TY  - RPRT
TI  - Optimizer-dependent training dynamics converge to the same one-third optimal data scaling
AU  - Hyunseok Lee
AU  - Mihir Basil
AU  - Yizhou Liu
AU  - Jeff Gore
PY  - 2026
UR  - https://arxiv.org/abs/2609.37745
ID  - 2609.37745
ER  -