This talk series presents my recent work on stabilizing column generation with learning-based deep dual-optimal inequalities. The goal is to improve the numerical behavior of decomposition-based exact optimization methods while keeping the resulting framework compatible with rigorous algorithmic guarantees and large-scale computational workflows.

I presented this topic at the 2025 INFORMS Annual Meeting and it is scheduled again at the 2026 Transportation Science & Logistics Conference. Across both venues, the emphasis is on how learning can support exact optimization in a way that remains transparent, modular, and useful for difficult routing and inventory problems.