _{3,30} and U15,10U _{15,10}, at most 5.88×10−65.88\times10^{-6} of the revenue. We also add Proposition C.1: each buyer's regret and IR violation are at most the solver gap.

W3/Q3: Expressiveness. Section 4.1 now states the result precisely: for any AMA with wi≡1w _i\equiv1, Theorem 3.2 constructs a continuous concave boost λˉ\bar\lambda that induces the same direct mechanisms, and by Theorem 4.4 an ICNN approximates λˉ\bar\lambda to any precision in ℓ∞\ell _\infty. The revised statement concerns the boost only, which is exactly what Theorems 3.2 and 4.4 imply. We fix wi=1w _i=1 solely for simplicity; learnable weights wi>0w _i>0 are a natural extension, since ASW remains concave in x\mathbf{x}.

W3/Q3: Novelty relative to [38] ([33] in the previous version of the manuscript). Only the training algorithm is inspired by [38]; the rest is new. (1) The unit-supply multi-buyer auction studied here is outside their model. (2) We introduce NAMs and prove that they lose no expressive power relative to AMAs, even with concave boosts (Theorem 3.2). (3) We parameterize the boost with ICNNs, which makes inference a convex program. (4) Our training algorithm follows the idea of [38] but is re-derived for learning NAMs. (5) Our outcome space is a polytope, so sampling needs reflection, for which we give a parallel approximation to Skorokhod reflection; [38] does not consider this.

W4: Scope of NAM's advantage on large nn. At n=15n=15, NAM is the only learned method that matches Itemwise Myerson: the best learned baseline reaches at most 80% of NAM on U15,10U _{15,10}, Beta15,10\mathrm{Beta} _{15,10} and I15,10I _{15,10}. This shows the advantage of NAM over learned baselines when the number of buyers is large.

Q4: Reflection used in experiments. We started from the exact Skorokhod reflection described in Section 4.2. It handles one boundary hit at a time, so it is hard to parallelize. To accelerate training, we approximate it: all violated constraints of the same family are reflected simultaneously, alternating between xij≥0x _{ij}\ge0 and ∑ixij≤1\sum _i x _{ij}\le1. This is possible because constraints within a family involve disjoint coordinates. The approximation gave training results similar to the exact Skorokhod reflection at a lower cost, and it is used in all reported experiments. Section 4.2 of the revised manuscript now describes it.

Minor. A footnote now states that missing standard deviations are treated as zero for comparison.

Thank you for your responses. This resolved my questions on reproducibility, I have raised my score.

Reviewer_Hi6JOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents a contribution to automated mechanism design using neural networks. In particular, it uses neural networks to design affine maximizer auctions.

These are a long-studied family of auction mechanisms, essentially affinely transformed versions of the famous VCG mechanism, which have nice properties in particular guaranteeing strategyproofness (no incentive to give untruthful bids). The key parameters of an affine maximizer are scalar weights per bidder (fixed to 1 in this paper, a reasonable choice for simplicity), and a scalar boost per possible outcome of the mechanism. The mechanism chooses whatever outcome maximizes weighted social welfare, plus the boost.

Several prior works use learning techniques to design affine maximizers. In these works, the affine maximizer is always defined by an explicit, finite range of possible auction outcomes, with a corresponding explicit boost per outcome. The novelty in the present work is to instead use a boost function which gives a boost for any possible auction outcome, encoded by a neural network that takes auction allocations as input.

The problem is that choosing the auction outcome now requires solving a maximization problem which includes an arbitrary neural network. The authors' solution is quite clever, one of those things that is obvious only in retrospect.

In Theorem 3.2, they show that it is without loss of generality when designing affine maximizers to consider boost functions that are concave in their inputs. Then, they represent the boost function using well-known techniques for input-convex neural networks, so even though there is a neural network involved, running the affine maximizer now requires solving a concave maximization problem, not an arbitrary one. What remains to optimize the mechanism as a whole is differentiating through this optimization problem, which is technically tricky but can be done. In experimental results the authors show that their approach scales well and beats previous AMA techniques on the usual benchmark of revenue under typical settings.

Critical strengths

The approach works and beats SOTA.

As simple as Theorem 3.2 is in retrospect, it's a delightful fact.

Experiments are in the Goldilocks zone, not excessive but do everything they need to establish the technique works.

This opens up new ways to optimize affine maximizer auctions. Affine maximizers are actually a very generally applicable family of mechanisms, and there are many uses beyond the stylized multi-item auction design problem studied here. As far as I can tell, as long as the outcome space for the problem is convex, the technique applied here should work directly.

Critical weaknesses

I don't think there are really critical weaknesses that would prevent publication. However, I do have some issues with the presentation.

Why draw a distinction between the AMA and the direct mechanism induced by the AMA? To me, this doesn't accomplish anything and adds a lot of slightly confusing material which I ultimately just needed to skip. It doesn't seem to me that Propositions 2.3 and 3.1 need to be given so much space. These are all AMAs and AMAs are DSIC. Maybe I am missing something.

Questions for authors

Are the Myerson auction revenues for your irregular distribution correct? Since it's irregular you need ironing, the true Myerson auction won't be the usual second price with reserve. It's not a meaningful problem for your results but probably worth checking.

I suspect the proof of Theorem 3.2 could be vastly simplified using the Fenchel conjugate. In particular, λNAM\lambda^{NAM} should be the double conjugate of the original λ\lambda. I haven't worked out the technicalities (because you are dealing with concave functions, I guess there have to be negative signs sprinkled everywhere, and one has to be a little careful with the domains of the functions) but I think this should work.

Differentiating through convex programs is something done very widely in deep learning, using various techniques (some of which have been reinvented multiple times), for various purposes. I think citing some of the existing work that does this, and clarifying where your particular technique fits in among all the many methods, would be helpful for the reader. The discussion in the appendix is very brief and I think it should be extended.

2 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer Hi6J

Thank you for the encouraging review and the insightful questions.

AMA vs. the induced direct mechanism. An AMA does not determine a unique direct mechanism: when the affine-welfare maximizer is not unique, different tie-breaking rules give different allocations and payments. Saying that "a NAM represents an AMA" would therefore be ambiguous. With the induced direct mechanism made explicit, Theorem 3.2 has a precise meaning: every direct mechanism induced by the AMA is also induced by the NAM. In addition, DSIC and IR are originally defined for direct mechanisms, so this also keeps the paper self-contained for readers less familiar with AMAs.

Q1 (Myerson under irregularity). Thank you for correcting this. Both Beta(1/2,1/2)\mathrm{Beta}(1/2,1/2) and II are irregular. For Beta, the ironing interval lies where the ironed virtual value is negative, so the optimal auction is still second price with the optimal reserve, and the reported values stay. For II, the ironed virtual value is a positive constant on the ironing interval, so ironing matters. We recomputed the ironed Myerson revenue in closed form and updated Table 2 and Appendix C.4: I2,2,I3,10,I3,30,I15,10I _{2,2},I _{3,10},I _{3,30},I _{15,10} become 0.441, 2.969, 8.906, 7.003.

Q2 (Fenchel conjugate). Thank you for this elegant observation. The biconjugate viewpoint is a promising route to a shorter proof of Theorem 3.2, and we will examine it in future work.

Q3 (differentiating through convex programs). We have extended the related-work discussion in Appendix A. It now covers more bi-level optimization methods, including those that differentiate through convex programs with first-order or second-order information, and places our method among them: our training algorithm is inspired by the covariance trick of Ma et al. [38] ([33] in the previous version of the manuscript), credited in the introduction, and uses only first-order information.

Thank you for your response and for making these changes. My view on the paper remains positive.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.