KJac: Constraint-Aware NAS for Learned Optimization

Abstract

This preprint presents KJac, a zero-training neural architecture search proxy for learned optimization in constrained optimization settings. The study evaluates Jacobian-based architecture scoring across linear projection, convex quadratic programming, QCQP-style constraints, and portfolio optimization tasks. It examines predictive correlation with final model performance, computational cost, feasibility-aware diagnostics, and architecture selection behavior. The work investigates the utility and limitations of Jacobian-based proxies for architecture search in constrained learned-optimization problems.