[Submitted on 31 Oct 2025]
Analysis of Spectral Momentum Optimizer
View PDFAbstract:This paper analyzes a spectral momentum optimizer for transformers that achieved worse performance (loss=9.85) compared to standard baselines like AdamW (loss=4.93). We identify key challenges in applying spectral methods to transformer optimization.
Submission history
[v1] Fri, 31 Oct 2025 06:24 UTC