Paper Readings

Notes from papers I actually sat with — architecture, training tricks, and the occasional thing that changed how I read the next one.

Transformer architecture

Attention Residuals: A Comprehensive Understanding

· 15-20 min read

This paper addresses a fundamental problem in training deep transformer models: uncontrolled hidden-state magnitude growth as model depth increases. The authors propose Attention Residuals (AttnRes), which replaces standard...

More to come.