acceptodds
Under review as a conference paper at ICLR 2027

Learning Collusion-Resistant Auctions via Coalition Regret Minimization

Abstract

Auctions serve as efficient frameworks for resource allocation and have profoundly influenced digital economies, but deriving optimal mechanisms for multi-bidder, multi-item auctions remains analytically intractable. Recently, machine learning (ML) has partially addressed this by designing revenue-maximizing auctions through individual regret minimization. However, these approaches primarily ensure strategy-proofness against single bidders, leaving mechanisms vulnerable to collusion, which often leads to substantial revenue loss. To address this, we introduce CoalitionNet, an ML-powered auction framework designed to minimize coalition regret. Specifically, it leverages utility externalities to evaluate joint deviations and trains a strong collusion generator for mechanism adversarial learning. Extensive experiments demonstrate that CoalitionNet achieves near-zero coalition regret while maintaining competitive revenue performance. Furthermore, it also demonstrates strong generalization across unseen coalition sizes as well as valuation distributions.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
automated mechanism design, adversarial training, multi-item 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_bx82OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CoalitionNet, a deep learning-based method for learning collusion-resistant auctions. CoalitionNet consists of a "collusion generator" and a "mechanism network". The mechanism network is trained to maximize the expected revenue and minimize the expected coalition regret. The collusion generator can generate candidate coalition as well as collusive bids, and is trained to maximize the expected coalition regret. Experiments show that CoalitionNet successfully learns auctions with low coalition regrets, while existing baselines that only minimize the individual regret fail.

Critical strengths
  • This paper integrates deep learning into collusion-resistant auction design, which is novel in my knowledge and meaningful.
  • The methodology is straightforward and well presented. The experiment shows the potential of CoalitionNet in learning collusion-resistant auctions.
Critical weaknesses

I have a few concerns about the soundness of experimental designs.

  1. There lack simple settings (e.g., 2 bidders and 1 item) where analytically optimal collusion-resistant auctions exist. Without comparisons in such settings, it's hard to evaluate whether CoalitionNet can approximate the optimal auctions in simple settings (which I also believe is a natural evaluation).
  2. There lack simple baselines (e.g., run optimal auctions to each item independently, similar to Itemwise Myerson auction) that can serve as collusion-resistant baselines in general settings (for general nn bidders and mm items). Without comparisons to such baselines, it's hard to evaluate whether CoalitionNet can outperform simple baselines and thus be practically useful in general settings.
  3. The paper does not describe how "coal. regret" is computed for a sampled v∈Rn×mv\in \mathbb{R} ^{n \times m}. In my understanding it can not be exactly computed with plausible computation load, so in experiments this value can only be computed heuristically. Therefore, the "coal. regret" is under estimated compared with the true coalition regrets. This is fine, but the paper should clearly state this, with description about how coalition regrets are empirically computed. Furthermore, if the computations of coalition regrets are similar to those used to train the "Collusion Generator", then it's likely that the under-estimation of CoalitionNet is more severe than other baselines.
Questions for authors
  1. In line 238, what is the role of ξ\xi in the methodology?
  2. In line 189-196, some questions about the "two critical challenges" claimed by the paper:
    1. Why does the paper claim "Combinatorial Search Space" as a challenge? I did not see how the CoalitionNet resolves this challenge. Sampling a coalition subset S⊂NS \subset N is a standard approach in my view either.
    2. I'm confused about what the second challenge "Gradient Vanishing in Joint Space" actually means. It seems to me that this challenge has not been acknowledged in existing works, and possibly the authors discovered this challenge in their researches. If so, it's better to provide empirical results to make the statement of this challenge more convincing.
  3. In Table 1, what dimension does "max. coal. regret" take "max" over?
  4. See my weaknesses above.
4 comments
AuthorsOpenReview ↗

Responses to W1 & W2: 2×1 Optimal Benchmark Validation and 5×5 Itemwise Myerson Comparison

We thank the reviewer for raising these insightful comments and apologize for any misunderstandings caused by the lack of clarity in our original manuscript. We have conducted the corresponding supplementary experiments and address your concerns point by point as follows.

