EVENT DOSSIER
The researchers proposed a Token-Level Advertising mechanism called Latent Advertiser Mixture Auction (LAMA), aimed at directly embedding advertisers’ influence into the content generation process. This mechanism requires advertisers to report local continuation values to induce specific next-word strategies, while the platform updates the assigned likelihoods through potential mixture decoding. The study demonstrated that LAMA satisfies the Markov DSIC and IR conditions, achieving nearly optimal KL-regularization benefits. Additionally, the team developed a learning-based implementation scheme that utilizes learned local advantages and the reported data required for online reconstruction of root values. In proof-of-concept experiments on real commercial search query segmentation, results showed that LAMA not only improved the platform’s overall benefits and revenue but also maintained user response quality.
提出一种名为 Latent Advertiser Mixture Auction (LAMA) 的 Token-Level Advertising 机制，将广告商影响力直接嵌入生成过程。该机制要求广告商报告局部延续值以诱导特定下一词策略，平台通过潜在混合解码并更新分配后验。研究证明 LAMA 满足 Markov DSIC 和 IR 条件，实现近最优 KL-正则化福利。此外开发了基于学习的实现方案，利用学习到的局部优势与根值在线重构所需报告。在真实商业搜索查询分割上的概念验证实验表明，LAMA 提升了平台福利与收入，同时保持用户响应质量。