Junyang Cai, a second-year Ph.D. student in computer science at the University of Southern California, shares his research on multitask representation learning for mixed-integer linear programming (MILP) in this video spotlight. Cai works with AI4OPT and is advised by Professor Bistra Dilkina.

His work introduces a two-step multitask learning framework designed to improve machine learning–guided MILP solvers. The framework first trains a shared network architecture to learn effective MILP embeddings across tasks, then fine-tunes task-specific layers for individual problem domains. Evaluated on benchmark problems such as combinatorial auctions and vertex cover, the model performs competitively on known instances and generalizes well to larger, unseen ones.

Cai’s research contributes to the development of foundation models for optimization, a central goal of AI4OPT.