Response to W1. We added a 2×12\times 1 scenario, where each bidder's valuation is drawn i.i.d. from U[0,1]\mathcal{U}[0,1]. The classical Myerson optimal auction for this case is a second-price auction (SPA) with reserve price r∗=0.5r^*=0.5. Analytically, its truthful expected revenue is 5/12≈0.41675/12 \approx 0.4167, while the canonical joint shading attack at 0.50.5 yields an attacked revenue of 3/8=0.3753/8 = 0.375 and a revenue drop of 1/24≈0.04171/24 \approx 0.0417. Here we measure three different methods, with each one evaluated over K=2,000K=2{,}000 valuation profiles v(k)∼U[0,1]2\boldsymbol{v}^{(k)} \sim \mathcal{U}[0,1]^2. For learned mechanisms, results are averaged over three random seeds:

  • Truthful revenue: The expected seller payment when all bidders report truthfully.
  • Attacked revenue: We use gradient ascent to identify the coalition C∗(v(k))C^*(\boldsymbol{v}^{(k)}) that achieves the largest utility gain on v(k)\boldsymbol{v}^{(k)} and record the resulting seller revenue.
Method (2×12\times 1)Truthful revenue ↑\uparrowAttacked revenue ↑\uparrowRevenue drop ↓\downarrow
Myerson0.41800.41800.37380.37380.04430.0443
RegretNet0.42320.42320.23170.23170.19150.1915
CoalitionNet (Ours)0.41730.41730.36200.36200.05540.0554

As shown in the results, CoalitionNet closely matches Myerson's truthful revenue while retaining 96.8% of SPA's attacked revenue, resulting in a revenue drop similar to the collusion-resistant mechanism. In contrast, RegretNet achieves high revenue under truthful bidding, but its attacked revenue drops sharply to 0.23170.2317 under group deviations, incurring a substantially larger revenue drop of 0.19150.1915. It shows that CoalitionNet can closely recover the classical results for collusion-resistant auctions, whereas learning-based mechanisms that only enforce individual IC fail to preserve revenue against collusive deviations.

Response to W2. We agree with the reviewer's comment that comparing against simple itemwise baselines is necessary, so we add Itemwise Myerson as a benchmark. It decomposes the multi-item auction into m=5m=5 independent single-item Myerson optimal auctions (r∗=0.5r^*=0.5) and we compare it against CoalitionNet in the 5×55\times 5 setting. Both methods are evaluated under the 5×55\times 5 additive valuation setting with valuations drawn i.i.d. from U[0,1]\mathcal{U}[0,1], using the exact same K=1,000K=1{,}000 shared evaluation profiles. To compute AvgCoal and MaxCoal (the mean and the per-profile maximum coalition regret across all 3131 non-empty coalitions, respectively), we approximate the worst-case deviation using 5050-step gradient ascent with 55 restarts.

Method (5×55\times 5)AvgCoal ↓\downarrowMaxCoal ↓\downarrowTruthful revenue ↑\uparrowAttacked revenue ↑\uparrowDrop ↓\downarrow
CoalitionNet (Ours)0.00630.00630.01330.01332.97132.97132.74772.74770.22350.2235
Itemwise Myerson (r∗=0.5r^*=0.5)0.17990.17990.95970.95973.37973.37972.42002.42000.95970.9597

As shown in the table above, although Itemwise Myerson achieves higher truthful revenue, it is vulnerable to coordinated bid shading across items: its attacked revenue falls sharply, resulting in a large revenue drop and high coalition regret. Meanwhile, CoalitionNet achieves higher attacked revenue than Itemwise Myerson while maintaining a substantially smaller revenue drop and lower coalition regret. It indicates that applying a single-item optimal auction independently to each item cannot withstand coordinated multi-item deviations, but CoalitionNet learns joint cross-item allocation and pricing rules that can preserve higher seller revenue under collusion.

