M+Adam Improves Low-Precision Training

AnimaAnandkumar · x · 2026-07-14

Recent findings show that for transformer models with **under 1B parameters** and training budgets ranging from **1x to 8x Chinchilla**, **M+Adam** consistently outperforms **AdamW** in low-precision master-weight settings. Key conclusions: - The benefits are most significant in **more aggressive low-precision training** scenarios. - Low-precision training requires the optimizer's update geometry to align with the floating-point grid. - M+Adam combines scale-aware progression and local correction by performing both **multiplicative updates** and **additive updates** within a single optimization step. - This method supports **end-to-end low-precision training**, eliminating the reliance on higher-precision master weights.

Related event: M+Adam Optimizer Improves Low-Precision Training(2 posts)→

Original post →

More from Research

Research channel →