FMIP: Multimodal Flow Matching for Mixed Integer Linear Programming

Abstract: Mixed-Integer Linear Programming (MILP) is a cornerstone of mathematical optimization, enabling the modeling of complex decision-making problems involving both integer and continuous variables. Despite its versatility, most MILP problems are NP-complete, making them challenging to solve in practice. Existing graph neural network (GNN)-based heuristics aim to reduce problem scale by predicting only the solutions on integer variables for a given instance, struggling to capture the intricate interplay between continuous and integer variables and lack sufficient representational power. To address these limitations, we propose FMIP, a novel multimodal flow-matching framework that models the joint distribution over integer and continuous variables in the mixed solution space of MILP. To enable more accurate and scalable heuristics, FMIP integrates a guidance mechanism to guide solution sampling under both objective function optimization and constraint satisfaction. We evaluate FMIP on seven standard MILP benchmarks. Our experiments show that FMIP improves solution quality by 50.04% on average over existing GNN-based predictive baselines. These results highlight FMIP's potential as a powerful new approach for developing learning based MILP solution strategy.