Response to W3. We agree with the reviewer that the original manuscript lacked a clear description of how coalition regret is computed, as well as the potential under-estimation issue. We fully agree that computing exact coalition regret requires taking the continuous supremum over joint misreports b~C∈[0,1]∣C∣×m\tilde{\boldsymbol{b}}_C \in [0,1]^{|C|\times m} across all 2n−12^n - 1 non-empty coalitions, which is computationally intractable in high-dimensional settings such as 5×55\times 5. Therefore, the reported coal. regret metrics in our manuscript are empirical approximations. We will explicitly state this limitation and add a dedicated description of the evaluation protocol in Section 5.1 of the revised manuscript.

Specifically, for each sampled valuation profile v(k)\boldsymbol{v}^{(k)}, we enumerate all 2n−12^n - 1 non-empty coalitions C⊆NC \subseteq N, hold non-coalition bids truthful (b~j=vj(k)\tilde{b}_j = v_j^{(k)} for all j∉Cj \notin C), and apply gradient ascent to directly optimize the coalition's total utility gain R^C(v(k))\hat{R}_C(\boldsymbol{v}^{(k)}) over the joint misreport b~C\tilde{\boldsymbol{b}}_C. Additionally, as demonstrated in Section 5.5 and Figure 5 of the manuscript, our learned/gradient-based attacks are close to the strongest possible deviations against exhaustive grid search in tractable small-scale settings. This empirical validation justifies using gradient-based search to approximate worst-case coalition deviations in larger settings. The gradient-ascent attack is applied identically and independently across CoalitionNet and all baseline mechanisms to ensure a fair empirical comparison.

Response to Q1. ξ\boldsymbol\xi is a stochastic exploration input, sampled as Gaussian noise and appended to each bidder token of the collusion generator. Its role is to produce diverse generator reports, allowing the generator to capture multiple distinct collusive bidding strategies for a given valuation profile.

Response to Q2a (Combinatorial Search Space). We apologize for the unclear exposition in the original manuscript, and we clarify below how CoalitionNet handles the 2n−12^n - 1 coalition space. CoalitionNet does not simply sample a random coalition subset CC and evaluate that single subset alone. For naive subset sampling, each forward pass inspects only one out of 2n−12^n - 1 coalitions, and the resulting utility signal can be noisy or misleading when non-benefiting members in CC cancel out gains or outside winners are excluded. Instead, our generator samples an active deviator set AA (composed of a prefix SS and a pivot bidder ii) whose bids are modified while all remaining bidders report truthfully. Crucially, because any bidder in C∖AC \setminus A stays truthful, this single modified bid profile is simultaneously feasible for all 2n−∣A∣2^{n-|A|} superset coalitions C⊇AC \supseteq A. Given the induced utility changes Δuj\Delta u_j, an external bidder k∉Ak \notin A increases the coalition's total utility if and only if Δuk>0\Delta u_k > 0. Consequently, rather than enumerating the 2n−∣A∣2^{n-|A|} supersets, the optimal superset is obtained analytically in O(n)O(n) time via FA(Δu)=max⁡C⊇A∑j∈CΔuj=∑j∈AΔuj+∑k∉A[Δuk]+.F_A(\Delta u)=\max_{C\supseteq A}\sum_{j\in C}\Delta u_j =\sum_{j\in A}\Delta u_j+\sum_{k\notin A}[\Delta u_k]_+. Thus, although the generator still samples prefixes and pivots to explore the space of active deviations, the analytic envelope eliminates the exponential enumeration over the 2n−∣A∣2^{n-|A|} coalition supersets for any given active deviation. We will clarify this distinction in Section 4.

