Paper Detail
Jiale Kang, Ziyin Yue, Zheng Zhan, Yangyi Huang, Weiyang Liu
While low-rank adaptation (LoRA) is widely used for parameter-efficient model adaptation, how to regularize its training dynamics for stable and effective optimization remains underexplored. Because LoRA initializes the up-projection to zero, its early optimization dynamics are largely governed by the down-projection. Building on this observation, we introduce Normalized Low-Rank Adaptation (NoRA), a simple yet effective method that normalizes the down-projection matrices during training. We further show that the same normalization can be applied only at initialization, improving standard LoRA without requiring repeated normalization throughout training. Across pretraining, supervised finetuning, and reinforcement learning, NoRA consistently accelerates convergence, improves performance and training stability, and mitigates catastrophic forgetting. These benefits require neither additional trainable parameters nor inference-time computation, making NoRA a simple and broadly applicable enhancement to LoRA.
No structured notes yet. Add `summary_sections`, `why_relevant`, `claim_impact`, or `next_action` in `papers.jsonl` to enrich this view.
No ranking explanation is available yet.
No tags.
@article{kang2026normalized,
title = {Normalized Low-Rank Adaptation},
author = {Jiale Kang and Ziyin Yue and Zheng Zhan and Yangyi Huang and Weiyang Liu},
year = {2026},
abstract = {While low-rank adaptation (LoRA) is widely used for parameter-efficient model adaptation, how to regularize its training dynamics for stable and effective optimization remains underexplored. Because LoRA initializes the up-projection to zero, its early optimization dynamics are largely governed by the down-projection. Building on this observation, we introduce Normalized Low-Rank Adaptation (NoRA), a simple yet effective method that normalizes the down-projection matrices during training. We fur},
url = {https://arxiv.org/abs/2608.31036},
keywords = {cs.LG, low-rank adaptation, LoRA, Normalized Low-Rank Adaptation, NoRA, down-projection, catastrophic forgetting, reinforcement learning, code available, huggingface daily},
eprint = {2608.31036},
archiveprefix = {arXiv},
}
{}