核心概念
A factorized framework combining an auction module and an LLM module to generate welfare-maximizing ad summaries in an incentive-compatible manner.
摘要
The content discusses a novel framework for running auctions to generate summaries of ads using a large language model (LLM). The key components are:
- Auction Module:
- Determines the relative prominence (allocation) of each ad in the summary based on bids and predicted click-through rates (pCTRs).
- Ensures incentive compatibility by having a monotonic allocation function and using Myerson's payment rule.
- LLM Module:
- Generates the ad summary based on the prominence allocation from the auction module.
- Satisfies a "faithfulness" property where the user attention to each ad is proportional to its prominence.
- pCTR Module:
- Provides unbiased estimates of the click-through rates for each ad given the prominence allocation.
The authors show that this factorized framework is without loss of generality and can achieve incentive compatibility. They also analyze the welfare-maximizing auction design for a specific case of "Dynamic Word Length Summaries".
Experiments on synthetic data demonstrate the feasibility and efficiency of the proposed framework compared to simpler baselines.
統計資料
The bids bi of each ad i follow a log-normal distribution LogNormal(0.5, 1).
The base click-through rate CTRi of each ad i is sampled from a uniform distribution Unif[0, 1].