Dynamic Ad Allocation in Large Language Model Responses Using Inference-Time FST-Constrained Auction
Abstract
A system dynamically integrates advertisements into large language model (LLM) responses using an auction mechanism conducted at inference time. During content generation, the system evaluates a pool of candidates that includes both advertiser-provided ad copy and an organically generated text continuation. To score each candidate, a finite state transducer (FST) constrains the LLM to evaluate the specific text, and the system extracts a length-normalized token log-probability to serve as a unified measure of contextual relevance. An auction score for each candidate combines this relevance measure with the advertiser’s bid, enabling a direct trade-off between revenue and the naturalness of the ad’s placement. The candidate with the highest score wins the auction, and the system inserts it into the LLM’s output. The process can operate as a sequential Vickrey auction, running at specific trigger points during generation and determining a clearing price for the winning ad. The system can also adjust a parameter that weights the relevance penalty after each ad insertion, making it progressively more difficult to show another ad. This training-free approach ensures ad copy fidelity while creating a method for ad allocation in conversational interfaces.
Keywords: large language model (LLM), ad allocation, finite state transducer (FST), inference-time auction, log-probability extraction, auction scoring, organic generation, candidate pooling, constrained evaluation, autoregressive generation
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
McColgan, Thomas, "Dynamic Ad Allocation in Large Language Model Responses Using Inference-Time FST-Constrained Auction", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12367