Reproduce “Understanding and Mitigating Training Instability in Large-Scale Recommender Systems: A Cascade Perspective” 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
Yuening Li, Zeyu Wang, Zhen Zhang, Jiayi Liu, Zhimeng Jiang, Yuji Roh, Weikai Xu, Mingyan Gao and 3 more
The authors frame training divergence in short-form video recommendation as a possible G-W-L cascade and test interventions targeting its stages. The combined treatment uses gradient controls, activation clipping, and Huber loss. On BC T1, it reports a 0.25% relative AUC increase over baseline. Aggregated measurements support interdependence, but do not establish the proposed sequence for every divergence event.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.