Don't Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference Paper • 2609.05275 • Published Sep 4 • 26