NoRA: a one-line normalization fix for LoRA boosts convergence, stability and fights forgetting

omarsar0 · x · 2026-09-07

Elvis Saravia highlights NoRA (Normalized Low-Rank Adaptation): a one-line change to LoRA that costs nothing and improves convergence, stability and forgetting.

Key insight: LoRA initializes the up-projection to zero, so early optimization is governed almost entirely by the down-projection — which tells you where to regularize.

Method: normalize the down-projection matrices during training; the authors also show normalizing once at initialization improves standard LoRA, the cheaper of the two options.

Results: benefits hold across pretraining, SFT and RL — faster convergence, better final performance, more stable training, less catastrophic forgetting. No added trainable parameters or inference-time compute, making it broadly applicable.

Related event: New Research Explains LoRA's Slow Convergence, Proposes One-Line NoRA Fix(3 posts)→

Original post →

More from Research

Research channel →