acceptodds
Under review as a conference paper at ICLR 2027

Incentive Compatibility Mechanism Design for the Auto-bidding Meta-Game

Abstract

In auto-bidding, advertisers no longer bid per impression but set a high-level target (e.g. tCPA), competing in an Auto-bidding Meta-Game (AMG) over reported targets. Running a truthful per-impression auction underneath does not make this meta-game truthful—an advertiser can still gain by misreporting its target. Within the platform-supported uniform-bidding interface, we formalize the AMG as a Myerson single-parameter environment and design a deployable Dominant-Strategy Incentive Compatible (DSIC) mechanism for it: with exact allocation information under first-price, truthfulness is restored by an end-of-day rebate that returns the advertiser's information rent and leaves the auction untouched. The truthful profile is also welfare-optimal under first-price. For other pricing rules the same payment construction is DSIC whenever the induced allocation is monotone; second-price numerical audits expose allocation nonmonotonicity and integration-sensitive deviation gains, so they do not validate that condition, though welfare optimality does not carry over. Deployment estimates the allocation curve offline; its signed settlement implements an approximately DSIC payment, not the exact rebate in general. We prove that uniform allocation error yields approximate DSIC. Under an additional uniform quadratic-growth condition, damped best responses track a neighborhood of truth; convergence to truth requires vanishing estimation error, not merely repeated estimation. We evaluate these theoretical predictions on a synthetic simulator across market sizes and value/conversion-rate distributions. An online A/B test on a deployed tCPA product gives an encouraging behavioral signal: a significant market-wide gain in cost, and significant gains in reported targets for the cohort it was designed to move.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Auto-bidding, Mechanism Design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_eArYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks how to charge advertisers so that they report their true values through an automated bidder's target-CPA interface. It adjusts payments at the end of the day using Myerson's formula, analyzes estimation error, and provides preliminary checks through simulations and an online discount experiment.

Critical strengths
  1. The approach changes only end-of-day payments while preserving the auction ranking. This is straightforward and potentially useful in practice. (pp. 4–5)
  2. The paper distinguishes the guarantees with exact allocation information from those under estimation error, and states the extra conditions needed for convergence. The proofs also handle stepwise allocation curves. (Theorems 1 and 3; Proposition 2)
Critical weaknesses
  1. It is unclear why the entire allocation curve needs to be learned. Under the paper's first-price setting and complete logs, the exact payment can be computed by summing the highest competing score for each won impression. What does the proposed approach add beyond this critical-price settlement? (pp. 3–6)
  2. The key experiments do not directly address the main concern. E2 does not test over-reporting. E3 uses the estimated allocation in both the value and payment terms, so truthful reporting is already optimal for this surrogate objective over a continuous report domain. The authors acknowledge these limitations, but evidence using actual utility is still lacking. (pp. 8–9)
Questions for authors
  1. Given the assumed logs, what prevents direct settlement at the critical price for each won impression? What additional problem does learning the full allocation curve solve?
  2. Do the deviation gains and dynamics still support the conclusions when E2 covers the full allowed report range and E3 uses the value of the actual allocation minus the estimated payment?
1 comment
AuthorsOpenReview ↗

§3.1 corrected and Appendix H added; E2 extended to over-reports; E3 re-run under actual utility

Weakness 1 and Question 1: why learn the whole curve rather than settle at critical prices.

You are right, and we have corrected the text. In the idealized model of §2---first-price ranking, fixed rivals, a complete candidate queue per impression, and an entry threshold that does not move when the advertiser changes its own report---the information rent has a closed form: writing m−i,j=max⁡k≠iτkqk,jm_{-i,j}=\max_{k\ne i}\tau_kq_{k,j} and xi,j=m−i,j/qi,jx_{i,j}=m_{-i,j}/q_{i,j}, the Myerson payment is exactly ∑j:xi,j<τim−i,j\sum_{j:x_{i,j}<\tau_i}m_{-i,j}, the sum of the critical scores of the impressions won. No curve need be fitted, and the settlement carries the same requirement we already impose on the rebate: applied after the auction and excluded from the pacing feedback. Nor is it limited to one realized day: sorting the thresholds once and accumulating along that order answers the payment at any report, so the exact route delivers the whole billing rule. Within the model your proposal is therefore correct and complete, and the submitted sentence "we must estimate GiG_i in order to compute the integral" was wrong as stated. We have replaced it (§3.1).

What the revision adds, in the new Appendix H, is why our own deployment estimates the curve regardless. The difficulty is not the arithmetic but the conditions the identity assumes, none of which our production setting supplies exactly.

Billing is per click and the factors entering qi,jq_{i,j} are themselves predicted, so the charge at a report the advertiser never submitted is a prediction rather than a sum over observed events: the impressions it would have won at that report have no realized clicks, and replaying the log does not recover them. The replay also reproduces the logged auction rather than the serving system around it---eligibility and targeting filters, traffic routing, frequency and quality controls, campaign-level overrides---all of which interact with a changed report in ways we neither model nor have measured, so exact arithmetic on replayed inputs is exact with respect to the replay and not with respect to production. The third obstacle is cost, and again not of the summation: assembling and refreshing a per-unit counterfactual queue every day sustains a few hundred units at our throughput, against the platform's active population.