Response to Q2b (Gradient Vanishing). We apologize for the lack of clarity regarding the gradient vanishing challenge and the absence of empirical substantiation in our original submission. Here, gradient vanishing refers to the vanishing of the attack objective's gradient with respect to the jointly modified bids b~A\tilde{\boldsymbol{b}}_A of the active deviators AA. Specifically, under a joint bid reduction, the sampled subset AA alone can experience a non-positive aggregate utility change even when an outside bidder benefits from the resulting allocation shift. As a result, the naive subset objective max⁡(0,UA)\max(0, U_A) saturates on the flat zero-gradient plateau of max⁡(0,⋅)\max(0, \cdot), preventing outside beneficiaries from providing any optimization signal to b~A\tilde{\boldsymbol{b}}_A. In contrast, our externality-guided objective max⁡(0,ΦA)\max(0, \Phi_A) retains the signed utility changes of members in AA and incorporates positive externalities from outside bidders, yielding non-zero gradients for the exact same joint report.

To quantify gradient vanishing empirically, we evaluate K=500K=500 held-out profiles in the 5×55\times 5 setting. For each active deviator size ∣A∣∈{1,2,3,4}\lvert A\rvert \in \{1,2,3,4\}, we randomly sample an active set AA initialized with moderate bid shading (b~A=0.5 vA\tilde{\boldsymbol{b}}_A = 0.5\,\boldsymbol{v}_A) and compute the ℓ2\ell_2 norm of the objective gradient with respect to the active bids, ∥∇b~AL∥2\|\nabla_{\tilde{\boldsymbol{b}}_A} \mathcal{L}\|_2. We define a profile as experiencing gradient vanishing if ∥∇b~AL∥2<10−4\|\nabla_{\tilde{\boldsymbol{b}}_A} \mathcal{L}\|_2 < 10^{-4}, and report the vanishing rate as the percentage of profiles satisfying this criterion.

BaselineVanishing rate (∣A∣=1/2/3/4\lvert A\rvert=1/2/3/4): naive max⁡(0,UA)\max(0, U_A) vs. max⁡(0,ΦA)\max(0, \Phi_A)
RegretNet94/99/91/55% vs. 16/13/8/5%
RegretFormer100/100/97/71% vs. 18/28/27/18%
AMenuNet100/97/83/44% vs. 85/66/35/20%

Across all three mechanisms, replacing naive max⁡(0,UA)\max(0, U_A) with max⁡(0,ΦA)\max(0, \Phi_A) significantly reduces the vanishing-gradient rate (e.g., from 94–99% down to 13–16% on RegretNet for ∣A∣=1,2\lvert A\rvert=1,2). These results empirically verify the critical role of ΦA\Phi_A in providing dense gradient signals for joint deviations. Indeed, this vanishing-gradient bottleneck is precisely what we encountered when initially attempting to directly extend RegretNet's paradigm to the coalition setting, which motivated our externality-guided formulation. We apologize for this misunderstanding and will revise the manuscript to clarify this motivation and include these empirical results.

Response to Q3. We are sorry for the unclear definition of Max coal. regret and Avg coal. regret in Table 1. Both metrics are empirical estimates computed over K=1,000K=1{,}000 sampled valuation profiles v(k)\boldsymbol{v}^{(k)} and all 2n−12^n - 1 non-empty coalitions C⊆NC \subseteq N, where each per-profile, per-coalition regret R^C(v(k))\hat{R}_C(\boldsymbol{v}^{(k)}) is approximated via gradient ascent as described in our response to W3:

