1 views collected around this technical thread.
This article presents MaRCA, a multi‑agent reinforcement‑learning framework that allocates computation resources across the full ad‑serving chain, modeling user value, compute cost, and action rewards to maximize ad revenue while keeping system load stable under fluctuating traffic.
2024 saw the advertising team achieve major breakthroughs in AI-generated ad creatives by introducing a multimodal reliable feedback network to improve image usability, releasing a large human-annotated dataset, and leveraging multimodal large language models for richer representation and more effective online/offline creative selection.
Xiaohongshu’s new “traffic‑optimal + dynamic decision” framework models splash‑screen ad allocation as a linear‑programming problem with volume guarantees, continuously adjusts weights via feedback, and pre‑computes cached decisions to preserve fast app startup, thereby boosting click‑through rates while meeting delivery commitments.
The paper proposes a teacher‑student architecture that uses feedback cascading and learning‑to‑rank models with ΔnDCG‑based loss and bid‑aware optimizations to align coarse and fine sorting stages, address sparse data, and improve recommendation consistency, achieving a 4.8% RPM lift in A/B tests.