Benchmarking

Reported wall-clock timings depend on the implementation and hardware used for benchmarking. The empirical evaluation focuses on convex quadratic programs in OSQP form and one MPC benchmark, leaving broader structured convex problems for future work. Training and testing use separate problem sizes, with larger sizes used for testing. The learned policies are trained separately for each problem family rather than as a single universal policy. The benchmark-suite experiments include Random QP, Portfolio Optimization, Lasso, SVM, and Control QP families.