  • Max coal. regret: The max is taken over the coalition dimension (all 2n−12^n - 1 non-empty coalitions ∅≠C⊆N\emptyset \neq C \subseteq N) for each sampled valuation profile v(k)\boldsymbol{v}^{(k)}, followed by taking the empirical expectation across the K=1,000K=1{,}000 profiles: Max coal. regret=1K∑k=1K(max⁡∅≠C⊆NR^C(v(k))).\text{Max coal. regret} = \frac{1}{K}\sum_{k=1}^{K} \left( \max_{\emptyset \neq C \subseteq N} \hat{R}_C(\boldsymbol{v}^{(k)}) \right).
  • Avg coal. regret: The avg takes the empirical mean across both the coalition dimension and the profile dimension: Avg coal. regret=1K(2n−1)∑k=1K∑∅≠C⊆NR^C(v(k)).\text{Avg coal. regret} = \frac{1}{K(2^n - 1)}\sum_{k=1}^{K} \sum_{\emptyset \neq C \subseteq N} \hat{R}_C(\boldsymbol{v}^{(k)}).

We will add these explicit definitions and equations to Section 5.1 and the caption of Table 1 in the revised manuscript.

Thank you for the authors' thorough response. Most of my main concerns are well addressed, and I decide to maintain my positive score.

Some comments:

  1. ICLR allows the authors to revise their manuscripts by submitting a new PDF. I would like to see that the authors implement their promised revision.
  2. Thank you for the further clarifications on the two challenges. The challenges and the proposed remedies really make sense to me, and I believe that the proposed remedies have independent value. It would be nice if the authors could conduct ablation studies to see how the ablated methods would perform if the challenge was not resolved. It would also be acceptable without additional ablation studies.
  3. For baselines, my initial concern is to compare CoalitionNet with baselines that are guaranteed to be collusion-resistant, not Itemwise Myerson. Itemwise Myerson is not collusion-resistant, so it is not the optimal baseline. I have not gone too far with revenue-optimal collusion-resistant single-item auctions, but I think "selling the item to the highest bidder with reserve price rr" should be a legal one, and I would like to see how CoalitionNet would compare with it (as well as other possible collusion-resistant baselines).
Reviewer_NbT5OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Learned auction mechanisms in the RegretNet family are trained to resist one bidder lying, but not groups. The paper's target is coalition regret: the largest joint gain a group can get by misreporting together while everyone else stays truthful. Enumerating coalitions is exponential, and plain gradient ascent on joint misreports is weak because collusion needs coordinated bid changes.

The proposed mechanism is a Transformer that outputs feasible allocations and payments, trained against a learned generator that proposes collusive bids. Instead of enumerating coalitions, coalition formation is modelled as a pivot move: pick a pivot bidder and a group that is already deviating, measure the utility each bidder receives when the pivot joins, and note that for given externalities the best coalition adds every bidder with a positive externality. The training score is the smaller of the gain around the pivot and the total gain. The theory is one proposition: summing the externality objective over pivots stays within a factor of the number of bidders of the true worst-case coalition regret. Experiments use five bidders and five items with additive values, against RegretNet, RegretFormer and AMenuNet.

Critical strengths
  1. The externality construction is a tractable surrogate for worst-case coalition regret, with a proven guarantee. Modelling coalition formation as a pivot move with positive-externality additions replaces an exponential search with a polynomial one, and the proposition bounds the surrogate by the number of bidders times the true worst-case regret. Coalition-robust training is otherwise infeasible at realistic sizes, and the guarantee means the surrogate is not merely heuristic.
  2. The empirical contrast is large and the revenue scale checks out. Attacked revenue stays at 2.75 for this mechanism against 0.09–0.38 for the baselines, and since the optimal revenue for this setting is about 3.36 — which AMenuNet matches — the cost of robustness is a real sacrifice of about 12% rather than an artefact of weak competitors. That is consequential because the standard objection to coalition-robust mechanisms is that they destroy revenue; this result shows the loss is bounded and modest.
Critical weaknesses

No ablation isolates the externality generator, so the method's premise is untested.

Questions for authors

NA

2 comments
AuthorsOpenReview ↗

Ablation Study: Validating the Externality Guidance

We thank the reviewer for pointing out the need for an ablation study. We provide an additional controlled experiment to isolate the contribution of the externality guidance ΦA\Phi_A.

