Token-Level Advertising
The paper proposes LAMA, an auction system that lets advertisers shape an AI model’s next-token choices during generation.
Advertisers report local continuation values, which are turned into advertiser-specific next-token policies. The platform decodes from a latent mixture of those policies while updating an allocation posterior. The authors claim the mechanism is incentive-compatible and individually rational in a Markov setting, with near-optimal KL-regularized welfare. Proof-of-concept tests on commercial-search query splits showed higher platform welfare and revenue while preserving response quality. ArXiv · AI/CL/LG's note
Advertisers report local continuation values, which are turned into advertiser-specific next-token policies. The platform decodes from a latent mixture of those policies while updating an allocation posterior. The authors claim the mechanism is incentive-compatible and individually rational in a Markov setting, with near-optimal KL-regularized welfare. Proof-of-concept tests on commercial-search query splits showed higher platform welfare and revenue while preserving response quality. ArXiv · AI/CL/LG's note
score 4