Statistical Benefits of Fine-Tuning from Pretrained Initialization in Diagonal Linear Networks Published: September 28, 2026Direct Link Previous Next