[2026-03-06] A recipe for scalable attention-based MLIPs: unlocking long-range accuracy with all-to-all node attention — Papers — Materialist