💡 This paper proposes the unified optimization framework MARS that combines preconditioned gradient methods with variance reduction via scaled stochastic recursive momentum, and demonstrates its signifi