DigiNews

Tech Watch by Johan Denoyer

← Back to articles

Aurora: A Leverage-Aware Optimizer for Rectangular Matrices

Quality: 8/10 Relevance: 9/10

Summary

Aurora introduces a leverage-aware optimizer for rectangular matrices that addresses Muon's row-normalization issues on tall matrices. It presents both a practical damped-iteration Aurora and a Riemannian variant, reports strong results on 1.1B pretraining and nanoGPT speedrun benchmarks, and releases open-source code.

🚀 Service construit par Johan Denoyer