These are reasons the exact-threshold route is unavailable to us, not claims that it is infeasible in general. Estimating GiG_i does not make the gap disappear; it puts the gap in a form the analysis can carry, since Theorem 3 bounds the IC slack by the uniform allocation error whatever its source.

Weakness 2 and Question 2, first part: E2 does not test over-reporting.

Fair, and now fixed. E2's grid previously ran over [0.02vi,vi][0.02v_i,v_i]; it now covers [0,Vsim][0,V_{\mathrm{sim}}] with the same markets, seeds, subsamples, estimator and integration, so only the search range changed. The measured deviation gain at n=104n=10^4 rises from 1.66%1.66\% to 4.45%4.45\% of gross truthful value, and the p95 of market means from 5.07%5.07\% to 15.88%15.88\%. The original under-report statistics reproduce exactly, so the comparison is paired.

Weakness 2 and Question 2, second part: E3 optimizes a surrogate under which truth is optimal by construction.

You correctly identified that optimizing a proxy built from the same monotone estimate cannot validate best responses under the actual allocation, and you asked what happens when the objective becomes the value of the actual allocation minus the estimated payment. We re-ran the dynamics with exactly that decision rule, now Appendix J: same markets, seeds, subsample stream, grid, damping and horizon, with the value term the only difference, so the two runs are paired market by market and the proxy branch reproduces Table 8. The terminal report error among truth-active advertisers is larger under actual utility in every one of the twelve configurations, by a median factor of 2.62.6; since that factor is unstable where the proxy error is near zero, the absolute values are the informative ones, and at N=50N=50 the error roughly doubles, from 0.0360.036–0.0410.041 to 0.0690.069–0.0900.090. Median shade stays within 0.980.98–1.001.00, so the gap is dispersion across advertisers rather than systematic shading. E3's near-truthful trajectory is therefore a property of its proxy objective and should not be read as evidence about behaviour under the real allocation.

Reviewer_bc2yOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies advertisers who report a target cost per acquisition to a platform’s auto-bidding algorithm rather than bidding directly on individual impressions. Advertisers have private per-conversion values and quasi-linear utilities, while the platform uses a restricted uniform-bidding interface with nonbinding budgets. The objective is to make truthful target reporting a dominant strategy through an allocation-preserving settlement, leaving the underlying auction unchanged.

The paper applies Myerson’s payment characterization to the induced allocation curve. Under its first-price model, the exact payment can be implemented through a rebate, and truthful reporting yields efficient allocation and individual rationality. It develops an offline replay estimator and shows that uniform allocation-estimation error implies approximate incentive compatibility. Additional report-displacement, welfare, and dynamic-convergence guarantees require quadratic-growth assumptions. Simulations examine shading and estimation sensitivity, while a production A/B test evaluates advertiser responses to a related discount scheme.

Critical strengths

This paper makes a solid contribution to the autobidding literature. To me, this problem looks theoretically natural.

The authors provide a comprehensive set of results for their model. They first characterize the Myerson-style condition for DSIC and the welfare-optimal implementation. The authors then study the estimation of the allocation and the impact of estimation error on IC and other guarantee.

The experiments (in particular online A/B test) provide a good complement to the theoretical contribution.

Overall, the paper is well-written. The results and analysis are clearly explained, and I enjoyed reading it.

Critical weaknesses

In this work, the authors consider a model without budget constraint (equivalently, the budgets are non-bidding). I understand this could be a necessary model simplification for the theoretical results. Nevertheless, since budget is a crucial component in the real-world autobidding system, it would be helpful if the authors could add some discussion about how their theoretical results and takeaway message changes when we incorporate the budget.

Compared with conceptual contribution and empirical findings, the technical contribution is relatively weak. Most theoretical results and their analysis are expected once the model is given. However, I still appreciate the authors' effort to establish them and present them nicely.

Questions for authors

NA

2 comments
AuthorsOpenReview ↗

binding budgets under pacing now discussed in Appendix L

Thank you for the careful reading and for the positive assessment.

Weakness 1: what changes when budgets bind.

We have added this discussion to Appendix L. Keeping the rest of the model fixed, a fixed exogenous cap has a short answer under pacing. The reported target bounds what the bidder pays per conversion and the pacer scales it by αi≤1\alpha_i\le1; writing βi=αiτi\beta_i=\alpha_i\tau_i and βi∗\beta_i^{\ast} for the largest coefficient the cap affords, the induced allocation is Gi(min⁡(τi,βi∗))G_i(\min(\tau_i,\beta_i^{\ast})). It follows the unconstrained curve while the budget is slack and is constant once the budget binds, because raising the target past βi∗\beta_i^{\ast} buys no additional traffic. It is therefore still non-decreasing in the advertiser's own report, Theorem 1 continues to apply, and truthful reporting remains optimal. On the flat part neither side gains: conversions, payment and advertiser payoff are all unchanged. That is why we did not develop the case further. It does leave truth weakly rather than uniquely optimal, which is the identification issue of Appendix I, to which a binding budget is simply a wider contributor.

