Introduction
The digital advertising landscape is currently dominated by auctions for static slots—fixed boxes on a webpage or specific positions in a search result. However, the rise of Large Language Models (LLMs) as primary interfaces for information retrieval and content creation necessitates a shift in how we think about "ad slots." In a generative environment, the item of value is not a static image but influence over the model's output. We introduce the Token Auction Model, a novel framework designed to handle auctions where multiple advertisers compete to influence a single, cohesive AI-generated creative.
The Token Auction Model
Traditional auction theory relies on explicit valuation functions (e.g., "I value this slot at $5"). LLMs, however, express preferences through token-by-token probability distributions. Our model treats each advertiser as an LLM "agent" that maps a sequence of tokens to a probability distribution over the next possible token.
The mechanism operates at the fundamental level of LLM generation:
- Token-by-Token Aggregation: Instead of selecting a single winner, the auction aggregates the preferred distributions of all bidding agents at every token step to produce a joint output distribution.
- Incentive Design: We define two critical properties for the aggregation function:
- Payment Monotonicity: An increase in an agent's bid must result in a final distribution that is "closer" to that agent's preferred distribution.
- Consistent Aggregation: Changes in one agent's bid should not arbitrarily disrupt the relative influence of other agents.
Aggregation Rules and Theoretical Insights
We explore two primary mathematical approaches to combining distributions: Linear Aggregation (a weighted average of token probabilities) and Log-linear Aggregation (minimizing KL-divergence).
A key theoretical contribution of this work is the Monotonicity Equivalence. We prove that satisfying basic incentive properties is mathematically equivalent to a monotonicity condition on the aggregation function. This allows for the design of "second-price" auction analogues even when the agents' internal valuation formulas are unknown. Furthermore, because state-of-the-art LLMs are stateless—predicting the next token based solely on the current prefix—the auction can proceed without requiring agents to plan entire sentences in advance.
Results and Discussion
To evaluate the mechanism, we simulated a co-marketing scenario between two fictitious entities: "Alpha Airlines" and "Beta Resort." By modifying a Google Bard model to access internal token distributions (logits) and using prompt tuning to create distinct agent personas, we tested how varying bid ratios (represented by the parameter $\lambda$) affected the output.
As shown in Table 1, the mechanism produces a smooth spectrum of content. At high $\lambda$ values, the output focuses exclusively on the airline. As $\lambda$ decreases, the LLM generates blended content (e.g., "Fly to Hawaii... and experience the beauty of the islands with... Beta Resorts"). At the lowest values, the focus shifts entirely to the resort. Notably, while Log-linear aggregation produces high-quality text, only the Linear Aggregation rule was found to be inherently monotone and thus incentive-compatible.
Limitations and Future Directions
While the Token Auction Model provides a robust theoretical framework, practical implementation faces challenges. Querying multiple LLMs at every token step introduces significant computational overhead. Additionally, the requirement for "white-box" access to model logits may be difficult in an ecosystem of proprietary APIs. Future research will explore applying this framework to other generative media, such as image and video segments, and investigating potential strategic manipulations of token distributions.
Summary updated