{"id":"arxiv/2611.08910","url":"https://arxiv.org/abs/2611.08910","tags":["math.NA","cs.LG"],"score":0,"title":"Task-Contribution Dynamics of Adam in Physics-Informed Neural Networks: Slow-Manifold Analysis and Gradient Modulation","domain":"math.NA","source":"arxiv","authors":["Enzhan Zheng","Xiaoliang Song"],"reviews":0,"subject":"math.NA","summary":"Physics-informed neural networks (PINNs) often suffer from imbalanced optimization of competing loss components, particularly when approximating solutions with high-frequency or multiscale features. Although existing balancing strategies have proven effective in alleviating this difficulty, the mechanisms governing the evolution of individual task contributions to Adam's parameter updates remain insufficiently understood. We introduce task-wise indicators to characterize this evolution and identify three distinct training regimes: smooth variation, intermittent spikes, and persistent…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-11-06","reactions":{}}