Generalization and Scaling of Neural Solvers

A key challenge in Neural Combinatorial Optimization is generalizing to instances that are larger or drawn from different distributions than those seen during training. I investigate curriculum learning, data augmentation, and inference-time search strategies to improve out-of-distribution robustness, and I benchmark learned solvers against exact and metaheuristic baselines (OR-Tools, LKH, Gurobi) on solution quality, optimality gap, and runtime.