GNNs break the Weisfeiler-Leman ceiling

🔥 What happened Researchers at the University of São Paulo introduced Message-Passing Graph Persistence (MGP), which folds message passing directly into graph filtrations for persistent homology. Until now, GNNs computed topological descriptors separately from representation learning; MGP merges the two. 💡 Why it matters GNNs are capped by the 1-WL test – they fundamentally cannot distinguish certain graph structures. MGP is provably strictly more expressive than prior PH-based descriptors and beats them on most benchmark datasets. For tasks like molecular classification, that's a real lever. ⚡ Our take Finally someone treats topology as more than an afterthought. Anyone still building GNNs without a topological layer in 2026 is leaving expressivity on the table – literally.