NoRA: Normalized LoRA Improves Convergence and Stability

The NoRA paper proposes normalizing the down-projection matrix to stabilize LoRA training, speeding up convergence and improving performance without extra parameters or added inference cost.

2026-09-01 ~ 2026-09-02 · 2 related posts

• Normalized LoRA Stabilizes Training Without Extra Cost — Jiale Kang · 2026-09-01
• NoRA: Normalized LoRA Boosts Convergence and Stability — burny_tech · 2026-09-02