Machine learning (ML) is increasingly advancing cryospheric science, offering new ways to extract information from the rapidly growing volume of satellite, airborne, and in situ observations and to accelerate numerical modeling. Deep learning has proven particularly effective for computer vision tasks such as calving front detection, glacier outline delineation, mapping of permafrost landforms, sea ice classification, and snow cover mapping. Beyond image analysis, ML-based emulators can reproduce the behavior of computationally expensive physical models. This enables large ensembles and the inversion of poorly constrained parameters. Hybrid and physics-informed approaches further promise to combine the flexibility of data-driven methods with physical consistency. More recently, foundation models pretrained on large Earth observation archives have raised expectations of strong transferability across regions and sensors, particularly for applications where labeled training data are scarce. However, their benefit for cryospheric applications has yet to be demonstrated, as they have not consistently outperformed task-specific architectures trained from scratch. Systematic evaluation on shared benchmark datasets is therefore essential to measure real progress. We invite contributions on all aspects of machine learning for the cryosphere, including but not limited to computer vision and remote sensing, model emulation and hybrid modeling, benchmark datasets, as well as trust and interpretability. We also welcome critical perspectives, such as barriers to adoption, computational costs, and cases where ML approaches did not outperform physics-based methods.