Reproduce “Expert-Guided Diffusion Planner for Auto-Bidding” 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
Yunshan Peng, Wenzheng Shu, Jiahao Sun, Yanxiang Zeng, Jinan Pang, Wentao Bai, Yunke Bai, Xialong Liu and 1 more
Kuaishou Technology researchers developed the Expert-Guided Diffusion Planner for Auto-bidding (EGDB), an auto-bidding framework that leverages expert trajectories with VAE-enhanced blended-forcing and an Expert-Guided Conditional Diffusion (EGCD) block, alongside accelerated sampling for real-time deployment. It achieved an 11.29% increase in conversions and a 12.36% growth in revenue in online A/B tests, while reducing inference resources by 25%.
Key insights
- Leveraging rich, personalized expert trajectories for guidance, rather than simple scalar returns, is crucial for generating optimal and structurally coherent bidding sequences in dynamic environments.
- Combining stable teacher-forcing with VAE-driven decode-forcing effectively mitigates exposure bias, allowing the model to not only imitate observed expert behavior but also generalize to and explore unseen optimal patterns.
- Employing a cross-attention mechanism within the diffusion model's architecture enables principled integration of both learned expert patterns and hard business constraints, ensuring adaptivity and compliance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.