Scion

Scion ("Training Deep Learning Models with Norm-Constrained LMOs") generalizes the Muon optimizer through a linear-minimization-oracle (Frank-Wolfe) view: it takes norm-constrained steps chosen per layer-geometry, unifying and extending recent geometry-aware optimizers. It has drawn third-party reimplementations and KAUST/EPFL follow-ups. ICML 2025.

Attribution note: Scion is the work of Volkan Cevher's LIONS lab at EPFL — not Martin Jaggi's MLO group — and it is not a SNAI/Apertus-branded artifact.