Response to W1. We conducted a paired ablation study in the 5×55 \times 5, U[0,1]\mathcal{U}[0,1] setting using three random seeds (42, 43, and 44). Both variants use the same training process and mechanism architecture. The only difference is whether the externality guidance ΦA\Phi_A is used for the collusion generator training.

For evaluation, we enumerate all 25−1=312^5 - 1 = 31 non-empty coalitions for each of K=1,000K = 1{,}000 profiles, and use five restarts with 50-step projected gradient ascent to maximize the true coalition utility UCU_C.

VariantΦA\Phi_AAvg. coalition regret ↓\downarrowMax coalition regret ↓\downarrowRevenue drop ↓\downarrow
CoalitionNet✓\checkmark0.0063 ± 0.00040.0133 ± 0.00060.2239 ± 0.0027
no_externality×\times0.1562 ± 0.00500.5122 ± 0.02250.9081 ± 0.0163

The results show that removing ΦA\Phi_A increases the average coalition regret from 0.00630.0063 to 0.15620.1562 and the maximum coalition regret from 0.01330.0133 to 0.51220.5122. The revenue drop also increases from 0.22390.2239 to 0.90810.9081. These results indicate that directly optimizing sampled coalition utility can miss profitable coalition deviations, whereas the externality-guided surrogate helps the collusion generator identify more harmful deviations and produces a stronger adversary during training. This performance gap isolates the contribution of ΦA\Phi_A and validates the externality guidance as a key component for detecting and suppressing coalition deviations. We will incorporate ablation results and the corresponding analysis into the revised manuscript.

Thanks for your response. The ablation experiment addresses my concern.

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

The paper introduces CoalitionNet, a new machine learning framework designed to solve the multi-bidder, multi-item auction mechanism design problem while providing robustness against bidder collusion. While recent deep learning approaches have made progress in designing revenue-maximizing auctions by minimizing individual agent regret, they remain strictly vulnerable to coordinated strategic behavior (collusion), which historically incurs massive platform revenue drops. CoalitionNet addresses this by shifting the optimization paradigm to minimize coalition regret via an adversarial learning process between the auction mechanism and a custom collusion generator. To circumvent the exponential computational overhead of searching through all possible bidder subsets, the authors introduce an externality-guided collusion generator that evaluates marginal utility gains to identify strategic joint deviations efficiently. Empirical evaluations show that CoalitionNet drives coalition regret near zero, sustains competitive revenue, and successfully generalizes across unseen valuation distributions and varied coalition sizes.

Critical strengths
  • evaluating joint deviations usually triggers an exponential search explosion over all possible bidder combinations (2N2^{N}). By introducing an externality-guided selection mechanism based on pivot bidders' marginal utilities, the framework achieves computational tractability without sacrificing performance.
  • the proposed framework does not just fit a single localized setting; the experiments demonstrate good generalization properties across both unseen coalition scales and entirely distinct valuation distributions, proving its viability under environmental drift.
Critical weaknesses
  • the externality-guided generator assumes a structured method of tracking marginal utility to find joint deviations. If real-world colluders use highly complex, asymmetric side-payment contracts or non-linear strategies that mask standard externalities, the adversarial generator might fail to accurately simulate and train against the true worst-case deviations.
  • training an adversarial setup introduces the classic risks of minimax optimization, such as mode collapse, oscillatory training dynamics, or failure to converge. there is a lack of detailed discussion on the stability techniques required to reliably train the framework.
Questions for authors

How does the collusion generator model the internal distribution of utility within a coalition? Can CoalitionNet defend against advanced bidding rings that utilize complex, asymmetric side-payment rules to incentivize joint deviations that do not look uniform on the surface?

