Reparameterized LLM Training via Orthogonal Equivalence Transformation | Read Paper on Bytez