GWO: A Gradient-Weighted Optimization Approach for Continual Learning in Large Language Models

Proposed GWO, a novel gradient-weighted optimization framework that dynamically adjusts parameter update magnitudes based on gradient importance scores. The method significantly reduces catastrophic forgetting while maintaining computational efficiency across sequential learning tasks.

提出了 GWO 梯度加权优化框架，根据梯度重要性分数动态调整参数更新幅度，在序列学习任务中显著降低灾难性遗忘并保持计算效率。