1 comment
AuthorsOpenReview ↗

Responses on Coalition Surplus, Side-Payment Robustness, and Training Stability

We thank the reviewer for the insightful comments regarding internal side-payments and adversarial training stability. We address each weakness and question in detail as follows:

Response to W1. In our problem formulation, bidders have quasilinear utilities, ui=vi⋅xi−piu_i = v_i \cdot x_i - p_i, so an internal side-payment is a monetary transfer among coalition members.

For example, since we have

Δuj=uj(vj;(b~C,v−C))−uj(vj;v),\quad \Delta u_j = u_j(\mathbf{v}_j; (\tilde{\mathbf{b}}_C, \mathbf{v}_{-C})) - u_j(\mathbf{v}_j; \mathbf{v}),

if internal transfers τj\tau_j satisfy ∑j∈Cτj≤0\sum_{j \in C} \tau_j \le 0 and make every member weakly better off, then

∑j∈CΔuj+∑j∈Cτj>0,\quad \sum_{j \in C} \Delta u_j + \sum_{j \in C} \tau_j > 0,

which implies ∑j∈CΔuj>0\sum_{j \in C} \Delta u_j > 0. Conversely, if a deviation gives the coalition ∑j∈CΔuj>0\sum_{j \in C} \Delta u_j > 0, this surplus can be redistributed through complex internal transfers so that every member is weakly better off. Thus, complex side-payments change only how surplus is allocated within a coalition; they do not determine whether the coalition’s deviation is profitable. Accordingly, CoalitionNet focuses on minimizing the coalition’s aggregate utility gain. When the aggregate gain is minimized, profitable space for internal side-payments is also minimized.

Response to W2. In auction design with adversarial learning, mode collapse typically means that the mechanism converges to a trivial no-allocation solution. We use two ways to prevent for this failure mode. First, the mechanism objective retains the truthful-revenue term, so a zero-allocation mechanism obtains nearly zero revenue and is not optimal. Second, we use a dynamic parameter λt\lambda_t for the coalition-regret penalty: λt\lambda_t is adjusted according to the mean coalition regret.

We have the following table which reports the mean training trace over seeds 42–44 at 500-iteration intervals:

Iter.RevenueAvg. coalition regretMax coalition regretλt\lambda_t
5002.6270.00620.015940.00
10001.7740.00070.002696.45
15001.8730.00060.002364.29
20002.3130.00080.003338.22
25002.6000.00160.005422.98
30002.7950.00300.008716.80
35002.8280.00320.008914.95
40002.8980.00340.009514.17
45002.8960.00340.009714.07
50002.9150.00360.010014.07

The trace exhibits transient adversarial fluctuations: revenue drops from 2.627 at iteration 500 to 1.774 at iteration 1,000, while λt\lambda_t increases sharply. These dynamics indicate that model collapse remains a possible risk in minimax training. However, collapse would correspond to a persistent drift toward zero revenue and a no-allocation mechanism. Under our two ways, revenue recovers to 2.915 by iteration 5,000, λt\lambda_t stabilizes around 14, and maximum coalition regret remains close to the target level. Thus, although we do not claim that collapse is impossible, none of the training seeds result in model collapse.

Response to Q1. CoalitionNet does not explicitly parameterize the internal distribution of side-payments. Instead, it optimizes aggregate coalition utility. Under bidders' quasilinear utility model and unrestricted internal monetary transfers, positive aggregate surplus is sufficient to compensate members who lose utility through transfers from members who gain. Thus, the generator can capture non-uniform bidding rings in which some active deviators sacrifice utility while another member obtains a larger gain, but the exact payment split is not needed to determine whether collusion is profitable or not. When aggregate coalition gain is near zero, nearly no profitable internal side-payment scheme exists.

Comments on the paper

1 comment

Dear Authors and Reviewers,

Thank you for supporting ICLR'27!

We are now entering the discussion phase. Please make the most of this time to exchange your views. I encourage each Reviewer, if you have not already done so, to read the other reviews and the Authors' responses. Please respond to the Authors and actively participate in the discussion to help clarify any remaining issues.

Thank you,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.