learning-to-optimize

This paper proposes Learning-to-Optimize as a missing architectural layer for AI-native networks, redefining optimization algorithms as offline knowledge generators to train surrogate models for efficient real-time inference in dynamic environments.

FunL2O presents a unified framework that uses LLM-driven program evolution to automate feature function design for learning-to-optimize methods, showing consistent improvements over hand-crafted features across continuous and discrete optimization tasks.