Abstract
We investigate a Bayesian mechanism design problem where a seller seeks to maximize revenue by selling an indivisible good to one of n buyers, incorporating potentially unreliable predictions (signals) of buyers' private values derived from a machine learning model. We propose a framework where these signals are sometimes reflective of buyers' true valuations but other times are hallucinations, which are uncorrelated with the buyers' true valuations. Our main contribution is a characterization of the optimal auction under this framework. Our characterization establishes a near-decomposition of how to treat types above and below the signal. For the one buyer case, the seller's optimal strategy is to post one of three fairly intuitive prices depending on the signal, which we call the "ignore", "follow" and "cap" actions.
Introduction
In modern digital markets, such as real-time advertising auctions, sellers leverage sophisticated machine learning (ML) models to predict bidder valuations. Traditional economic theory typically models prediction error as statistical noise (e.g., Gaussian noise), where the signal is assumed to be in the proximity of the truth. However, modern ML systems—particularly generative models—often fail through "hallucinations." In these instances, a model may produce a confident but entirely irrelevant output when encountering novel contexts.
This phenomenon introduces a significant technical challenge for auction designers. When a signal has a high probability of being exactly correct, the seller's posterior belief contains a "point mass" (an atom) at the signal value. Because standard auction theory, such as Myerson’s 1981 framework, relies on smooth probability densities, these spikes in probability cause traditional models to break down. This paper provides the mathematical framework necessary to design optimal auctions in this "randomly wrong" regime.
The Hallucination Framework
The authors model the ML prediction $s$ as a binary state: with probability $1-\gamma$, the signal is perfectly accurate ($s = v$); otherwise, it is a hallucination drawn independently from the prior distribution. To solve for the optimal auction, the research moves beyond standard Myerson techniques, employing the "ironing" formalism developed by Monteiro and Svaiter (2010).
A key technical contribution of this work is proving that in this hallucination-prone setting, the complex optimization problem admits a closed-form solution. The authors focus specifically on signal-revealing direct mechanisms, where the seller publicly discloses the ML signal. This approach is motivated by the need for "credibility"; if a seller uses a private, unverifiable signal to set prices, bidders may perceive the auction as unfair or untrustworthy.
Optimal Pricing Strategies
In traditional noise models, the optimal response to a signal is "shrinkage"—pulling the price slightly toward the average. In the hallucination model, however, the optimal response is qualitatively different and based on specific regimes. For a single bidder, the optimal price follows a four-regime pattern based on the signal's magnitude:
- Ignore: If the signal is very low, it is likely a hallucination. The seller should ignore it and stick to the standard monopoly price.
- Follow: For intermediate signals, the prediction is likely accurate. The seller sets the price equal to the signal.
- Cap: For moderately high signals, the risk of a hallucination (and thus losing the sale entirely) is too high. The seller "caps" the price below the signal to hedge against error.
- Follow Again: For extremely high signals, the potential payoff is so significant that it justifies the risk of a hallucination; the seller returns to following the signal.
This structure is supported by a property the authors call "near-decomposition," which allows the mathematical treatment of values below and above the signal to be solved almost independently.
Results and Performance
While the paper is primarily theoretical, the authors evaluated their mechanisms using numerical experiments on standard probability distributions, including Uniform, Beta, Exponential, and Lognormal distributions. Key findings include:
- Optimal Characterization: The paper provides the first closed-form characterization of "ironed virtual values" for distributions with atoms at the signal.
- High Efficiency: In multi-bidder simulations, the proposed "k-uncapped eager auction"—which adjusts reserves to account for hallucinations—achieves at least 97% of the theoretically optimal revenue.
- Robustness in Heavy Tails: For heavy-tailed distributions (common in ad auctions), the mechanism improves revenue by 15% to 20% compared to naive benchmarks that either treat the signal as 100% reliable or ignore it entirely.
- Qualitative Superiority: The "ignore/follow/cap" structure remains robust even when small amounts of Gaussian noise are added to the hallucination model, suggesting it is a superior mental model for modern ML-assisted pricing.
Conclusion
This research establishes a rigorous foundation for auction design when signals are prone to hallucinations rather than just simple noise. By characterizing the optimal auction and providing intuitive pricing regimes, the authors offer a path forward for integrating "black box" ML predictions into credible, revenue-maximizing market mechanisms. Future work may extend this framework to scenarios where the hallucination probability ($\gamma$) is unknown or fluctuates over time.
Summary updated