Reproduce “Constraint-Aware Generative Re-ranking for Multi-Objective Optimization in Advertising Feeds” 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
Chenfei Li, Hantao Zhao, Weixi Yao, Ruiming Huang, Rongrong Lu, Geng Tian, Dongying Kong
The Constraint-Aware Generative Re-ranking (CGR) framework from Bilibili integrates multi-objective optimization and strict business constraints directly into a unified autoregressive model for advertising feeds. This approach reduces inference latency by over 85% and improves Revenue Per Mille by 6.8% and Click-Through Rate by 4.9% in online A/B tests.
Key insights
- Explicitly leveraging the bounded nature of advertising constraints (e.g., few ads per page) allows for transforming a factorial combinatorial optimization problem into a tractable, bounded autoregressive decoding task.
- Unifying sequence generation, reward estimation, and direct constraint handling within a single neural network significantly reduces inference latency and obviates the need for separate evaluator networks or external post-processing.
- Theoretical guarantees demonstrate that the bounded decoding procedure can achieve global optimality for the constrained optimization problem by exhaustively enumerating the tractable feasible set.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.