Erwin

Erwin: A Tree-based Hierarchical Transformer for Large-scale Physical Systems (ICML 2025) — ball-tree partitioning gives linear-time attention on large physical systems spanning cosmology, molecular dynamics, and particle fluids. AMLab UvA-led (Zhdanov, van de Meent); Welling coauthors with a CuspAI dual affiliation, but the paper is AMLab/UvA-branded and counts as UvA. Adoption signals: an independent follow-up with natively trainable sparse attention and author-line integrations with DeepSeek NSA and Transolver at the ICML 2025 long-context-FM workshop.