We derive Riemannian metrics from pretrained EBMs to compute data-aware geodesics. Our approach outperforms standard methods across datasets, offering a scalable solution for learning data geometry in high-dimensional spaces.

We introduce OceanBench, a global benchmark with curated data and standardized evaluation tracks to advance reproducible, data-driven short-range ocean forecasting.

A scalable EEG foundation model leveraging 60,000+ hours of data, adaptable to any electrode setup, offering ready-to-use embeddings and state-of-the-art performance across diverse tasks.