A more consequential limitation is the advertiser model itself. Some units in the production test lowered their reported target after receiving a discount, which the utility-maximizer objective does not explain under any budget regime. Finding a model that captures how automated bidding agents actually respond, and designing the corresponding mechanism, is our main direction for future work.

Weakness 2: the technical contribution is relatively weak.

We agree that most of the theoretical results follow once the model is set up. Our view is close to yours: the work is in the modeling choice and in what it takes to deploy the construction, rather than in the derivations.

Thanks for your response. I am satisfied with both the paper and the response. I remain positive about the work and like its conceptual contribution.

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

The paper studies the problem of strategic target reporting by advertisers using auto-bidding systems. The paper proposes an allocation-preserving payment correction based on the Myerson payment identity, implemented as a rebate under exact first-price allocation information. It bounds incentive errors under estimated allocation curves and presents synthetic experiments alongside a production discount test

Critical strengths
  1. The paper clearly separates incentives over reported targets from incentives in individual auctions. This provides a useful framework for explaining why the underlying auction alone does not resolve target manipulation

  2. The exact mechanism and estimated implementation are carefully distinguished. In particular, the paper explains when settlement is a nonnegative rebate and when estimation can require a signed transfer

  3. The robustness analysis separates approximate incentive compatibility from stronger claims about report convergence. This avoids treating monotonicity or repeated estimation as sufficient for advertisers to converge to their true values

Critical weaknesses
  1. The central exact construction is an application of the classical single-parameter payment identity. The distinctive contribution therefore depends heavily on estimating and deploying the correction, where the paper provides more limited validation

  2. Experiment E3 optimizes a proxy utility that uses the estimated allocation for both value and payment. Because the estimated payment is constructed from that same monotone allocation, near-truthful behavior does not validate best responses under the actual allocation

  3. The report-convergence guarantees require uniform quadratic growth, which fails when the true value lies inside a flat interval of the allocation curve. Such intervals naturally occur in finite-impression markets, so these guarantees need stronger justification for the evaluated setting

4.The production test uses a discount channel that does not preserve allocation exactly, and private values are unobserved. Moreover, subgroup membership depends on utilization measured during the test, which limits a causal interpretation of the reported subgroup gains

Questions for authors

NA

1 comment
AuthorsOpenReview ↗

Appendix I measures where quadratic growth fails; Appendix J re-runs E3 under actual utility

Weakness 2: E3's proxy utility does not validate best responses under the actual allocation.

We agree, and the revision tests exactly that. Appendix J re-runs the E3 dynamics with advertisers best-responding to viGi(τ)−P^i(τ)v_iG_i(\tau)-\widehat P_i(\tau), the value of the actual allocation against the estimated payment, paired market by market with the original run so that the value term is the only difference; the proxy branch reproduces Table 8 node for node. Terminal report error among truth-active advertisers is larger under actual utility in all twelve configurations, by a median factor of 2.62.6, and at N=50N=50 it roughly doubles, from 0.0360.036–0.0410.041 to 0.0690.069–0.0900.090. Near-truthful behaviour in E3 is therefore a property of the proxy objective, as you said, and we no longer read it as evidence about the actual allocation.

Weakness 3: quadratic growth fails inside flat intervals.

This is correct and we have measured it. Appendix I reports exact best-response geometry across four value laws, four conversion-rate laws and three impression counts, 5,760 advertiser records in total. Every one of them has a non-unique best response, so no positive μ\mu exists at any of them; 4,800 of the report-interior types lie strictly inside a flat stretch of the allocation. This follows from counting discrete impressions, not from a particular distribution.

We also quantify what it costs. At M=20,000M=20{,}000 the median active advertiser can move its report by 0.0230.023–0.086%0.086\% of value without changing payoff, but the upper tail matters more for incentives: p95 width reaches 13.0%13.0\% under capped exponential values, and up to 18.3%18.3\% of active advertisers can deviate by a full 1%1\% at no cost (Table 9). More traffic narrows these intervals without removing them: for lognormal values the 1%1\% lossless rate falls from 17.2%17.2\% to 6.8%6.8\%, a paired reduction of 10.310.3 points (Table 10). Nonuniqueness bounds how precisely reports can be identified; it is not a profitable manipulation, since every report in the interval earns exactly the truthful payoff, so exact DSIC is unaffected. We therefore keep report convergence as a conditional extension with its growth assumption stated explicitly; the appendix delimits where the assumption fails rather than arguing it back into place.

Weakness 4: the production test.

We accept this criticism and have nothing to set against it: the discount is delivered through an existing billing-correction channel rather than our rebate, private values are unobserved, and eligibility is re-evaluated daily, so we claim nothing causal for the tCPA and budget deltas. The cost delta compares the randomized groups in full and is not subject to the subgroup point.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.