Reproduce “Generative Optimization for Incentivized Advertising with Global Level Constraints” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Gege Chen, Ning Luo, Hao Jiang, Da Li, Wenzheng Shu, Teng Sha, Yanxiang Zeng, Wenxin Tai and 2 more

Researchers from Kuaishou Technology and the University of Electronic Science and Technology of China developed GOAL, a generative optimization framework that models continuous incentive allocation for advertising while strictly enforcing global ROI constraints. The system achieved a 2.184% increase in ROI and a 2.559% lift in revenue in online A/B tests compared to a production baseline.

Key insights

  • Generative sequence models can effectively capture complex, non-Markovian user behaviors and generate precise continuous actions for incentive optimization, outperforming traditional MDP-based RL.
  • Conditioning the generative policy on a learned Lagrange multiplier and using a constraint-aware Mixture-of-Experts enables dynamic adaptation to varying global ROI constraints from a single trained model.
  • A data-driven incentive tokenization strategy combined with weighted next-token prediction effectively handles the wide range and heavy-tailed distribution of continuous incentive values, improving precision.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.