Deep learning models frequently struggle with long-tailed recognition, where data distribution is highly imbalanced across classes. A common strategy involves a two-stage decoupling paradigm, separating initial representation learning from subsequent classifier retraining. During the crucial classifier retraining phase, adaptive norm rescaling is a popular technique that adjusts per-class weight norms via parameter regularization. This regularization, however, inherently introduces additional hyperparameters, complicating the optimization process. Researchers have introduced a novel solution: a "Monotonic Adaptive Norm Rescaling" (MANR) approach specifically designed for hyperparameter-friendly optimization in long-tailed recognition scenarios1. This method aims to simplify the training workflow by minimizing the need for extensive manual tuning of these new hyperparameters. By reducing the dependency on intricate parameter adjustments, MANR promises to enhance the efficiency and stability of long-tailed recognition models. Practitioners addressing highly skewed datasets can therefore anticipate a more streamlined and robust training methodology, accelerating the development of high-performing deep learning systems.