OSQP

Changing the relaxation parameter does not alter the factorization, making it attractive for frequent online adaptation. Under fixed rho, learned relaxation policies outperform baseline OSQP across five benchmark families in both iteration count and runtime. OSQP already adapts a diagonal penalty matrix heuristically while keeping the relaxation value fixed at 1.6I. Under adaptive rho, learned Gamma_k can reduce iterations while runtime depends on interactions with OSQP's rho-update heuristic. In OSQP-like solvers, changing the penalty parameter may trigger costly matrix refactorization.