The goal of multi-task learning is to learn diverse tasks within a single
unified network. As each task has its own unique objective function, conflicts
emerge during training, resulting in negative transfer among them. Earlier
research identified these conflicting gradients in shared